Datasets:

KenLuo

/

RC-Sum-PLOS

like 0

10.1371/journal.pntd.0000432

WR279,396, a Third Generation Aminoglycoside Ointment for the Treatment of Leishmania major Cutaneous Leishmaniasis: A Phase 2, Randomized, Double Blind, Placebo Controlled Study

Cutaneous leishmaniasis (CL) is a disfiguring disease that confronts clinicians with a quandary: leave patients untreated or engage in a complex or toxic treatment. Topical treatment of CL offers a practical and safe option. Accordingly, the treatment of CL with WR279,396, a formulation of paromomycin and gentamicin in a hydrophilic base, was investigated in a phase 2 clinical study in Tunisia and France. A phase 2, randomized, double blind, vehicle-controlled study was conducted to assess the safety and efficacy of topical WR279,396 when applied twice a day for 20 days as treatment for parasitologically confirmed CL. The study protocol established the primary efficacy end point as complete clinical response (CCR) defined as 50% or greater reduction in the ulceration size of an index lesion by day 50 (D50) followed by complete re-epithelialization by D100, and no relapse through D180. Ninety-two subjects were randomized. Leishmania major was identified in 66 of 68 isolates typed (97%). In the intent-to-treat population, 47 of 50 WR279,396 treated participants (94%) met the definition of CCR, compared with 30 of 42 vehicle-placebo participants (71%) [p = 0.0045]. Erythema occurred in 30% and 24% of participants receiving WR279,396 and placebo, respectively [p = 0.64]. There was no clinical or laboratory evidence of systemic toxicity. Application of WR279,396 for 20 days was found to be safe and effective in treating L. major CL, and offers great potential as a new, simple, easily applicable, and inexpensive topical therapy for this neglected disease. ClinicalTrials.gov NCT00703924

Cutaneous leishmaniasis is due to a small parasite (Leishmania) that creates disfiguring sores, and affects more than one million persons (mainly children) each year. Treating lesions with a cream—instead of with injections as currently done—would greatly improve the well-being of affected patients. No cream formulation that would be efficient and would not create important skin irritation has been identified yet. Here, we tested a new cream formulation (WR279,396) containing paromomycin and gentamicin, two members of a well-known family of antibacterial antibiotics (aminoglycosides). Injectable paromomycin is efficient in other forms of the disease (visceral leishmaniasis). This was a carefully monitored study (phase 2) involving mainly children in Tunisia and France. The cream was applied twice a day for 20 days. The proportion of patients treated with the paromomycin-containing cream (active formulation) that cured (94%) was higher than that observed (71%) in patients treated with a cream that did not contain the active product (placebo formulation). Local irritation affected less than one-third of the patients and was usually mild. This new cream formulation was safe and effective in treating cutaneous leishmaniasis, thereby providing a new, simple, easily applicable, and inexpensive treatment for this neglected disease.

The incidence of cutaneous leishmaniasis (CL) globally is 1.0–1.5 million cases annually [1]. There are several available therapeutic options, but none is optimal [2]. In Tunisia, the standard treatment is with intralesional injections of pentavalent antimonials [2] the recommended drugs used for the treatment of both visceral leishmaniasis and cutaneous leishmaniasis, first introduced 60 years ago. Intralesional injections are painful, and they are difficult to administer to children, to patients with multiple lesions, or when lesions are located on the extremities [3],[4]. In such cases systemic antimony is often administered, despite its cost (US$50–200 per course), variable efficacy [2],[5], and potential or frank toxicity [2],[6]. Topical therapy of CL is an approach that is potentially efficient, practical and safe [7], yet a product fulfilling all those requirements has not yet been identified [8]. The aminoglycoside paromomycin is the most studied compound as a potential topical treatment for CL [9], and parenterally it is being aggressively pursued as a highly effective treatment for human visceral leishmaniasis [10]. As a class, aminoglycosides accumulate in lysosomes [11] where Leishmania multiply, and offer the potential to be highly effective therapeutics. In topical preparations, paromomycin is a component of two antileishmanial products currently available outside the International Conference of Harmonization zone (ICH for USA, EC, and Japan). The first, developed by El-On [12] and marketed in Israel as Leshcutan, contains 15% paromomycin and 12% methyl-benzenthonium chloride (MBCL) in white soft paraffin. It has demonstrated good efficacy in treating CL [13],[14] but its usefulness is hampered by increased incidence of dermatologic irritation attributable to the MBCL [13],[14],[15],[16],[17]. The second formulation combines paromomycin (no MBCL) with urea and while non-irritating its efficacy remains largely undistinguished, with reported cure rates little better than placebo in Iran (47% v. 44%) and Tunisia (27% v. 18%) [18],[19],[20]. WR279,396 is a hydrophilic formulation of paromomycin 15% plus a second aminoglycoside (gentamicin 0.5%), that was developed in order to exploit the antileishmanial efficacy of the aminoglycosides while eliminating the potential for the skin irritation caused by MBCL. In this preparation, the addition of gentamicin has been shown to increase the antileishmanial efficacy of paromomycin in rodents [21]. In a Balb/c mouse model of CL, WR279,396 cured lesions caused by L. major (MON-4), L. amazonensis, L. mexicana, and L. panamensis strains in 100% of the mice without subsequent relapse [21]. These results were recently confirmed in a C57Bl/6 L. major MON-26 model [22]. In a pilot study in humans in the New World (L. panamensis), WR279,396 was well tolerated and shortened cure time, but had no effect on the overall cure rate at six months [23]. Herein, we report the results of a recently completed phase 2, randomized, double blind, vehicle-controlled trial in Tunisia and France to assess the efficacy and safety of topical WR279,396 administered twice a day for 20 days for the treatment of patients with CL caused by L. major. Eligible participants were from the Sidi Bouzid governorate (Central Tunisia), where L. major MON-25 is endemic, and travelers returning to Paris from L. major-endemic areas in North and Sahelian Africa, who had skin lesions that were suspected to be CL. Criteria for inclusion were age between 5 and 75 years, the presence of parasitologically confirmed CL, lesions that were primarily ulcerative (i.e., not purely verrucous or nodular) and measured ≥1 cm2 and ≤5 cm2. Criteria for exclusion were history of known or suspected hypersensitivity or idiosyncratic reactions to aminoglycosides; previous use of antileishmanial drugs (within 3 months) or nephrotoxic or ototoxic drugs; prior diagnosis of leishmaniasis; more than 5 lesions, or a lesion in the face that in the opinion of the attending dermatologist could potentially cause significant disfigurement; significant medical problems as determined by history or laboratory studies; breast feeding and pregnancy. Participants also had to have normal Romberg tests and no relevant findings on baseline audiometry. In cases where the participant presented more than one lesion, investigators treated all lesions as per protocol with the same blinded study treatment as the index lesion. The study was a phase 2, randomized, double blind, vehicle-controlled, multi-center trial. Participants were randomized in a 1∶1 allocation ratio to receive either WR279,396 or placebo-vehicle, each of which was applied twice daily for 20 days and covered with an occlusive dressing (Tegaderm, 3M Laboratory, Saint Paul, MN). Investigators, who were blinded to whether participants received WR279,396 or placebo-vehicle, evaluated lesions for clinical response on D20 (i.e., the end of the treatment period), D50 (i.e., 30 days after the conclusion of treatment), D100, and D180. A sequence of genuine random numbers for the randomization procedure was obtained from the “fourmilab.ch/hotbits” website by a member of the Department of Chemical Information, Walter Reed Army Institute of Research, Silver Spring, Maryland and purged of duplicates. The random numbers are generated by a process which takes advantage of the inherent uncertainty in the quantum mechanical laws of nature. Specifically, they are generated by timing successive pairs of radioactive decays detected by a Geiger-Müller tube interfaced to a computer. This process is better than the pseudo-random number algorithms typically used in computer programs. The randomization of the study drugs was done by an independent group, Fischer BioServices, Rockville, Maryland a contractor to The U.S. Army Medical Research Acquisition Activity (USAMRAA), Ft. Detrick, Maryland. The study protocol established the primary efficacy end point as complete clinical response (CCR), defined as complete reepithelialization (i.e., length×width of ulceration = 0×0) of the index lesion by D50 or a >50% reepithelialization by D50 followed by complete reepithelialization on or before D100 with no relapse ever having occurred from D50 through D180. Relapse was defined as an increase in the area of ulceration relative to the previous measurement. Participants who did not complete the 180-day period of observation were considered to have failed to achieve CCR because relapse could not be fully assessed. The index lesion was defined as the uppermost, primarily ulcerative, parasitologically positive lesion on the body (excluding the ears) or, if two lesions were equally uppermost, the left uppermost primary ulcerative lesion. The secondary endpoint was the safety and tolerance of WR279,396. The primary performing Institutions were the Institut Pasteur in Tunis, Tunisia, and the Medical Center Institut Pasteur, in Paris, France. Investigators measured all lesions in two perpendicular directions and took photographs at the following time points: prior to therapy, at the end of therapy (D20), and at 30 days (D50), 80 days (D100), and 6 months (D180) after the end of therapy. Medical personnel applied study drug (i.e., placebo-vehicle or WR279,396) twice daily for 20 days to all CL lesions present at baseline at a dose of 0.05 ml per 1 cm2 of CL lesion at a primary health facility in Tunisia and at the Medical Center of the Institut Pasteur in Paris. Each CL lesion was cleaned with soap and water and sterile 0.9% saline, and then dried using sterile USP Type VII Gauze sponges before application of study drug. Next, medical personnel dispensed study drug directly onto the ulcer from a pre-loaded 1 ml syringe without a needle, and spread drug over the ulcer using the finger of a disposable glove so as to penetrate even under the ulcer's borders. Study drug was to remain undisturbed (i.e., not wiped off and not wetted) for 4 hours after each application, so the adhesive polyurethane film dressing, Tegaderm, was applied over the top of the lesion following drug application. Investigators observed each participant for 30 minutes after application of study drug. Lesions and surrounding skin were evaluated for pain, erythema, and edema each day that the topical creams were administered and at follow-up study visits. The participants were also observed and questioned daily for the occurrence of systemic side effects (e.g., vertigo, tinnitus) using a standardized questionnaire. Diminished hearing was verified with the Danplex S42 audiometer (GN Otometrics, Maarkaervej 2A, DK-2630, Taastrup, Denmark). Clinical and laboratory evidence of side effects was determined on D10 and D20 by changes from baseline in serum creatinine, hearing, and Romberg tests. A Digmatic Caliper, Mitutoyo Corporation, model No. CD-6CS with a resolution of 0.01 mm and an accuracy of ±0.002 mm was used to measured lesions size. Lesions were measured by a trained investigator that followed a Study Specific Procedure (SSP-279396-01-003) in two perpendicular directions; in its greatest dimension, and at 90 degrees to the first measurement. Patients were not given incentives to come back for follow-up visits; patients were actively followed-up. Before entry into the study, investigators obtained written informed consent from all participants or, for pediatric participants, from parents/guardians. Comparison of WR279,396 to placebo-vehicle was justified for several reasons: CL caused by L. major is self-limiting and heals without treatment after several months. Furthermore, the trial allowed participants whose condition worsened to withdraw from the study and receive standard therapy. In addition vehicle application provided the following advantages: (i) protection against bacterial infection by keeping the lesion(s) clean and occluded; (ii) direct access to the medical team that performed the medical history, physical exam, dermatology exam, and laboratory test ; and (iii) complete parasitological diagnosis. The study protocol (Principal Investigator Dr. Max Grögl), case report form, and SOPs were approved in the United States by the Walter Reed Army Institute of Research, Scientific Research Committee. A second level review of the protocol, consent form and all amendments was conducted by the Human Subjects Research Review Board (HSRRB), Commanding General, U.S. Army Medical Research and Materiel Command (USAMRMC), the Medical Ethical Committee of the Institut Pasteur de Tunis, Tunisia, and the Consultation Committee for the Protection of Individuals in Biomedical Research at Hospital Tarnier-Cochin, Paris, France. The study was conducted in accordance with Good Clinical Practice (GCP) under an Investigational New Drug (IND) application submitted to FDA. The Direction de la Pharmacie et des Médicaments, Ministère de la Santé Publique, Tunisia, and the Agence Française de Sécurité Sanitaire des Produits de Santé were informed of the trial. This study was conducted in accordance with ethical principles that have their origins in the Declaration of Helsinki and the Belmont Report. The Quality Assurance Office of the U.S. Army Medical Materiel Development Activity monitored the study. WR279,396 is an off-white to yellowish, thick cream containing 15% (w/w) paromomycin-sulfate (Farmitalia) and 0.5% (w/w) gentamicin-sulfate (Schering) as active components. Study drugs were manufactured by the University of Iowa, College of Pharmacy under Good Manufacturing Practice (GMPs). The placebo consisted of the vehicle without the active components and trace amounts of coloring agents to match the appearance and maintain the blind. Each lesion to be evaluated for efficacy was aspirated and/or scraped and/or biopsied. Proof of infection was documented through either the demonstration of motile promastigotes in aspirate cultures or the microscopic identification of Leishmania amastigotes in material obtained from CL lesions. Iso-enzyme [24] and/or PCR [25] analysis of the parasites isolated from the CL lesions was completed after study treatment had been started. Iso-enzyme and PCR analyses were carried out according to published protocols [24],[25]. The protocol calculated a sample size of 50 participants per group with 80 percent power and a Type I error rate of 5 percent to detect a 30 percent difference in the proportion of participants achieving CCR, assuming a CCR proportion of 35 percent in the placebo-vehicle group and 65 percent in WR279,396 participants, with a 5% expected rate of loss to follow-up. Analyses included all randomized participants under the intention-to-treat principle and the randomization was coordinated between the two clinical sites. Continuous data were compared using the Wilcoxon rank-sum test and categorical data were compared using the Fisher's exact test. StatXact version 7 (Cytel Software Corporation, Cambridge, MA) was used to calculate 95% exact confidence intervals (CIs) of the difference in the proportion achieving CCR with the option to compute a CI on the difference of two binomial proportions based on the standardized statistic and inverting two one-sided tests. A log-rank test was used to compare time to reepithelialization without relapse. Because the trial collected data on clinical response only at discrete time-points, namely the D20, D50, D100, and D180 visits, the time-to-event analysis grouped reepithelialization times according to the visit at which investigators observed the event. To adjust for baseline differences, a linear model for the proportion of participants achieving CCR was fit for each baseline variable of interest with covariates for treatment group and the baseline variable. To examine whether the effect of WR279,396 varied between subgroups, we calculated the Breslow-Day test for homogeneity of the odds ratio Between March 2003 and January 2005, 142 participants (27 in Paris and 115 in Sidi Bouzid, Tunisia) were screened, of whom 92 (10 in Paris and 82 in Sidi Bouzid, Tunisia) underwent randomization; 50 were assigned to the WR279,396 group and 42 to the placebo-vehicle control group (Figure 1). The study was conducted over at least one entire leishmaniasis season. Figure 1 presents the distribution of participants from screening until study completion in the two treatment groups for both sites. Forty-nine of 50 participants randomized to WR279,396 and 41 of 42 participants randomized to placebo-vehicle completed the study. All participants had lesions that were parasitologically confirmed by smear, culture or both. Iso-enzyme testing of 18 isolates (8 isolates from the Paris site and 10 from Tunisia) identified L. major in 17 participants and L. infantum in one participant, who was from the French site. L. major isolates from the Paris site were MON-74, MON-26, MON-25, -, and all L. major isolates from Tunisia were MON-25. Fifty isolates from Tunisia were tested using PCR, which identified L. major in 49 participants and L. tropica in one participant. In total, L. major was identified in 66 of 68 isolates typed (97%). With one exception, applications of study drugs were conducted according to the protocol. In this one case, treatment was stopped after only 12 applications (6 days) due to skin irritation and conjunctivitis that resulted from inadvertent contact of study drug to the eye while sleeping. However, this participant's lesion rapidly improved without any subsequent therapy allowing follow-up evaluations to be conducted as per protocol. All 92 participants received study drug (either placebo-vehicle or WR279,396). Except for 2 participants, who withdrew voluntarily from the study during or following the 20-day treatment period to receive alternative therapy (Figure 1), no participant was lost to follow-up, and all major end-points were accessible for all. Overall, the two treatment groups were similar in baseline demographics and disease characteristics (Table 1). A greater proportion of placebo-vehicle participants were 18 or older compared to WR279,396 participants. Because participants at the Medical Center Institut Pasteur in Paris contracted CL while traveling, a greater proportion of participants were 18 years or older compared to Tunisian participants. Fifty-four participants had a single lesion at baseline. The distributions of lesion area, both of the index lesion and of all baseline lesions, were roughly equivalent between the groups. Forty percent of participants in each treatment group had the index lesion above the belt, and all but two WR279,396 participants had the index lesion on the limbs. The median number of days before treatment since participants first noticed a baseline CL lesion/papule was 62 in both groups. Two participants were confirmed to be infected with a non-L. major species. One was infected with L. infantum and the other with L. tropica. Self-healing occurs less frequently with both species than with L. major. Both participants were in the WR279,396 group and achieved CCR, although the participant infected with L. infantum received only 6 days of treatment. We explored CCR in subgroups defined by baseline characteristics including number (1 and >1), index lesion area (<100 and ≥100 mm2), location (upper and lower body), and age (<60 and ≥60 days) of lesions. After adjustment singly for each baseline factor, the statistical benefit of WR279,396 over placebo-vehicle remained (Table 3). Adjustment for age less than 18 years, however, noticeably lessened the estimate of the treatment effect, with an adjusted difference in the proportion of CCR of 16 percent (95% CI: 1, 30). The higher proportion of participants under 18 years achieving CCR combined with the greater proportion of WR279,396 participants less than 18 was responsible for this. When modeling lesion number, index lesion area, and lesion age continuously, only adjustment for index lesion area differed appreciably from the model where area was modeled categorically. Adjusting for index lesion area as a continuous covariate resulted in an estimated difference in the proportion of CCR of 19 percent (95% CI: 4, 34). No baseline factor appeared to modify the effect of WR279,396, as indicated by Breslow-Day homogeneity tests, all of which were above 0.20. Topical administration of WR279,396 was generally safe and well tolerated (Table 4). No death occurred during this clinical trial, and the only serious adverse event was an arm fracture unrelated to study medication. Overall, the number of participants experiencing adverse events was comparable, with roughly a quarter of participants in each group experiencing an adverse event. The most commonly reported event was erythema at the site of application, which occurred in 30 percent of participants who received WR279,396 and 24 percent of participants who received placebo-vehicle with onset within 30 minutes of application (p = 0.64). Mild pain within 30 minutes of application was reported in roughly 14 percent of participants in each group. No participant had an increase from baseline serum creatinine following administration of study drug (D10 and D20). Only mild increases and decreases in hearing acuity from baseline were detected on audiometry, occurring with similar frequency in both groups (28% and 21% in WR279,396 and placebo-vehicle, respectively; p = 0.63). There was no report of vertigo and no abnormal Romberg test result in participants who received WR279,396. For a neglected disease like leishmaniasis, the development of a GMP formulation that is safe and efficacious is a step in the right direction. There is a general lack of safe, effective, and affordable pharmaceuticals worldwide to treat or prevent neglected diseases that disproportionately cause high mortality and morbidity among the world's poor in the developing world [26]. Of the many examples of neglected diseases, L. major CL is perhaps one disease that should have numerous good solutions by now, yet the internationally accepted standard treatment remains largely tied to antimony, even for all the problems associated with its use. For decades, clinicians caring for patients with CL have been confronted with a difficult choice: either leave patients untreated (a common proposal for patients with five or fewer uncomplicated lesions due to L. major), or engage in a complex or toxic treatment for this disfiguring, but non life-threatening disease. In this study, we found that WR279,396 was well-tolerated and induced complete clinical response (CCR) in a significantly greater percentage of participants compared to placebo-vehicle in participants with CL due to L. major. These results raise a strong possibility that we may exit from this old quandary of how best to manage patients with L. major CL. Of the 49 participants treated in the WR279,396 arm, only 1 (2%) failed to achieve complete reepithelialization of his lesion in less than 2 months. If the efficacy of WR279,396 can be reproduced in subsequent phase 3 clinical studies, complex therapeutic decisions in L. major CL may become the exception rather than the rule. The time-to-event analysis raised two interesting observations: First, the response seen with placebo-vehicle was markedly higher than the response reported in placebo-treated participants in a previous paromomycin-urea trial performed at the same site in 1995 (71% versus 32%) [19]. Thus, an intrinsic efficacy of the vehicle of WR279,396 on CL ulcerations may account for part of this difference, and for the unexpectedly high placebo cure rate in the trial reported here. Second, and consistent with our earlier studies evaluating WR279,396, during the 20-day drug application period (between D1 and D20), the mean ulceration area for WR279,396-treated participants decreased at a slower rate than in placebo-vehicle treated participants (Figure 2). This initial transient slowing in ulceration closure was not totally unexpected. The natural progression of the healing process in CL entails a decrease in the depth of the ulceration as the parasite load decreases followed by a reduction in the ulceration width as re-epithelialization progresses. Thus, in treated participants, the non-improvement at day 20 of the mean ulceration area may be linked to the inflammatory response as parasites are killed by WR279,396. This slower decrease of ulceration area, limited to the 20-day drug application phase followed by a significant acceleration in healing after D20, bore no negative clinical impact. Indeed, only 1 participant in each group requested to be withdrawn before the major end-point evaluation at D50. Finally, and perhaps paradoxically, in most patients, the fact that reepithelialization started after the end of the 20-day application period may actually favor compliance with this treatment schedule should the drug become widely available. WR279,396 continued to demonstrate an excellent safety profile with very few local and no systemic adverse events observed. This trend was similar to our previous experience with this product compiled from pre-clinical, Phase 1, and two earlier phase 2 studies in the New World [23]. Importantly, this topical preparation containing two aminoglycosides displayed no detectable renal or VIIIth cranial nerve toxicity. These safety observations are in accord with the findings from a recent study of intramuscular (IM) paromomycin for visceral leishmaniasis in India, which also showed no clinically significant kidney or VIIIth nerve toxicity yet the systemic exposure in that IM study was much greater than from WR279,396 applied topically [27]. In addition to the promising efficacy observed in this study against L. major MON-25, data collected thus far indicate that WR279,396 will likely be broadly effective against a wider variety of leishmania species. Several key observations support such optimism. First WR279,396 was efficient not only in L. major MON-26 [22] , and in L. major MON-4, but also in L. amazonensis, L. mexicana, and L. panamensis in infected mice [21]. Second WR279,396 was active in L. panamensis in humans in Colombia [23]. Third, L. tropica is very sensitive to paromomycin in vitro [28]. And fourth, but not least, in an L. tropica focus of Turkey, the paromomycin+MBCL formulation induced a 37.5% cure rate at 4 weeks, suboptimal, but significantly higher than the 0% cure rate in oral ketoconazole-treated patients [29],[30]. The application of WR279,396 for 20 days was found to be safe and effective (94% cure rate) in treating L. major CL, and offers great potential as a new, simple, easily applicable, and inexpensive topical therapy for this neglected disease [31]. L. major CL in North Africa, Sahelian Africa, and the Middle East, involves tens to hundreds of thousands of people each year, many of whom are children [32]. Health systems are often unable to cope with these epidemics. In this context, a simple, straightforward treatment is crucial.

10.1371/journal.pbio.1001282

Convergent Evolution of Escape from Hepaciviral Antagonism in Primates

The ability to mount an interferon response on sensing viral infection is a critical component of mammalian innate immunity. Several viruses directly antagonize viral sensing pathways to block activation of the host immune response. Here, we show that recurrent viral antagonism has shaped the evolution of the host protein MAVS—a crucial component of the viral-sensing pathway in primates. From sequencing and phylogenetic analyses of MAVS from 21 simian primates, we found that MAVS has evolved under strong positive selection. We focused on how this positive selection has shaped MAVS' susceptibility to Hepatitis C virus (HCV). We functionally tested MAVS proteins from diverse primate species for their ability to resist antagonism by HCV, which uses its protease NS3/4A to cleave human MAVS. We found that MAVS from multiple primates are resistant to inhibition by the HCV protease. This resistance maps to single changes within the protease cleavage site in MAVS, which protect MAVS from getting cleaved by the HCV protease. Remarkably, most of these changes have been independently acquired at a single residue 506 that evolved under positive selection. We show that “escape” mutations lower affinity of the NS3 protease for MAVS and allow it to better restrict HCV replication. We further show that NS3 proteases from all other primate hepaciviruses, including the highly divergent GBV-A and GBV-C viruses, are functionally similar to HCV. We conclude that convergent evolution at residue 506 in multiple primates has resulted in escape from antagonism by hepaciviruses. Our study provides a model whereby insights into the ancient history of viral infections in primates can be gained using extant host and virus genes. Our analyses also provide a means by which primates might clear infections by extant hepaciviruses like HCV.

Hepatitis C virus (HCV) causes chronic liver disease and is estimated to infect 170 million people worldwide. HCV is able to establish a persistent infection in part by inhibiting the innate immune response. It does so by using its protease, NS3, to cleave the host's antiviral factor MAVS, which normally activates the interferon response. Using an assay that measures MAVS activity, we found that multiple primate species contain a version of MAVS that is resistant to HCV antagonism. Surprisingly, most of these primates have independently converged on changes in the same amino acid residue of MAVS to escape cleavage by the HCV protease. We found that the HCV protease has lower binding affinity for these resistant MAVS variants, which consequently are more effective at restricting HCV infection. Using a combination of phylogenetic and functional analyses of proteases from other HCV-related viruses, we infer that ancestral primates were likely exposed to and adapted to HCV-like viruses. One consequence of this adaptation is that changes that have given rise to extant MAVS variants may now provide protection from modern-day viruses.

Among the myriad of antiviral mechanisms employed by mammalian cells, the ability to sense viral RNA has emerged as a critical component of innate immunity. Viral RNA is detected in the cytoplasm by sensor proteins RIG-I and MDA-5 [1]–[3]. Both these sensors act through a common downstream effector Mitochondrial antiviral signaling (MAVS) (also known as IPS-1, VISA, and Cardif), which in turn mounts an interferon response (schematic of MAVS pathway in Figure S1) [4]–[7]. Given the importance of the RNA virus-sensing pathway in the antiviral response, it is not surprising that several highly diverse classes of viruses have evolved ways to inhibit multiple steps of the viral-sensing pathway. In particular, several viruses encode antagonists of MAVS function [5],[8]–[13]. Hepatitis C virus (HCV) encodes a protease NS3, which in concert with its cofactor NS4A cleaves human MAVS [5],[8]. HCV is a single-stranded positive sense RNA virus that belongs to the Flaviviridae class of viruses that causes chronic liver disease and is estimated to infect 170 million people globally (about 3% of the world's population) [14],[15]. HCV is known to naturally infect only humans, although chimpanzees can also be experimentally infected with it. GB viruses GBV-A, GBV-B, and GBV-C are other HCV-related viruses belonging to the Flaviviridae class that infect primates, although it is not clear whether they are pathogenic to their hosts. We refer to all GB viruses as hepaciviruses in this article, although some GB viruses have also previously been referred to as Pegiviruses (the International Committee on Taxonomy of Viruses has not yet formally assigned GB viruses to any genus) [15]. Recently, HCV-like viruses have also been found in non-primate mammalian species, specifically in bats and dogs [16],[17]. We investigated the functional consequences of MAVS evolution, putatively driven by antagonism from ancient viruses (paleoviruses). Viral antagonism can impose persistent selective pressure on host antiviral factors like MAVS, resulting in positive selection (i.e., accumulation of an excess of nonsynonymous changes relative to synonymous changes over evolutionary time). Positive selection has been seen previously in many antiviral factors characterized in primate genomes [18]–[24]. Positive selection largely reflects adaptations to past viral infections, with adaptive changes providing beneficial consequences for the host to overcome viral antagonism. However, the resulting adaptive changes can also influence resistance or susceptibility to present-day viruses. For example, adaptive changes at key residues in the antiviral gene Protein Kinase R (PKR), likely driven by ancient viruses, are important determinants of PKR's ability to resist antagonism by present-day poxviruses [18],[25]. Thus, antiviral genes evolving under positive selection are good candidates to be genetic determinants of resistance or susceptibility to present-day viruses. Here, we sought to determine whether MAVS has adaptively evolved during primate history and whether this positive selection has consequences for its resistance against HCV antagonism. We found that MAVS has evolved under strong positive selection in primates. We further found that one consequence of this selection is that MAVS from multiple primate species has independently become resistant to HCV antagonism. Remarkably, most of these primate MAVS variants are resistant due to changes at the same residue, and we dissect the functional consequences of this change using a combination of biochemistry and virology. Finally, based on functional analysis of other extant HCV-like viruses, we infer that these adaptive “escape” changes in MAVS were likely driven by ancient hepaciviruses, potentially providing paleoviral insights into hepaciviral infections over the course of primate evolution. We sequenced and analyzed MAVS cDNA from a total of 21 simian primate species representing nearly 40 million years of simian primate evolution (Figure 1A) [26]. The phylogeny constructed with MAVS sequences was in agreement with the recently published well-supported phylogeny of primates [26]. We found that many residues in MAVS were highly conserved throughout the course of 35 million years of primate evolution, reflecting the constraint to maintain MAVS function via purifying selection. However, using maximum-likelihood analyses [27],[28], we found that MAVS has evolved under strong positive selection in primates. Specifically, models that allow the rate of nonsynonymous (dN) changes to be greater than that for synonymous (dS) changes are statistically a much better fit to the data than models that disallow positive selection (p<0.001, Table 1). Although there appears to be considerable variation in dN/dS ratios across numerous branches of the MAVS phylogeny (Figure 1A), according to branch-site random effects likelihood (REL) analysis, no single branch has statistically significant evidence of positive selection (p>0.05) [29]. Indeed, models that allow dN/dS to be variable along branches do not have a statistically higher likelihood than models that account for one common dN/dS ratio across the entire phylogeny (p>0.05, Table 1) [27],[28], suggesting that positive selection is not confined to a few primate lineages but is rather pervasive throughout the primate phylogeny. In addition to inferences about adaptive evolution across the entire MAVS protein, we can infer positive selection at the single amino acid resolution based on recurrent nonsynonymous mutations at a single position across multiple lineages (dN/dS>1 at a single residue). According to maximum likelihood analysis using PAML [27],[28], 10 residues have high posterior probabilities of having dN/dS greater than 1 (>0.9). Nine of these 10 residues are also supported by an independent random effects likelihood (REL) analysis (Figure 1B, Table S1) [30]. These findings of positive selection are consistent with the hypothesis that MAVS has encountered and adapted to viral antagonists throughout primate history. Given the rapid divergence of primate MAVS via positive selection, we investigated whether MAVS alleles from different primate species harbored functional variation in terms of their susceptibility to antagonism by extant viruses. For instance, if primates have encountered and adapted to escape antagonists similar to HCV protease, then MAVS from some of these primate species might be resistant to inhibition by HCV NS3/4A. We tested this possibility using a facile assay to measure HCV NS3/4A inhibition of MAVS activity. Over-expression of human MAVS in cultured cells is sufficient to activate the interferon response, which can be assayed by the IFN-β promoter-driven firefly luciferase reporter. We cloned MAVS from 20 different primate species into expression vectors with a strong CMV promoter and transiently transfected them into human-derived 293T cells. We found that MAVS from all primate species was capable of inducing IFN-β-driven luciferase expression, despite relying on the signaling machinery of human cells (Figure 2A, Figure S2). Consistent with previous observations [5],[8], we observed that co-expression of HCV NS3/4A protease robustly inhibited human MAVS signaling to the IFN-β promoter (Figure 2A). This inhibition is specific to the HCV protease-mediated proteolysis at residue 508, since introducing a C508R mutation in human MAVS blocks this inhibition, as shown previously (Figure 2A) [8]. We found that HCV NS3/4A also inhibited signaling driven by MAVS from most primates. However, MAVS proteins from four species—olive baboon, rhesus macaque, spider monkey, and dusky titi monkey—were capable of significant IFN-β induction even in the presence of the HCV protease (Figure 2A). Except for olive baboon and rhesus macaque, all the “resistant” species are phylogenetically well separated from each other by “susceptible” species (Figure 1A). Thus, resistance to inhibition by HCV protease appears to have been independently acquired at least three different times. We next investigated the molecular basis of MAVS resistance to HCV NS3/4A inhibition. We found that MAVS from olive baboon has the same cysteine to arginine change at position 508 as in our positive control MAVS with C508R mutation (Figure 2B), disrupting the protease cleavage site and thus providing an explanation for its escape from HCV protease inhibition. However, the cysteine at 508 is conserved in other “resistant” MAVS' (Figure 2B), suggesting that genetic changes at different residue(s) must be responsible for their resistance to HCV protease. By examining the MAVS protein sequences from rhesus macaque, spider monkey, and dusky titi monkey, we found that changes at only a single residue, 506, can parsimoniously explain MAVS escape from HCV protease antagonism in these variants (Figure 2B). While most primate MAVS proteins have an ancestral valine residue at position 506, the three HCV protease “resistant” MAVS' have all independently altered this residue to either glycine or alanine. We therefore considered the remarkable possibility that MAVS has converged on the same escape strategy multiple times in primate evolution. Consistent with escape from viral antagonism being responsible for this convergent evolution, residue 506 is one of nine MAVS residues with high statistical support of having evolved under positive selection (Figure 1B, Table 1). To assess whether changes at position 506 are causal for resistance to HCV NS3/4A protease, we swapped the evolved glycine and alanine residues in rhesus macaque and spider monkey MAVS, respectively, back to the ancestral valine (G506V and A506V, Figure 3A). We found that both of these swapped MAVS variants were capable of IFN-β induction and were now susceptible to HCV NS3/4A inhibition (Figure 3A). Conversely, changing the ancestral valine in human and talapoin MAVS to glycine (V506G) and to alanine (V506A) in tamarin monkey MAVS confers significant resistance to NS3/4A protease antagonism (Figure 3A). Taken together, these data demonstrate that changes at residue 506 in MAVS are necessary and sufficient to explain most of the variation in resistance and susceptibility to the HCV NS3/4A protease among primate MAVS proteins. However, not all changes appear to be equally protective. For instance, Allen Swamp Monkey has a methionine instead of the ancestral valine at position 506 but is still susceptible to inhibition in the luciferase assay (Figure 2B, Figure 3A) (but see below). Residue 506 occupies the P3 position of the NS3/4A protease cleavage site, only two residues away from where the protease cleaves MAVS (Figure 2). Previous studies have demonstrated that this position is an important determinant of HCV polyprotein cleavage by NS3/4A protease, with valine being the preferred residue at P3 [31],[32]. However, functional effects of variation at the P3 position for MAVS cleavage were previously unknown. We investigated whether “resistant” changes at residue 506 protect MAVS from cleavage by the viral protease (Figure 3B). We found that “resistant” rhesus macaque and spider monkey MAVS' are not cleaved by the NS3/4A protease (Figure 3B). However, single amino acid substitutions back to ancestral valine render both rhesus macaque (G506V) and spider monkey (A506V) MAVS' susceptible to cleavage. Conversely, V506G changes in human and talapoin MAVS and a V506A change in tamarin monkey MAVS render these previously susceptible variants to now become resistant to NS3/4A protease cleavage (Figure 3B). Intriguingly, we also found that Allen's swamp monkey MAVS, which possesses a methionine at position 506, becomes more susceptible to cleavage when the methionine is changed back to the ancestral valine (M506V) (Figure 2B, Figure 3B), suggesting that methionine at 506 does offer protection even though this was less evident from the IFN-β induction assay (Figure 2A, Figure 3A). Thus, cleavage susceptibility to HCV NS3/4A protease explains the variation we saw in the IFN-β induction assay (Figure 2A, Figure 3A). To investigate the mechanism underlying cleavage resistance, we tested the importance of residue 506 (position P3) for stable complex formation between NS3/4A and MAVS. To measure stable interaction between NS3/4A and MAVS, we used the catalytically inactive protease to prevent cleavage and release of MAVS from a MAVS/NS3/4A complex. As the protease catalytic activity is needed to cleave the protease polypeptide into NS3 and its cofactor 4A, which is required for proper NS3 function, we used a single chain protease construct of NS3 that contains an active site mutation to prevent MAVS cleavage, but contains the amino acids of NS4A required for maximal protease activity fused to the amino terminus of the NS3 protease (sc-NS3 (S1165A)) [33]. We observed reduced binding of the HCV protease to human MAVS when the “susceptible” valine at residue 506 is mutated to “resistant” glycine (Figure 3C). These data suggest that reduced binding between MAVS and the protease contributes to protection from cleavage of MAVS variants with “resistant” changes at position 506. Next, we tested whether resistance to cleavage by the HCV protease allows MAVS to restrict HCV replication with greater efficiency during an infection. Cell lines that harbor subgenomic HCV replicons provide a good model system to study HCV replication [34]. Importantly, replication of this HCV genome is susceptible to IFN-induced host responses [35]. We therefore used a Huh7 cell line that harbors a HCV type 1b–derived replicon to test whether cleavage-resistant variants of MAVS are better able to restrict HCV replication [34]. Transient transfection of wildtype human MAVS confers a nearly 2-fold higher protection against HCV replication in cells (Figure 4). Moreover, transfection with cleavage-resistant human MAVS (V506G) or rhesus MAVS confers an additional 2-fold higher restriction of HCV replication, which is equivalent to that achieved by complete loss of MAVS cleavage due to mutation at the P1 position (C508Y) (Figure 4). The actual level of resistance is likely underestimated in these experiments since not all cells exposed to HCV replication were expressing MAVS, which was transiently transfected. We conclude from these experiments that resistant variants of MAVS are better able to restrict HCV replication, with changes at residue 506 (P3 position) allowing MAVS to achieve the same level of HCV restriction as the non-cleavable MAVS with mutation at the P1 position. Statistical support of positive selection based on convergent evolution in four independent lineages suggests that changes at residue 506 represent adaptations driven by viral antagonists. Functional analysis demonstrates that changes at this residue now protect MAVS against cleavage by protease from an extant virus HCV, allowing MAVS to better inhibit HCV replication. Taken together, these data are consistent with the hypothesis that changes at residue 506 represent adaptations to antagonism by proteases from HCV-like viruses. To support this hypothesis, we wished to test proteases from viruses that infect non-human primates that might antagonize MAVS at the same position as the HCV protease. The GBV-B virus is closely related to HCV and has been tentatively assigned to the Hepacivirus genus. And although its natural host is unknown, it has been shown to infect tamarins and marmosets [36],[37]. More distantly related to HCV are the GBV-A and GBV-C (also known as hepatitis G) viruses, which are known to naturally infect primates. GBV-A infects New World monkeys while GBV-C is found in humans and chimps. These two viruses have tentatively been assigned to a new Pegivirus genus, as a sister genus to the Hepacivirus genus within the Flaviviridae family (although this designation has not yet been formally accepted by the International Committee on Taxonomy of Viruses) [37]–[42]. We refer to the Hepacivirus/Pegivirus genera as hepaciviruses here, for convenience and because of their shared properties we observed (Figure 5). All GB viruses encode the NS3 protease. We therefore tested the ability of NS3 proteases from these viruses to antagonize human MAVS. Consistent with our hypothesis, we found that the NS3 protease from both GBV-A and GBV-C strongly inhibited signaling by human MAVS (Figure 5B). Consistent with previous reports, NS3 protease from GBV-B also inhibited human MAVS, while proteases from the more distantly related pestivirus bovine viral diarrhea virus (BVDV) and flavivirus yellow fever virus (YFV) did not antagonize MAVS (Figure 5) [43]. Thus, the property of antagonizing MAVS is shared and exclusive to the group of viruses encompassing the HCV and GB viruses. We next tested the effect of positively selected changes at residue 506 on the ability of the GB viruses to inhibit MAVS. Remarkably, changes at residue 506 away from ancestral valine in MAVS allow it to signal in the presence of NS3 proteases from all GB viruses (Figure 6, Figure S3). However, there are some species- and virus-specific differences. For example, methionine at residue 506 in Allen's swamp monkey MAVS appears to provide significant protection against protease from GBV-B virus compared to other GB viruses and HCV (Figure 2A, Figure 6). Also, alanine at residue 506 in red-mustached tamarin monkey did not provide a much higher protection against antagonism by GB viruses (compared to HCV, Figure 2A) than the ancestral valine. These idiosyncratic differences likely reflect slightly different properties of the HCV and GB virus proteases. However, our overall data suggest that the ability of NS3 proteases to antagonize MAVS, and to be impeded by the same adaptive changes in MAVS, is a phylogenetically discrete characteristic exclusive to HCV and the three GBV viruses, and likely inherited from a common ancestor. Importantly, these data demonstrate that non-human primates are susceptible to infection by viruses that antagonize MAVS in a manner similar to HCV. To infer the minimum age of viruses that drove the evolution at residue 506, we wanted to date these adaptive changes. We took advantage of the phylogenetically well-characterized macaque lineage, which includes rhesus macaque with the “resistant” change at position 506 [44]. We sequenced the C-terminus of MAVS from six different Macaque species and found that all of them, including the most ancestral species, the barbary macaque M. sylvanus, possess the “resistant” glycine instead of the ancestral valine at position 506 (Figure 7). Furthermore, we found no polymorphisms at this position within 20 rhesus macaque samples, suggesting that the change to glycine is likely a fixed change within the species. Since the macaque lineage separated 5–8 million years ago from that of baboons [44], which retain the ancestral valine at position 506, our data suggest that the progenitor macaque lineage encountered and adapted to a virus that antagonized MAVS at least 5–8 million years ago. In this study, we have shown that positive selection in MAVS of multiple primate species has independently resulted in its escape from antagonism by hepaciviral proteases. Including baboons that have acquired a change at residue 508, we found that 5 out of the 20 primates that we characterized possess MAVS resistant to hepaciviral antagonism. Remarkably, MAVS resistance in four of these species is a product of independently acquired changes at the same single residue 506, which allow escape from hepaciviral NS3 protease recognition and cleavage. HCV cleaves MAVS in order to block the interferon response. The importance of this strategy employed by HCV is highlighted by the fact that interferon treatment is often clinically used to successfully cure HCV infection. Furthermore, a polymorphism near a type III interferon gene is associated with spontaneous and treatment-induced clearance of HCV [45],[46]. Thus, given the importance of the interferon response, we hypothesize that primates that possess MAVS resistant to cleavage by NS3 protease should be better at clearing hepaciviral infections than species with susceptible MAVS. Consistent with this idea, we found that resistant variants of MAVS are better able to restrict HCV replication in cell culture. Furthermore, although sampling of hepaciviruses from primates has been sparse, thus far hepaciviruses have only been found in primates with “susceptible” MAVS, such as human, chimp, common marmoset, red-chested mustached tamarin, and owl monkey [38]. In contrast, hepaciviruses were assayed for but not found in spider monkeys, which have “resistant” MAVS [38]. Our study provides functional evidence that implicates host genetics in determining the outcome of diverse hepaciviral infections even between closely related primate species. The remarkable case of convergent evolution at residue 506 is indicative of adaptive evolution, likely driven by ancient viruses (paleoviruses). Based on the phylogenetic and the functional evidence using extant viruses as surrogates, we can attempt to infer the nature of these paleoviruses, although our “indirect paleovirology” is limited in its ability to formally prove the existence of such paleoviruses. Functional evidence that changes at residue 506 provide protection from antagonism by HCV is consistent with the hypothesis of ancient hepaciviruses being the causative agents of the evolution at residue 506. Since HCV per se is a human-specific virus, it is unlikely to have been responsible for MAVS evolution in non-human primates. Instead, we sought to find viruses with similar antagonistic properties as HCV. We found that NS3 proteases from GBV-A and GBV-C viruses, which naturally infect non-human primates, not only share the ability to antagonize MAVS but are also impeded by the same evolutionary changes at residue 506 (GBV-B protease also behaves in a similar manner to the other GB virus proteases and while it can experimentally infect some New World primates, its natural host is not known). These data suggest that despite the high degree of divergence, all hepaciviruses are capable of antagonizing MAVS. This shared property must have been present in their common ancestor and subsequently inherited by all extinct and present-day hepaciviruses. Furthermore, this common ancestor must have been present tens of millions of years ago. GBV-C viruses that infect humans and chimps are thought to have co-diverged with their ancestors 7 million years ago [47]. Similarly, GBV-A viruses are thought to have co-speciated with their hosts over the course of last 20 million years [47],[48]. Divergence of GBV-C from GBV-A and from their common ancestor from HCV presumably occurred even earlier. Taken together, these data suggest that hepaciviruses capable of antagonizing MAVS are an extremely ancient group of viruses that are old enough and distributed widely enough to be responsible for driving MAVS evolution at residue 506. Consistent with the ancient presence of hepaciviruses, we dated the change at residue 506 in the macaque lineage to 5–8 million years ago. Our findings imply that HCV and GB viruses represent the latest in a continuum of hepaciviral infections that have plagued primates for millions of years. Finally, it is interesting to note that MAVS in three Old World monkeys (OWM) has independently evolved resistance to hepaciviral antagonism despite the fact that hepaciviruses have not yet been detected in OWM. Multiple lines of evidence outlined above support the hypothesis that ancient hepaciviruses were responsible for driving evolution in MAVS. However, it is also formally possible that a different class of viral antagonists unrelated to the hepaciviral NS3 proteases drove the evolution at residue 506. Although it is not formally possible to rule out such an alternative, there is little theoretical or empirical evidence in support of this. Residue 506 is not an “Achilles' heel” of MAVS that would make it an especially attractive target for antagonism by other viruses. There does not appear to be anything special about residue 506, except that it sits within a stretch of residues that corresponds to the cleavage site for hepaciviral protease NS3. Consistent with this idea, all other known viral antagonists of MAVS, including proteases from other viruses, interact with different regions of MAVS [9]–[12]. Furthermore, other residues under positive selection in MAVS are spread throughout the length of the protein, suggesting that MAVS can be antagonized along multiple surfaces and not just at or near residue 506. Taken together, these reasons make it unlikely that other viral antagonists besides hepaciviral proteases have converged to drive the evolution at residue 506. We have previously proposed that virus-driven adaptive evolution of host antiviral factors can be used as a tool in paleovirology, the study of ancient viruses and their impact of host evolution [49],[50]. The reasoning behind this idea is that positive selection in host antiviral factors essentially represents viral “footprints” that can reveal the action of ancient viruses. An inherent limitation of our “indirect” paleovirology approach is that it cannot formally rule out the possibility that either an unrelated virus or another selective pressure altogether was responsible for the observed evolution. Thus, in the case of evolution at residue 506 in MAVS, we cannot formally rule out alternate selection scenarios. The discovery of endogenized hepaciviral “fossils” in primate genomes might be an important confirmation of ancient hepaciviruses and their ability to infect primates. While such viral “fossils” would provide direct evidence for existence of paleoviruses, they would not necessarily reveal the functional consequences of these viruses on host evolution. Thus, “fossil”-based and “indirect” paleovirology should be viewed as complementary approaches, both of which are important to determine the identity of paleoviruses and their impact on host genomes. One limitation of using adaptive evolution in host factors to study paleovirology that has been previously discussed [50] is the difficulty of deciphering the history of a single type of virus. For example, protein kinase R (PKR) is a broad-spectrum antiviral factor that is antagonized by a multitude of diverse viruses including ssRNA and dsDNA viruses, and often at the same domain. This is reflected in the large number of residues in PKR that are evolving under positive selection [18]. Such overlapping antagonism might obscure the history of individual viruses. Moreover, every subsequent round of adaptation might overwrite the paleoviral record of all previous infections. In contrast, despite MAVS acting against a broad panel of RNA viruses, there are only nine residues evolving under positive selection over primate evolution. This suggests that only a few viruses directly antagonize MAVS to drive its positive selection. This limited antagonism makes it less likely that more than one virus will converge on the same residue. Thus, studying proteins such as MAVS with sparse positive selection widely distributed along its length offers the opportunity to make paleoviral inferences with greater specificity. Our study shows that the nature of the viral antagonist also contributes to the specificity of the paleoviral record. For example, it is surprising that the diverse hepaciviruses we tested have not dramatically altered the substrate specificity of their NS3 proteases over a long period of evolution. One explanation—supported by recent data from genomic viral “fossils” [51]—is that long-term virus evolution is much slower than expected due to functional constraint [50]. Indeed, the NS3 proteases are highly constrained because besides cleaving MAVS, they are also responsible for cleaving the viral polypeptide at multiple locations. Thus, given that paleovirology relies on the use of extant viruses as surrogates of their ancient counterparts to test the function of positively selected changes in host factors, viral “footprints” left by functionally constrained antagonists like proteases are particularly well suited for paleovirology [11],[12]. Functional dissection of the interaction between residue 506 in MAVS and HCV provides a model to study paleovirology using extant viruses and host immune factors. This model can be applied to the study of other residues in MAVS that have evolved under positive selection in primates, which we suspect also reflect a history of antagonism by other viruses. For example, coxsackievirus uses its 3C protease to cleave human MAVS between residues 148 and 149 (Figure S4) [12]. Remarkably, residue 149 has evolved under one of the strongest signatures of positive selection in MAVS (Figure 1B, Table S1). It is also interesting to note that residue 423, which forms the P5 position within the cleavage site for Hepatitis A virus protease, is also evolving under positive selection (Figure 1B, Table S1, Figure S4). Future experiments will be needed to establish the functional importance of residues 149 and 423. Nevertheless, the fact that the cleavage sites for all known protease antagonists of MAVS contain residues evolving under positive selection suggests that the correlation is unlikely to be simply a coincidence (Figure S4). A particularly intriguing case of positive selection at residue 60 in MAVS also promises to reveal paleoviral insights (Figure 1B). This residue has independently acquired changes in six primates. In every instance, the ancestral asparagine has changed to serine (Figure S5). This selective change likely reflects the pressure to escape from some viral antagonism by evolving away from the ancestral asparagine, while continuing to maintain MAVS function, which might be particularly sensitive to changes at residue 60. Residue 60 occurs within the highly conserved CARD domain, which is utilized to directly interact with viral RNA sensors RIG-I and MDA-5 [6]. Interestingly, MAVS in colobus monkey is polymorphic for both asparagine and serine, suggesting that the selective pressure driving the change might be very recent. It will be interesting to functionally test whether serine at residue 60 provides protection from one of the known viral antagonists of MAVS. Alternatively, the dramatic convergent evolution at residue 60 might help discover a new viral antagonist of MAVS. Isolation of total RNA was done as described previously [18]. Primate MAVS cDNAs were isolated by RT-PCR using 50 ng of total RNA as template and TA-cloned into pCR2.1 (Invitrogen). Primer MP-11, which sits upstream of the start site, and primer MP-12, which is downstream of the stop codon, were used for RT-PCR (Table S2). MP-34, along with MP-11 and MP-12, was used for sequencing the cDNAs. MAVS cDNA sequences obtained in this study are in the process of being submitted to GenBank. Common marmoset sequence was obtained from the UCSC Genome Browser. Primate MAVS cDNAs were prepared using 250 ng total RNA as template and oligo dT primer (Superscript III, Invitrogen). cDNAs were subsequently used as a template for PCR (Phusion, NEB). MAVS cDNAs were cloned into pFLAG-CMV-2 between HindIII and BglII downstream of and in frame with FLAG tag. Single amino acid changing point mutations were introduced in human MAVS as follows: MP-51, a forward primer upstream of the MAVS cDNA in pFLAG-CMV-2, was used in combination with a reverse primer that included the desired change. A forward primer with the desired change was then used with MP-52, a reverse primer downstream of MAVS cDNA in pFLAG-CMV-2. The resulting products from these two PCRs were used as a template for stitch PCR with MP-51 and MP-52 primers. The resulting PCR product, which included the desired change in MAVS, was ligated into pFLAG-CMV-2 between HindIII and BglII in frame with FLAG tag. The remaining single amino acid changes in MAVS were introduced using similar stitching strategy, but desired amino acid changes were achieved by stitching together N- and C-terminal MAVS of different but closely related species. NS3/4A from GBV-B, GBV-C, HCV, and GBV-A was cloned into pcDNA3.1/Hygro(−) between XbaI and HindIII downstream of and in frame with HA tag (cloned between NheI and XbaI). GBV-A NS3/4A was synthesized (Genscript Inc., Piscataway, New Jersey) using strain A-lab sequence. Other plasmids used in this study were: pJC453–IFN-β promoter luciferase firefly reporter construct (kindly provided by Zhijian Chen), pRL-TK–renilla transfection control construct (Promega Inc.), pcDNA6-BVDVpro (kindly provided by Kui Li), pcDNA6-YFVpro (kindly provided by Kui Li), Human MAVS (C508Y) (previously described [52]), and pcDNA-Myc-sc-NS3 (S1165A) (previously described [33]). Amino acid sequence alignment was performed using ClustalW. This amino acid alignment was used as input for DNA sequence alignment in PAL2NAL. Aligned DNA sequences were subsequently used for all analyses done in PAML software. The well-supported primate phylogeny (Figure 1A) [26] was used for analyses done in PAML [28]. Maximum-likelihood analysis to determine whether MAVS is evolving under positive selection and to determine which residues are evolving under positive selection was done using the sites models in codeml program of PAML [28]. Sites models allow the dN/dS ratio to vary among residues. Branch models, which allow the dN/dS ratio to vary among different branches of the phylogeny, were used to determine dN/dS values for different lineages. Two-ratio tests to detect episodic selection were performed by comparing a likelihood model that allows dN/dS to vary across phylogeny to model with dN/dS fixed at 1. REL and branch-site REL analysis were performed using a web-based implementation of HyPhy package (www.datamonkey.org). Branch-site REL is based on likelihood ratio tests that identify all lineages with a proportion of sites that are evolving with dN/dS>1, without making a priori assumptions about which lineages these are in the phylogeny [29]. 5×104 HEK293 cells were plated per well in 96-well plates and transfected on the next day with LT1 (Mirus) transfection reagent. The following plasmids were co-transfected in triplicates: MAVS plasmid (2.5 ng), luciferase reporter construct pCJ453 (50 ng), transfection control pRL-TK construct (10 ng), none or one of the viral protease encoding plasmids (pJC1461—2.5 ng, pMP125—2.5 ng, pMP122—10 ng, pMP123—2.5 ng, pMP126—10 ng), and pFLAG-CMV-2 (enough to bring final DNA concentration equal to 150 ng). One triplicate set of wells was transfected without MAVS plasmid and one set without any DNA. 24 h after transfection, Dual-Glo Luciferase Assay System (Promega) was used to lyse the cells and measure luciferase firefly and renilla activity with a luminometer. Amount of firefly or renilla activity in wells transfected without any DNA was subtracted as background. Firefly activity was divided by the renilla activity in the same well to control for transfection efficiency. This firefly/renilla ratio from cells expressing MAVS was divided by the firefly/renilla ratio from wells without MAVS to calculate IFN-β fold induction, which was then averaged across the set of three wells. It was not possible to get enough lysate from a single well of a 96-well plate for immunoblot analysis. We therefore scaled the entire transfection process by 4-fold in 24-well plates. Therefore, 20×105 cells were plated per well in 96-well plates and transfected with 4 times as much LT1 transfection reagent. Concentration of all the plasmids was also increased 4-fold. 24 h after transfection, cells were resuspended in 36 µl NTE buffer (10 mM Tris—pH 8, 1 mM EDTA, 50 mM NaCl) containing 0.18 mg protease inhibitor cocktail Complete Mini (Roche). Cells were then lysed on ice for 10 min in 44 µl of NP-40 buffer (1% NP-40, 0.2% sodium-deoxycholate, 0.12 M NaCl, 20 mM Tris—pH 8) containing 0.22 mg protease inhibitor cocktail, 2.4×10−6 mol DTT, and 1.2×10−6 mol PMSF. Lysate was spun at maximum speed on tabletop centrifuge for 10 min. 75 µl of supernatant was mixed with 19 µl of 5× SDS sample buffer with 10% β-ME. Samples were boiled for 10 min and then 5–10 µl was used for Western blot analysis. Mouse monoclonal anti-FLAG M2 antibody (Sigma-Aldrich) was used to detect FLAG-tagged MAVS, and mouse monoclonal anti-HA.11 antibody (Covance) was used to detect HA-tagged HCV NS3/4A. Rabbit polyclonal anti-beta Actin antibody (Abcam) was used to detect actin. HRP-linked anti-mouse and anti-rabbit IgG antibodies (GE Healthcare) were used as secondary anti-bodies. Primary anti-bodies were used at 1∶1,000 dilution at 4° overnight and secondary anti-bodies were used at 1∶10,000 dilution for 1 h at room temperature. SuperSignal West Dura ECL substrate (Thermo Scientific) was used for detecting HRP. For co-immunoprecipitation analysis, HEK 293 cells were transfected with empty vector or pcDNA-Myc-sc-NS3 (S1165A) protease [33] that contains aa 21–32 of HCV NS4A fused to aa 1026–1206 of the HCV NS3 protease with the protease active site mutation S1165A, as well as empty vector or MAVS-expressing constructs. Cells were lysed in 1% Triton X-100, 150 mm NaCl, and 25 mM Tris-Cl pH 7.5. Coimmunoprecipitation was performed using FLAG M2-agarose beads (Sigma-Aldrich) followed by immunoblot analysis using the following antibodies: anti-myc (Abcam), anti-Flag M2-Peroxidase (Sigma-Aldrich), and anti-tubulin (Sigma-Aldrich). Huh7-K2040 is a Huh7-based human hepatocyte cell line that contains the self-replicating subgenomic replicon from HCV 1b [34]. Huh7-K2040 is transfected with control vector or plasmid expressing the indicated MAVS construct for 36 h. Cell lysates were analyzed by immunoblot with antiserum specific to HCV NS5A [53], tubulin (Sigma-Aldrich), and FLAG M2 (Sigma-Aldrich) for detection of the various MAVS constructs. HCV protein was further detected using the hyperimmune serum from an HCV-infected patient that recognizes HCV proteins NS3, NS4B, and NS5A [35]. Protein densitometry was determined using the NIH ImageJ program. To derive fold restriction, NS5A (rabbit polyclonal)/tubulin ratio for each sample was divided by the NS5A(rabbit polyclonal)/tubulin ratio from the vector alone control.

10.1371/journal.ppat.1006592

The macrophage cytoskeleton acts as a contact sensor upon interaction with Entamoeba histolytica to trigger IL-1β secretion

Entamoeba histolytica (Eh) is the causative agent of amebiasis, one of the major causes of dysentery-related morbidity worldwide. Recent studies have underlined the importance of the intercellular junction between Eh and host cells as a determinant in the pathogenesis of amebiasis. Despite the fact that direct contact and ligation between Eh surface Gal-lectin and EhCP-A5 with macrophage α5β1 integrin are absolute requirements for NLRP3 inflammasome activation and IL-1β release, many other undefined molecular events and downstream signaling occur at the interface of Eh and macrophage. In this study, we investigated the molecular events at the intercellular junction that lead to recognition of Eh through modulation of the macrophage cytoskeleton. Upon Eh contact with macrophages key cytoskeletal-associated proteins were rapidly post-translationally modified only with live Eh but not with soluble Eh proteins or fragments. Eh ligation with macrophages rapidly activated caspase-6 dependent cleavage of the cytoskeletal proteins talin, Pyk2 and paxillin and caused robust release of the pro-inflammatory cytokine, IL-1β. Macrophage cytoskeletal cleavages were dependent on Eh cysteine proteinases EhCP-A1 and EhCP-A4 but not EhCP-A5 based on pharmacological blockade of Eh enzyme inhibitors and EhCP-A5 deficient parasites. These results unravel a model where the intercellular junction between macrophages and Eh form an area of highly interacting proteins that implicate the macrophage cytoskeleton as a sensor for Eh contact that leads downstream to subsequent inflammatory immune responses.

The protozoan parasite Entamoeba histolytica can establish an enteric infection in human hosts that leads to symptoms ranging from diarrhea to abscesses in the liver and the brain. Host susceptibility to amebic infection is in part determined by the quality and potency of the host immune response that occurs once the parasite overcomes the mucus bilayers and colonic epithelial barriers, and invades underlying tissues. At the cellular level, one of the key events that shape the inflammatory response occurs during direct parasite interaction with host macrophages via surface proteins. The ensuing cascades of intracellular signaling events have only partly been uncovered. Interestingly, only direct interaction between live parasites and macrophages, as opposed to soluble factors or dead parasites, is a prerequisite to the generation of a prompt raging pro-inflammatory response. We have sought to further elucidate the mechanisms by which macrophages distinguish live parasites and found that the macrophage cell skeleton undergoes rapid significant alteration upon Eh contact. Furthermore, we uncovered a previously unknown role for two Eh enzymes in triggering macrophage pro-inflammatory responses. Through this work, we gain a better understanding of the molecular interactions that occur at the macrophage-ameba interface that regulate host inflammatory responses.

Amebiasis is caused by the protozoan parasite Entamoeba histolytica (Eh) and is estimated to affect approximately 50 million people worldwide [1]. It is a major cause of mortality and morbidity, particularly in children of developing countries. Eh carriers remain asymptomatic in most cases, but in approximately 10% of infected individuals, Eh invades the intestinal tissues and triggers symptoms such as diarrhea, dysentery and colitis [2]. Eh can also migrate to other soft tissue organs, notably the liver where it causes abscesses, and can lead to death. The reasons that lead to the development of symptomatic infection are not fully defined but host genetic makeup, quality of the immune response and Eh expression of virulence factors are thought to be contributing factors. For example, polymorphisms in the leptin gene, as well as expression of certain HLA class II alleles, have been associated with increased vulnerability to amebiasis [3–6]. At the parasite level, expression of many virulence factors by Eh have been shown to modulate the host inflammatory response as well as the ability to invade soft tissue organs outside the gut [7, 8]. Eh binding to the mucus layers and host cells is mediated by its major surface component, the Gal/GalNAc lectin (Gal-lectin) [9, 10]. This 170kDa cell surface molecule has been shown to trigger innate immune responses and is a major target of the adaptive immune response [11–16]. Through homology search of the Eh genome, at least 50 genes encoding for cysteine peptidases (CP) have been identified [17, 18]. Of these, EhCP-A1, EhCP-A2 and EhCP-A5 are the most highly expressed in axenically cultured Eh isolate HM-1:IMSS and account for the majority of cysteine proteinase activity in vitro [17, 19]. EhCP-A5 is expressed on the cell surface, EhCP-A2 localizes to the internal and external cell membrane and EhCP-A1 localizes to intracellular vesicles [20–22]. These CPs have been shown to play a role in the pathogenesis of amebiasis [20, 23–25]. Moreover, as these CPs are required for Eh life cycle and are key virulence factors, they represent attractive pharmaceutical targets [26]. The immune response is also a major determinant of the capacity of Eh to cause disease [2, 7, 8]. Upon invasion of underlying gut tissue, Eh are met by gut resident macrophages. These cells have been shown to secrete high amounts of pro-inflammatory cytokines such as TNF-α and IFN-γ and possess amebicidal activity through the release of nitric oxide [11, 27]. TNF-α secretion has been associated with increased diarrheal disease, whereas nitric oxide and IFN-γ are protective [28]. We have shown that macrophages only secrete IL-1β upon direct contact with Eh [16, 29] that was dependent on initial adhesion by Eh Gal-lectin and engagement of macrophage α5β1 integrin by Eh cysteine protease 5 (EhCP-A5) RGD sequences that activated the NLRP3 inflammasome. Thus, there is growing evidence that signaling events initiated at the intercellular junction between Eh and macrophages shape the ensuing immune response that are central in determining host susceptibility. We hypothesize that the macrophage cytoskeleton acts as a sensor that can distinguish the threat posed by invasion of live pathogens within the underlying gut tissue as opposed to soluble factors that bind to host immune cell receptors. In this study, we sought to uncover the molecular mechanisms that allow macrophage recognition of invasive Eh. We found that live Eh in contact with macrophages induces caspase-6-dependent cleavage of cytoskeletal proteins triggered by EhCP-A1 and EhCP-A4 and culminates in IL-1β secretion by macrophages. These studies demonstrate a novel macrophage cellular pathway, as well as a novel function for parasite CPs in triggering IL-1β secretion upon contact with Eh. Moreover, this study further strengthens the hypothesis that direct contact between macrophages and Eh is a key event that initiates multiple cellular mechanisms at the intercellular junction that culminates in a potent pro-inflammatory response. Recent studies [16, 29, 30] have established that direct contact between host cells and Eh mediated through the Gal-lectin and engagement of host cell integrins by EhCP-A5 at the intercellular junction is a critical initiation step in eliciting a robust pro-inflammatory response. These events require an intact actin cytoskeleton for recruitment of the NLRP3 inflammasome to contact sites as well as caspase-1 activation and IL-1β secretion [29]. Eh have been shown to distinguish between live and dead cells and adapt their capacity to interact with the host membrane, as underlined by the fact that the former triggers trogocytosis preferentially whereas the latter induces phagocytosis of host cells [31, 32]. These findings suggest that dynamic remodeling of the host cell cytoskeleton is probable upon direct contact with Eh and may have functional implications in terms of host cell, as well as parasite responses. To establish whether Eh induces cytoskeletal remodeling in macrophages upon direct contact, macrophages were incubated with Eh for 10 min, and the cytoskeletal proteins actin and tubulin were visualized by confocal microscopy. As predicted, both tubulin and actin were heavily concentrated at the macrophage-Eh contact site (Fig 1). Furthermore, microtubules were rapidly polarized towards the Eh contact site at the macrophage microtubule-organizing center. These results show that the cytoskeleton dynamically responded to Eh attachment at the macrophage intercellular junction. Cytoskeletal-associated proteins are intricately and extensively regulated by post-translational modification to regulate cytoskeletal dynamics during many cellular processes such as adhesion, cell death as well as cell-specific functions such as cytokine secretion. Previous studies have shown that Eh in contact with liver sinuisoidal endothelial cells triggers the breakdown of actin stress fibers, and relocalization of cytoskeletal-associated proteins [33, 34]. In the setting of macrophage-Eh interactions, phosphorylated paxillin and α5β1 integrin have been shown to concentrate at sites of contact between macrophages and Eh [29]. We investigated whether other macrophage cytoskeletal proteins were subjected to post-translational modifications and found that upon Eh contact with macrophages the degradation of at least three key cytoskeletal-associated proteins, talin, Pyk2 and paxillin occurred in a time- and dose-dependent manner (Fig 2A and 2B). Pyk2 is a signaling kinase mostly expressed in immune cells that triggers signaling downstream of integrins [35]. Paxillin and talin both act as scaffold proteins downstream of integrin signaling and play a key role in cytoskeletal remodeling [36, 37]. Appearance on western blot of a 53-kDa paxillin cleavage product occurred within 2 min suggesting that this is an early event during macrophage interaction with Eh. Although we originally postulated that paxillin would be phosphorylated upon contact with Eh, we were unable to assess the phosphorylation status of paxillin by western blot, as it was rapidly degraded. Interestingly, although Eh also engage integrins on the surface of T84 human colonic cells cytoskeletal cleavage was minimal, as appearance of paxillin cleavage products (~55 kDa) were only observed with high concentrations of Eh (Fig 2C and 2D). The 53-kDa and lower bands on the other hand, are similar to the main cleavage product observed in macrophages, but are only present with high concentrations of Eh (Fig 2D). As Pyk2 is not expressed at detectable levels in T84 colonic cells cleavage of its ortholog Focal Adhesion Kinase (FAK) was assessed and it was not degraded. We have previously shown that Eh engages the α5β1 integrin at the surface of macrophages through EhCP-A5 RGD sequences to promote NLRP3 inflammasome activation. Although many of the cytoskeletal proteins cleaved upon contact with Eh were found downstream of integrins, pre-treatment of macrophages with the RGD integrin α5β1 inhibitory peptide did not inhibit cleavage of paxillin using high concentrations of Eh (Fig 2E compared to Fig 2B). Furthermore, cleavage of macrophage cytoskeletal proteins was similar in WT BMM as well as in Asc-/- BMM that lack the Nlrp3 adapter ASC, and in Nlrp3-/- BMMs (Fig 2F). These results show that macrophage interaction with Eh triggers rapid and potent cleavage of cytoskeletal-associated proteins that is independent of the previously characterized EhCP-A5/α5β1 integrin interaction [29]. Paxillin has been shown to be susceptible to proteolytic cleavage by calpains and caspase-3 in the context of adhesion regulation and apoptosis [38–40]. Similarly, calpain-mediated cleavage of talin and Pyk2 is necessary for formation and disassembly of focal adhesions in osteoclasts [41]. Furthermore, both talin and paxillin have been shown to be cleaved by calpains in Jurkat T-cells incubated with Eh [42]. Accordingly, we next investigated whether cleavage of these cytoskeletal proteins were mediated by proteolytic cleavage by calpains or caspases. To do this, macrophages were pre-treatment with the pan-caspase inhibitor Z-VAD-fmk and in response to Eh, inhibited the cleavage of talin, Pyk2 and paxillin in a dose-dependent manner (Fig 3A). Through the use of a caspase inhibitor panel, we next identified the specific caspase responsible for cleavage of these proteins as caspase-6 (Fig 3B). None of the other inhibitors specific for caspases-1, -2, -3, -5, -8, and -9 inhibited Eh-induced cleavage of cytoskeletal-associated proteins. Consistent with these findings we found that caspase-6 was rapidly activated in Eh contacted macrophages, as shown by the appearance of the cleaved form of caspase-6 and degradation of one of its substrates, lamin (Fig 3C). Predictably, cleavage of macrophage cytoskeletal proteins was inhibited in a dose-dependent manner by the caspase-6-specific inhibitor Z-VEID (Fig 3D). Similarly, cleavage and activation of caspase-6 and its substrate lamin were also inhibited with increasing concentrations of Z-VEID upon macrophage stimulation with Eh (Fig 3E). To confirm specificity for caspase-6 involvement in paxillin degradation, we silenced caspase-6 by siRNA in THP-1 cells and used scramble siRNA as a control, followed by incubation with Eh for 2 min. As predicted, silencing caspase-6 inhibited the degradation of paxillin cleavage products (proteins >55 kDa) with increasing concentration of caspase-6 siRNA (Fig 3F). Cleavage of lamin was used as a positive control that was also inhibited from degradation with increasing concentration of caspase-6 siRNA as compared to the negative and scrambled siRNA controls. This data, together with the inhibition of paxillin by treatment with Z-VEID, suggests that Eh-induced cytoskeletal cleavage in macrophages was dependent on caspase-6 activation. To determine whether paxillin is a direct substrate of caspase-6, we assessed whether caspase-6 was able to cleave paxillin in vitro by incubating paxillin immunoprecipitated from THP-1 cells with recombinant human caspase-6. We found that recombinant caspase-6 was able to cleave immunoprecipitated paxillin, and that this cleavage was inhibited by treatment with Z-VEID (Fig 3G). The bottom band corresponds to mouse IgG from the immunoprecipitate, which cross-reacts with the secondary antibody used for western blot and served as a negative control for caspase-6 cleavage. A similar pattern was observed with the known caspase-6 substrate lamin (Fig 3G). To assess whether calpains also participated in the cleavage of the cytoskeletal protein paxillin, macrophages were pre-treated with increasing concentrations of E-64 that inhibits cysteine proteases as well as calpains, and it did not inhibit cleavage upon incubation with Eh (Fig 3H). We next investigated the Eh requirements for triggering cleavage of the cytoskeletal-associated proteins. Gal-lectin is the major surface adhesin expressed by Eh and is necessary for binding macrophages [10, 43]. Predictably, inhibited Eh binding to macrophages with galactose completely abrogated cleavage of the cytoskeletal proteins talin, Pyk2 and paxillin whereas glucose as an osmotic control had no inhibitory effect (Fig 4A). We next evaluated whether stimulation of macrophages with subcellular fractions of Eh was capable of triggering cleavage of the cytoskeletal-associated proteins, as to further assess and identify the Eh virulence factors responsible for this mechanism. Cleavage of talin, Pyk2 and paxillin was only observed upon stimulation of macrophages with live Eh but not with fresh lysates of whole Eh, the membrane fraction or the cytoplasmic fraction derived from equivalent numbers of Eh (Fig 4B), or with increasing concentrations of Eh excreted/secreted soluble proteins (SP) (Fig 4C). These results indicate that contact with live Eh to macrophages was critical and necessary to initiate cytoskeletal cleavage. This could be due to the necessity of the dynamic engagement of multiple receptors during contact between Eh and macrophages, as supported by our observations for EhCP-A5 and Gal-lectin [16, 29]. Similarly, the process of trogocytosis by Eh has been shown to preferentially occur upon contact with live host cells [31, 32]. To assess whether EhCP activity was necessary for cleavage of macrophage proteins, Eh were pre-treated overnight with the CP inhibitor E-64. Not surprisingly, E-64 treated Eh completely inhibited cleavage of cytoskeletal-associated proteins, indicating a role for Eh cysteine proteinase(s) in triggering caspase-6 activation and cleavage of paxillin (Fig 5A). To determine whether EhCP-A5, a cell-surface CP that has been shown to directly bind integrins on the surface of epithelial cells and macrophages [29, 30], plys a role in this process we exposed macrophages to EhCP-A5 deficient Eh and they were efficient at triggering the cleavage of paxillin that was inhibited in E64 treated EhCP-A5 parasites (Fig 5B). This mechanism thus appears to be independent of the previously identified interaction between EhCP-A5 and α5β1. Based on these results, we next set out to identify the Eh cysteine proteases responsible for activating caspase-6 and subsequent cleavage of macrophage cytoskeletal proteins. Inhibitors to EhCP-A1 and EhCP-A4 (WRR483 and WRR605, respectively) have been carefully designed and synthetized to specifically inhibit their activity both in vitro and in vivo [20, 44]. EhCP-A1, along with EhCP-A5, is unique as they are not found in non-invasive E. dispar [19]. EhCP-A1 has been shown to be highly transcribed in vitro and released by cultured trophozoites [17, 23, 45]. Inhibition of EhCP-A1 with the WRR483 inhibitor has been shown to reduce Eh invasion in human intestinal xenografts in SCID mice [20]. EhCP-A4 transcription levels, on the other hand, are low in vitro, however it has been shown to be one of the most up-regulated CPs during Eh cecal infection in mice [17, 46]. Fittingly, administration of WRR605 to Eh-infected mice has been shown to significantly reduce Eh burden and decrease inflammation [44], however the underlying molecular mechanisms remain unknown. To determine whether EhCP-A1, EhCP-A4, or both, were involved in mediating cleavage of macrophage cytoskeletal-associated proteins, we pre-treated Eh with either one of the synthetic inhibitors WRR483 and WRR605, or both, prior to incubation with macrophages. Pre-treatment of Eh with these inhibitors individually partially inhibited cleavage of talin, Pyk2 and paxillin in macrophages (Fig 5C). Whereas WRR483 did not inhibit the cleavage of any cytoskeletal protein, WRR605 partially rescued the cleavage of talin and Pyk2 and modestly paxillin. Interestingly, pre-treatment with both inhibitors had an additive effect and was similar to E64 treated Eh indicating that both EhCP-A1 and EhCP-A4 participated in cytoskeletal cleavage. To confirm the specificity of the WRR483 and WRR483 inhibitors, we treated Eh trophozoites with either WRR483, WRR605, both, or E-64 as a control with known substrates to EhCP-A1 (Z-RR) and EhCP-A4 (Z-VVR). Degradation of Z-RR occurred in a linear fashion over time (Fig 5D, top panel). Enzymatic activity in the presence of the inhibitors is displayed in the bottom panel (Fig 5D). Differences in the units in the velocity and specific activity graphs are due to differences in the fluorogenic nature of the substrates used. As expected, the WRR483 inhibitor completely inhibited the degradation of the Z-RR substrate, and to a level similar to E-64, whereas WRR605 has a minimal effect (Fig 5D). Degradation of the known EhCP-A4 substrate (Z-VVR), on the other hand, was inhibited by both WRR483 and WRR605 (Fig 5D). Together, they had an additive effect that was similar to the E-64 control. Whether this represents inhibition of EhCP-A4 by WRR483, or whether EhCP-A1 has specificity for the Z-VVR substrate remains to be determined, although the additive effect of both inhibitors suggest the later. Nevertheless, given that complete inhibition of paxillin cleavage is not seen with either inhibitor alone, but rather in an additive fashion when both inhibitors are present, suggest that both cysteine proteases are involved in triggering macrophage cytoskeletal-associated protein cleavage. The relative contribution of each cysteine protease remains to be further investigated with the use of recombinant proteins. To quantify the involvement of EhCP-A1 and EhCP-A4 upstream of caspase-6 activation, we assessed caspase-6 activation by flow cytometry using a caspase-6-FLICA probe, which detects active caspase-6. To do this THP-1 macrophages were either left untreated or exposed to Eh, or treated with various EhCP inhibitors (E-64, WRR483, WRR605, or both WRR483 and WRR605), or with staurosporine (STS) as a positive control for caspase-6 activation. Inhibition with E64 or with EhCP-A1 and EhCP-A4 in combination had the greatest effect on preventing caspase-6 activation (Fig 6A). We also assessed localization of caspase-6 with the FLICA probe by confocal microscopy. As envisaged, activated caspase-6 was strongly localized to regions within macrophages that were polarized towards Eh (Fig 6B). This polarization was disrupted in interactions with WRR483/605-treated Eh. STS was included as a positive control that shows diffuse caspase-6 activation within the macrophage cytoplasm (Fig 6B). Previous studies have shown that EhCP-A1 is contained in cytoplasmic vesicles of trophozoites, whereas EhCP-A4 localizes to intracellular vesicles, the nucleus and perinuclear endosplasmic reticulum [20, 44]. EhCP-A1 is released upon stimulation of trophozoites with mucin whereas EhCP-A4 is released upon contact of target cells [44, 47]. We sought to determine whether these cysteine proteinases are expressed on the Eh cell surface and whether they polarized to the macrophage-Eh interface. Cell-surface localization of these two amebic cysteine proteinases upon contact with macrophages was assessed by staining non-permeabilized cells. We found that in the absence of macrophages, EhCP-A1 is diffusely distributed in a punctate-like pattern on the surface of Eh. Upon contact with macrophages, however, EhCP-A1 polarized to the macrophage contact point (Fig 7). EhCP-A4 displays a similar diffuse pattern in the absence of macrophages, and similarly polarizes to the macrophage junction upon contact. Quantification of the polarization of EhCP-A1 and EhCP-A4 is included on the right-hand side (Fig 7). Positive values indicate asymmetrical staining in the axis towards the macrophages, whereas null values are indicative of a symmetrical distribution of the staining patterns throughout Eh. Therefore, EhCP-A1 and EhCP-A4 staining shows polarization towards the macrophage contact interface. As caspase-6-mediated cleavage of cytoskeletal proteins of macrophages is an early event upon contact with Eh, we next determined whether it impacted pro-inflammatory cytokine secretion. We hypothesized that caspase-6-mediated cleavage of cytoskeletal proteins may participate to the macrophage inflammatory response upon contact with Eh. To determine this, we first measured IL-1β pro-inflammatory cytokine secretion in macrophages pre-treated with the caspase-6 inhibitor, or DMSO control, for 60 min followed by incubation with Eh for 60 min. Pre-treatment of macrophages with Z-VEID-fmk significantly (p<0.0001) inhibited IL-1β release upon contact with Eh (Fig 8A). To evaluate the contribution of cell death in active secretion of IL-1β as opposed to passive release following cell death, we also measured LDH levels in media as an indicator of cell death. LDH levels corresponded to up to (35%) of LDH levels from the positive lysis control, indicating that a portion of macrophages was killed by Eh following incubation though the majority was still intact (Fig 8B). Z-VEID pre-treatment of macrophages prior to Eh incubation did not significantly alter LDH release (Fig 8B). Therefore, caspase-6 activation does not seem to significantly contribute to cell death within the first 60 min. To quantify the contribution and upstream involvement of EhCP-A1 and EhCP-A4 to IL-1β secretion, we incubated macrophages with Eh pre-treated with WRR483, WRR605, or both. IL-1β secretion by macrophages was significantly decreased when stimulated with either inhibitor (p<0.0001), and more so when used in combination (Fig 8A). LDH release was significantly inhibited by WRR483 and WRR605 (p<0.003), or both (p<0.0002) indicating these CPs may contribute to macrophage cell death, although this may be mediated through a mechanism independent of caspase-6 (Fig 8B). With the use of a highly sensitive Luminex human focused 13-plex-discovery assay we determined that no other cytokine other than IL-1β was secreted at detectable levels, or dependent on caspase-6 (Fig 8C). IL-2, IL-4, IL-6, IL-12 and IL-13 secretion was undetectable by Luminex upon stimulation of THP-1 by Eh trophozoites. IL-5, GM-CSF, IFN-γ, MCP-1 and TNF-α were secreted at very low levels (less than 20pg/ml). We further confirmed a role for caspase-6 in THP-1-mediated secretion of IL-1β upon stimulation by Eh. Consistent with results obtained with the Z-VEID inhibitor, we found that silencing of caspase-6 by siRNA decreased IL-1β secretion in a concentration-dependent manner (Fig 8D). Taken together, these data show that EhCP-A1, EhCP-A4, and caspase-6 participate in the pro-inflammatory response of macrophages to Eh. In this study, we determined the molecular events that take place at the intercellular junction between macrophages and Eh. We observed cytoskeletal rearrangement consistent with Eh engagement of cell-surface receptors, where microtubules reorient towards the target Eh, as well as actin accumulation at the contact point. This is reminiscent of many immune cell interactions such as the immune synapse and the phagocytic cup where receptor engagement, including integrins, induces microtubule polarization towards the target as well as actin accumulation at the site of contact [48, 49]. This polarization could be a means to deliver key signaling complexes at the contact site, including the NLRP3 inflammasome which has been shown to co-localize with actin at the contact site [29]. The α5β1 integrin as well as phosphorylated paxillin was also shown to localize at the contact site indicating downstream signaling occurred upon interaction [29]. We thus sought to further investigate post-translational regulation of key cytoskeletal-associated proteins upon macrophage contact with Eh. While we hypothesized that these proteins would be subjected to phosphorylation, what we observed was that these proteins were rapidly subjected to cleavage. Cleavage of talin, Pyk2 and paxillin by calpains or caspase-3 has been reported in the context of adhesion disassembly and apoptosis [40, 41, 50–52]. We cannot, however, preclude the fact that phosphorylation may precede cleavage and actually make these proteins susceptible to cleavage by inducing conformational change [38, 53]. This is supported by previous studies showing that serine phosphorylation of paxillin promotes its cleavage during cell migration [54]. Interference with the host cell cytoskeleton may be a mechanism by which Eh ingest host cells, either by trogocytosis or by phagocytosis. Trogocytosis is observed rapidly upon contact with live host cells that exhibit cell membrane deformability [31, 32]. Modification of the host cell cytoskeleton upon contact may represent the early molecular mechanisms allowing for cell deformability and subsequent ingestion of membrane portions by Eh. The process of trogocytosis, however, has not yet been reported in macrophages. Our data unveil a new role for caspase-6 in the alteration of cytoskeletal-associated proteins, as well as in the macrophage inflammatory response to Eh. We show that cleavage of talin, Pyk2 and paxillin upon contact with Eh was mediated by caspase-6 activation as inhibition of caspase-6 by Z-VEID prevented their cleavage. In the latter, caspase-6 was shown to concentrate proximally to points of contact with Eh. Although caspase-6 is classified as an effector caspase along with caspase-3 and caspase-7, there is increasing evidence that its role extends beyond apoptosis [55]. A role for caspase-6 in B-cell activation and differentiation, and more recently in macrophage activation, has been shown [56, 57]. Moreover, caspase-6 is ubiquitously expressed and the highest levels of pro-caspase-6 are observed in fetal and adult colon [58]. Therefore, further studies in the role of caspase-6 in gut inflammatory responses are warranted to assess whether this mechanism is unique to amebiasis or whether caspase-6 plays a broader role in triggering cytokine secretion and cytoskeletal remodeling. The caspase-6-mediated cleavage of cytoskeletal-associated proteins was only observed when macrophages were incubated with live Eh. Similar results were observed with NLRP3 inflammasome activation upon ligation of EhCP-A5 and Gal-lectin. Soluble EhCP-A5 or Gal-lectin alone was not sufficient to induce NLRP3 and caspase-1 activation and subsequent IL-1β release [29]. This likely reflects the importance that Eh factors to be delivered to the intercellular junction in a directed manner allowing for saturation of macrophage receptors to trigger downstream signaling cascades. We propose that this mechanism allows macrophages to distinguish between interactions of live invading Eh in comparison to sampling of soluble components of the gut lumen. The escalating pro-inflammatory response ensuing detection of live pathogens thus reflects the tailoring of the reaction against tissue-invading pathogens. The requirement for direct cell contact may also reflect the necessity of Eh to attach to host cells to release intracellular contents at the intercellular junction. In fact, EhCP-A1 and EhCP-A4, which we show are necessary for triggering cleavage of cytoskeletal-associated proteins, localize to intracellular vesicles in Eh [20, 44]. EhCP-A1 localizes to large cytoplasmic vesicles whereas EhCP-A4 is contained in acidic vesicles in Eh, yet has been shown to be released in culture medium as well as in the microenvironment in vivo. EhCP-A4 undergoes autocatalytic activation at acidic pH, provided by acidic vesicles, yet its optimal proteolytic activity is observed at physiological pH [44]. By staining the surface of non-permeabilized cells, we showed that both Eh cysteine proteinases are found at the cell surface, and polarized to the macrophage contact site. This supports a model by which these cysteine proteinases are delivered at the junction site in a defined manner, and have a role in triggering the macrophage inflammatory response. Future studies on the dynamics of EhCP-A1- and EhCP-A4-containing vesicles upon contact with host cells are necessary to uncover the mechanism of release in the microenvironment. Our data unveil new roles for both EhCP-A1 and EhCP-A4 in eliciting pro-inflammatory responses by macrophages. Interestingly, EhCP-A4 expression is significantly higher in HM-1:IMSS than in the less virulent Rahman strain, whereas EhCP-A1 and EhCP-A5 transcription levels are similar between those two strains [45]. Similarly, EhCP-A1 release in medium is significant only for HM-1:IMSS compared to E. dispar and the less virulent L6 Eh clone [23]. EhCP-A1 expression has been shown to increase almost twofold following invasion in a mouse model of amebic colitis, whereas EhCP-A5 expression was unchanged [46]. Contrary to our initial hypothesis that EhCP-A5 would have a central role in regulating key cytoskeletal-associated proteins as it binds to macrophage integrin, it was not necessary for triggering cleavage of talin, Pyk2 or paxillin. Rather, the finding of another mechanism that relies on direct contact between macrophage and Eh, and that culminates in IL-1β release further reinforces that the intercellular junction is a hot spot for molecular events that shape the subsequent pro-inflammatory response during amebiasis. We propose an integrated model (Fig 9) by which contact between live Eh and macrophages is first established by binding via the Gal-lectin at the Eh cell surface as well as EhCP-A5 binding to macrophage α5β1 integrin. We hypothesize that contact mediates the release of EhCP-A1 and EhCP-A4 from Eh in a polarized fashion at the intercellular junction. The EhCP-A5/integrin axis mediates NLRP3 inflammasome activation whereas EhCP-A1 and EhCP-A4 mediate caspase-6 activation and subsequent cytoskeletal-associated protein cleavage. Together, these pathways culminate in potent IL-1β release from macrophages. The molecular mechanisms linking some of these events remain to be addressed. For example, the requirements for EhCP-A1 and EhCP-A4 release and interaction with macrophages as well as the signaling events that culminate in caspase-6 activation remain unknown. The consequence of post-translational modification of cytoskeletal-associated proteins in remodeling of the macrophage cytoskeleton upon contact with Eh also needs to be examined. Nonetheless, this study underlines the fact that the intercellular junction is a hotspot for initiation of the inflammatory response to Eh and further investigation will unravel mechanistic pathways that will provide a better understanding of macrophage and amebic biology. E-64 was obtained from Sigma-Aldrich. Caspase inhibitors were obtained from Enzo Life Sciences. Anti-paxillin (clone 349) was purchased from BD Biosciences. Anti-Pyk2 and anti-GAPDH were acquired from EMD Millipore. Anti-talin was obtained from Santa Cruz Biotechnologies. Anti-caspase-6 and anti-lamin were acquired from Cell Signaling Technology. Anti-mouse and anti-rabbit conjugated to HRP were from Jackson ImmunoResearch. Cell Tracker Blue, Cell Tracker Orange, A647-conjugated phalloidin and A488-conjugated anti-tubulin were obtained from Life Technologies. Recombinant human caspase-6 and the Z-VVR-AMC substrate were purchased from Enzo Life Sciences. The Z-Arg-Arg-pNA.2 HCl (Z-RR) substrate was purchased from Bachem. The FAM-FLICA caspase-6 assay kit was purchased from ImmunoChemistry Technologies. Transfection was carried out with the INTERFERin (Polypus) and caspase-6 siRNA (Santa Cruz Biotechnologies) or scramble siRNA (GE Dharmacon). Antibodies to EhCP1 and EhCP4, and WRR483 and WRR605 inhibitors were a gift from Dr. Sharon Reed, University of California, San Diego). Human IL-1β secretion in cell culture was quantified by ELISA (R&D Systems). LDH levels were measured using the CytoTox-ONE homogeneous membrane integrity assay (Promega). Eh trophozoites of the highly virulent HM-1:IMSS strain were grown axenically in TYI-S-33 medium supplemented with 100U/ml penicillin and 100μg/ml streptomycin sulfate as described previously [59]. EhCP-A5-deficient trophozoites were a gift from Dr. David Mirelman (Weizmann Institute of Science, Rehovot, Israel) and identically cultured. Virulence was maintained by regular sub passage in gerbil livers as previously described [60]. Trophozoites were harvested during log-phase growth by centrifugation at 200 x g at 4°C for 5 minutes and resuspended in RPMI. The THP-1 monocytic cell line (ATCC, Manassas, VA) was cultured in RPMI with 10% FBS, 10mM HEPES, 50μM 2-mercaptoethanol, 100U/ml penicillin and 100μg/ml streptomycin sulfate in a humidified incubator with 5% CO2. THP-1 cells were differentiated into macrophages by seeding 8 x 105 cells per well in 12-well plates in complete medium supplemented with 50ng/ml phorbol-12-myristate-13-acetate (PMA) overnight. Bone marrow derived macrophages (BMM) were cultured from bone marrow cells of C57BL/6 mice and cultured for 6 days in complete medium supplemented with 30% L929-cell supernatant. Cells were then plated at 5 x 105 cells per well in 12-well plates in complete medium. On the day of experiment, BMDM were treated with 1μg/ml LPS for 3.5h prior to stimulation with Eh. The T84 epithelial cell line (ATCC, Manassas, VA) was cultured in DMEM/F-12 with 10% FBS, 10mM HEPES, 100U/ml penicillin and 100μg/ml streptomycin sulfate. For experiments, cells were seeded at 3 x 105 cells per well in 12-well plates and grown to a confluent monolayer. For stimulation with Eh, cells were incubated at 37°C with in serum-free RPMI for the indicated times. Eh were washed off with cold PBS and cells were lysed in lysis buffer (1% Triton X-100, 20mM Tris, 100mM NaCl, 1mM EDTA, 200mM orthovanadate, sodium fluoride, 0.1% SDS, PMSF, leupeptin, aprotinin, and protease inhibitor cocktail). For inhibitor treatment of macrophages, PMA-differentiated THP-1 cells were pre-treated with inhibitors for 1h at 37°C in serum-free RPMI prior to incubation with Eh. For inhibition of adhesion, Eh were incubated with 55mM D-galactose, or glucose as control, for 5 min prior to incubation with THP-1 macrophages. For irreversible inhibition of total cysteine proteinase activity of Eh, Eh were incubated overnight in medium with 100μM of E-64, as previously described [61]. For inhibition of specific EhCP activity, Eh were incubated with 20μM of the EhCP-A1 inhibitor (WRR483), 20μM of the EhCP-A4 inhibitor (WRR605), or both, for 30 min and then washed prior to incubation with THP-1 macrophages. To evaluate enzymatic specificity of the WRR483 and WRR605 inhibitors, known EhCP-A1 and EhCP-A4 substrates (Z-Val-Val-Arg and Z-Arg-Arg, respectively) were incubated for 0 to 20 minutes at 37°C with Eh trophozoites pre-treated with either WRR483, WRR605, both, or E-64 as described above. Cleavage of the Z-VVR-AMC fluorogenic substrate was detected at the 460nm wavelength, and cleavage of the chromogenic Z-RR-pNA.2 HCl substrate was detected at 405nm. For in vitro cleavage of proteins by caspase-6, THP-1 cells were lysed with lysis buffer as described above, and 10μg of lysate was incubated with anti-paxillin, anti-lamin or anti-mouse IgG for 15 min and 1.5 h with Protein A/G plus agarose. The immunoprecipitate was then incubated with 2 units of recombinant caspase-6 for 16 h at 37°C. Immunoprecipitates were then resuspended in reducing sample buffer and boiled prior to loading onto SDS-PAGE gel. THP-1 cells were differentiated overnight with PMA prior to transfection as described above. Cells were transfected with caspase-6 siRNA, or scramble siRNA as a control, with the INTERFERin reagent, as per the manufacturer’s protocol. A total of 8uL of INTERFERin reagent was used for each transfection with caspase-6 siRNA, at the indicated concentrations, or scrambled siRNA as the control. Media was replaced with complete RPMI 24 h later. Eh stimulation was performed 48 h following siRNA transfection for 2 min for analysis by Western blot, or 60 min for assessment of IL-1β secretion. Equal amounts of lysate were loaded on SDS-PAGE gel followed by transfer onto polyvinylidene difluoride (PVDF) membrane and blocking in 5% skim milk. Western blots were done using the indicated primary antibody and appropriate HRP-conjugated secondary antibody, and visualized with either SuperSignal Chemiluminescence Reagents (Pierce Biotechnology) or ChemiLucent ECL detection (EMD Millipore). When sequential Western blots were performed, PVDF membranes were incubated with stripping buffer (25mM Glycine, 1% SDS, pH 2.0) for 30 min at room temperature, followed by extensive washing and blocking with 5% skim milk. None of the antibodies used cross-reacted with proteins in Eh lysate, as verified by Western blot. For visualization of Eh contact with THP-1, THP-1 cells were PMA-differentiated and seeded on glass coverslips overnight. Eh were stained with Cell Tracker Blue according to the manufacturer’s protocol prior to incubation with THP-1 cells for 15 min at 37°C. After incubation, cells were gently washed with PBS, fixed with 4% paraformaldehyde for 10 min, permeabilized with 0.1% NP-40 in PBS followed by staining with Alexa-488-conjugated anti-tubulin, or isotype control, and with Alexa-647-conjugated phalloidin. For visualization of caspase-6 activation in THP-1 cells following stimulation with Eh, Eh were stained with Cell Tracker Orange according to the manufacturer’s protocol prior to incubation with PMA-differentiated THP-1 cells for 15 min at 37°C. Slides were fixed and permeabilized as described above. Slides were stained with the FAM-FLICA caspase-6 probe as per the manufacturer’s protocol for 60 min, and with DAPI for 15 min for visualization of nuclei. Coverslips were then washed with the provided apoptosis washing buffer and mounted onto microscope slides with ProLong antifade (Thermofisher) and visualized using the Olympus IX81 FV1000 Fluoview Laser Scanning Confocal. For staining of EhCP-A1 and EhCP-A4, THP-1 were seeded onto glass coverslips and incubated with Eh as described above. Slides were fixed in 4% paraformaldehyde for 15 min but not permeabilized to detect only extracellular surface-bound EhCPs. Following extensive washing post-fixation with 0.2% Tween 20 in PBS, slides were blocked with 5% normal donkey serum and subsequently incubated overnight with primary antibodies against EhCP-A1 and EhCP-A4 at 4°C. Slides were then washed and incubated for 60 min with the appropriate secondary antibodies (Donkey anti-Rabbit 1:250 and donkey anti-Chicken IgY 1:250) and DAPI at room temperature. Quantification of the abundance of activated caspase-6 at the THP-1/Eh interface was done in ImageJ as previously published30. Quantification of EhCP polarization was performed in ImageJ by measuring the integrated density of the front half of Eh and rear facing half, expressing this as a ratio adjusted for total area size. Figures for confocal images were created in Adobe Photoshop CS5. PMA-differentiated THP-1 cells were harvested with TrypLE (ThermoFisher Scientific) for 5 min at 37°C followed by washing with serum-containing media. 106 THP-1 cells were incubated with Eh (10:1) and stained for 60 min with the FAM-FLICA caspase-6 reagent as per the manufacturer’s protocol. One sample of cells was also treated with STS for 3 h as a positive control for caspase-6 activation. A sample with Eh cells only was used to establish exclusion gates based on forward and side scatter and FAM-FLICA caspase-6-stained Eh showed no non-specific staining of trophozoites. Samples were washed with the washing buffer provided in the FAM-FLICA caspase-6 assay kit and fixed with 4% paraformaldehyde prior to acquisition with the BD FACSCanto Flow Cytometer. Data analysis was done with FlowJo. PMA-differentiated THP-1 cells were incubated with Eh (10:1) in serum-free media for 60 min at 37°C. Supernatants were collected, centrifuged at 3000 x g for 5 min to spin out cells, and transferred to new eppendorf tubes. Quantification of IL-1β and LDH levels for each sample were measured by ELISA as per the manufacturer’s protocol. Samples were also analyzed by Luminex addressable laser bead-based immunoassay (Human Focused 13-Plex Discovery Assay; Eve Technologies, Calgary, AB). C57BL/6 mice were obtained from Charles River. Asc-/- mice, bred onto a C57BL/6 background, were obtained from Drs. Muruve and Beck (University of Calgary). Femurs and tibia from these mice were used for the growth of bone-marrow-derived macrophages (BMM). The Health Sciences Animal Care Committee from the University of Calgary, have examined the animal care and treatment protocol (AC17-0017) and approved the experimental procedures proposed and certifies with the applicant that the care and treatment of animals used was in accordance with the principles outlined in the most recent policies on the “Guide to the Care and Use of Experimental Animals” by The Canadian Council on Animal Care. Experiments are representative of at least three independent experiments unless otherwise noted. Statistical significance between groups was assessed by Student’s t-test whereby p<0.01 was considered significant. For quantification of confocal images, a minimum of 6 images was used for each condition.

10.1371/journal.ppat.1002059

Crystal Structure and Functional Analysis of the SARS-Coronavirus RNA Cap 2′-O-Methyltransferase nsp10/nsp16 Complex

Cellular and viral S-adenosylmethionine-dependent methyltransferases are involved in many regulated processes such as metabolism, detoxification, signal transduction, chromatin remodeling, nucleic acid processing, and mRNA capping. The Severe Acute Respiratory Syndrome coronavirus nsp16 protein is a S-adenosylmethionine-dependent (nucleoside-2′-O)-methyltransferase only active in the presence of its activating partner nsp10. We report the nsp10/nsp16 complex structure at 2.0 Å resolution, which shows nsp10 bound to nsp16 through a ∼930 Å2 surface area in nsp10. Functional assays identify key residues involved in nsp10/nsp16 association, and in RNA binding or catalysis, the latter likely through a SN2-like mechanism. We present two other crystal structures, the inhibitor Sinefungin bound in the S-adenosylmethionine binding pocket and the tighter complex nsp10(Y96F)/nsp16, providing the first structural insight into the regulation of RNA capping enzymes in (+)RNA viruses.

A novel coronavirus emerged in 2003 and was identified as the etiological agent of the deadly disease called Severe Acute Respiratory Syndrome. This coronavirus replicates and transcribes its giant genome using sixteen non-structural proteins (nsp1-16). Viral RNAs are capped to ensure stability, efficient translation, and evading the innate immunity system of the host cell. The nsp16 protein is a RNA cap modifying enzyme only active in the presence of its activating partner nsp10. We have crystallized the nsp10/16 complex and report its crystal structure at atomic resolution. Nsp10 binds to nsp16 through a ∼930 Å2 activation surface area in nsp10, and the resulting complex exhibits RNA cap (nucleoside-2′-O)-methyltransferase activity. We have performed mutational and functional assays to identify key residues involved in catalysis and/or in RNA binding, and in the association of nsp10 to nsp16. We present two additional crystal structures, that of the known inhibitor Sinefungin bound in the SAM binding pocket, and that of a tighter complex made of the mutant nsp10(Y96F) bound to nsp16. Our study provides a basis for antiviral drug design as well as the first structural insight into the regulation of RNA capping enzymes in (+)RNA viruses.

Most eukaryotic cellular and viral mRNAs are modified by the addition of a polyadenine tail at the 3′- terminal and a cap structure at the 5′-terminal. The RNA cap protects mRNA from degradation by 5′ exoribonucleases, ensures efficient mRNA translation, and prevents recognition of viral RNA via innate immunity mechanisms[1], [2], [3], [4]. The RNA cap is made of an N7-methylated guanine nucleotide connected through a 5′-5′ triphosphate bridge to the first transcribed nucleotide, generally an adenine. Through 2′-O methylation of the latter, this cap-0 structure (7MeGpppA…) may be converted into a cap-1 structure (7MeGpppA2′-O-Me…). In the eukaryotic cell, the cap is added co-transcriptionally in the nucleus by three sequential enzymatic reactions[1], [5]: (i) an RNA triphosphatase (RTPase) removes the 5′ γ-phosphate group of the nascent mRNA; (ii) a guanylyltransferase (GTase), dubbed capping enzyme, catalyses the attachment of GMP to the 5′-diphosphate mRNA; and (iii) an S-adenosylmethionine (SAM)-dependent (N7-guanine)-methyltransferase (N7MTase) methylates the cap onto the N7-guanine, releasing S-adenosylhomocysteine (SAH). In general, a SAM-dependent (nucleoside-2′-O-)-methyltransferase (2′-O-MTase) further intervenes, in higher eukaryotes, to yield a cap-1 structure. The viral RNA capping machinery is structurally and mechanistically diverse, and RNA viruses often deviate from the paradigmic eukaryotic mRNA capping scheme. For example, alphaviruses methylate GTP onto the N7-guanine before the presumed attachment of 7MeGMP to the nascent viral 5′-diphosphate mRNA[6]. In the case of single-stranded negative-sense (-)RNA viruses, such as the vesicular stomatitis virus, the L polymerase attaches GDP rather than GMP to a nascent viral 5′-monophosphate mRNA, covalently linked to the viral capping enzyme[7]. Other viruses, such as influenza virus capture a short capped RNA oligonucleotide from host cell mRNAs and use it as an RNA synthesis primer. This process is known as « cap snatching »[8]. In 2003, a novel coronavirus named Severe Acute Respiratory Syndrome coronavirus (SARS-CoV[9]) was responsible for the first viral pandemic of the new millennium with ∼8000 cases globally and a 10 % case-fatality rate. Coronaviruses encode an unusually large membrane-associated RNA replication/transcription machinery comprising at least sixteen proteins (nsp1-to-16)[10]. For SARS-CoV, the RNA cap structure likely corresponds to a cap-1 type[11], [12], [13]. As in many other (+)RNA viruses, the RTPase activity is presumably embedded in the RNA helicase nsp13, whereas the GTase remains elusive. RNA cap 2′-O-MTase activity was first discovered in the feline coronavirus (FCoV) nsp16[14]. Shortly after, SARS-CoV nsp14 was shown to methylate RNA caps in their N7-guanine position[15]. Curiously, although closely homologous to that of FCoV, recombinant SARS-CoV nsp16 alone was devoid of enzymatic activity. It was demonstrated[16], [17], [18], [19] that nsp10 interacts with nsp16, conferring 2′-O-MTase activity to nsp16 on N7-methyl guanine RNA caps selectively[16]. The latter selectivity implies that RNA cap methylation obeys an ordered sequence of events during which nsp14-mediated N7-guanine methylation precedes nsp10/nsp16 RNA 2′-O methylation. Nsp10 is a double zinc finger protein of 148 residues whose crystal structure is known[20], [21]. Together with nsp4, nsp5, nsp12, nsp14, and nsp16, nsp10 has been found to be essential in the assembly of a functional replication/transcription complex[22]. Drawing on these observations, nsp10 has been proposed to play pleiotropic roles in viral RNA synthesis[23] and polyprotein processing through interaction with the main protease nsp5[24]. SAM-dependent MTases belong to a large class of enzymes present in all life forms. These enzymes catalyze the transfer of the SAM methyl group to a wide spectrum of methyl acceptors, indicating that a common chemical reaction is used on a variable active-site environment able to activate the methyl acceptor atom. Although SAM-dependent MTases share little sequence identity, 2′O-MTases exhibit a KDKE catalytic tetrad and a very conserved folding made of a seven-stranded β-sheet surrounded by one to three helices on each side[25], always similar to the paradigmatic catechol-O-MTase[26]. The SAM binding site general location is conserved, suggesting that evolutionary pressure on the MTase fold has maintained the same SAM-binding region whilst accommodating the versatile chemistry of the methyltransfer reaction. Structural and functional studies of viral MTases involved in RNA capping is an expanding research area, since these enzymes show unexpected diversity relative to their cellular counterparts, and thus constitute attractive antiviral targets. Crystal structures of viral RNA cap MTases exist for only three viral families, namely Poxviridae, Reoviridae, and Flaviviridae. The Vaccinia virus VP39 crystal structure was the first to be elucidated in 1996[27]. The structure of this DNA virus RNA 2′-O-MTase revealed a conserved MTase fold similar to that of RrmJ (also named FtsJ), the canonical reference folding for RNA cap MTases[26]. More recently, the crystal structure of a second Vaccinia virus N7-guanine RNA cap MTase domain (D1) was determined in complex with its activator protein D12[28]. The study revealed that D12 also bears an MTase fold, but has lost catalytic capability due to truncation of its SAM binding site. In turn, Reoviridae provided the first RNA cap MTase structures at 3.6 Å resolution as forming part of the reovirus core[29]. Another RNA cap machinery was more recently described for the non-turreted orbivirus Bluetongue virus VP4 protein at 2.5 Å resolution[30], which revealed a three-domain protein, with a “head” guanylyltransferase domain, a central N7-guanine MTase, and a “bottom” 2′-O-MTase domain. This architecture illustrates the sequence of three out of the four chemical reactions involved in RNA capping described above. Regarding (+)RNA viruses, MTase structural information at the atomic level is only available for a single genus. The flavivirus N-terminus domain (residues 1–265) of the NS5 RNA-dependent RNA polymerase harbors an RrmJ fold with an N-terminus extension able to accommodate RNA cap structures[31], [32]. This enzyme carries both N7-guanine MTase and 2′-O-MTase activities on a single domain with one shared active site[33]. Homologous domains have been crystallized for a number of flaviviruses, revealing a conserved fold and activity[34], suggesting that MTases might represent interesting targets for drug design. No other (+)RNA virus RNA cap MTase crystal structures have as yet been defined. In 2003, the identification of the 2′-O-MTase signature sequence in the SARS-CoV genome added nsp16 to the list of putative targets for antiviral drugs[35]. Several compounds have been shown to inhibit viral MTases, such as the co-product of the MTase reaction SAH, Sinefungin, and aurintricarboxylic acid (ATA)[14], [36], [37], [38], [39]. In this paper, we report the crystal structure of the SARS-CoV 2′-O-MTase nsp16 in complex with its activator, the zinc finger protein nsp10, at 2.0 Å resolution, in conjunction with mutagenesis experiments, binding and activity assays. These results lay down the structural basis for the nsp10 function as an activator of nsp16-mediated 2′-O-MTase. We identify residues playing key roles in the nsp10/nsp16 interaction, as well as other residues involved in 2′-O-MTase catalysis and RNA binding. We also report the crystal structure of the nsp10/nsp16 complex bound to the inhibitor Sinefungin. Comparison with known cellular SAM binding sites points to the nsp16 nucleobase binding pocket as a possible target for the design of selective antiviral molecules. We observed that purified nsp16 was unstable in solution, impeding crystallogenesis. Yeast double-hybrid and co-immunoprecipitation experiments on purified SARS-CoV nsp10 and nsp16 have uncovered the reciprocal interaction of these two proteins[16], [17], [18]. Indeed, SARS-CoV nsp16 exhibits 2′-O-MTase activity only when complemented with SARS-CoV nsp10, raising the interesting possibility that nsp10 acted as a scaffold for nsp16. Co-expression of nsp10 and nsp16 using a bi-cistronic prokaryotic expression vector facilitated affinity chromatography purification and crystallization of the complex[16], [40]. Crystals diffracted to ∼1.9 Å. The position of the nsp10 protein was determined using molecular replacement with the SARS-CoV nsp10 protein structure[20] as a search model. Strong peaks in both the residual and anomalous Fourier maps confirmed the presence of two zinc ions. Nsp16 was well defined by its electron density except for two flexible loops (residues 19–35 and 135–137) with high B factors and weak or missing electron density. These loops are solvent-exposed at each side of the putative RNA-binding groove (see below). Structure determination data and refinement statistics are reported in Table 1. The heterodimer can be conveniently viewed as nsp16 sitting on top of a nsp10 monomer (Fig. 1A). The nsp10 overall structure in the complex remains essentially unchanged relative to published structures of nsp10 alone, with its N-terminus comprising two α-helices, a central β-sheet domain, and a C-terminus domain containing various loops and helices (see[20], [21], Fig. 1B). Comparison with existing crystal structures of nsp10 using DaliLite[41] rendered nsp10 atomic coordinates very similar to those of nsp10 in our nsp10/nsp16 complex. The average RMSD is about 0.77 Å in 118 residues (PDB codes 2FYG, 2G9T and 2GA6[20], [21]). This indicates that neither significant conformational change nor surface modification occurs in nsp10 when binding to nsp16. The nsp10 structural Zn2+ ions are not directly involved in the nsp10/nsp16 interface (Fig. 1A). Nsp16 adopts a canonical SAM-MT fold (Figs. 1B, 2A and B), as defined initially for the catechol O-MTase[25]. The seven-stranded β-sheet MTase fold has been described as having a secondary structure topology defining two binding domains, one for SAM and the other for the methyl acceptor substrate (Fig. 2A). The nsp16 topology matches those of dengue virus NS5 N-terminal domain and of vaccinia virus VP39 MTases[27], [31]. Nsp16 lacks several elements of the canonical MTase fold, such as helices B and C (Fig. 2B). Electron density corresponding to one molecule of S-adenosylhomocysteine (SAH), the co-product of the methylation reaction, was identified in the putative SAM-binding site (Figs. 1A and 3A). Neither SAM nor SAH was added to the purification or crystallization buffers, therefore it must have been captured from the medium by nsp16 during bacterial growth. The SAH molecule is found with its adenine in an anti conformation and the ribose pucker in a southern (2′-endo/3′-exo) conformation. All the residues involved in SAM/SAH binding are absolutely conserved in coronavirus np16s (Fig. S1). Binding specificity for SAM/SAH is achieved by holding distal SAM/SAH carboxylic and amino groups through five hydrogen bonds (G81, N43, Y47, G71, and D130) (Fig. 3A and Fig. S2A). The ribose moiety is held by three hydrogen bonds involving Y132, G73, and D99. As in the case of other MTases[25], the SAH binding cleft is globally positively charged. However, an aspartic acid (D99) acts as the ribose-sensing residue with its side chain carboxyl making strong hydrogen bonds with both ribose hydroxyls (Fig. S2A). Binding of the adenine base involves few contacts. The nucleobase occupies a loose hydrophobic pocket engaging two hydrogen bonds of moderate strength with side chain and main chain atoms of conserved residues D114 and C115, respectively. Soaking the crystals into a Sinefungin-containing buffer captured this MTase inhibitor in the SAH binding site almost perfectly superimposable on SAH (Fig. 3B and Fig. S2B). Binding involved the same residues and contacts as SAH. Inhibition of the MTase reaction by Sinefungin therefore probably occurs competitively. The Sinefungin amino group quasi-isosteric to the donated SAM methyl group indicates a cavity where the 2′-hydroxyl of the capped RNA is expected to bind. Lining this empty substrate cavity are the residues proposed to be involved in the catalytic reaction: K46, D130, K170, and E203[16]. Alanine substitutions in the catalytic tetrad (K46, D130, K170, or E203) almost completely block 2′-O-MTase activity without jeopardizing binding to nsp10 (Table 2, and [16]). Several SAM-binding residues (N43, G73, D99 and Y132, Fig. S2A) were substituted by alanine. Although they conserve their specific nsp10 binding properties, indicating that they are correctly folded, they all show a drastically reduced MTase activity (Table 2), validating the structural description of the nsp10/nsp16/SAH ternary complex. We recently reported[16] that nsp10/nsp16 MTase activity requires Mg2+. Although the crystallization buffer contains Mg2+, we were unable to locate any such cation in the nsp16 active site. In enzyme activity assays, the Mg2+ ion can be substituted by Mn2+ or Ca2+, but not Zn2+ (data not shown, see also[16]). A peak of electron density presumably corresponding to Mg2+ is localized onto nsp16, distant from the SAH-binding cavity. The Mg2+ coordination mode is through six first-shell water molecules in an octahedral geometry (Fig. 4). Binding via water molecules, involves T58 and S188 side chain hydroxyls and the main chain carbonyl of E276. Since there are no carboxylic acids involved in binding this cation, it was suspected that its presence resulted from the crystallization procedure[42], with no biological relevance. However, the T58A, T58N, T58E and S188A substitutions show 43, 70, 99 and 72% loss of activity, respectively (Table 2), with no significant effect on the stability of the nsp10/nsp16 complex except for T58E whose association was 54 % that of wild-type. These residues are located on three distinct structural elements at the C-terminus of helix Z (T58), the N-terminus of β6 (S188), and in the central part of helix A3 (E276), respectively (Fig. 4 and Fig. S3). The cation may thus hold these elements together. The nsp10/nsp16 complex absolutely requires an N7-methyl guanine capped RNA substrate to exhibit MTase activity[16]. The structural basis for the preferential binding to methylated N7-guanine versus non-methylated caps has been elucidated in four cases, those of VP39[27], eIF4E[43], CBC[44], and PB2[45] proteins (PDB codes 1AV6, 1EJ1, 1H2T, and 2VQZ, respectively) bound to cap analogues or capped RNAs. In these cases, the methylated base specificity is achieved through increased binding energy resulting from the stacking of the N7-methyl guanine between parallel aromatic residues of the cap binding protein. The presence of the methyl group greatly enhances π-π stacking, providing a dominant effect over unmethylated guanine[46]. Despite numerous attempts, cap analogues (m7GpppA, GpppA, m7GpppG, GpppG) and short capped RNA substrates (m7GpppA(C)n) could neither be co-crystallized with nsp10/nsp16 nor soaked and bound onto preformed nsp10/nsp16 crystals. However, the atomic coordinates of the N7-methyl guanine RNA oligomer in complex with VP39[47] provided data from which a model of RNA binding to the nsp16 protein was derived. SAM molecules identified in both structures were superimposed, and the VP39-bound RNA was positioned onto the nsp16 structure. After minimal manual adjustments not exceeding 5 Å, the VP39 RNA was a reasonably good fit into an nsp16 hydrophobic groove radiating from the catalytic site (Fig. 5), establishing very few contacts with nsp10. We note that the protein side diametrically opposite to the proposed hydrophobic RNA binding groove is highly positively charged (not shown), an observation that may account for the difficulty of achieving experimental RNA binding in the proposed RNA binding site. In the absence of robust data to guide docking of the guanine cap, the m7Gpp cap structure was not positioned in the structure but two possible N7-methylated cap guanine binding areas are indicated by arrows (Fig. 5). The first transcribed nucleotide together with its ribose receiving the methyl group fit well in the active site (Fig. 5, panel B) as predicted in the proposed mechanism. The same holds for the immediately preceding three nucleotides. The base of the first transcribed nucleotide may be held by contact with P134 and Y132, bending the extending RNA cap structure. Accordingly, the substitution of Y132 greatly depresses MTase activity (Table 2). We also note that Y132 is located in the vicinity of a highly mobile loop (residues 135-138) not always visible in our crystal structures suggesting that this loop may move in order to wrap the triphosphate moiety of the RNA cap and/or the RNA cap itself. The solvent exposed side chain of Y30 may also participate in RNA binding. In the model, the highly mobile side chain of Y30 was flipped out in an alternative conformation in order to open the groove. In that position, Y30 should specifically contact the third transcribed nucleotide. Our mutagenesis data confirms the importance of Y30 since its replacement with either Ala or Phe severely impairs MTase activity without affecting the interaction with nsp10 (Table 2). All nsp10 secondary structure elements but helices 2 and 5 contact nsp16 (Fig. 1). The nsp10 contact points can be viewed as 5 small patches A to E (residues 40–47, 57–59, 69–72, 77–80, and 93–96, respectively, Fig. 6A). In turn, these five patches contact most of the nsp16 SAM-binding structural elements in 4 areas, I to IV (Fig. 6B, Fig. S3)), mainly involving β2, β3, αA, αZ, and for area IV, the loop connecting helices A2 and A3 at the C-terminus (Figs. 1 and 6). In total, the interface of the heterodimer involves 53 residues, 23 and 30 from nsp10 and nsp16, respectively. In nsp10, a single residue (Asn10, at the edge of the interaction surface) is not conserved out of 23 (4.3 %), whereas in nsp16 there are 8 non-conserved residues out of 30 (26.7 %) (Fig. S1). The interface has a buried surface area of 1820 Å2, with nsp10 contributing to 930 Å2 and nsp16 to 890 Å2. Four nsp10 patches included in the 5 interaction patches identified here were recently mapped using reverse yeast two-hybrid methods coupled to bioluminescence resonance energy transfer and in vitro pull-down assays (see[18] and below). To probe the observed crystal structure of the interface further, we engineered 5 new nsp10 alanine mutants (N40A, L45A, T58A, G69A, and H80A, see Table S1) sitting in patches A, B, C and D. Whereas T58A, G69A and H80A showed limited effect on nsp16 binding, N40A reduced it to 64% of wild type affinity, and L45A almost abrogated it. The crystal structure indicates that the nsp10 interface proposed by Lugari et al.[18], is a correct and conservative estimation, as the interface also includes L45 belonging to patch A. We also confirm the positive co-relation of the detected nsp10/nsp16 interaction with MTase activity. In no instance can nsp16 be active in the absence of nsp10/16 complex formation. The Y96 position is of particular interest. Alanine substitution (Y96A) abrogates interaction whereas a phenylalanine (Y96F) increases both interaction and MTase activity[18]. In order to understand how residue 96 plays such a pivotal role, we determined at 2.0 Å resolution the crystal structure of this nsp10(Y96F)/nsp16 complex (Table 1). Strikingly, the absence of the hydroxyl group does not alter the topology of the interface. Wild-type and Y96F residues superimpose without significant difference at all atomic positions (not shown). Either Y96 or F96 is in direct contact with nsp16 helix αZ, which carries the catalytic residue K46. Detailed surface analysis using PISA indicates that the position of K46 in Y96F nsp10 is identical to that of K46 in wild-type nsp10, ruling out a better alignment of catalytic residues of the Y96F mutant. The nsp10(Y96F)/nsp16 differs from wild-type nsp10/nsp16 in the SAH binding site, though. When compared to that of wild-type, the SAH occupancy is much lower (∼0.3 versus ∼1), leading to poor density definition. SAH is known to be a fairly good inhibitor of the methylation reaction. Therefore, a lower binding affinity might translate into less end-product inhibition, and account for the observed increased activity. We measured the affinity of nsp16 for SAH using fluorescence spectroscopy, but no significant differences were found (not shown). Likewise, the MTase inhibition pattern by SAH was identical for wild-type and nsp10(Y96F)/nsp16 (not shown). We therefore infer that the previously observed ∼10-fold increased stability of the heterodimer[18] may be responsible for the increased activity. A more hydrophobic character of the interaction may appear upon the loss of the tyrosine hydroxyl which, in the wild-type protein, was not engaged in any polar contact. We thus attribute the increased activity of the nsp10(Y96F)/nsp16 complex relative to wild-type to a stronger equilibrium association of nsp10(Y96F) with nsp16 than that of wild-type nsp10 with nsp16. Mutation analysis was also conducted on nsp16 residues presumably involved in the interface and interfacial activation (Tables 2 and S2). Several mutants (V78A, V104A, L244A, M247A) in patches II, III and IV completely disrupt the nsp10/nsp16 complex and annihilate nsp16 MTase activity. Interestingly, we also identified nsp16 mutants still interacting with nsp10, but with a strongly reduced 2′-O-MTase activity (I40A, M41A, V44A, T48A, Q87A, D106A) suggesting, that these mutations in the nsp10/nsp16 interface may alter the fine positioning of catalytic residues without any significant effect on nsp10 binding. Accordingly, most of these mutants are localized in αZ helix of patch I which contains the K46 catalytic residue. On the other hand, patch II and III mutants tend to have more mitigated phenotypes, yielding to full-blown interaction with only about half of the expected activity. Finally, patch IV mutants were totally inactive. Using all mutants reported in Table 2, a plot (Fig. 6C) of interaction versus activity shows that the nsp10/nsp16 interaction is strictly required to obtain significant nsp16 MTase activity. The SARS-CoV RNA cap 2′-O-MTase is a heterodimer comprising SARS-CoV nsp10 and nsp16. When bound to nsp10, nsp16 is active as a type-0 RNA cap-dependent 2′-O-MTase, ie., active only when the cap guanine is methylated at its N7 position[16]. The nsp10/nsp16 crystal structure shows that nsp16 adopts a typical fold of the S-adenosylmethionine-dependent methyltransferase family as defined initially for the catechol O-MTase[25]. A good alignment (170°) is found between the SAH sulfur atom, a water molecule present in both SAH- and sinefungin-bound nsp16 structures, and the K46 ε-amino group (Fig. 3A and B). This geometry provides interesting hints for a catalytic mechanism, as the positions of the catalytic residues (K46, D130, K170, E203) match spatially those of the vaccinia virus VP39 2′-O-MTase[47]. At the initial stage of the reaction the 2′-hydroxyl of the capped RNA substrate would occupy the position of the water molecule. In turn, E203 and K170 decrease the pKa of the K46 ε-amino group that becomes a deprotonated general base (-NH2) able to activate the RNA 2′-hydroxyl at neutral pH. In VP39, K175 has been identified as the general base catalyst[47] with a pKa depressed by ∼ 2 pH units by the neighbouring D138 and R209 residues[48]. These findings indeed suggest a related mechanism: once K46 has activated the 2′-hydroxyl group, the 2′-oxygen would produce an in line attack through a SN2-like mechanism onto the electrophilic SAM methyl group. The methyl group would pass through a pentavalent intermediate with the 2′-O and sulfur at apical positions. D130 is positioned to stabilize the transient positive charge on the donated methyl atom of SAM before the sulfur recovers a neutral electric charge during SAH generation (Fig. 3B). Unlike most SAM-dependent MTases, the SARS-CoV nsp10/nsp16 enzyme requires a divalent cation, either magnesium, manganese or calcium[25]. We have found that this cation does not reside in the active site. Instead, the cation is coordinated through water molecules by three residues located on three distinct structural elements. It is thus possible that one divalent cation, presumably Mg2+, present in the host cell at millimolar levels, plays a structural role in holding these three nsp16 structural elements together and so regulate the enzyme activity. It is intriguing that T58A is more active than T58N or T58E that can still bind the water that chelates to the metal. Alternatively, it is possible that divalent cations such as Mg2+ or Ca2+ act as a phosphodiester charge shield to allow RNA binding in the hydrophobic binding groove[49]. The main regulation mechanism of nsp16 is through its physical association with nsp10. Nsp16 is unstable in solution, and nsp10 acts as a scaffold for nsp16, yielding a stable dimer active as an RNA cap-dependent (nucleoside-2′-O)-MTase. The complex is assembled through a ∼890 Å2 contact surface in nsp16, an area typically in the intermediate zone differentiating strongly from weakly associated dimers[50]. This finding is consistent with a Kd estimated at ∼0.8 µM[18] that qualifies the nsp10/nsp16 complex as a rather weak heterodimer. The nsp10 interaction surface identified in the crystal structure was confirmed by site-directed mutagenesis and overlaps that previously identified by indirect methods[18]. Remarkably, the nsp10 surface in the nsp10/nsp16 complex is essentially identical to that of uncomplexed nsp10 crystallized alone by others[20], [21](Fig. S4). It is therefore reasonable to see this heterodimer as a non-permanent species which would tolerate nsp10 or nsp16 engaging in interactions with other partners. This notion is actually in line with the involvement of nsp10 in a network of protein-protein interactions that we and others have proposed[17], [19]. Donaldson et al.[23] have engineered mutations in nsp10 using reverse genetics. Out of eight mutations that turned out to be in the nsp10/nsp16 interface (this work), five, two and one rendered lethal, debilitated, and viable phenotypes, respectively[23]. Interestingly, the nsp10(Q65E) mutant providing a temperature-sensitive phenotype[22], [23] does not map in the nsp10/nsp16 interface, confirming that nsp10 has a pleïotropic role. Our mutagenesis analysis shows that the formation of an nsp10/nsp16 complex is a pre-requisite for MTase activity (Fig. 6C) indicating that physical association of nsp10 and nsp16 is essential to activate nsp16 2′-O-MTase activity and foster efficient virus replication. We note that most interface mutants exhibit a severe loss of their 2′-O-MTase activity, whereas the apparent association affinity is often only modestly affected. That minor changes in the interface translate into potent effects is also dramatically illustrated by the Y96F mutation, where the loss of a single hydroxyl provokes a significant change in affinity[18]. Remarkably, it is not the most active complex that was selected in nature, since the nsp10(Y96F)/nsp16 complex is both more stable and more active than the wild-type heterodimer (this work and[18]). This is yet another observation hinting at the involvement of nsp10 in protein-protein interaction networks including other partners than nsp16, such as nsp5 and nsp14[17], [19]. In most other coronaviruses, the nsp10 residue at position 96 is a phenylalanine. It would be interesting to determine whether this polymorphism is relevant to the SARS-CoV pathogenicity at any (direct or indirect) level, or if compensating polymorphisms in other coronaviral nsp10 (or nsp16) restore a weaker nsp10/nsp16 association equivalent to that of the SARS-CoV pair. Since a bona fide viral RNA cap is key in evading the host cell innate immunity[4], [51], a minimal level of 2′-O-MTase activity would be expected to be critical to virus survival. MTase activation through dimerisation of two viral protein partners has already been reported in the case of the vaccinia virus D1/D12 N7-guanine MTase[28]. However, the activating D12 subunit does not contact the D1 subunit through a homologous surface mainly defined by canonical αA and αZ helices. Rather, the D1/D12 activation surface would be located at a 90° clockwise rotation relative to the nsp10/nsp16 interface depicted in Fig. 1A. In the case of dengue virus, the bi-functional N7-guanine and 2′-O-MTase is part of the N-terminus of the dengue NS5 protein. Based on reverse genetic data and modeling[52], the MTase domain would be associated with the Pol domain through an interface topologically similar to that of nsp10/nsp16, i.e., involving mainly helices αA, αZ and strands β2 and β3 as depicted in Fig. 1A. We have previously shown that the nsp10/nsp16 is only active as N7-guanine methylated capped RNA, implying that RNA cap methylation obeys to an ordered sequence of events where nsp14-mediated N7-guanine methylation precedes nsp10/nsp16 RNA 2′-O methylation[16]. In the absence of data regarding the RNA substrate, we built a model of RNA binding based on that of the vaccinia virus VP39 ternary complex structure. Interestingly, our model proposes that the RNA interacts only with nsp16 residues, in keeping with what was recently suggested based on RNA binding assays[14]. Although the position of the cap structure on the nsp16 surface remains to be determined, our model suggests a well-defined position for the ribose of the first transcribed nucleotide in the active site. In agreement with mutagenesis analysis, the model also suggests that the transcribed RNA 5′-end stacks between Y132 and Y30. Furthermore, this model is consistent with the observation that coronavirus MTase requires RNA substrates of at least 3 transcribed nucleotides in length[14]. It is also worth to know that a comparison of nsp16 and VP39 electrostatic surfaces reveals that the putative RNA-binding groove of nsp16 is mostly hydrophobic, whereas the VP39 RNA-binding groove is positively charged. This variation would imply a change in the nature of the RNA/protein interaction. Viral MTases are increasingly evaluated as potential drug design targets[34], [37], [53]. We have crystallized the inhibitor Sinefungin with the nsp10/nsp16 complex. Sinefungin exhibits an IC50 of 0.74 µM, 16-fold lower than that of SAH as reported by Bouvet et al.[16] using purified nsp10/nsp16. Analysis of the structure suggests a likely mechanism of action that also accounts for the observed inhibitory effect of this drug. We note that the adenine nucleobase does not fit snugly into its binding pocket, raising interest regarding structure-based drug design. Preliminary examination of eukaryotic non-viral MTase structures from main classes as defined in Martin and McMillan[25] indicates that the SAH adenine is bound tighter in any of the latter enzymes than in the nsp16 SAM-binding site, indicating a possible breach to achieve anti-coronavirus selectivity with a small molecule inhibitor of nsp16. In conclusion, the crystal structures presented here extend our general understanding of the mechanism and regulation of viral RNA cap MTases in (+)RNA viruses, and point to both the nsp10/nsp16 interface and the substrate binding sites as putative antiviral targets. Both nsp10 and nsp16 were expressed from the same dual expression vector pmCOX [16]. Nsp10 had a N-terminal strep-tag (WSHPQFEK), and nsp16 a N-terminal hexa-histidine tag. The purification and crystallogenesis of the nsp10/nsp16 complex was performed as described in [40]. Typical crystals of the wild-type nsp10/nsp16 appear in hanging drops after 24 h at 20°C in 0.1 M CHES pH 9, 1.52 M MgCl2 hexahydrate. Crystals (a = 68.53 Å, b = 184.74 Å, c = 129.01 Å, C2221) contain one nsp10/nsp16 complex per asymmetric unit, with a solvent content of 70 % and Vm of 4.17 Å3/Da. Crystals of nsp10(Y96F)/nsp16 were grown in 67 mM CHES pH 8.5, 0.99 M MgCl2 hexahydrate, 33 mM Tris-HCl, 8.3 % PEG 8000. Both crystallization conditions yielded crystals diffracting to 1.9 Å when exposed to synchroton radiation at the ID14-1 beamline of the European Synchrotron Radiation Facility, Grenoble, France. Crystals were cryo-cooled in the same buffer supplemented with 15 % glycerol. Crystal soaking was performed in the same buffer supplemented with 5 mM SAH or Sinefungin during 24 h. The position of the nsp10 protein was unambiguously determined by molecular replacement using the program PHASER[54] with the nsp10 protein (2FYG), as search probe[20]. Strong peaks in both the residual and anomalous Fourier maps confirmed the presence of two Zinc ions at the expected positions within the nsp10 protein, thus giving confidence in the validity of the MR solution. Phases calculated from this partial model were combined with SAD phases from the Zn atoms using PHASER. To ameliorate the resulting low quality density map, phases were improved with PARROT[55]. An initial model, comprising both nsp10 and nsp16, was automatic built by successive use of BUCCANEER[56] and ARP/wARP[57]. The resulting model was subject to several cycles of manual rebuilding using COOT [58] and refinement with REFMAC [59]. The protein structure model could be built, except the strep and hexahistidine tags. In nsp16, density was too weak for the mobile, solvent exposed nsp16 loop 136–139, Y30 (see “Results”), and 2 and 6 residues in N- and C-terminus, respectively. Likewise, nsp10 solvent exposed 9 and 8 residues in N- and C-terminus were missing, respectively. Overall, the chain traces are unambiguous, with clear electron density including for a single SAH residue bound to the nsp16 protein. Solvent accessible surfaces were calculated using program AREAIMOL [60] with a 1.7 Å radius sphere as the probe (Table 1) and values rounded to the nearest 5 Å2. Conformational differences were analyzed using the DynDom server (http://www.cmp.uea.ac.uk/dyndom/main.jsp). Figures were created using PYMOL (http://www.pymol.org). The coordinates of the wild-type/SAH, mutant, and wild-type/Sinefungin structures have been deposited at the Protein Data bank under PDB codes 2XYQ, 2XYV, and 2XYR, respectively. The modeling of the RNA cap structure in the nsp10/nsp16 complex structure is derived from the analysis of the structure of the vaccinia virus methyltransferase VP39 crystallized in complex with a capped RNA and a S-Adenosylhomocysteine[47] (SAH) (pdb code: 1AV6). The two structures are manually aligned using COOT[58] based on the position of SAH binding sites, as well as SAH, and Sinefungin (SFG) molecules. The RNA binding site of VP39 is only partly overlapping that of nsp16 whilst the shape of the cavity is similar; thus local adjustments necessary to accommodate the RNA molecule in its binding groove were done manually using COOT. The side chain of tyrosine 30 of nsp16initially pointed to the putative RNA binding site, preventing any bona fide modeling. In order to fit the RNA molecule in the cavity, an alternative conformation was sought for this side chain. The second most common conformation for the tyrosine side chain was selected. Due to the biochemistry data and surface electrostatic analysis, it is not possible to describe with certainty the final position of the cap, thus the cap was removed and replaced by arrows symbolizing possible positions. No other modification was performed on the RNA, the Sinefungin molecule or the nsp16 structure. The SARS-CoV nsp10 and nsp16-coding sequences were amplified by RT-PCR from the genome of SARS-CoV Frankfurt-1 (accession number AY291315) as previously described[16]. The nsp10 and nsp16 genes (encoding residues 4231–4369, 5903–6429, and 6776–7073 of replicase pp1ab) were cloned into a Gateway modified dual-promotor expression plasmid and in the gateway pDest 14 expression vector. In this backbone, SARS CoV nsp10 can be expressed under a tet promoter and encodes a protein in fusion with a N-terminal strep tag, whereas nsp16 is expressed under a T7 promoter and encodes a protein in fusion with a N-terminal hexahistidine tag. The mutants were generated by PCR using the Quickchange site–directed mutagenesis kit (Stratagene), according to the manufacturer's instructions. AdoMet and cap analogs GpppA and 7MeGpppA were purchased from New England BioLabs, the[3H]-AdoMet was purchased from Perkin Elmer and Sinefungin (adenosylornithine) from Sigma-Aldrich. E. coli C41 (DE3) cells (Avidis SA, France), containing the pLysS plasmid (Novagen), were transformed with nsp10 or nsp16 cloned in pDest14, or nsp10/nsp16 cloned in pmCox, and grown in 2YT medium supplemented with appropriate antibiotics. The expression of strep-tagged nsp10 or 6His-tagged nsp16 mutants was induced (DO600 = 0.6) by adding 50 µM IPTG, and the expression of the nsp10/nsp16 complex by adding 50 µM IPTG and 200 µg/L of anhydrotetracycline. After an incubation for a 16 h at 24°C, the cell were pellets, frozen and resuspended in lysis buffer (50 mM HEPES, pH 7.5, 300 mM NaCl, 5 mM MgSO4, 5 mM β-mercaptoethanol (only for nsp10) supplemented with 1 mM PMSF, 40 mM imidazole, 10 µg/ml DNase I, and 0.5% Triton X-100. After sonication and clarification, proteins were purified either by IMAC (HisPurTM Cobalt Resin; Thermo Scientific) chromatography[16] (nsp10 mutants and nsp16 mutants), and the nsp10/nsp16 complex was purified by using Strep-Tactin sepharose (IBA Biotagnology) as previously described[16]. All purified proteins were analyzed by SDS-PAGE. The binding of wild-type nsp10 to mutant nsp16, and that of mutant nsp10 to wild-type nsp16 was quantified using ImageJ as described[16]. MTase activity assays were performed in 40 mM Tris-HCl, pH 8.0, 5 mM DTT, 1 mM MgCl2, 2 µM 7MeGpppAC5 or GpppAC5, 10 µM AdoMet, and 0.03 µCi/µl [3H]AdoMet (GE Healthcare). Short capped RNAs (7MeGpppAC5, GpppAC5, were synthesized in vitro using bacteriophage T7 DNA primase and were purified by high-performance liquid chromatography (HPLC) as previously described[61]. In the standard assay, nsp10 and nsp16 were added at final concentrations of 600 nM, and 200 nM, respectively, and the amount of 3H-CH3 transferred onto 7MeGpppAC5 substrates was determined by filter binding assay as previously described[16].

10.1371/journal.pgen.1005049

Replicative DNA Polymerase δ but Not ε Proofreads Errors in Cis and in Trans

It is now well established that in yeast, and likely most eukaryotic organisms, initial DNA replication of the leading strand is by DNA polymerase ε and of the lagging strand by DNA polymerase δ. However, the role of Pol δ in replication of the leading strand is uncertain. In this work, we use a reporter system in Saccharomyces cerevisiae to measure mutation rates at specific base pairs in order to determine the effect of heterozygous or homozygous proofreading-defective mutants of either Pol ε or Pol δ in diploid strains. We find that wild-type Pol ε molecules cannot proofread errors created by proofreading-defective Pol ε molecules, whereas Pol δ can not only proofread errors created by proofreading-defective Pol δ molecules, but can also proofread errors created by Pol ε-defective molecules. These results suggest that any interruption in DNA synthesis on the leading strand is likely to result in completion by Pol δ and also explain the higher mutation rates observed in Pol δ-proofreading mutants compared to Pol ε-proofreading defective mutants. For strains reverting via AT→GC, TA→GC, CG→AT, and GC→AT mutations, we find in addition a strong effect of gene orientation on mutation rate in proofreading-defective strains and demonstrate that much of this orientation dependence is due to differential efficiencies of mispair elongation. We also find that a 3′-terminal 8 oxoG, unlike a 3′-terminal G, is efficiently extended opposite an A and is not subject to proofreading. Proofreading mutations have been shown to result in tumor formation in both mice and humans; the results presented here can help explain the properties exhibited by those proofreading mutants.

Many DNA polymerases are able to proofread their errors: after incorporation of a wrong base, the resulting mispair invokes an exonuclease activity of the polymerase that removes the mispaired base and allows replication to continue. Elimination of the proofreading activity thus results in much higher mutation rates. We demonstrate that the two major replicative DNA polymerases in yeast, Pol δ and Pol ε, have different proofreading abilities. In diploid cells, Pol ε is not able to proofread errors created by other Pol ε molecules, whereas Pol δ can proofread not only errors created by other Pol δ molecules but also errors created by Pol ε molecules. We also find that mispaired bases not corrected by proofreading have much different likelihoods of being extended, depending on the particular base-base mismatch. In humans, defects in Pol δ or Pol ε proofreading can lead to cancer, and these results help explain the formation of those tumors and the finding that Pol ε mutants seem to be found as frequently, or more so, in human tumors as Pol δ mutants.

Unlike prokaryotes, eukaryotic cells have multiple DNA polymerases involved in chromosomal replication. It was first demonstrated in Saccharomyces cerevisiae [1] and then in human cells [2] that Pol α, Pol δ, and Pol ε were necessary for normal replication. It was subsequently found that two of these polymerases, Pol δ and Pol ε, had 3′ to 5′ exonuclease proofreading activities that could be inactivated to yield proofreading defective enzymes [3–5]. The Pol α-primase complex initiates DNA replication with short RNA primers followed by limited elongation by Pol α; this initiation takes place for each Okazaki fragment and is likely the case for initial initiation of the leading strand as well [6]. Using the two proofreading mutants and analysis of various mutational spectra, it was proposed that leading and lagging strands of replication were each replicated primarily by only one of the two polymerases, Pol δ and Pol ε, [7–9]. At that point, it was not possible to determine which of the polymerases was responsible for each of the replication strands. The use of mutations in each of the DNA polymerases that decrease their fidelity has proven very useful in analyzing their roles in replication. It was suggested that Pol δ, but not Pol ε, could proofread errors created by Pol α [10], supporting a model in which lagging strand synthesis was performed by Pol δ. Mutator alleles of Pol ε were consistent with its role in leading strand synthesis [11], and mutator alleles of Pol δ showed its activity in lagging strand synthesis [12]. A later genome-wide analysis using a Pol δ mutator allele again demonstrated that most Pol δ errors were on the lagging strand [13]. Therefore a current model of replication in yeast is that the lagging strand is replicated by Pol δ and the leading strand by Pol ε. The fact that a similar differentiation is observed in the very distantly related yeast Schizosaccharomyces pombe has led to the suggestion that this model is likely true for at least most eukaryotes [14]. One major issue in understanding yeast DNA replication has been the extent to which the leading strand is replicated only by Pol ε. It was found that the catalytic activity of Pol ε is not essential [15], demonstrating that Pol δ is in some cases able to replicate both strands. In addition, it has been consistently found that proofreading defective alleles of Pol ε have a much weaker mutator phenotype than do proofreading defective alleles of Pol δ [6–8,16–18]. Such results have led to proposals that Pol δ could replicate the leading strand under conditions of dysfunction [19] or could be part of an alternative fork formed after stalling on the leading strand [20]. However, the most comprehensive model of Pol δ involvement in leading strand replication was proposed by Pavlov and Shcherbakova based on an extensive survey of the literature and some of their unpublished work [6]. Their model also has initial synthesis of the leading strand by Pol ε and the lagging strand by Pol δ. They envision a variety of different possibilities for an interruption on the leading strand, including incorporation of an incorrect nucleotide that would be difficult to extend, a lesion on the leading strand, collision with RNA polymerase, or spontaneous dissociation of Pol ε. In any of those cases, they propose that reinitiation would be done by Pol δ and not Pol ε [6]. In addition to proofreading, an extremely important system for maintaining fidelity of replication is the mismatch repair system (MMR). The mismatches that result from incorporation of mispaired bases are recognized in eukaryotes by homologues of the bacterial protein MutS, generally MutSα, a heterodimer of Msh2 and Msh6, and MutSβ, a heterodimer of Msh2 and Msh3 [21–23]. Base-base mismatches are recognized almost entirely by MutSα, although there is evidence for recognition of some base-base mispairs by MutSβ [24]. Insertion and deletion mismatches are recognized by both MutSα and MutSβ, with small loops preferentially recognized by MutSα and larger loops preferentially recognized by MutSβ [21–23] with an additional preference of MutSα for repair of insertion loops and repair of deletion loops by MutSβ [25]. Recognition by MutS homologues is followed by interaction with homologues of MutL, usually MutLα in eukaryotes [21–23]. The newly replicated DNA is excised, followed by resynthesis. The method of MMR strand discrimination is still not completely known, but is likely a result of the endonuclease activity of the MutL homologues and interaction of MMR proteins with PCNA [26–28] as well as the incorporation of ribonucleotides into the newly synthesized strand [29,30]. Given the role of proofreading in eliminating mispairs and the role of MMR in repair of mismatches, one might predict that the two systems would function in the same pathway and would exhibit synergistic interactions, and that is what has been observed [3,6,7,17,18]. Although haploid strains of S. cerevisiae defective either in MMR or Pol δ proofreading grow well and are viable, the double mutant is not viable, whereas the corresponding double mutant with a Pol ε proofreading defect is viable [3]. This latter result seems to indicate that Pol ε proofreading plays a lesser role in replication than does Pol δ proofreading. In addition to proofreading and MMR, a major determinant of replication fidelity is the accuracy of the polymerase itself and its ability to extend mispaired nucleotides [31]. DNA polymerases can vary substantially in their fidelity; both Pol δ and Pol ε are relatively accurate even in the absence of proofreading [32,33]. If a mispair is formed and not corrected by proofreading, there is a wide variability in how well various mispairs can be extended—as measured, for example, in vitro with Taq DNA polymerase or E. coli Pol I Klenow fragment [34,35]. The difference in extension efficiency can be explained at least in part by the structure of the mispairs [36]. It has proven difficult to study proofreading in vivo. Due to the strong synergism with MMR, any measurement of proofreading in the presence of MMR reveals only those mispairs that escaped MMR. However, proofreading defects coupled with an absence of MMR give mutation rates so high that the resulting strains are sick, even as diploids. The situation is even more complex if one is interested in analyzing the role of each replicative polymerase, as the direction of replication of a given assay region is frequently not known, and thus the mispair leading to mutation is indeterminate. In this work, we make use of a collection of trp5 mutants, each of which can revert to wild type via only one given base pair change [37]. Those trp5 alleles are placed in a region with a dependable origin of replication so that for each strain it is known which strand is replicated on the leading strand and which on the lagging strand [37]. In order to examine proofreading in the absence of MMR, we use diploid strains that are hemizygous for the trp5 mutations. Our results are consistent with replication of the lagging strand by Pol δ and initial replication of the leading strand by Pol ε. However, we find that Pol δ can proofread errors on both leading and lagging strands, including errors created by other proofreading-defective Pol δ molecules, whereas a Pol ε molecule is only able to proofread its own errors. It has been difficult to reconcile the much greater mutator phenotype of Pol δ-proofreading deficient strains compared to Pol ε-proofreading deficient strains with a model of replication in which each of the DNA polymerases is primarily responsible for the replication of one strand of DNA. Our results explain this discrepancy by demonstrating Pol δ proofreading of Pol ε errors on the leading strand. Our finding that reversion rates of diploid strains heterozygous for either proofreading deficiency are similar is consistent with reports that Pol ε-proofreading deficient human cells have a disposition toward tumor formation at least as strong as that of Pol δ-proofreading deficient cells. We find large differences in extension efficiencies of various base-base mismatches and additionally find that an 8-oxoG-A mismatch is extended very efficiently and is not subject to proofreading. The demonstration of greatly varying mismatch extension biases in vivo can potentially help explain the striking differences in tumor spectra between mammalian cells deficient in Pol δ proofreading compared to Pol ε. Various methods have been used to examine the effects of proofreading and to distinguish replication of the leading and lagging strands. We decided to make use of strains containing mutations in an essential codon of the TRP5 gene because of the mutation specificity and low background of spontaneous reversions [37]. Although this assay permits analysis of base pair mismatches in only one sequence context, it does allow the study of specific mispairs in ways not previously possible. Because we were interested in studying the effect of proofreading on replication, it was necessary to use strains deficient in MMR, as there is a strong synergism between MMR and proofreading [3,6,7]; in addition, MMR shows replication strand bias, further complicating analysis in the presence of MMR [38,39]. For experiments involving defective MMR, in contrast to almost all previous studies of proofreading we used strains deficient in MutSα (msh6) since the overall mutational burden is less in such strains compared to a total deficiency in MMR [18] and there is at most only a slight effect of MutSβ on base pair substitution mutagenesis [24]. Our assumption in using strains deficient only in MutSα was that they would be somewhat healthier than those devoid of all MMR, and that assumption has recently been shown to be true for pol2–4 strains [40]. Because haploid cells deficient in both Pol δ proofreading and MMR are not viable [6], diploid strains were necessary for all of our studies. We used the well-studied pol2–4 allele to create Pol ε polymerases devoid of proofreading [4], and the pol3–5DV allele to disrupt proofreading in Pol δ [10,41–44]. A strain of opposite mating type, deleted for the TRP5 gene, was crossed with various trp5 mutant strains, creating hemizygous trp5 mutants whose reversion could be measured by plating on media lacking tryptophan. As shown in Fig. 1A, each of the trp5 alleles is present in both forward and reverse orientations near a dependable origin of replication (ARS306), such that a given DNA sequence will be replicated on the leading strand in one orientation and on the lagging strand in the other orientation [37]. Because of the extremely low reversion rates in wild-type strains, and unlike previously used assays to study proofreading, the observed increase in reversion rates in proofreading-defective strains is due almost entirely to the mispair created on the strand replicated by the proofreading-defective polymerase. A specific example with the trp5-G148T allele is illustrated in Fig. 1B. The trp5-G148T allele reverts to wild-type via a T-A to G-C mutation which would occur by incorporation of either a T-C or A-G mispair during replication. For most spontaneous revertants, there would be no way to ascertain which mispair had occurred. However, if one assumes a model in which Pol ε, whose catalytic subunit is encoded by the POL2 gene, replicates the leading strand and Pol δ, whose catalytic subunit is encoded by the POL3 gene, replicates the lagging strand, the identity of the mispair inducing the reversion can be determined. For strains in the F orientation, reversion in Pol ε-proofreading defective strains would be due to T-C mispairs, while reversion in Pol δ-proofreading defective strains would be due to A-G mispairs. The opposite would be true in strains with the R orientation. We first measured the rates of spontaneous reversion of the hemizygous trp5-G148T allele to the Trp+ phenotype in various genetic backgrounds. In general, the single-mutant reversion rates were not distinguishable from each other or from wild-type, in part due to the very low reversion rates and correspondingly large Confidence Intervals (Fig. 2A and S1 Table). It has been known for many years that defects in proofreading and MMR were synergistic [3,7], and the reversion rates of strains deficient in both MMR and one proofreading activity (Fig. 2A and S1 Table) strongly demonstrate that fact. With one exception (the trp5-G148T msh6 pol2–4 R strain) the double-mutant reversion rates were one to two orders of magnitude higher than any of the single-mutant reversion rates. Although defects in Pol ε proofreading (pol2–4) generally result in a lower mutation rate than defects in Pol δ (pol3–01 or pol3–5DV) [7,43], we observed that in msh6 strains with the F orientation, the mutation rates were approximately the same (compare msh6 pol2–4 F with msh6 pol3–5DV F in Fig. 2A), whereas they were vastly different in msh6 strains with the R orientation. Based on previously reported results, we would have expected that the reversion rates for the msh6 pol3–5DV strains to have been much higher than the msh6 pol2–4 strains in either orientation. Because of the unexpected differences in reversion rates of the double mutants, we repeated the experiments in a second set of diploid strains, containing the hemizygous trp5-A149C allele (Fig. 2B and S2 Table). Reversions of the trp5-A149C allele occur via the same mismatches as for the above trp5-G148T allele, but the bases on the primer and template strands are switched. The pattern of reversion rates in the trp5-A149C strains was very similar to that of the trp5-G148T strains (compare Fig. 2A to Fig. 2B) with wild-type or single-mutant strains having low and generally similar reversion rates and double-mutant strains having much higher reversion rates. Thus in two independent sets of strains, we observed striking effects of the orientation of the TRP5 marker gene on reversion rates due to proofreading defects; for pol2–4 msh6 strains, reversion rates were much higher in the F orientation and for pol3–5DV msh6 strains reversion rates were much higher in the R orientation. The effect of TRP5 orientation on reversion rates of the double mutant strains was both striking and unexpected. These results were uncovered due to the novelty of this assay; previous assays for spontaneous mutations in proofreading mutants have used either forward mutation rates in genes such as CAN1 or URA3 that give many different types of mutations [3,7,8,10,16,18,40,41,43,45–50] or reversion assays that usually involve slippage in simple repeats to give frameshifts in the his7–2 or hom3–10 alleles or various alleles of LYS2 [3,7,10,16–18,41,43,45–49,51,52]. Unless the replication direction of a marker gene is known, there is little information to be gained by measuring mutation frequencies of an inverted copy of the gene. With only two exceptions, none of the mutation assays referred to above examined orientation of the marker gene for spontaneous mutation. Those two exceptions were analysis of mutation spectra in a URA3 gene in both orientations near a defined origin of replication in the chromosome [7] and of the SUP4-o gene in both orientations on a yeast plasmid [8]. In both cases, the mutational spectra were different in the two orientations, which was used as an argument that the two polymerases replicated different strands, but there was no analysis of any differences in mutation rates in the two orientations [7,8]. As discussed above, assuming that the leading strand was replicated by Pol ε and the lagging strand by Pol δ [11,12], it is possible to infer which mispair was increased in each strain. For example, in the trp5-G148T strain in the F orientation, a defect in Pol2 proofreading should increase the number of T-C mispairs, but should have no effect on the number of A-G mispairs (see Fig. 1B). For each double mutant strain in Fig. 2, we could associate a reversion rate with the particular mispair that should have led to the reversion event as shown in Fig. 3. We then did two comparisons. For a given mispair, we compared the reversion rate for that mispair in a pol3–5DV strain to the same mispair in a pol2–4 strain. For example, the trp5-G148T pol3–5DV F reversion rate (presumably occurring via A-G mispairs) is 29-fold greater than the trp5-G148T pol2–4 R reversion rate (also dominated by A-G mispairs) (Fig. 3). These comparisons show that for the same mispair, in MMR-deficient strains, the reversion rate of the pol3–5DV strain is greater than the reversion rate in a pol2–4 strain, consistent with previous results showing a greater mutator effect in Pol δ proofreading-defective strains compared to Pol ε proofreading-deficient strains. We next compared the reversion rates due to one mispair compared to the other mispair in strains with the same proofreading defect. For example, the reversion rate of a trp5-G148T F pol2–4 strain (with increased T-C mispairs) is 17-fold greater than the reversion rate of the trp5-G148T R pol2–4 strain (with increased A-G mispairs) (Fig. 3). In every case, the reversion rate due to increased T-C or C-T mispairs was greater than that due to increased A-G or G-A mispairs. That result suggested that either T-C mispairs were more readily formed than A-G mispairs, or that they were more easily extended, and thus more susceptible to proofreading, than A-G mispairs. This analysis helps explain the difference in mutational spectra observed in strains defective in either Pol δ or Pol ε proofreading: for at least certain mismatches, the frequency of forming and extending one mismatch is much greater than forming and extending the complementary mismatch. However, these experiments cannot distinguish whether the difference is due to likelihood of formation of a given mispair or the relative efficiency of extending a given mispair. Comparing the effect of a Pol δ proofreading defect to a Pol ε proofreading defect for the same mispair revealed that the reversion rate for pol3–5DV msh6 mutants was 19- to 29-fold higher than for pol2–4 msh6 mutants in trp5-G148T strains and approximately 2- to 7-fold higher in trp5-A149C strains (Fig. 3), which might suggest that Pol δ was less accurate than Pol ε in the absence of proofreading. However, an analysis of in vitro activity suggests that is not the case and that the accuracy of the proofreading defective enzymes is similar, with that of Pol ε being slightly less in certain instances [33]. If the inherent accuracies of Pol δ and Pol ε without proofreading activity are similar, an alternative explanation for the much higher reversion rates of pol3–5DV msh6 strains compared to pol2–4 msh6 strains would be that Pol δ could proofread Pol ε errors, but not vice-versa. An initial step was to ask whether a wild-type polymerase molecule was able to proofread errors created by a proofreading-defective molecule of the same type. We answered this question by constructing diploid strains deficient in MMR, but only heterozygous for proofreading defects (msh6/msh6 pol2–4/POL2 abbreviated pol2–4± or msh6/msh6 pol3–5DV/POL3, abbreviated pol3-5DV±). In heterozygous strains, we would expect half of the polymerase molecules to be mutant and half to be wild-type. If errors produced by the mutant polymerase were not subject to subsequent proofreading by wild-type polymerases in the heterozygous strains, we would expect the reversion rate of the heterozygous strains to be one half that of homozygous proofreading defective strains. The results are shown in Fig. 4 and S3 Table and compared in Table 1. The difference between pol2–4± and pol3-5DV± strains is striking. For Pol δ, the difference between heterozygous and homozygous strains ranged from 12- to 46-fold (Table 1). This result strongly suggests that wild-type Pol δ molecules could proofread the errors created by the proofreading-defective Pol δ molecules, as the presence of one wild-type POL3 allele reduces reversion rates by over an order of magnitude. This conclusion is also consistent with the suggestion that Pol δ can proofread Pol α errors [10]. Proofreading of Pol α errors by Pol δ would imply that DNA strands created by Pol α, but with some error of replication, are subject to proofreading by Pol δ, as well as extension by Pol δ. Similarly, cis-proofreading of Pol δ errors by Pol δ would imply that a DNA strand synthesized by Pol δ but containing errors in replication would be subject to proofreading by Pol δ, as well as extension by Pol δ. Strains that were heterozygous or homozygous for Pol ε proofreading (msh6/msh6 pol2–4/POL2 or msh6/msh6 pol2–4/pol2–4) were very similar in reversion rates. The increase between strains that are heterozygous in Pol ε proofreading (pol2–4±) and those that are homozygous was approximately two-fold for each strain (Table 1). One issue is whether that difference is statistically significant. Although it is standard to assume that 95% Confidence Intervals for two measurements should not overlap in order for the difference to be significant, in fact 95% Confidence Intervals that do not overlap are significant at the P = 0.005 level and it is 83% Confidence Intervals that are significant at the P = 0.05 level if they do not overlap [53]. Consequently, 83% Confidence Intervals were calculated for the various strains defective in Pol ε proofreading. Those results are given in S3 Table and show that three of the four comparisons in Table 1 are significantly different. A recent measurement of CAN1 mutation rates in diploids that were pol2–4/pol2–4 or pol2–4/POL2 also found a 2-fold difference [49]. A 2-fold difference is what would be expected if a given region were replicated either by a proofreading-defective or proofreading-competent molecule and there was no compensating effect of the wild-type molecules on the proofreading-defective molecules. If the reversion rates of the pol2–4 and pol2–4± strains were considered to be not significantly different, the implication would be that the pol2–4 allele was dominant over the wild-type allele. In either case, there is no evidence of any cis-proofreading by Pol ε. We have a set of 12 trp5 haploid strains, with 6 different alleles of TRP5, each in both orientations relative to the ARS306 origin of replication [37]. An analysis of the remaining trp5 strains defective in MMR and heterozygous in one of the proofreading mutants was performed and the results shown in Fig. 5 with the data given in S4 Table. The trp5-G148A and trp5-A149G alleles are in some ways analogous to the trp5-G148T and trp5-A149C alleles as they also revert via complementary mutations, AT→GC and GC→AT respectively (Fig. 6). However, unlike the situation with the trp5-G148T and trp5-A149C alleles, the complementary mispairs in the trp5-A149C strain are in the strain of opposite orientation compared to the mispairs in the trp5-G148A strain (compare Fig. 3 and 6). With both sets of strains, there are orientation biases in reversion rates and the biases are opposite in msh6 pol2-4± strains compared to msh6 pol3-5DV± strains (Fig. 5). The reversion rates for these strains were analyzed in Fig. 6 in a manner similar to that shown in Fig. 3 for the trp5-G148T and trp5-A149C strains. In contrast to the strains with homozygous proofreading deficiencies in Fig. 3, the relative reversion rates for a given mispair in pol3-5DV± compared to pol2–4± strains are much more similar, with the reversion rates in some pol2–4± strains being higher than for the equivalent mispair in pol3-5DV± strains (Fig. 6). In every case the loss of proofreading for a T-G mispair causes a higher reversion rate than the loss of proofreading for an A-C mispair. Thus it appears that either T-G mispairs are formed at a higher rate than A-C mispairs, or they are more easily extended. The situation with the trp5-G148C and trp5-A149T strains is quite different. With those strains, there is a very low reversion rate even in the absence of MMR and a partial proofreading defect (Fig. 5). In all of those strains, reversion is due to the mispairing of identical bases: G-G or C-C and A-A or T-T respectively. Transformation of cells by single-stranded oligonucleotides (oligos) in which a permanent change is made to either chromosomal or plasmid DNA by introduction of oligos into the cell has been studied extensively in three systems: E. coli, mammalian cells, and yeast. In E. coli, numerous experiments from multiple labs support a mechanism in which oligos anneal to single-stranded DNA at the replication fork and serve as primers for continued replication, with oligos annealing to the lagging strand being considerably more efficient than when annealing to the leading strand of replication [54–62]. Mechanistic studies of oligo transformation in mammalian cells are more difficult than in E. coli. However, multiple labs have shown that oligo transformation is associated with cellular replication [63–65], that it is more efficient in S phase [66,67], that the oligo is incorporated into the genome, likely during replication [68], and that evidence suggests that the transforming oligos do so by serving as primers for replication [69–71]. Transformation in both E. coli and mammalian cells is inhibited by MMR, in agreement with the association of oligo transformation and replication [56–58,60,61,63,69,72–81]. In yeast, we have shown that oligos transform more efficiently when directed to the lagging strand [25,39,82–84], that transformation is inhibited by MMR by removing oligo sequences creating MMR-recognized mispairs [25,39,82–84], that oligos transform by incorporation [83], and that the 5’ end of transforming oligos is usually removed by a process partially dependent on Rad27, suggesting removal as part of Okazaki-like processing [84]. We also showed that in normally growing cells, MMR specifically removes oligo sequences that are part of mispairs, but that if oligo sequences escape MMR recognition and survive past S phase, MMR no longer can distinguish between the oligo and chromosomal sequences [83]. There remain two questions about oligo transformation in yeast: how oligo-directed replication could occur on the leading strand and whether transformation might generally occur in single-stranded gaps remaining in the G2 cell cycle phase. Work from the Marians lab has shown in vitro in E. coli that there can be “lesion skipping” on the leading strand that can result in repriming of replication [85,86]. Many years ago, it was found in UV-irradiated yeast cells that on both the leading and lagging strands short single-stranded gaps were observed that were proposed to be the result of repriming events [87]. Proposals that Pol δ could replicate the leading strand under conditions of dysfunction [19] or could be part of an alternative fork formed after stalling on the leading strand [20] would also suggest some type of repriming event on what was the leading strand. A very recent study of in vitro yeast replication showed that Pol ε is tightly associated with the CMG helicase during leading strand synthesis but that it can periodically cycle on and off PCNA-DNA [88]. An analysis of that work suggested that such cycling could provide access to a mismatched primer for extrinsic proofreading [89]. An oligo bound to the leading strand might appear much like a normal replicative end exposed by a cycling off of the Pol ε-CMG complex. Although we cannot rule out the possibility of transformation occurring in single-stranded gaps left in G2, we consider that possibility unlikely as a general mechanism. Our cells are undamaged and growing in rich medium before transformation. It seems unlikely that there would be sufficient single-stranded gaps in the particular region to be transformed to account for the high transformation frequencies we have observed in some cases [39]. It is also not clear why in G2 there would be five-fold or more single-stranded gaps on what used to be the lagging strand compared to what used to be the leading strand of replication. The very active involvement of MMR and of Rad27 also seem more compatible with a process occurring during replication rather than post-replicationally. Therefore we consider it likely that in yeast, as appears to be the case in E. coli and mammalian cells, oligos transform by annealing to a single-stranded region at the replication fork, with a strong preference for lagging strand, and then serve as pseudo-Okazaki primers for replication. If oligos can serve as primers for replication, it might be possible to transform strains with oligos that create a mispair necessary for reversion at their very 3′ end as indicated in Fig. 7 assuming the mispair was extended rather than being proofread. We tested this hypothesis by transformation using Oligo 148C with a 3′ C that would create a T-C mispair necessary for reversion of the trp5-G148T allele (Fig. 7). The results are given in Table 2. Because the transformation results in the table are relative to transformation with an oligo creating a mispair internal to the oligo, low transformation in these experiments indicates either the removal of the 3′ terminal mispair necessary for reversion of the strain, or failure to extend the mispair. When Oligo 148C was transformed into strains in the R orientation, which would put the oligo on the lagging strand, we obtained a relatively low number of transformants in an msh6 strain. That number did not increase in pol2–4 strains, but increased about 6-fold in the msh6 pol3-5DV± strain and about 30-fold in the msh6 pol3–5DV strain. In strains with the F orientation, transformation of the msh6 strain is even lower, as the oligo would anneal to the leading strand. There is little if any increase in the msh6 pol2–4 strain or the msh6 pol3-5DV± strain but a large increase (~30-fold) in the msh6 pol3–5DV strain. When we attempted to perform the same experiment with Oligo 148G, creating a G-A mispair on the opposite strand from Oligo148C, we obtained essentially no revertants in any background. We performed the same type of oligo transformation experiment in trp5-A149C strains, using Oligos 149A and 149T. As illustrated in Fig. 7, these oligos produce the same mismatches for extension as in the trp5-G148T strains, but with the opposite base as primer in the mispair. The results of these experiments are given in Table 2. As in the trp5-G148T strains, there is little transformation of msh6 or msh6 pol2–4 strains. In contrast to the results with Oligo 148G, there is measurable transformation of Oligo 149A in msh6 pol3–5DV± strains and substantial transformation in msh6 pol3–5DV strains, even when directed to the leading strand. It thus appears, at least in this sequence context, that extension of an A in a G-A mispair is much more likely than the G in an adjacent G-A mispair. Oligo 149T gives robust transformation in msh6 pol3–5DV strains, similar to transformation with Oligo 148C. These experiments help make several important points. None of the oligos (Oligo 148C, Oligo 148G, Oligo 149A, or Oligo 149T) showed much transformation in msh6 cells, but 3 of the oligos (all except Oligo 148G) showed significant transformation in msh6 pol3–5DV cells, when targeted to either the lagging or leading strand. Those results demonstrate that the lack of transformation in msh6 cells is due to proofreading of the 3′ terminal mismatch by Pol δ, as elimination of Pol δ proofreading is sufficient to enable robust transformation by the oligos. In addition, the effect of Pol δ proofreading is observed, whether the oligos are targeted to the lagging or leading strand. Because incorporation of the mismatch created by the 3′ terminal base of the oligos is necessary for transformation, these results also strongly suggest that the oligos must be serving as primers for continued DNA synthesis for it is difficult to propose another mechanism that could explain oligo transformation with a 3′-terminal mismatch. There is marked variability in the efficiency by which the oligos are able to transform. Oligo 148G gave essentially no transformants in any strain (although the same oligo with a modified G at the end gave robust transformation, see below). Oligos 148C and 149T gave the highest levels of transformation, while Oligo 149A gave markedly lower levels of transformation. These relative transformation efficiencies are similar to the differences in reversion rates seen in the trp5-G148T and trp5-A149C strains (Figs. 1 and 2; S1 and S2 Tables). In those double mutant strains, the lowest reversion rates were due to G-A mispairs in which G was on the primer strand; reversion rates due to A-G mispairs with the A on the primer strand were also very low. In all cases, reversion rates due to T-C or C-T mismatches were much higher. In our analysis of the orientation effects on reversion rates, we were not able to discriminate between reversion rate effects due to different frequencies of formation of certain mispairs or differences in elongation frequencies (see above). However, these results with oligo transformation suggest that the biased reversion rates are due at least in part to differential frequencies in elongation of various mispairs. Our oligo results show that at least in this sequence context a G paired opposite an A is very rarely elongated so that no matter how frequently such a mispair might be formed, it would rarely be extended. These results also are in line with in vitro mismatch extension experiments (see Discussion). In contrast to the effects of Pol δ proofreading on oligo transformation, we observed no effect of elimination of Pol ε proofreading, whether oligos were targeted to either the leading or lagging strand. These results indicate that, no matter what mechanism is responsible for oligo transformation, it is Pol δ alone that interacts with, and elongates, the oligo. If one accepts a model in which oligos transform by priming at the replication fork, these results would suggest that any replication restart due to oligo priming on the leading strand would be extended by Pol δ and not Pol ε, in line with a model of replication restart on the leading strand being due to Pol δ [6]. We also examined oligo transformation in strains heterozygous for Pol δ proofreading (msh6/msh6 pol3–5DV/POL3 or msh6 pol3-5DV±). As can be seen in Table 2, in most cases the difference between transformation in msh6 pol3–5DV strains was significantly greater than in msh6 pol3-5DV± strains, and in 3 cases was 20–50 fold greater. Those differences between pol3-5DV± and pol3–5DV strains are similar to the differences in reversion rates shown in Table 1 and are consistent with cis-proofreading by wild-type Pol δ. These oligo transformation experiments can also help explain how one could understand proofreading by wild-type Pol δ of errors made by proofreading-defective Pol δ molecules. When a proofreading-defective Pol δ molecule inserts a mispaired base, presumably there is some frequency at which the polymerase will extend the mispair; when that happens, the mispaired base is no longer susceptible to proofreading. Frequently, one would assume that the mispaired base would stall the polymerase synthesis and in the absence of the ability to proofread, might cause a release of the polymerase, exposing the mispaired base to other exonucleases in the cell. A wild-type Pol δ molecule could bind to the primer-DNA substrate and either extend, or more likely proofread, the mispair. A proofreading-defective Pol δ molecule could interact with the substrate, either extending the mispair, or disassociating. The oligos with 3′ mispairs mimic a dissociated primer-DNA complex. In pol3-5DV± strains, if a proofreading-defective Pol δ molecule would usually extend the mispair when it interacted with the 3′ mispair, one would expect that the difference in extension frequencies between pol3-5DV± and pol3–5DV strains would be 2-fold. The fact that it is much greater suggests that many of the proofreading-defective polymerase interactions are not productive, allowing more chances for the mispair to be proofread by the wild-type Pol δ. This same scenario in vivo could explain how mispairs could be cis-proofread and also why there is such a large difference in reversion rates in pol3-5DV± and pol3–5DV strains. It is known that incorporation of oxidatively damaged nucleotides can lead to mutations [90] and that MMR can recognize 8-oxoG-A mispairs [83,91–93]. It is not known to what extent an 8-oxoG-A mispair due to incorporated 8-oxoGTP might be subject to proofreading. We therefore used Oligo 148oxoG that creates an oxoG-A mispair at the 3′ end of the oligo in the trp5-G148T strains. The results of those transformations are given in Table 2. The results with Oligo 148oxoG are very different from those observed with any of the other oligos. There are a substantial number of transformants in msh6 strains of both orientations. However, there is not a significantly greater number of revertants in any proofreading-defective strain, suggesting that the oxoG-A mispair is not subject to proofreading. Because the Oligo 148oxoG has exactly the same sequence as the Oligo 148G except for the modified 3′ terminal base, these results support the conclusion that the extremely low numbers of Oligo 148G transformants are due to failure to extend the G-A mispair and not due to low formation of the G-A mispair. There is also no difference in Oligo 148oxoG transformants in msh6 pol3-5DV± compared to msh6 pol3–5DV strains in contrast to the differences in those strains observed with the other oligos. Those results suggest that there is no inherent defect in elongation ability of the proofreading defective Pol δ enzyme. Our results presented above support a model in which 1) proofreading errors are usually corrected by MMR 2) in the absence of proofreading the incorporation of a mispaired base strongly depends on the efficiency of its extension by DNA polymerase; 3) upon insertion of a mispaired base, proofreading-defective DNA polymerase molecules will either extend the mispair, or failing extension will dissociate from the primer end allowing proofreading of the mispaired base by other DNA polymerase molecules; 4) DNA Pol ε is not able to proofread 3′ mispairs created by other DNA polymerase molecules including other Pol ε molecules; and 5) DNA Pol δ can proofread 3′ mispairs on both the lagging or leading strand. Not only does this model explain our results in yeast, but it can also help explain many features of tumor formation in mammals due to DNA polymerases defective in proofreading function. Some of the earliest work on proofreading and MMR in yeast found a multiplicative relationship between mutants defective in proofreading and MMR and those results were interpreted as demonstrating serial action of proofreading and MMR [3,7,94]. Our single-mutant rates, given in Fig. 2 and S1 and S2 Tables, are so low and have such large Confidence Intervals that they cannot be used in such calculations, but the double mutant rates are sufficiently high that they suggest synergism and not multiplicativity and thus seem at odds with the previous results. The two assays used by Morrison and Sugino [3,7,94] were forward mutation rate measurements in URA3 and reversion of the his7–2 frameshift allele. In both of those assays, the wild-type mutation rate was much higher than in our assay, and the mutation rate in the absence of MMR was increased by, in one of their haploid analyses, 41-fold in the URA3 assay and 150-fold in the his7–2 assay (Table 1 in [7]). It is thus very likely that the underlying mutation rates observed in those assays reflected errors not due to DNA polymerase proofreading defects. In contrast with the previous assays, we know in our trp5 reversion assay not only what base pair mutation is made, but in the case of proofreading mutants what particular base-base mispair led to the reversion event. In the case of the single proofreading mutants, we know that a proofreading error leading to a reversion event and thus creating a base-base mispair should be corrected by MMR, and the slight increase in reversion rate is almost certainly due to random escape from MMR, as no repair system will be 100% efficient. The amount of escape would presumably depend in part on the particular mispair and sequence context, as MMR repair of base-base mispairs is sequence dependent [95]. The small increase in reversion rates observed in the msh6 mutants is somewhat more complex. It is likely that some of the reversion events are due not to a failure in proofreading but rather to mispairs that results from damaged DNA, as we have previously demonstrated [96,97]. What appears to be a higher mutation rate in the msh6 A149C strain compared to the msh6 G148T strain, for example, could be due to mispairs involving an 8-oxoG and not due to proofreading errors. Such errors should not be increased in strains that would be defective in proofreading. Therefore in any proofreading-defective MMR-defective double mutant, we would expect to see a large increase in reversion rate due to the failure of MMR to repair proofreading errors, and that is what is observed. In the absence of MMR, the probability of a base pair mutation is a function of the probabilities of misinsertion of a base, its removal by proofreading, and elongation from the mismatched base pair. As noted above, we found a very strong orientation dependence in MMR-deficient, proofreading-defective, strains with four of the six different trp5 mutations (trp5-G148T, trp5-A149C, trp5-G148A, and trp5-A149G; Figs. 2 and 5). However, our reversion data did not allow us to discriminate between mispairs that are formed at a high rate and mispairs that are easily extended. For example, a mispair that was easily formed, but very poorly extended, would likely contribute little to the overall reversion rate. Our oligo transformation experiments, however, allow us to analyze efficiencies of mispair elongation. Using oligos that formed terminal A-G or C-T mispairs, we found that transformation efficiencies in msh6 pol3–5DV strains were quite variable depending on the particular mispair and that the variability correlated with the variability in reversion rates in msh6 pol3–5DV strains. Thus our results with oligos containing 3′ mismatches suggest that at least part of the reason for orientation bias in reversion rates was due to differential extension rates from various mispairs. Although as discussed above our oligo results are likely to reflect extension only by Pol δ, the fact that we see similar orientation bias with Pol ε proofreading mutants (Figs. 2 and 5), strongly suggests that Pol ε has similar elongation biases. Our oligo experiments studied only a subset of possible base mispairs—those for which we had reversion data in homozygously-deficient proofreading strains. As shown in Fig. 5, we found evidence using heterozygously-deficient proofreading strains that other mispairs also showed biased orientation effects. We think it likely that those biases could also be explained by differential mispair elongation efficiencies. It has been difficult to devise experiments that would measure mispair extension within the context of a chromosome, as both proofreading and MMR are very effective at eliminating extensions of mispaired bases. However, there have been some in vitro measurements of mispair extension. Even those measurements are complicated by the demonstrated sequence effects on mispair extension [35] and the necessity to use DNA polymerases devoid of proofreading activity. For the E. coli exonuclease-deficient Klenow fragment of Polymerase I, it was found with two exceptions that in each sequence context, extension of mispairs with identical base pairs was the least favored of all combinations [35], in line with the low reversion rates of the trp5-G148C and trp5-A149T strains observed in Fig. 5. The two mispairs that were least favored of all were extension of G opposite template A and extension of A opposite template G [35], again agreeing with the failure of Oligo 148G to transform trp5-G148T strains and the relative low transformation of trp5-A149C strains with Oligo 149A (Table 2). A similar pattern of mispair extension was observed with Taq DNA polymerase [34] and AMV reverse transcriptase and Drosophila Pol α [98]. In these publications, the mispair most efficiently extended was primer G against template T [35,98], and of all of our heterozygous reversion rates, two of the three highest were mispairs in which primer G against template T would have been on the strand with a heterozygous proofreading deficiency (trp5-G148A F msh6 pol3-5DV± and trp5-G148A R msh6 pol2–4±, Fig. 5). Thus the existing in vitro data show clear differences for DNA polymerase extension of different 3′ terminal mispairs. Our in vivo results, including both reversion rates of different trp5 mutants and our oligo transformation experiments, show biases that are consistent with the in vitro data demonstrating differential extension efficiencies of various mispairs. In strains proficient in proofreading, these differential extension frequencies are unlikely to be evident; however, in proofreading-defective strains, mispair extension bias is likely to be much more important, and underappreciated. Several studies have shown that the mutation spectra of Pol ε and Pol δ proofreading deficient strains differ [7,8,17]; one would expect that mispairs less likely to be extended would be most susceptible to trans-proofreading. Therefore we propose that differential mispair extension frequencies can explain not only the biased reversion rates we have found, but, more generally, the differences in mutation spectra observed in strains deficient in Pol δ compared to Pol ε proofreading. Many base pair mutations are likely due to mispairings involving a damaged base. Indeed, we speculate that the higher spontaneous mutation rates in the trp5-A149C wild type and msh6 strains compared to the equivalent trp5-G149T strains (Fig. 1) is due to oxidative damage of the template G in the trp5-A149C strains; we have shown that increased endogenous oxidative damage leads to greatly increased reversion rates of this mutation [96]. It is also known that incorporation of oxidized nucleotides represents a mutagenic threat to organisms [90,99] and we have previously shown that yeast can use exogenously added 8-oxoGTP and mutagenically insert it into the genome [83]. We also showed that MMR greatly inhibits the incorporation of 8-oxoGTP into the chromosome [83]. However, it has not been clear how well an 8-oxoG-A mispair could be proofread. Our interest in using Oligo 148oxoG for transformation is that it mimics the incorporation of 8-oxoGTP into the DNA and thus allows analysis of processes acting on the 8-oxoG-A mispair. In contrast to the lack of transformants with Oligo 148G, we obtained large numbers of transformants using Oligo 148oxoG, containing an 8-oxoG at the 3′ end rather than a G. As seen in Table 2, we find that the 8-oxoG-A mispair is essentially not recognized by proofreading, as there is a large number of transformants in msh6 strains, and the number is not increased in proofreading-defective strains. Thus not only is the 8-oxoG-A mispair extended well, in stark contrast to the lack of extension of the G-A mispair, but it is not recognized by proofreading. It now appears well established that in yeast, and likely most eukaryotes, the leading and lagging strands of replication are usually replicated by different DNA polymerases as suggested nearly two decades ago [7] and more recently demonstrated in detail [11–13]. With that understanding, it has been difficult to explain why there is a much greater increase in the mutation rates of strains with a proofreading deficiency in Pol δ compared to Pol ε [6–8,16–18]. That result is even more surprising given evidence that MutSα function is more efficient on the lagging strand which would be replicated by Pol δ [38] and that MMR in general appears to balance the fidelity of replication of leading and lagging strands [100]. Our analysis of reversion rates in homozygously versus heterozygously proofreading-deficient strains indicated that wild-type Pol δ polymerases could cis-proofread, whereas wild-type Pol ε polymerases could not proofread errors created by proofreading-deficient Pol ε molecules. Our oligo transformation experiments were consistent with the reversion analysis: wild-type Pol δ molecules prevented transformation by oligos, even in the presence of proofreading-defective Pol δ molecules. Elimination of all Pol ε proofreading activity made no difference in oligo transformation, even if the transforming oligos were targeted to the leading strand of replication. Although as stated above we cannot conclusively rule out that oligo transformation events might take place post-replicationally, the oligo transformation experiments are consistent with trans-proofreading by Pol δ, as there was insignificant transformation by oligos directed to the leading strand unless Pol δ proofreading was inactivated. It should be noted that an earlier analysis of Pol δ and Pol ε replication proposed that stalled leading strand replication would be continued by Pol δ [6]. How does a polymerase error become susceptible to proofreading by a different polymerase molecule? Because of the large size of the DNA polymerase molecules, it is likely that an error would have to cause a replication stall followed by at least partial release of the polymerase before a 3′ mispair could be exposed to a different polymerase molecule. The lagging strand is discontinuously replicated so it is not surprising that Pol δ molecules could proofread errors created by proofreading-defective Pol δ molecules, particularly as there is already evidence that Pol δ can proofread Pol α errors [10]. The leading strand is normally synthesized continuously; as noted above there is recent evidence suggesting that a Pol ε-CMG helicase complex can periodically cycle on and off PCNA-DNA and thus expose a mispair to extrinsic proofreading [88]. However, there is no known mechanism in which a different Pol ε molecule could be brought in to proofread, which is consistent with our reversion analysis demonstrating lack of Pol ε cis-proofreading. Our oligo transformation experiments suggest an additional possibility: that Pol δ molecules could proofread errors on the leading strand. Given the possibility of Pol δ proofreading of Pol ε errors, we can examine our reversion data for evidence of such trans-proofreading and how extensive it might be. The increase in reversion rate in msh6 pol2–4 strains by orders of magnitude over either single mutant indicates that a large number of Pol ε errors are not subject to Pol δ proofreading. Those reversion events must be due to errors by the proofreading-defective Pol ε polymerase that did not cause polymerase dissociation but were then extended by the polymerase. For Pol δ, there is a large increase in reversion rates, not only of msh6 pol3–5DV strains compared to either single mutant, but of msh6 pol3–5DV strains compared to msh6 pol3-5DV± strains (Table 1). The reversion rate in the msh6 pol3–5DV strains is due to the error rate of the pol3–5DV enzyme. Assuming that half of the replication in the msh6 pol3-5DV± strains is done by the wild-type Pol δ and half by the proofreading defective Pol δ, we would expect the reversion rate in the msh6 pol3-5DV± strains to be one half that of the msh6 pol3–5DV strains. The fact that the difference is 12- to 46-fold indicates that more than 90% of the time, a polymerase error results in a polymerase dissociation event that allows proofreading by a wild-type Pol δ enzyme. We can then assume that the reversion rate in a pol3-5DV± strain is indicative of the error rate due to molecules that do not dissociate but continue replication from the mispair. The reversion rate in msh6 pol2–4 or pol2–4± strains would be a combination of errors created by Pol ε proofreading-defective molecules that did not dissociate minus errors that were created and then proofread by the wild-type Pol δ molecules. Table 3 shows a comparison of the msh6 pol2–4± reversion rates compared to those of the msh6 pol3-5DV± strains for orientations that would have the same mispaired bases for reversion. (For example, the trp5-G148T msh6 pol2–4± F strain would be expected to revert via extension of a mismatched C opposite T on the leading strand, whereas the trp5-G148T msh6 pol2–4± R strain would be expected to revert via extension of a mismatched C opposite T on the lagging strand as illustrated in Fig. 3.) In these comparisons, the pol2–4± reversion rates are generally similar to the pol3-5DV± reversion rates. Given our assumption that each of these rates is due mainly to mispairs that are extended by the initiating polymerase, these results indicate that the underlying inaccuracy and tendency to extend mispairs of each polymerase is roughly similar. However, if one does the same comparison with the completely homozygous msh6 pol2–4 and msh6 pol3–5DV strains for which we have data, the reversion rate for each mispaired base configuration is much higher with the pol3–5DV mutant than the pol2–4 mutant (Table 3). As noted above, there is no reason to think that Pol ε is inherently more accurate, especially since in vitro results suggest that if anything Pol ε is slightly less accurate [33]; therefore the lower reversion rate for msh6 pol2–4 strains on identical mismatches compared to msh6 pol3–5DV suggests that there is quite substantial Pol δ proofreading of Pol ε errors. These results also indicate that when a DNA polymerase incorporates a mispair that it is unable to proofread, there is a high probability of polymerase dissociation from the template. Some results in the literature have been used to suggest that there could be functional redundancy of Pol δ and Pol ε proofreading activities so that Pol ε proofreading might, for example, be able to correct Pol δ errors. For example, a study of proofreading and MMR using a frameshift reversion assay found reversion rates of triple mutants (pol2–4 pol3–01 msh2) were not higher than that of double mutants and found two hotspots in mutation spectra of pol2–4 pol3–01 double mutants that were not present in either single mutant [17]. Based on those results, the authors suggested “The presence of these hotspots only in the double mutant is consistent with the functional redundancy between the Pol δ and Pol ε exonuclease activities as deduced from mutation rate measurements.” [17]. Comparison of mutation rates in very sick strains is quite problematic. Our double mutant strains that were only partially deficient in MMR (msh6) were quite sick; the strains used in the cited experiments were completely deficient in MMR (msh2). The phenomenon of “error catastrophe” and saturation of MMR was shown in E. coli in 1996, for example [101] and is likely observed in those results, as the measured mutation rate of a pol2–4 pol3–01 msh2 strain is actually lower than that of a pol2–4 pol3–01 strain [17]. Because of those high mutation rates one would assume that MMR would be reduced due to saturation in the pol2–4 pol3–01 strain. Therefore hotspots that would appear in the double mutant strain could be due to mispairs that were less well recognized by MMR and would be more likely to escape in the partially MMR-defective environment of the double mutant strain. In addition, it is now known that there are suppressors that can arise in mutants that are defective in MMR and proofreading [40,48], which is one reason we have been careful to use multiple isolates in our experiments. A proposed functional redundancy of the proofreading exonucleases would be counter to the finding that Pol ε cannot even cis-proofread its own errors, and more importantly could not explain the much higher mutation rates that have been consistently observed in Pol δ proofreading-defective strains compared to Pol ε proofreading-defective strains. Our results demonstrating trans-proofreading by Pol δ are also consistent with experiments that show that MMR appears to use Pol δ for resynthesis of DNA on either replication strand after mismatch excision [18,22,23,41,102–105]. In MMR, extension of the newly excised primer strand would be analogous to extension of one of our oligos. The disassociation of the polymerase from certain mispairs could also explain the replication checkpoint activation seen in Pol δ proofreading defective strains [46]; once a poorly-extended mispair is incorporated in a Pol δ proofreading defective strain, replication would be inhibited. Because of the low mutation rate of a pol2–4 mutant in their assay [46], it was not possible to determine if there were a replication checkpoint activation due to Pol ε proofreading defective mutations, but our results would suggest that checkpoint activation would not be observed, due to proofreading of the Pol ε errors by wild-type Pol δ molecules. In order to explain experimental results that were not consistent with a model of replication in which Pol δ was responsible for all lagging strand synthesis and Pol ε was responsible for all leading strand synthesis, Pavlov and Shcherbakova proposed that lagging strand synthesis was performed by Pol δ, but that leading strand synthesis, although begun by Pol ε, was completed after any interruption by Pol δ [6]. Their model proposes that the lower mutation rate of strains lacking Pol ε proofreading relative to those lacking Pol δ proofreading can be explained by the fact that switching of leading strand synthesis to Pol δ is the rule, and “the majority of the genome replication involves copying of both DNA strands by Pol δ” [6]. That model, however, is inconsistent with recent whole genome sequence analysis of replication in yeast by fidelity mutants of Pol α, Pol δ, and Pol ε indicating that most leading strand synthesis must be done by Pol ε [106]. Our model differs from that of Pavlov and Shcherbakova in that Pol δ synthesis on the leading strand could be accompanied by proofreading of Pol ε errors, thus substantially reducing the amount of leading strand synthesis by Pol δ necessary to explain the differential mutation rates observed in proofreading-defective strains. Thus our model can explain the considerably higher mutation rate of strains deficient in Pol δ proofreading compared to Pol ε proofreading, the lower viability of such Pol δ strains, but also the observation that each polymerase is responsible for most replication of only one strand of DNA. We chose to analyze proofreading mutations in the absence of MMR due to the known synergism of proofreading mutations and MMR and also because of the very low reversion rates in our strains, even when defective in proofreading. In principle, one would expect that any errors normally corrected by proofreading would be repaired by MMR. However, the fact that defects in proofreading alone do show increased mutation rates [7] is an indication that some of the excess replication errors manage to escape MMR. Recently, mutations in Pol ε and Pol δ in human endometrial and colorectal cancers have been found that appear to be pathogenic [49,107–110]. Although many of the mutations appear to be in domains that would affect proofreading, one of the more common mutations appears to affect fidelity as well as proofreading [49]. In general, these mutations appear to be heterozygous, inherited dominantly although somatic mutations are also seen, present in MMR proficient tumors, and the cancer spectrum of Pol δ and Pol ε mutations appears to be different [107–109]. Proofreading mutations have been made in the Pol δ and Pol ε polymerases of mice and studied in vivo. In contrast to the results seen in human tumors mentioned above, there is no tumor phenotype of heterozygous proofreading mutations in either Pol δ or Pol ε, but a robust tumor phenotype for homozygous mutations [111,112]. The tumor phenotype of the two proofreading mutations was perhaps even more distinct than that observed in humans [111]. The mutation rates of proofreading defects in each polymerase was measured and the mutator phenotype of a Pol ε defect was found to be greater than that of a Pol δ defect; the mutation rate of mice with homozygous defects in both Pol δ and Pol ε proofreading was not measurably greater than that of either single defect [111]. In contrast to the tumor phenotype, there was found to be an increased mutation rate in heterozygous defects in proofreading of either Pol δ or Pol ε compared to wild-type mutation rates with the heterozygous Pol ε defect giving a larger effect than that of Pol δ [111]. The above results seen in mice and humans seem puzzling in light of previous yeast work, in which the conclusion has been that defects in Pol δ proofreading are much more mutagenic than defects in Pol ε proofreading—although those conclusions were based almost entirely on homozygous proofreading defects. However, many of those results are compatible with our findings. Most mutations due to proofreading errors, particularly those that would escape MMR, would be expected to be base pair mutations, and that in fact is what is observed in mice [111]. Oncogenic and tumor suppressor mutations due to base pair mutations would be expected to be at least somewhat sequence specific and the marked orientation biases observed in our pol2–4 and pol3–5DV strains indicate that the probability of a given base pair mutation could be strongly dependent on which polymerase was proofreading defective and the orientation of replication of that gene in a given tissue. Therefore the differences in tumor spectra are perhaps not so surprising. The relative prevalence of Pol ε mutations compared to Pol δ mutations in human tumors is one of the most striking differences compared to what would have been expected from previous yeast work. Most of the observed human proofreading mutations are heterozygous and as Table 3 indicates, even in the absence of MMR, the reversion rate of some of our trp5 strains is higher or about the same level in pol2–4± strains compared to pol3-5DV± strains. It has been found that MMR due to MutSα is more efficient on the lagging strand than the leading strand, at least in yeast [38], which would tend to reduce even further the relative mutational bias in Pol δ mutants compared to Pol ε mutants. It is also possible that there could be some selective pressure for second site mutations to moderate the error rate of either proofreading polymerase as has been observed in yeast [40,48], particularly in homozygously-deficient animals. The fact that heterozygous defects in proofreading can lead to tumors in humans, but not in mice, is similar to findings with other genes and may be reflective of the much longer lifespan of humans than mice. For example, mice homozygously-deficient in Msh6 show a strong tumor phenotype, but there is little increase in heterozygous mice [113]. In humans inheriting heterozygous MSH6 mutations, there is a significant increase in various types of tumors [114]. The genotypes of all strains used in these experiments can be found in S5 Table. All haploid strains containing a TRP5 point mutation were derivatives of the strains previously published [37]. For creation of diploid strains that would be hemizygous for the trp5 point mutations, we used a haploid strain of opposite mating type, BY4741 [115] that shared parentage with our strains, but contained complementary markers. We further modified BY4741 by restoring the strain to Leu+ and making an exact deletion of the TRP5 gene by delitto perfetto [116], creating GCY2122 (S5 Table). The MSH6 gene was deleted by transformation with a PCR fragment generated from the MSH6 gene deletion described in [117] or from a strain containing an MSH6 deletion created with a loxP-kanMX-loxP fragment [118]. When a second allele of MSH6 was to be deleted in a diploid strain, a PCR fragment obtained from a strain containing an MSH6 deletion made by insertion of the loxLE-hphNT1-loxRE fragment contained in pZC3 [119] was used. pol2–4 haploid mutants were created by transformation with plasmid YIpJB1 as described [4]. pol3–5DV haploid mutants were created by transformation with EagI-digested pY19 [43], selecting for Ura+ cells. Cells were subsequently selected for URA3 loss and screening for strains containing the pol3–5DV mutation. In order to create msh6 pol3–5DV haploid strains, pol3–5DV cells were first transformed with pBL304, a plasmid containing POL3 on a URA3 CEN plasmid, which was constructed by Peter Burgers and is described in [7]. The MSH6 gene in such strains could subsequently be deleted with the strain maintaining viability. Diploid strains were constructed by mating of two haploid strains followed by selection on synthetic dextrose (SD) medium lacking methionine and leucine [120]. Diploid msh6 pol3–5DV strains were constructed by mating of the two haploids, one being MSH6, and the other containing the POL3 plasmid pBL304 rescuing the msh6 pol3–5DV genotype, followed by deletion of the second MSH6 allele with a hphNT1 marker [119] followed by selection for loss of the POL3 plasmid. Oligos used for transformation were gel purified (Eurofins MWG Operon); the sequences are listed in S6 Table. Reversion analysis was performed as described [96]. Reversion rates and Confidence Intervals were calculated [96] using the program Salvador [121–123]. When multiple reversion experiments were done for a given genotype, the median value was used for subsequent analysis. The reversion rates of heterozygous and homozygous proofreading-deficient strains (Table 2) were considered to be different if the 83% confidence levels did not overlap [53]. Oligo transformation was essentially as described previously [25,82–84]. An overnight culture of a strain was diluted 1:50 in YPAD [120], incubated with shaking at 30° to an OD600 of 1.3–1.5, washed twice with cold H2O, and once with cold 1 M sorbitol. After the final centrifugation, all solution was removed from the cells and a volume of cold 1 M sorbitol equal to that of the cell pellet added to resuspend the cells. For a typical transformation, 200 pmol of a Trp oligo was added to 200 μl of this cell suspension in a 2-mm gap electroporation cuvette, and the mixture electroporated at 1.55 kV, 200 Ω, and 25 μF (BTX Harvard Apparatus ECM 630). Immediately after electroporation, the cell suspension was added to 5 ml of YPAD, and the cells incubated at 30° with shaking for 2 h. Cells were then centrifuged, washed with H2O, and plated on SD medium lacking tryptophan [120] to select transformants. In order to control for some of the variability we observed in transformation efficiencies, one portion of each cell suspension was electroporated with the Trpwt40 oligo, which reverts all of the strains via a centrally-located mismatched base and is thus not subject to any proofreading effect.

10.1371/journal.pcbi.1005930

Biophysical network modeling of the dLGN circuit: Effects of cortical feedback on spatial response properties of relay cells

Despite half-a-century of research since the seminal work of Hubel and Wiesel, the role of the dorsal lateral geniculate nucleus (dLGN) in shaping the visual signals is not properly understood. Placed on route from retina to primary visual cortex in the early visual pathway, a striking feature of the dLGN circuit is that both the relay cells (RCs) and interneurons (INs) not only receive feedforward input from retinal ganglion cells, but also a prominent feedback from cells in layer 6 of visual cortex. This feedback has been proposed to affect synchronicity and other temporal properties of the RC firing. It has also been seen to affect spatial properties such as the center-surround antagonism of thalamic receptive fields, i.e., the suppression of the response to very large stimuli compared to smaller, more optimal stimuli. Here we explore the spatial effects of cortical feedback on the RC response by means of a a comprehensive network model with biophysically detailed, single-compartment and multicompartment neuron models of RCs, INs and a population of orientation-selective layer 6 simple cells, consisting of pyramidal cells (PY). We have considered two different arrangements of synaptic feedback from the ON and OFF zones in the visual cortex to the dLGN: phase-reversed (‘push-pull’) and phase-matched (‘push-push’), as well as different spatial extents of the corticothalamic projection pattern. Our simulation results support that a phase-reversed arrangement provides a more effective way for cortical feedback to provide the increased center-surround antagonism seen in experiments both for flashing spots and, even more prominently, for patch gratings. This implies that ON-center RCs receive direct excitation from OFF-dominated cortical cells and indirect inhibitory feedback from ON-dominated cortical cells. The increased center-surround antagonism in the model is accompanied by spatial focusing, i.e., the maximum RC response occurs for smaller stimuli when feedback is present.

The functional role of the dorsal lateral geniculate nucleus (dLGN), placed on route from retina to primary visual cortex in the early visual pathway, is still poorly understood. A striking feature of the dLGN circuit is that dLGN cells not only receive feedforward input from the retina, but also a prominent feedback from cells in the visual cortex. It has been seen in experiments that cortical feedback modifies the spatial properties of dLGN cells in response to visual stimuli. In particular, it has been shown to increase the center-surround antagonism for flashing-spot and patch-grating visual stimuli, i.e., the suppression of responses to very large stimuli compared to smaller stimuli. Here we investigate the putative mechanisms behind this feature by means of a comprehensive network model of biophysically detailed neuron models for RCs and INs in the dLGN and orientation-selective cortical cells providing the feedback. Our results support that the experimentally observed feedback effects may be due to a phase-reversed (‘push-pull’) arrangement of the cortical feedback where ON-symmetry RCs receive (indirect) inhibitory feedback from ON-dominated cortical cell and excitation from OFF-dominated cortical cells, and vice versa for OFF-symmetry RCs.

Visual signals from the retina pass through the dorsal geniculate nucleus (dLGN), the visual part of thalamus, on the way to the visual cortex. However, this is not simply a one-way flow of information: cortical cells feed back to both relay cells (RCs) and interneurons (INs) in the dLGN and thus shape the transfer of visual information in the circuit [1–6]. Although there is no broad consensus about the effects of cortical feedback on sensory processing, there are many experimental studies that provide insight into its potential roles [7–20]. For example, cortical feedback has been observed to switch the response mode of RCs between tonic and burst modes [21, 22] and to synchronize the firing patterns of groups of dLGN cells [17]. Further, the studies have reported both enhanced and reduced responses of dLGN neurons from cortical feedback, and the functional role of cortical feedback is still debated [3, 23, 24]. One line of inquiry has addressed the question of how cortical feedback modulates the receptive-field properties of RCs. Cortical feedback was early shown to affect the length tuning of RC responses [12], and a series of studies from Sillito and co-workers have investigated how cortical feedback influences the RC responses to flashing spots and patch gratings, i.e., circular patches of drifting gratings [4, 13, 15, 16, 18, 19]. Retinal ganglion cells (GCs) provide the feedforward input to the dLGN circuit, and the receptive fields of both GCs and RCs have a roughly circular shape where an excitatory center is surrounded by an inhibitory surround [25–27]. For a flashing-spot stimulus the maximum response occurs for a spot centered on the receptive field which exactly covers the receptive-field center [27]. When the spot size is gradually increased to also stimulate the inhibitory surround, the response is gradually reduced until the entire surround is also covered. This phenomenon is referred to as center-surround suppression, and it is known that such suppression is increased for RCs compared to the GCs that provide the dominant feedforward input [27]. A part of this increased suppression likely stems from feedforward mechanisms in the dLGN circuit, i.e., a broad feedforward retinal input to LGN interneurons, in turn providing increased feedforward surround inhibition to the RCs [27, 28]. Increased center-surround suppression implies that the neurons are less responsive to broad visual stimuli and instead more tuned to narrow stimuli or sharp spatial variations in the visual scene. Thus dynamical tuning of this suppression may be a mechanism for the nervous system to adapt to changing light conditions and viewing demands to create an efficient representation of the stimulus [29]. Although the receptive fields of dLGN cells appear largely determined by the feedforward retinogeniculate input, corticothalamic feedback has been shown to increase the inhibitory surround, i.e., increase the suppression to very large stimuli [4, 12, 13, 15, 16, 19, 30]. Other studies have reported enhanced responses of dLGN neurons [18, 30, 31] when using smaller stimuli. Interestingly, cortical feedback has been experimentally observed to increase the surround suppression both for flashing spots [32] and patch gratings [4, 19], though, the increase has been found to be larger for patch gratings [2, 4]. The topic of the present modeling study is to investigate what aspects of the thalamocortical loop, and in particular what type of cortical feedback pattern, may underlie these observed changes in RC center-surround antagonism. While the use of computational modeling to study the effect of cortical feedback on visual processing is not new, previous projects have investigated feedback effects on the temporal processing of RCs [33–38]. Modeling studies of spatial aspects have to our knowledge been limited to relatively simple firing-rate models [39, 40] where, for example, dLGN INs have not been explicitly included. The focus in [39] was on exploring cortical feedback effects on observed effects of RC responses to discontinuity in orientations in gratings in bipartite stimuli. In [40] the extended DOG (eDOG) model was introduced, allowing for analytical explorations of effects of cortical feedback in certain settings, i.e., with certain combinations of excitatory and (indirect) inhibitory feedback from ON- and OFF-center cortical cells onto RCs. In that study a preliminary use-case showed that a phase-reversed (‘push-pull’) arrangement of cortical feedback where ON-center RCs receive direct excitation from OFF-driven cortical cells and balanced indirect inhibitory feedback from ON-driven cortical cells, may provide increased center-surround antagonism. Here we instead consider a biophysically detailed model where RCs and INs, as well as orientation-selective layer-6 pyramidal cortical cells (PYs), are explicitly included. The model is an extension of a recently developed network model of the feedforward part of the dLGN circuit [41]. The neuron models include a host of Hodgkin-Huxley type active conductances [42–44], and an important feature is the multicompartment IN model that incorporates both axonal and triadic inhibition of RCs [45]. Another key element of our model circuit is the explicit incorporation of both ON-symmetry and OFF-symmetry cells which, unlike for the rate-based eDOG model [40], allows exploration of a wide range of putative synaptic patterns for the feedback from cortical cells to RCs and INs, i.e., both same symmetry (ON to ON, OFF to OFF) and cross-symmetry (ON to OFF, OFF to ON). By comparing results from a wide range of feedback patterns, we find that our results support that a phase-reversed arrangement of the cortical feedback seems most effective in increasing the center-surround antagonism observed both for flashing spots and, even more significantly, for patch gratings. The core of the network model comprises two-dimensional grids of synaptically connected dLGN and cortical neurons of ON and OFF receptive-field arrangements. The network is driven by dLGN neurons that receive spikes encoding visual input from the retina. The network includes populations of retinal ganglion cells (GC), dLGN RCs and INs, and PYs of layer 6 in the primary visual cortex (Fig 1). Each layer is scaled to span a monocular patch of 10 deg × 10 deg in the visual field and contains 10 × 10 neurons of each symmetry type (ON/OFF), except in the case of dLGN INs for which there are 25 per symmetry type (20% of the total number of dLGN cells [46]). Based on the wiring rules of the cat dLGN, it has been estimated that a 1 deg × 1 deg patch of the dLGN contains about 10 RCs of one symmetry type on average at an eccentricity of 7 deg [47]. Thus, one simulated RC in our model would correspond to about 10 RCs of the cat dLGN. In the tuning of the model, we have chosen model parameters giving GC and RC responses similar to the cat experiments described in [27, 28]. Here the recordings were done on cells with receptive fields centered in areas of the visual field some distance away from the center of gaze (area centralis in cat). Retinal GCs have a circularly symmetric center-surround receptive field that is inherited by dLGN RCs through one-to-one excitatory synapses as shown for cells of the ON and OFF pathways in Fig 1. In these receptive fields, the center and surround present an antagonistic push-pull arrangement [48]. A bright stimulus confined to the center of the ON-cell receptive field or a dark stimulus placed on the surround of the receptive field evoke a depolarization of the ON cell. By contrast, an ON cell is hyperpolarized by projecting either a dark stimulus to the center of the receptive field or a bright stimulus to the surround. The opposite behavior applies for OFF-center cells. The feedforward elements of the dLGN are the same as in [41]. LGN INs receive input from four retinal ganglion cells via the triadic synapses and the proximal IN dendrites. RCs receive axonal inhibition through the IN axon and triadic inhibition by the IN dendrites at the triadic synapses, resulting in fast inhibition. The cortical populations of PYs receive strong input from an elongated area of three RCs of the same symmetry and weak input from an adjacent row of three RCs of the opposite symmetry. PYs come in two different orientation-selectivity variants: horizontally-selective or vertically-selective. Further, each of these two cortical populations also come with ON and OFF symmetry making a total of four distinct cortical populations. This is a simplified representation of the thalamocortical loop as it neglects that the strongest thalamic input to primary visual cortex arrives in layer 4 while the feedback inputs to dLGN cells come from cells in layer 6. The models for the dLGN and cortical neurons are all biophysically detailed in the sense that they include a variety of Hodgkin-Huxley type active conductances explicitly reproducing generation of action potentials. The GC spiking mechanism is not modeled explicitly, instead this input is modeled by means of phenomenological filter models as in [41]. Conductance-based synapses were assumed, i.e., I syn ( t ) = w f syn ( t - t s - t Δ ) ( V - E syn ) θ ( t - t s - t Δ ) , (10) for a presynaptic spike arriving at ts. Here the weight w is the maximal conductance of the postsynaptic receptors and Esyn is the reversal potential. fsyn is the temporal envelope of the synaptic conductance modeled as the difference between two exponential functions specified by time constants τrise and τdecay (Eqs. 6.4–6.6 in [56]). tΔ is the conduction time delay from the generation of the presynaptic spike to the initiation of the postsynaptic response and was set to a fixed value of 1 ms for all synaptic connections. Action potentials of RCs, INs and PYs were detected by upward somatic voltage crossings at −10 mV. While AMPA receptors mediate all excitatory connections in this model, GABAA receptors mediate all inhibitory synaptic interactions. Parameters of synaptic connections are shown in Table 3. Parameters of retinogeniculate and intrathalamic connections remain similar to those presented in [41]. An exception is the GC input to the IN part of the triad, for which we reduced the synaptic weight to compensate for the added excitatory input from corticothalamic connections not present in the previous model [41]. The results are divided into two distinct parts. In the first part results for the feedforward part of the circuit is presented, mainly to validate the model against previous findings in the literature. The studies of the effects of cortical feedback are presented in the second part where the feedforward circuit explored in the first section is used as a starting point. Before studying the effects of cortical feedback on the RC response specifically, we describe the feedforward response of the different cell types in the network model when the cortical feedback is deactivated, i.e., corticothalamic synapses from PYs to dLGN relay cells (RCs) and interneurons (dLGN INs) are disconnected. In this situation the RC response is driven only by excitation from its GC afferents and feedforward inhibition from INs. After exploring above the feedforward response of the different cell types in the network model, we now move on to investigate how cortical feedback to the dLGN circuit affects the spatial response properties of RC cells. This will depend on the detailed corticothalamic connectivity pattern which is not yet experimentally fully resolved. In the next sections, we thus present simulation results for the different alternatives considered in Fig 3. In the present paper we have developed a mechanistic network model of the thalamocortical system with explicit representations of LGN cells (relay cells (RCs) and interneurons (INs)) and orientation-selective layer 6 simple cells placed on two-dimensional spatial grids. The LGN and cortical cells are represented by biophysical neuron models based on the cable equation and Hodgkin-Huxley type active conductances. The input of the model is provided by retinal ganglion cells (GCs) implemented by means of descriptive filter models. The main focus of the study has been exploration of the effects of cortical feedback on the spatial responses of RCs to flashing-spot and patch-grating stimuli as this has received substantial experimental attention [2, 4, 19, 32]. Comparison of our simulation results with previous experimental findings supports the notion that a ‘push-pull’ (phase-reversed) organization of cortical feedback [62], i.e., ON-center RCs receive direct (monosynaptic) excitatory feedback from OFF-dominated cortical cells and indirect inhibitory feedback from ON-dominated cortical cells, provides a dual effect that simultaneously amplifies excitatory responses in the receptive-field center and inhibitory responses in the receptive-field surround of RCs [18, 83]. As a result, the center-surround antagonism of RCs is amplified by cortical feedback and the maximum RC response occurs for reduced stimulus sizes. The combination of these two effects, excitatory in the receptive-field center [18] and inhibitory in the receptive-field surround [4, 19], may be understood as complementary functions that dynamically sharpens the spatial focus of the receptive field and increase their spatial resolution. The main results from our model study were the area-response curves for flashing-spot and patch-grating stimuli, a commonly used measure of visual responses for cells in the early stages of the visual system [2, 4, 18, 19, 27, 28, 40, 80, 81]. We first considered the case with a rough balance between excitatory and inhibitory feedback so that the main effect of cortical feedback is on the shape of the area-response curves, not the magnitude (Figs 10 and 14). With a phase-reversed feedback arrangement a clear feedback-induced increase in surround suppression is observed both for flashing spots and patch gratings (Fig 10), as quantified by the center-surround antagonism coefficient α (Eq 12) (Table 4). Such a feedback-induced increase of surround suppression has been observed in experiments with both flashing spots [32] and patch gratings [4, 19], although the effect appears more significant for patch gratings [2, 4]. Our model results gave a larger increase of surround suppression for the patch-grating stimulus, but not as prominent as the increase reported by Sillito et al. [4]. With the same choice of parameters, a phase-matched feedback arrangement resulted in very little change in surround suppression for both types of stimulus (Fig 14). Increased surround suppression implies that RC cells in relative terms become more responsive to small stimuli and, thus, the cell more selective in spatial tuning. An additional effect of the phase-reversed feedback is the shrinking of the stimulus size giving the maximum responses in the area-response curves, clearly observed for the phase-reversed feedback, but largely absent for phase-matched feedback (Figs 10 and 14). We next did a parameter sweep, i.e., investigated the effects of cortical feedback on the RC area-response curves for a wide range of different synaptic weights between PYs and dLGN neurons and for the different spatial feedback kernels (1 × 1 and 2 × 2) (Figs 16–19). The results for our two key area-response curve measures, the stimulus diameters giving the largest response and the center-surround antagonism coefficient α, were summarized in Fig 20. A first observation was that both for flashing-spot and patch-grating stimuli, the phase-reversed and phase-matched cases gave very different dependency of the center-surround suppression, i.e., center-surround antagonism coefficient α, on synaptic weights (Fig 20A). For the phase-reversed case, high values of the center-surround antagonism coefficient were achieved by those parameter combinations that exert both strong excitation and (indirect) inhibition to the RC (towards the bottom right corner). Here the ON-center inhibition and the OFF-center excitation both contribute to increasing the surround suppression. Thus large values of the surround suppression can be achieved even when excitatory and inhibitory effects are roughly balanced [18, 83]. In contrast, for the phase-matched case, feedback-induced increases in the center-surround antagonism coefficient α required the inhibition to dominate the excitation. This reflects that the effects of ON-center inhibition and ON-center excitation in the feedback tend to cancel each other out. This is in accordance with the observation in Figs 10 and 14 where the area-response curve for the ‘inhibition-only’ case was seen to represent an intermediate case between the phase-reversed and phase-matched situations. When comparing the different spatial feedback patterns for the phase-reversed case, the 2 × 2 feedback pattern was seen to be more effective in increasing surround suppression in the RC response than the 1 × 1. Incidentally, a spatially widespread feedback pattern has been suggested by anatomical studies of the innervation pattern of corticothalamic axons in the dLGN [63]. For flashing-spot stimuli only small variations in the diameters producing the maximal RC response were observed when varying the synaptic weights (Fig 20B). However, for patch-grating stimuli a reduction was observed in the maximum-response diameter was observed when one or both types of cortical feedback were present. Other modeling studies have also investigated the effect of cortical feedback on spatial processing of RCs with different stimulus patterns [39, 40]. The focus in [39] was on exploring the role of cortical feedback in modulating RC responses to discontinuity in orientations in gratings in bipartite stimuli. In [40] the extended DOG (eDOG) model was introduced, allowing for analytical explorations of effects of cortical feedback in certain settings, i.e., with certain combinations of excitatory and (indirect) inhibitory feedback from ON- and OFF-center cortical cells onto RCs. There a preliminary use-case showed that a phase-reversed (‘push-pull’) arrangement of cortical feedback where ON-center RCs receive direct excitation from OFF-driven cortical cells and balanced indirect inhibitory feedback from ON-driven cortical cells, may provide increased center-surround antagonism. Our biophysical model and the above-discussed firing-rate models represent opposite extremes in terms of biological detail in LGN circuit models [86]. Models at an intermediate complexity level where the cells are modeled as integrate-and-fire neurons have also been used to explore cortical feedback effects on LGN cell [33–36]. However, these have focused on temporal response properties such as feedback-induced spike synchronization [35], long-lasting correlations [36] and effects of feedback on visual latency [33], not the spatial properties which has been the main topic here. An obvious next application of the present model would be to explore temporal response properties of LGN cells and, in particular, how these are affected by various types of cortical feedback. One line of inquiry would be to explore the relative roles of feedforward and feedback connections in shaping the temporal receptive fields of LGN cells, analogous to the questions addressed by the firing-rate models in [37] and [38]. Another line of research would be on studying spike synchronicity and correlations as addressed earlier with integrate-and-fire models [35, 36]. A third line could be to explore in detail how the temporal structure of the PSTH, and in particular the ‘interval histogram’ of RC spikes, is affected by feedback [34]. In addition to feedback from cortex, both RCs and INs receive inhibitory feedback from neurons in the thalamic reticular nucleus (TRN) [5]. TRN neurons are thought to play a key role in the process of sleep spindle oscillations generated within the thalamic circuitry [42, 43]. The TRN also contributes to the control of visual attention and awareness [87], but the effects on procession of visual signals remain poorly understood [88]. TRN neurons do not receive direct input from the retina as LGN INs, instead they receive feedforward visual signals from collaterals of geniculocortical axons. TRN neurons also receive cortical feedback through corticothalamic axons, and their synapses on RCs are situated in close proximity to those of corticothalamic axons [1]. Given this organization of synaptic connections and its position within the network, TRN cells are likely to influence the transfer of visual information in a different manner than LGN INs. Modeling studies exploring the putative role of TRN neurons on visual processing have already been pursued [89], and the present biophysical model could be extended to include also such neurons when more is known about these neurons and their possible role in visual processing. The present model assumes static synapses while a number of studies have demonstrated short-term plasticity in different synapses of the thalamocortical circuit, i.e., short-term depression at the retinogeniculate [90, 91] and geniculocortical [92, 93] synapses, as well as in the feedback connection from cortex to INs [94]. In contrast, the feedback connection from cortex to RCs appears to be facilitating [90, 95]. Such plasticity opens up for an even richer dynamical repertoire of the circuit, and would be an interesting topic for a future study using the present model with static synapses as a starting point. In particular, it would be interesting to explore if short-term synaptic plasticity could affect our prediction that phase-reversed cortical feedback is the most effective mechanism for increasing center-surround antagonism. dLGN cells have two different response modes, burst and tonic, suggested to relate of the animal [5, 96, 97]. Modulatory inputs from other parts of the brain may switch between these modes by shifting the baseline membrane potentials of RCs and INs. Tonic firing has been suggested to be more suitable for transferring visual information because it avoids nonlinear distortions created during burst firing, while burst firing was suggested to be best suited as an ‘alarm clock’, i.e., rapid stimulus detection [5]. Recent studies have demonstrated, however, that thalamic bursts can also contribute to sensory processing [98–101]. In the current study, our RC and IN models were based on data from dLGN neurons that rested on relatively depolarized membrane potentials, -60 mV and -63 mV, respectively, and fired predominantly in the tonic mode (Fig 2). An exploration of the functional roles of the two firing modes, and putative switches between them, would be another natural extension of the present work. The present model of primary visual cortex is obviously simplified. Cells in layer 4 of cortex are the main targets of projections from RCs, while the feedback from cortex to dLGN comes from cells in layer 6. Even though there are also projections from RCs to layer-6 cells, there are likely cross-layer processing in cortex that affects the thalamocortical feedback loop and difficult to capture by a single-layer cortex model. Despite the model simplicity, the pyramidal-cell receptive fields produced by our network model (Figs 4, 12 and 15) are nevertheless seen to resemble the receptive fields of simple cells which also has been observed in layer 6 of cat visual cortex [102]. Thus the error introduced by our simplified cortical network model could be modest for the present application, but this needs further exploration when thalamocortical models including more comprehensive cortical circuitry becomes available. Further, there are several neural mechanisms that our simplified model of cortical orientation tuning does not account for, such as recurrent cortical excitation or horizontal inhibitory connections [58, 103–105], which can amplify a weak orientation bias. Although the area-response curves of cortical cells to the patch grating in Figs 8 and 9 showed a marked difference for gratings at preferred and non-preferred orientations, stimuli presented at non-preferred orientations did not suppress cortical response to the background rate as observed experimentally in some cells [106]. A stronger orientation selectivity of the cortical cells would likely affect the feedback-induced changes in RC response, but how, and to what extent, remains to be explored. While one option for extending the present model would be to add more neuron types to a single-layer cortex model, it might be tempting to aim to connect the present biophysically detailed model for the dLGN circuit with an equally detailed model for the primary visual cortex. However, at present such models are lacking, and a comprehensive model based on biophysical neuron models including both the dLGN and, say, V1 would anyway be computationally extremely demanding. An alternative could be to instead model V1 dynamics with simpler neuron models such as the Potjans-Diesmann network model based on integrate-and-fire neurons [107]. Experimental studies of cortical feedback effects on response properties in the dLGN have been ongoing for at least 40 years (see, e.g., [7]). However, a recurring challenge has been to reversibly remove cortical feedback in a controlled manner to compare physiological responses of dLGN cells with and without cortical feedback. Both cooling [11] and pharmacological manipulations [18] have been used. However, the advent of optogenetics now offers unprecedented opportunities for highly-controlled activation or deactivation of individual cell types. In [108] the role of layer-6 cells in providing gain control for the visual responses in the upper layers of mouse visual cortex was studied by such techniques. A similar study where visual responses of dLGN cells are measured while the corticothalamic cells in layer 6 are selectively activated or deactivated by photostimulation, would be most welcome for testing predictions of the present model.

10.1371/journal.pntd.0002785

A Dual Drug Sensitive L. major Induces Protection without Lesion in C57BL/6 Mice

Leishmaniasis is a major health problem in some endemic areas and yet, no vaccine is available against any form of the disease. Historically, leishmanization (LZ) which is an inoculation of individual with live Leishmania, is the most effective control measure at least against cutaneous leishmaniasis (CL). Due to various reasons, LZ is not used today. Several live attenuated Leishmania have been developed but their use is limited. Previously, we developed a transgenic strain of L. major that harbors two suicide genes tk and cd genes (lmtkcd+/+) for use as a challenge strain in vaccine studies. These genes render the parasite susceptible to Ganciclovir (GCV) and 5-flurocytosine (5-FC). The dual drug sensitive strain of L. major was developed using gene targeting technology using a modified Herpes Simplex Virus thymidine kinase gene (hsv-tk) sensitive to Ganciclovir antibiotic and Saccharomyces cerevisae cytosine deaminase gene (cd sensitive to 5-flurocytosine) that were stably introduced into L. major chromosome. BALB/c mice inoculated with lmtkcd+/+ developed lesions which upon treatment with GCV and 5-FC completely healed. In the current study, the transgenic lmtkcd+/+strain was assessed as a live vaccine model to determine the time necessary to develop a protective immune response. C57BL/6 mice were inoculated with the transgenic lmtkcd+/+strain, and treated at the time of inoculation (day0) or at day 8 after inoculation. Immunized animals were challenged with wild-type L. major, and complete protection was induced in mice that were treated at day 8. The results show that in contrast to leishmanization, in group of mice inoculated with a dual sensitive L. major development and persistence of lesion is not necessary to induce Th1 response and protection.

Leishmaniasis is still a major health problem in some endemic foci, yet no vaccine is available against any form of leishmaniasis. It is a general belief that recovery from cutaneous leishmaniasis (CL) is accompanied with long life protection. An inoculation of live pathogenic L. major into healthy individuals to induce lesion similar to CL is called Leishmanization (LZ). Historically LZ showed to be the most effective control tool against CL. One of the drawbacks and reason for discontinuation of LZ was lesion development, which rarely lasts long. Treatment of CL is not an easy task. One line of development of an effective vaccine against leishmaniasis, a transgenic strain of L. major harboring two suicide genes tk and cd genes (lmtkcd+/+), was developed and previously checked in BALB/c mice. In this study, C57BL/6 mice were inoculated with transgenic lmtkcd+/+strain; the rate of protection, parasite burden and the type of immune response were checked, and the results showed that complete protection induced by inoculation of lmtkcd+/+strain if treatment is initiated on day 8 post inoculation.

Cutaneous leishmaniasis (CL) manifests as a localized self-healing lesion(s) that in rare cases develops to a non-healing lesion. If non-healing lesions develop, they are extremely difficult to treat with current therapies [1]. Control measures for leishmaniasis such as vector and/or réservoir control are not always practical, especially in remote endemic areas with limited resources. Efficacy of available drugs for leishmaniasis especially for CL is not acceptable and resistant is emerging [2], [3], [4], [5], [6]. Leishmanization (LZ) involves inoculating of individuals with live virulent Leishmania major to induce a single lesion that mimics a natural infection but with the lesion located at a predetermined site. Upon healing, the leishmanized individuals are protected against natural infection. LZ has been shown to be the most effective control measure at least against CL but the practice has been discontinued except on a limited scale in Uzbekistan. Primarily this is due to the development of chronic lesions that require medical intervention [7], [8], [9]. Despite ample evidence that development of an effective vaccine against leishmaniasis is possible there is still no vaccine available against any form of human leishmaniasis [10], [11], [12], [13]. One approach is to derive attenuated live vaccine strains of Leishmania through genetic manipulation to develop a parasite strain which has no virulence or a limited pathogenicity. A number of genetically manipulated Leishmania strains have been developed and studied in animal models with controversial results [14], [15], [16], [17], [18]. Previously, we developed a transgenic strain of L. major (tk+/+–cd+/+)[lmtkcd+/+] harboring two suicide genes tk and cd genes that confer susceptibility to GCV and 5-FC,.as a challenge strain for vaccine studies. When BALB/c mice were inoculated in the flank with lmtkcd+/+, lesions developed at the site of inoculation, upon treatment with GCV and 5-FC complete healing occurred [16]. To extend these studies lmtkcd+/+ was used to determine whether persistent infection is required for induction of a protective immune response against subsequent L major infection. The lmtkcd+/+ promastigotes were inoculated into C57BL/6 mice and the inoculated mice were treated at set times with GCV to clear the infection. The mice were then challenged with wild type L major. Long term (3 months) complete protection against challenge with wild type L major was achieved with as little as 8 days vaccination time demonstrating that persistent infection is not required for complete protection. The ethical committee; Institutional Animal Care and Research Advisory Committee of Pasteur Institute of Iran, Education Office dated January, 2008, based on the Specific National Ethical Guidelines for Biomedical Research issued by the Research and Technology Deputy of Ministry of Health and Medicinal Education of Iran, issued in 2005, approved the protocol. The L. major promastigotes (MHOM/IR/76/ER) used and from which the transgenic lmtkcd+/+parasites were derived, this L. major is the same isolate which was used for mass leishmanization, preparation of old world experimental vaccine and the Leishmania used for the skin test. Promastigotes were cultured in M199 medium (Life Technologies, Inc.) supplemented with 10% heat inactivated fetal calf serum (Gibco BRL) and 25 mM HEPES (Gibco BRL), pH 7 at 26°C. The parasite virulence was maintained by passage in BALB/c mouse. Female C57BL/6 mice, 6–8 week-old were purchased from the Animal Breeding Facility Centre (ABFC) of Pasteur Institute, Karaj, Iran. The animals were maintained in the animal facility of the Pasteur Institute of Tehran. The experiments were carried out according to the guidelines of Ethic Committee for Human use of Laboratory Animals, Pasteur Institute, Tehran, Iran. Mice were inoculated subcutaneously (SC) at the right hind footpad with 2×106 stationary phase promastigotes of either L. major (MHOM/IR/76/ER) wild type (WT) or the transgenic lmtkcd+/+parasites in 50 µl PBS. The mice inoculated with lmtkcd+/+were divided into 3 groups and treated with a combination of GCV and 5Fcyt, 100 mg/Kg, intra-peritoneally (IP) either at the time of parasite inoculation (day 0), at day 8 after inoculation or for the control group which was left untreated. The dosage of the drugs used in this study was based on our previous study (17). The lmtkcd+/+ inoculated groups were challenged in the left footpads with 2×106 virulent WT L. major SC at 3 weeks after the end of the treatment period. The lesion development was recorded by weekly measurement of the footpad thickness at the site of inoculation using a metric caliper up to 12 weeks after inoculation. Parasite burden was quantified once at week 10 after inoculation of the mice with either L. major wild type or with lmtkcd+/+ and again 5 weeks after the challenge with wild type L. major (2–5 mice per group). The parasite burden in the spleen and draining lymph nodes were determined using limiting dilution analysis. To enhance sensitivity, 2-fold dilutions of the samples (up to 1/100) were used. Delayed-type hypersensitivity (DTH) reaction was checked prior to challenge by injection of freeze-thawed (FT) Leishmania major (2×106 promastigotes in 50 µl per injection) into the contralateral uninfected hind footpad. FT L. major promastigotes were prepared by repeating a freeze (−196°C)/thaw (37°C) cycle ten times. Footpad swelling was measured using a metric caliper at 24, 48 and 72 h after injection. Three mice from each group were sacrificed before and at 5 weeks after challenge inoculation, spleens were removed and cells cultured in complete RPMI-1640 medium in the presence or absence of 20 µg/well of Soluble Leishmania Antigens (SLA, 107 Leishmania promastigotes/ml equal to 100 µg/ml) or Concavalin A (ConA;10 µg/ml) or without stimulation as a control. The levels of IFN-γ and IL-4 at weeks 5 and 10 post inoculation with lmtkcd+/+ or WT L.major and 5 weeks after challenge were determined in the supernatant collected from spleen cell culture (5 mice per group). Briefly, single spleen cell suspension was prepared, cultured and re-stimulated either with SLA (100 µg/ml) or Con A (10 µg/ml). The supernatant was collected at 72 h. Then, the levels of IFN-γ and IL-4 were titrated using ELISA method according to the manufacturer's instruction (Bender Medsystems, Gmbh, Austria). The sensitivity of the ELISA kits was 3 pg/ml for IL-4 and 7.5 pg/ml for IFN- γ. At week 5 after challenge, different groups of mice were tail bled and the levels of anti-Leishmania IgG1 and IgG2a Abs were checked by ELISA. All experiments were done in triplicates and the data was expressed as means ± S.E.M. The data was analyzed by one-way ANOVA followed by Tukey's test using SPSS V.13 software. P value<0.05 was considered as statistically significant. C57BL/6 mice were inoculated SC with live wild type (WT) L. major parasites or lmtkcd+/+ parasites and were either left untreated or treated with GCV/5-FCy at day 0 or day 8. Lesion development was followed by the measurement of footpad thickness. Following challenge with L. major, the protection rate and the immune responses generated were assessed. C57BL/6 mice inoculated with lmtkcd+/+ or WT parasites and left untreated developed a similar lesion size which was cured around week 8–9. In contrast, no lesion was developed in the group of mice which was inoculated with lmtkcd+/+ and received GCV/5-FCyt treatment at day 0 or day 8. The group of mice inoculated with WT L. major which was treated with GCV/5-FCyt developed a lesion similar to the untreated group of mice (Fig. 1A). The draining lymph nodes (LN) and spleen parasite burden was measured at week 10 post-inoculation (5 mice/group). The results showed no difference in the number of parasite in spleen and LN's in groups of mice inoculated with WT L. major and the group which was inoculated with lmtkcd+/+ and received no treatment, the parasite burden of spleen at week 10 after inoculation is presented in Fig. 1B and only parasite burden of spleen at week 5 after challenge with WT L. major is presented in Fig. 2B. At weeks 5, 10 post-inoculation and week 5 post challenge mice (5 per group) were sacrificed and spleens were removed. A single cell suspension of spleen was prepared and cultured in the presence of either SLA (100 µg/ml), Con A (10 µg/ml) or without additional stimulation, lymphocyte transformation test (LTT) was done at 72 hours and the results showed a significantly (p<0.05) stronger LTT in group of mice with history of L. major infection and the group which was inoculated with lmtkcd+/+parasites and treated on day 8 than the group of mice inoculated with lmtkcd+/+parasites and treated on day 0 (Fig. 1C). The supernatants were collected and the levels of IFN-γ were titrated (Fig. 1D). Similar levels of IFN-γ were produced in spleen cells of group of mice inoculated with WT L. major and the group of mice inoculated with lmtkcd+/+. The level of IL-4 production was low and similar in group of mice inoculated with wild-type L. major or inoculated with lmtkcd+/+ at week 16 post infection (data not shown). To assess whether groups of C57BL/6 mice inoculated with lmtkcd+/+ parasites are protected against WT L. major challenge, at week 5–6 post inoculation (3 weeks after the end of treatment upon commencing time), the groups of mice which received lmtkcd+/+ and were treated on day 0 or 8 were challenged with L. major. As well, a group of mice which had healed spontaneously after L. major infection and a group of naïve mice were inoculated with L. major as controls. The results showed that the group of mice which was inoculated with lmtkcd+/+ parasites and treated with GCV/5-Fcyt on day 8 and then challenged with WT at week 6, did not develop any lesion or swelling similar to the group of mice challenged with L. major after previously self-healing lesion. In contrast, the group of mice which was inoculated with lmtkcd+/+ and treated at the same time (Day 0) with GCV/5-Fcyt and the group of naïve mice inoculated with L. major developed lesions (Fig. 2A). The parasite burden was quantified in draining LN at week 5 post-challenge with L. major, as shown in Fig. 2B. The number of parasites isolated from the group of mice which was inoculated with lmtkcd+/+ and treated at day 8 with GCV/5-FCyt and the group of mice which had previously self-healed following L. major infection was significantly (p<0.05) lower than the group which was inoculated with lmtkcd+/+ and treated at the same time (day 0) and the group of naïve mice which were inoculated with L. major for the first time. The number of parasites was very low in the groups of mice inoculated with either lmtkcd+/+ and treated at day 8 or inoculated with lmtkcd+/+ and not treated or the group of mice with history of L. major infection or the group of mice which were inoculated with lmtkcd+/+ and treated at day 0, no significant difference was seen between the number of parasite in these groups. DTH was done in different group of mice by injection of freeze-thawed (FT) Leishmania major (2×106 promastigotes in 50 µl per injection) into the contra lateral uninfected hind footpad. The results are presented in Fig. 2C, a similar strong DTH response is seen in group of mice inoculated with WT L. major, or inoculated with lmtkcd+/+ and treated with GCV/5-FCy on day 8 or left untreated, a low DTH response was seen in groups of mice inoculated with lmtkcd+/+ and treated with GCV/5-FCy on day 0 or uninfected naïve mice. At week 10 after inoculation (before challenge) and 5 weeks after challenge, the splenocytes were cultured, stimulated in vitro with either SLA (100 µg/ml), or Con A (105 µg/ml), or left unstimulated. LTT was done and the culture supernatants were collected at 72 hours and the level of IFN-γ and IL-4 was titrated using ELISA method. A significantly (p<0.05) stronger LTT was seen in mice with history of L. major infection and the group which was inoculated with lmtkcd+/+ parasites and treated on day 8 than the group of mice inoculated with lmtkcd+/+ parasites and treated on day 0 (data not shown). The level of IFN-γ was significantly higher in groups of mice inoculated with WT L. major or inoculated with lmtkcd+/+ and treated with GCV/5-FCy on day 8 or left untreated (Fig. 2D). The level of IL-4 was similar in all the groups (Fig. 2E). Serum samples were collected at 5 weeks after challenge, the results are presented in Fig. 2F, as shown a significantly (P = 0.002) higher anti-L.major IgG antibodies were seen in the group of mice with history of L. major lesion or group of mice inoculated with lmtkcd++ and treated with GCV/5-FCy on day 8, in comparison with the group of naïve mice or group of mice inoculated with lmtkcd+/+ and treated with GCV/5-FCy on day 0. IgG1 and IgG2a showed a significant (P = 0.001) increase after challenge compared to before challenge in all the groups and no significant difference was seen between the groups. Cutaneous leishmaniasis manifests as a self-healing skin lesion(s) in exposed parts of the body, the healing process for lesions depends upon the Leishmania species involved and the host immune response. Usually healing takes up to 2 years, but CL might not be cured for several years with currently available treatments. Choices of therapeutic treatments for CL are limited and not always effective, often requiring multiple injections, introduce side effects and control measure tools are not always practical and successful [1], [2], [3], [6], [19], [20], [21]. It is well established that individuals with a history of CL are protected against development of further CL lesion. CL lesion(s) development is accompanied by the induction of strong immune response shown by in vivo and in vitro tests (9, 21). Despite many studies on leishmaniasis, immunological surrogate marker(s) of protection is not well defined in human leishmaniasis [9], [10], [22]. There is ample evidence to suggest that development of an effective vaccine against leishmaniasis is possible, but so far no vaccine is available against any form of leishmaniasis. The results of phase 3 clinical trials using crude Leishmania as vaccine were not promising [4], [12], [23], [24]. It has been shown that in vitro CD4+/CD8+ T-cell responses to live Leishmania major are significantly stronger than responses to dead parasites [25]. The only successful protective measure against CL has been shown to be leishmanization. One of the major drawbacks of LZ is the development of a lesion which might not heal during the expected time period and not respond to treatment [7], [9], [10]. Research have therefore focused on developing a Leishmania strain which upon inoculation does not induce a lesion or induces a lesion with limited pathogenicity, but at the same time maintains immunogenicity and as such induce protection in which the leishmanized individuals upon natural infection induce no lesion or even a limited fast healing lesion. In this regard attenuated and genetically manipulated Leishmania were developed and showed to induce protection in murine model of leishmaniasis [4], [15], [16], [26], [27]. Co-inoculation of Leishmania with CpG ODN showed to reduce the pathogenicity, but yet no Leishmania preparation reached to human use [28], [29]. Previously, the same group developed a recombinant double drug sensitive strain of lmtkcd+/+ by integration of a genetically engineered HSV tk gene to confer sensitivity to GCV, and the S. cerevisiae cd gene to induce sensitivity to 5-fluorocytosine. Inoculation of BALB/c mice with lmtkcd+/+ induces lesion similar to WT L. major, but the lesion was controllable by treatment with GCV/5-FCyt [16]. BALB/c mice does not mimic human CL so in the current study, C57BL/6 strain which is not a perfect model of human CL but more mimic the disease is used. Leishmanization which is an inoculation of virulent L. major in a predetermined part of the susceptible individuals, LZ induces a lesion similar to natural infection, protection against further multiple lesions is usually developed upon cure of the lesion caused by LZ and so far LZ showed to be the most effective preventive measure against CL. The main drawback of LZ is development of lesion [16]. Using drug sensitive Leishmania mimic natural infection similar to LZ and at the same time due to sensitivity of Leishmania to approved drugs assures a controllable lesion. As it is presented in Fig. 1, C57BL/6 mice inoculated with L. major lmtkcd+/+ showed a lesion similar to WT L. major (Fig. 1A, Fig. 2A) with no difference in parasite burden (Fig. 1B, Fig. 2B). A very low number of Leishmania parasite is detected in the group of mice inoculated with lmtkcd+/+ and treated with GCV/5-FCyt, A small number of Leishmania was detected in spleen of C57BL/6 mice long after recovery from L. major infection (unpublished data). A similar Th1 response was induced shown by LTT (Fig. 1C), DTH (Fig. 2C) and the cytokine levels of IFN-γ (Fig. 1D, Fig. 2D) and IL-4 (Fig. 2E) in groups of mice inoculated with WT and group of mice inoculated with lmtkcd+/+and treated with GCV and 5-FCyt on day 8 or left untreated, although in the group of mice inoculated with lmtkcd+/+and treated with GCV/5-FCyt on day 8, no lesion was developed at the site of inoculation but the reason for small increase in the size of footpad swelling is due to a slight inflammation which induced at the site of inoculation. Upon challenge with L. major, no lesion was developed and strong protection was seen similar to the group of mice cured from L. major infection (Fig. 2 A). The results showed that despite of no lesion development which was due to under control of recombinant L. major with ganciclovir and 5-Flourocytosin, strong Th1 immune response and protection against WT L. major was induced.

10.1371/journal.pntd.0004585

Diagnostic Accuracy of Antigen 5-Based ELISAs for Human Cystic Echinococcosis

Clinical diagnosis and follow up of cystic echinococcosis (CE) are based on imaging complemented by serology. Several immunodiagnostic tests are commercially available, but the development of new tools is still needed to overcome the lack of standardization of the target antigen, generally consisting of a crude extract of Echinococcus granulosus hydatid cyst fluid. In a previous work, we described a chromatographic method for the preparation of a highly enriched Antigen 5 fraction from hydatid cyst fluid. The high reactivity of patient sera against this preparation prompted us to evaluate further this antigen for the serodiagnosis of CE on a larger cohort of samples. A total of 327 sera from CE patients with heterogeneous conditions for cyst stage, cyst number, organ localization, drug therapy, and surgical intervention, together with 253 sera from healthy controls, were first analyzed by an ELISA based on the Ag5 preparation in two different experimental setups and, in parallel, by a commercial ELISA routinely used in clinical laboratories for CE serodiagnosis. The Ag5 ELISAs revealed different sensitivity (88.3% vs 95.3%) without significant differences in specificity (94.1% vs 92.5%), for the two setups, respectively. Moreover, possible relationships between the Ag5 ELISA absorbance results and clinical variables were investigated. Chi squared test, bivariate logistic regression and multiple regression analyses highlighted differences in the serology reactivity according to pharmacological treatment, cyst activity, and cyst number. The two Ag5 ELISAs revealed different performances depending on the setup. The good diagnostic sensitivity and the high reliability of the Ag5 preparation method make this antigen a promising candidate for the serodiagnosis of CE. Further studies will be needed to evaluate the ability of our test to provide useful information on specific CE clinical traits.

Cystic echinococcosis is a neglected disease caused by the larval stage of the tapeworm Echinococcus granulosus complex affecting both humans and livestock. The disease is considered one of the world’s major zoonoses, and represents a public health problem. Clinical diagnosis and follow-up is mainly based on imaging, while serology should complement imaging-based diagnosis when imaging features are unclear. However, current commercial immunoassays lack satisfactory sensitivity and specificity. They are mostly based on crude antigen preparations of E. granulosus hydatid cyst fluid, a heterogeneous mixture containing molecules of both parasite and host origin, thus standardization is also an issue. Ag5 is one of the most immunogenic proteins present in the hydatid cyst fluid. In a previous work, we described a method enabling the preparation of a highly enriched Ag5 fraction. Here, we present the evaluation of the diagnostic performances of this preparation in two ELISA setups, using a large number of human sera. The influence of several clinical variables on the performance of the tests was also assessed. The results obtained by the Ag5 ELISAs, combined with the robustness of the Ag5 preparation method, make this antigen a promising candidate for the serodiagnosis of CE.

Cystic echinococcosis (CE) is a neglected zoonotic disease caused by the larval form of the tapeworm Echinococcus granulosus complex. The definitive hosts are dogs and other canids, while sheep and other livestock are the natural intermediate hosts; humans are occasional intermediate hosts. Intermediate hosts can be infected by ingestion of food and water contaminated with the parasite eggs eliminated with the feces of infected dogs. The early phase of infection is generally asymptomatic. Small, well encapsulated, viable cysts or old cysts with pseudosolid content typically do not induce major pathology, and patients may remain asymptomatic for years or even permanently. This is likely the reason why almost 50% of CE patients recorded in the Italian Hospital Discharge Records have been diagnosed accidentally during investigations for other diseases, and 57% of cases are people over 60 years old [1]. CE has many important economic effects, the most evident and tangible of which is the cost of expensive medical treatment for human cases; moreover, there is also a strong negative impact on the economy due to the large diffusion of CE among livestock [2–3]. Currently, CE can be treated according to four different approaches: surgery, percutaneous techniques, chemotherapy with benzimidazoles, and with a “watch and wait” approach for inactive cysts. Unfortunately, 20%–40% of the patients respond only temporarily to chemotherapy, and revert to their previous stage (mainly CE2 and CE3b) after the end of treatment [4]. The most affected regions include the Mediterranean, Eastern Europe, parts of South America, parts of Africa, and Central Asia/Western China. In Italy, the average annual incidence rate of hospital cases (AIh) between 2001 and 2012 was 1.6/105 inhabitants [1]. At present, no marker of cyst viability and therapy efficacy exists, and serology may remain positive for years even after successful therapy. As a consequence, long-term follow-up with imaging is required for the clinical management of patients. It is therefore important to invest in innovative technologies that facilitate the monitoring and control of this infectious disease in humans and in farm animals. Currently, CE diagnosis in humans is mostly based on imaging techniques [5], and the clinical approach is based on the WHO international classification of ultrasound images according to the stage of the cyst: CE1, CE2, CE3b (active), CE3a (transitional), and CE4 and CE5 (inactive) [6–8]. Ultrasonography (US), due to the relatively low cost and size of the equipment, is easily transportable in remote resource-poor areas, provides a useful tool for screening, clinical diagnosis, and cyst monitoring [9–11]. However, serology has an important role in supporting the diagnosis of CE, since serological tests are generally cheap, quick, and require less trained and specialized personnel for result interpretation. This is particularly true in areas where US expertise in the diagnosis of CE is scant and/or not easily accessible, and when lesions do not show pathognomonic signs of a parasitic origin, such as young CE1 cysts or inactive CE4–CE5 cysts. Unfortunately, these stages are also those with a broader differential diagnosis (e.g. with simple cysts, neoplastic lesions) whose serology results are also difficult to interpret [12–13]. Commercially available immunoassays are mostly based on hydatid cyst fluid (HCF), collected from infected animals. However, the heterogeneity of this preparation negatively impacts on the sensitivity and specificity of the tests. Many purified native, recombinant, or synthetic antigen preparations have been tested in the last decade, although with controversial results [14–16]. This is most likely due to the poor inter-laboratory reproducibility of antigenic preparations that often rely on outdated methodologies, improperly defined as “purifications”. This adds to the use of different panels of sera, as well as to the lack of clinical characterization and appropriate classification/grouping of sera, used for validation [17–19]. Among HCF antigens, Antigen 5 (Ag5) and Antigen B (AgB) are the most abundant and immunogenic proteins, whose role in the life cycle of the cestode has been assessed only partially [20]. In a recent work [21], our research group reported a straightforward, robust and reproducible chromatography method that enables the preparation of a HCF fraction highly enriched in native Ag5. This highly enriched antigen demonstrated a strong reactivity, both in western blotting and ELISA formats, when tested on a limited panel of sera from CE patients, encouraging a more extensive examination of its diagnostic potential. Here we present a large-scale study (327 cases and 253 controls) aimed to evaluate the diagnostic accuracy of two different Ag5 ELISA setups compared with that of a commercially available ELISA widely employed for routine diagnostic. We also investigated the association between readings of the Ag5-based ELISAs with selected clinical variables of the patients. When the three assays were compared for sensitivity and specificity, the Ag5 ELISA had significantly higher sensitivity. Moreover, the performances of both Ag5 ELISA setups were statistically associated with clinical variables known to influence serology results. These data support the use of this Ag5-based preparation in highly sensitive diagnostic tests and prompt further investigation on its use in the follow-up of CE patients. A written informed consent to use of leftover serum after routine serology for research was obtained from patients at the time of sample collection. The study was approved by the ethics committees of IRCCS San Matteo Hospital Foundation, Pavia, Italy, Prot N. 20150004877, for sera from CE patients, and of the local health authority of Sassari (ASL N. 1, Sassari), Prot N. 1123/L, for sera from healthy controls. A prospective study was performed on sera from patients with hepatic and extra-hepatic CE, collected at the Department of Infectious Diseases of the IRCCS San Matteo Hospital Foundation, Pavia, Italy, where the WHO Collaborating Centre for Clinical Management of Cystic Echinococcosis is based. Sera from healthy subjects, collected at the Sassari Hospital Blood Donor Center, Sassari, Italy, were used as a control group. At the time of serum collection, all patients with abdominal cysts were diagnosed by US and CE cysts were classified according to the World Health Organization–Informal Working Group on Echinococcosis (WHO-IWGE) standardized US classification, by a clinician with long standing experience in the US diagnosis of CE, as part of the routine diagnostic procedures. This classification groups cysts in six stages based on a biological/dynamic approach: active (CE1 and CE2), transitional (CE3a and CE3b) and inactive (CE4 and CE5) cysts. However, CE3b are actually biologically active [8]. On the other hand, from a serological point of view, CE3 cysts of both stages often show comparable results, indicating that biological activity at the time of serum collection may not immediately influence serological responses in these stages [13]. Therefore, for the purpose of analysis, sera from patients with active (CE1, CE2 and CE3b) and transitional (CE3a) cysts were grouped together. Patients having multiple cysts in different stages were assigned to the group of the most active cyst, independently from its hepatic or extra-hepatic localization. HCF crude samples were collected in two different Sardinian slaughterhouses (CE/IT2383M, Tula, Sassari and CE/IT2078M, Lula, Nuoro). Fluid was aspirated from liver and lung cysts found in infected sheep. The hydatid fluid was centrifuged at 1000 g at 4°C and the supernatant stored at -80°C. Enriched Ag5 was obtained as described previously [21]. Briefly, after desalting and concentration, aliquots of sheep HCF were fractionated by Fast Protein Liquid Chromatography (FPLC) on a Superdex-200 column (10/300 GL, GE Healthcare, Uppsala, Sweden). The fractions of interest were pooled and their protein content was evaluated by tandem mass spectrometry on a Q-TOF hybrid mass spectrometer equipped with a nano lock Z-spray source and coupled on-line with a nanoAcquity chromatography system (Waters, Manchester, UK) to verify the quality of the preparation. Sera were tested in duplicate in the parasitology diagnostic laboratory of the IRCCS San Matteo Hospital Foundation, Pavia, Italy, by laboratory personnel with long standing experience in diagnostic parasitology, using a commercial ELISA test (RIDASCREEN Echinococcus IgG, R-Biopharm, Darmstadt, Germany), for the detection of Echinococcus specific total IgG, according to manufacturer’s instructions. Tests were read at 450nm in a spectrophotometer, and a Sample Index (SI) was calculated and interpreted for each serum according to manufacturer’s instructions; ELISA was considered positive for SI >1.1, negative for SI <0.9, and border line for 0.9 ≤SI ≤1.1. However, for statistical evaluations, borderline results were classified together with negatives. Readers were blind to the results of the other tests. Sera were tested with the Ag5 ELISAs based on our Ag5 enriched preparation [21], following two alternative setups, A and B, in the laboratory of Porto Conte Ricerche, Alghero, Italy. Briefly, for setup A, microplates (Nunc-Maxisorp Immunoplate, Waltham, Massachusetts, USA) were coated with 100 μL/well of a 100 ng/mL antigen solution in phosphate buffered saline (PBS). After blocking and washings, sera were added at 1:500 dilution in 2% bovine serum albumin in PBS-0.05% tween-20 (BSA in PBS-0.05%T) and incubated at 37°C for 1 hour. For setup B, microplates were coated with 100 μL/well of a 50 ng/mL antigen solution in PBS, and sera were added at 1:200 dilution. In both cases, secondary antibody (horseradish peroxidase conjugated anti-human IgG, Sigma-Aldrich, St. Louis, MO, USA) was diluted 1:100,000 in 2% BSA in PBS-0.05%T and incubated at 37°C for 1 hour. Finally, the substrate (3,3',5,5'-Tetramethylbenzidine Liquid Substrate, Supersensitive, Sigma) was added. The absorbance was read at 620 nm after 1 hour incubation using a Tecan Sunrise (Tecan Group, Ltd., Männedorf, Switzerland) microplate reader. All sera were tested in duplicate. In order to compare results obtained from different plates, a Sample Ratio (SR) was calculated according to the following formula: SR=Sample mean−Negative control meanPositive control mean−Negative control mean where the negative control was a pool of fifteen healthy donors and the positive control was the Working Standard Anti-Echinococcus Serum, Human (NIBSC, Potters Bar, England). Reproducibility among different Ag5 batches was evaluated for both setup A and setup B on three different, independent preparations. Ten sera (high, medium and low positive control samples) were tested in duplicate and results were evaluated by calculating the mean coefficient of variation (CV) among the three tests, for each setup. All measurements were carried out in parallel by two experienced research fellows. Readers were blind to the results of the other tests. Data analysis was performed with MedCalc Statistical Software version 15.2.2 (MedCalc Software bvba, Ostend, Belgium; http://www.medcalc.org; 2015). A receiver-operator characteristic analysis (ROC) [22–23] was performed to determine a cut-off value for each Ag5 based test. The standard error and the area under the curve were calculated according to DeLong et al. [24]. Levels of sensitivity were plotted against levels of 100 minus specificity at each cut-off point on a ROC curve. Threshold values used were those associated with the highest Youden index J [25]. In order to calculate the best ELISA cut-off values, and to improve sensitivity on active-transitional cysts, that are generally seropositive with HCF-based tests (while patients with CE4 and CE5 cysts and post-surgical patients have most commonly negative or low results), ROC curves were built by using SR values from patients with CE1, CE2, CE3a and CE3b as positive group (171 sera) and healthy controls as negative group (253 sera). The area under the ROC curve (AUC) was used to define the antigen discriminatory power (between subjects with active-transitional cysts and subjects with inactive cysts or without the disease). A p-value <0.05 was considered statistically significant. McNemar test was performed, on all 580 sera, to compare the sensitivities of the two in-house Ag5 setups and the commercial assays. Differences in SR or SI values between groups were analyzed by Kruskal-Wallis test, for the three ELISAs, independently; when more than two groups were analyzed, after Kruskal-Wallis test, pairwise multiple comparisons were evaluated by Conover test with Bonferroni correction [26–27]. Inter-rater agreement test was used to evaluate the agreement between the gold standard (US) and the in-house ELISAs, and the results were expressed by Kappa (K) statistic, with 95% confidence interval [28]. Moreover, three statistical analyses were performed on the results of sera from CE patients to assess the effect of the cyst stage, the number of cysts and the previous treatment with albendazole (ABZ), potentially affecting the assay performance of Ag5 ELISAs [13,18]. While cyst localization is an important variable and its influence on ELISA results is potentially interesting, it was not evaluated given the low number of patients with extra hepatic cysts. Chi squared test was applied to compare the sensitivity of the two Ag5 ELISA setups within subgroups of patients, classified according to their clinical variables. Then, the same variables were further evaluated by a bivariate logistic regression, to consider their influence on test performances and calculate an odd ratio (OR) for each pair of variables. Finally, all the examined clinical variables were concurrently analyzed by a multiple regression, to evaluate their relationship on the diagnostic result. A p-value <0.05 was considered statistically significant. A total of 580 blood sera were collected from June 2008 to June 2012, including 283 from patients with CE cysts, 44 during follow-up after surgery, and 253 from healthy subjects. As summarized in Table 1, 295 sera were collected from patients with hepatic cysts (90.2%), one serum from a pulmonary case (0.3%), and the remaining 31 sera (9.5%) were from patients with other localizations (including peritoneum, kidney, and leg); single cysts were found in 146 patients (44.6%), whilst 2 or more cysts were found in the other 137 subjects. The reproducibility among Ag5 lots was assessed in both setups, demonstrating the reliable performances of our Ag5 preparation. Specifically, setup A had a CV of 9.7%, and setup B had a CV of 11.3%. All sera were analyzed by the commercial assay RIDASCREEN and by the two experimental Ag5 ELISA setups, from November 2011 to September 2012. Then, Ag5 ELISAs were evaluated by ROC curves (Fig 1 and Table 2) to define optimal cut-off values for data analysis. The areas under the ROC curve (AUC) were 0.962 and 0.978 for setup A and setup B, respectively. Serological results and their statistical significance are summarized in Tables 3 and S1. At the best cut off value (0.261 and 0.120 for setup A and setup B, respectively), the two Ag5 ELISA setups showed different sensitivity. In particular, Ag5 ELISA setup B revealed an overall sensitivity higher than both Ag5 setup A and RIDASCREEN test, whilst Ag5 setup A showed similar sensitivity to the commercial kit. More in detail, Ag5 setup B results displayed statistically significant differences when CE3b, CE4 and post surgery patients were examined. These differences persisted when grouping patients as active-transitional and inactive. Concerning the control sera, a higher number of donors tested positive in Ag5 setup B (7.5%), followed by Ag5 setup A (5.9%) and RIDASCREEN (1.6%). Hence, the Ag5 ELISA, especially in setup B, revealed a higher sensitivity and a lower specificity than the commercial RIDASCREEN test. The comparison among the Ag5 ELISAs described in this work and the commercial ELISA is also plotted in Fig 2. Although it should be noted that the ELISA values are not directly comparable due to the difference in OD normalization, Kruskal-Wallis test on SR or SI values confirmed that all the three ELISAs were able to discriminate between patients and healthy controls (Fig 2A, 2B and 2C); statistically different results were also obtained with the three ELISAs, when patients were grouped taking into account the active-transitional versus the inactive stages of CE (Fig 2D, 2E and 2F). Finally, none of the three methods was able to completely discriminate among any of the CE groups and post surgery follow-up patients (Fig 2G, 2H and 2I); however, pairwise comparisons of the subgroups highlighted some differences. Both CE1 and CE2 cysts were different from CE5 in the three assays; CE3a and CE3b were always comparable, but behaved differently in the three tests. The inactive cysts showed important differences in all the tests: CE4 revealed SR or SI values statistically different from CE3a, CE3b and CE5 in Ag5 setup A and RIDASCREEN ELISAs, whilst it differed only from CE5 in Ag5 setup B. CE5, on the contrary, performed differently from all the other groups (except for post surgery patients) in all the ELISAs. Finally, post surgery patients showed dissimilar behavior in all the tests, although with a higher divergence from active-transitional groups for RIDASCREEN results. Further, the Ag5 setup A and the commercial kit, provided a wider range of antibody levels, whilst for Ag5 setup B results were concentrated in a narrower range. The agreement between the gold standard method (US) and our Ag5 ELISAs was evaluated by inter-rater agreement (Kappa) test. When considering patients with CE1, CE2, CE3a and CE3b, kappa (K) value was 0.828, with a standard error of 0.0279 (setup A), or 0.869, with a standard error of 0.0243 (setup B), confirming the excellent agreement between the imaging diagnosis and the ELISA results. When CE4 and CE5 patients were also included in the test, this agreement was poorer, with a K value of 0.718, with a standard error of 0.0295 (setup A), or 0.795, with a standard error of 0.0262 (setup B). This is not surprising, since these patients, due to the inactivity of the cysts, are often negative to ELISA. In addition to the cyst stage, other major clinical variables such as the number of cysts and the previous ABZ treatment were taken into account using Chi squared test to assess the ability of the Ag5 ELISAs to discriminate among CE patients with different clinical traits. Results are summarized in Table 4. For both Ag5 ELISA setups, this test showed a statistical significance (p-value <0.05) for the stratification of patients in terms of single vs multiple cysts, active-transitional vs inactive cysts and for current or past ABZ treatment vs no ABZ. More in detail, pharmacologically treated patients gave positive results to both Ag5 setups more frequently than untreated patients; further, patients in the active or transitional stage had higher positivity rates than patients in the inactive stages; finally, patients with more than one cyst were positive to Ag5 ELISAs more frequently than patients with one cyst. All the above-mentioned clinical variables were evaluated by a bivariate logistic regression. The statistical significance persisted, with fairly high values of odds ratios, when grouping patients according to active-transitional vs inactive cysts and for ABZ treatment vs no ABZ, but it became only borderline significant for Ag5 setup A, when single vs multiple cysts were considered. The interdependence of test results from the pharmacological treatment and the activity of the cysts was confirmed for both Ag5 ELISA setups by multiple regression; the effect of the number of cysts, instead, was only borderline significant for setup A (p-value = 0.055), remaining significant for setup B. CE is a public health and economic issue, concerning both humans and farm animals, and requires an early and unambiguous diagnosis. Imaging techniques remain the most reliable method for an accurate diagnosis. Serological tests are required for diagnosis confirmation in doubtful cases, but their current sensitivity and specificity are unsatisfactory, while their value in the monitoring of patients during follow-up is very limited. The development of robust and stage-specific serological tests is therefore still needed. Currently, commercially available serological kits are based on western blotting, hemagglutination, and ELISA, and mostly use HCF as target antigen, a complex mixture of host and parasite electrolytes, proteins, nitrogenous waste products, carbohydrates and lipids. Its composition is known to vary, often significantly, from cyst to cyst [21,29–30]. As a consequence, sensitivity and specificity are very heterogeneous across tests. Technological improvements have provided increasingly reliable antigens and tools [14,16,18,31–35], but their performances are still suboptimal and their production is often expensive or patented, limiting their use in the most affected regions, which are often developing countries that cannot afford the appropriate facilities [17]. Ag5 and AgB are reported to be the most abundant and immunogenic proteins in the cyst fluid [20]. After an initial growing interest in the use of Ag5 for diagnostics, the focus has moved mostly towards AgB, and its subunits. However, it is reported that a proportion of CE patients with active cysts do not develop a detectable humoral response against AgB [21,36]. Ag5 cross-reactivity issues, as well as low sensitivity and specificity, have been discussed in many papers [16,37–38], and they are probably the main reason for the decline in Ag5 use for CE immunodiagnosis. Part of the cross-reactivity was associated with the presence of phosphorylcholine bound to the Ag5 38 kDa subunit [39–40]. On the other hand, Ag5 protein shares 96.7% and 85.5% identity with the homologous sequences of Taenia solium and Echinococcus multilocularis, respectively, and it is inevitable that the same epitopes are present on these proteins. However, interestingly, Ahn and coworkers [30] showed that Ag5 seems to be immunoreactive in every stage of the pathology, as opposed to AgB, whose proteoforms revealed a reduced antibody capturing activity in CE1, CE4 and CE5 stages. Further studies using sera from patients with other relevant parasitoses are needed to assess the behavior of our Ag5-based ELISAs and to evaluate the value of Ag5-based assays for patient follow-up. Concerning the low sensitivity and specificity reported in previous works, it must be underlined that all the experimental studies concerning Ag5 date back to decades ago, when the analytical techniques themselves suffered from low sensitivity. Therefore, it is likely that the low diagnostic performance reported so far for tests based on Ag5 could be explained with the high heterogeneity of the antigen preparations used at that time. Our results show that Ag5 is a sensitive antigen and further studies using sera from patients with non-CE solid lesions are warranted to evaluate and optimize the cut-off value of the ELISAs when higher sensitivity is needed in the differential diagnosis of inactive cysts. In a previous work [21], a chromatographic method for the reproducible preparation of native Ag5 from different HCF sources was described, enabling the production of a protein fraction highly enriched in Ag5, as verified by mass spectrometry. Preliminary ELISA experiments on a limited panel of CE sera revealed the high reactivity of this antigen. Therefore, we were prompted to evaluate the diagnostic performance of this antigen preparation on a substantial number of CE patients and healthy control sera. ROC curves generated by both Ag5 ELISA setups using CE1, CE2, CE3a and CE3b sera demonstrated its good sensitivity and specificity. When comparing the diagnostic accuracy of these Ag5-based ELISAs with a commercial kit routinely used in clinical laboratories, the excellent performances for Ag5 setup B outperformed those of the commercial test. Nevertheless, specificity was higher for the commercial kit. When we evaluated the effect of clinical variables on Ag5 ELISA results, we found that patients with more than one cyst, and/or in the active or transitional stage, and/or who received drug therapy, were positive to Ag5 test more frequently than the other patients. The bivariate logistic regression and the multiple regression both highlighted an effect due to the pharmacological treatment and to the cyst activity, while the number of cysts maintained a statistical significance only when setup B was used, confirming the importance of these variables as reported in other previous works [13,18,35]. The biological mechanism at the basis of the influence of ABZ treatment on serology results was attributed to high levels of ABZ in cyst fluid causing the germinal laminated membranes to become more permeable, inducing a leakage of their antigenic contents in the blood stream. In turn, the leakage of parasite antigens triggers and sustains increased concentration of circulating antibodies[41–42]. Concerning the effect of the number of cysts and the stage of the disease, no experimental study has demonstrated the biological mechanisms underlying these serology patterns, however it is likely that the loss of cyst wall integrity during the evolution of the cyst and the presence of a large antigenic mass when multiple cysts are present may explain the observed serology behavior. In conclusion, the described serological assay, combining robustness, sensitivity, and easiness of execution, with the low cost, high reproducibility and rapidity of the Ag5 preparation method, makes this antigen a promising candidate for the serodiagnosis of CE especially in the setup B. Moreover, to our knowledge, this is the first report on the influence of the pharmacological treatment, the cyst stage, and the number of cysts on the results of an Ag5-based ELISA test.

10.1371/journal.ppat.1003675

Loss of the TGFβ-Activating Integrin αvβ8 on Dendritic Cells Protects Mice from Chronic Intestinal Parasitic Infection via Control of Type 2 Immunity

Chronic intestinal parasite infection is a major global health problem, but mechanisms that promote chronicity are poorly understood. Here we describe a novel cellular and molecular pathway involved in the development of chronic intestinal parasite infection. We show that, early during development of chronic infection with the murine intestinal parasite Trichuris muris, TGFβ signalling in CD4+ T-cells is induced and that antibody-mediated inhibition of TGFβ function results in protection from infection. Mechanistically, we find that enhanced TGFβ signalling in CD4+ T-cells during infection involves expression of the TGFβ-activating integrin αvβ8 by dendritic cells (DCs), which we have previously shown is highly expressed by a subset of DCs in the intestine. Importantly, mice lacking integrin αvβ8 on DCs were completely resistant to chronic infection with T. muris, indicating an important functional role for integrin αvβ8-mediated TGFβ activation in promoting chronic infection. Protection from infection was dependent on CD4+ T-cells, but appeared independent of Foxp3+ Tregs. Instead, mice lacking integrin αvβ8 expression on DCs displayed an early increase in production of the protective type 2 cytokine IL-13 by CD4+ T-cells, and inhibition of this increase by crossing mice to IL-4 knockout mice restored parasite infection. Our results therefore provide novel insights into how type 2 immunity is controlled in the intestine, and may help contribute to development of new therapies aimed at promoting expulsion of gut helminths.

Infection with intestinal parasitic worms is a major global health problem, with billions of people infected world-wide. Often these worms (known as helminths) develop a long-lasting chronic infection, due to failure of the host to mount the correct type of immune response that would normally expel the parasite. However, how the immune system is controlled leading to chronic helminth infection is not well understood. Here we identify a novel pathway of importance in the development of chronic helminth infection. Using a model parasite which infects mice, we find that a protein called transforming growth factor beta (TGFβ signals to T-cells early during the development of chronic infection and that blocking this signal protects mice from infection. We have also uncovered a key pathway and cell type that controls TGFβ function during development of chronic infection. When a protein called integrin αvβ8 is absent from dendritic cells of the immune system, TGFβ is no longer activated to signal to T-cells and mice are able to mount a protective (type 2) immune response resulting in worm expulsion. Our findings therefore provide new insights into how chronic infections develop and identify potential molecular targets for the prevention of chronic helminth infection.

Gastrointestinal parasitic helminth infections are extremely prevalent, affecting nearly one quarter of the world population. Development of chronic infection, defined as the presence of adult worms in the host, results in severe morbidity and health problems and has been heavily linked with promotion of poverty in affected regions [1]. Current treatments involve the use of anti-helminthic drugs to kill the parasite, but this does not prevent rapid re-infection with worms and encounters problems with drug resistance. As infections with these intestinal parasites are usually chronic, it is likely that helminths are able to influence the immune system to prevent their expulsion. Therefore, understanding the cellular and molecular pathways that regulate the immune response during helminth infection will be crucial in identifying novel therapeutic targets for these poorly managed infections. A key cytokine that plays a multi-functional role in controlling immune responses is transforming growth factor beta (TGFβ) [2]. TGFβ can affect many different cell types, with data highlighting a crucial role for TGFβ in regulation of CD4+ T-cells, both dampening and promoting effector responses depending on the context of the immune response [3], [4]. Importantly, although many cells can produce TGFβ, it is always made as an inactive complex that must be activated to produce biological function [5]. Thus, activation of TGFβ is a key regulatory step in controlling the function of TGFβ in the immune system. Given its importance in regulating diverse T-cell responses, it is not surprising that TGFβ plays a crucial role in the maintenance of immune homeostasis and prevention of autoimmunity. Thus, mice lacking TGFβ receptors in T-cells develop multi-organ inflammatory disease [6], [7] and lack of TGFβ production by T-cells results in autoimmunity and colitis [8]. Interestingly, recent data has implicated TGFβ-like molecules produced by helminths in regulating immune responses during parasite infection [9]. However, the function of TGFβ during helminth infection and how it is regulated to control immune responses to intestinal parasites is poorly understood. Here we show that mice infected with the intestinal parasite Trichuris muris, a homologue of the human pathogen Trichuris trichuria [10], display enhanced TGFβ signalling in CD4+ T-cells early during infection and that antibody-mediated blockade of TGFβ significantly reduces worm burden during the development of a chronic infection.. We find that integrin αvβ8 expressed by dendritic cells (DCs), which we have previously shown to be a key pathway in activating TGFβ during intestinal homeostasis [11], [12], is required for early induction of TGFβ signalling in CD4+ T-cells during development of chronic helminth infection. Importantly, mice lacking integrin αvβ8 expression on DCs are completely protected from chronic infection, with this protection resulting from a specific early upregulation of a Th2-type immune response. Our results therefore provide novel insights into regulatory mechanisms of importance during chronic gastrointestinal parasite infection, and may help contribute to the development of new therapies aimed at promoting expulsion of helminth infection. Development of a chronic parasite infection is believed to result from an inappropriate suppression of host immunity, although the exact molecular mechanisms governing these pathways remain unclear. Given the fundamental importance of CD4+ T-cells in regulating parasite infection and the key role for TGFβ in regulating many aspects of T-cell biology, we analysed TGFβ signalling in T-cells during development of a chronic infection with the helminth Trichuris muris. In C57BL/6 mice receiving 30 T. muris eggs, a dose shown previously to induce a chronic infection [13], we observed a specific increase in phosphorylation of Smad 2/3 (pSmad2/3) in mLN CD4+ T-cells, which is the initial signalling event triggered by engagement of TGFβ with its receptor [14]. This increase in TGFβ signalling was observed as early as day 3 post-infection, and was still evident at day 7 post-infection (Figure 1A and B), before returning to levels seen in uninfected mice by day 14 post-infection (Figure 1B). Similar early increases in CD4+ T-cell pSmad2/3 were also observed in cells taken from the lamina propria of the parasite's niche, the caecum and proximal colon (Figure S1 in Text S1). These data indicate that TGFβ signalling in CD4+ T-cells is an early hallmark of chronic T. muris infection. To directly examine the functional importance of TGFβ in the development of a chronic T. muris infection, we injected C57BL/6 mice with a TGFβ function-blocking antibody before and during infection. Interestingly, mice receiving TGFβ function-blocking antibody were significantly protected from worm infection (Figure 1C). Thus, our data indicate that, during development of chronic infection, TGFβ plays an important role in promoting infection by the intestinal parasite T. muris. We next sought to determine the mechanisms responsible for enhanced TGFβ signalling and function during T. muris infection. One potential explanation for enhanced TGFβ signalling observed in CD4+ T-cells is enhanced activation of host latent TGFβ during infection. We have recently identified integrin αvβ8, expressed by DCs, as a key activator of latent TGFβ in the intestine during immune homeostasis [11], [12]. Thus, to determine the importance of this pathway in promoting TGFβ signalling in CD4+ T-cells during T. muris infection, we analysed T-cell responses in C57BL/6 control mice and mice lacking integrin αvβ8 on DCs (Itgb8 (CD11c-Cre) mice) [11] during infection. Interestingly, the increase in TGFβ signalling observed in CD4+ T-cells early during T.muris infection was significantly reduced in Itgb8 (CD11c-Cre) mice, with pSmad2/3 levels remaining similar to those observed in uninfected mice during the first week of infection (Figure 2A and B). This integrin αvβ8-dependent induction of Smad2/3 phosphorylation was confirmed by Western blot analysis for pSmad2/3 in purified CD4+ T-cells from infected mice (Figure 2C). In contrast, we did not observe any differences in pSmad2/3 induction in dendritic cells between control and Itgb8 (CD11c-Cre) mice (Figure S2A and B in Text S1), indicating that the integrin αvβ8-mediated TGFβ activation does not trigger autocrine TGFβ signalling in DCs during early infection. To directly test whether DCs produced enhanced levels of active TGFβ via expression of integrin αvβ8 during T. muris infection, we isolated DCs from control and Itgb8 (CD11c-Cre) mice and measured their ability to activate TGFβ using an established active TGFβ reporter cell line [15]. Indeed, we observed an enhanced ability of intestinal DC activation to produce active TGFβ early during the development of chronic T. muris infection, which was completely absent in DCs lacking expression of integrin αvβ8 (Figure 2D). Thus, during development of chronic T. muris infection, enhanced TGFβ activation by integrin αvβ8 on DCs is important in triggering TGFβ signalling pathways in CD4+ T-cells. To determine whether TGFβ activation by integrin αvβ8 on DCs was functionally important during development of chronic infection with T. muris, we analysed worm numbers in control and Itgb8 (CD11c-Cre) mice infected with a chronic dose of T. muris eggs. Strikingly, Itgb8 (CD11c-Cre) mice were completely protected from chronic infection by T. muris at day 35 post-infection, with mice showing protection as early as day 14 post-infection (Figure 2E). Indeed, protection from infection observed in Itgb8 (CD11c-Cre) mice was even more pronounced than that observed using antibody-mediated blockade of TGFβ function (Figure 1C). It has been reported that expression of CD11c-Cre may drive recombination in a subset of CD4+ CD11clo activated T-cells [16], and we have previously reported that integrin αvβ8 is expressed by CD4+ T-cells [11]. Thus, to test whether protection from infection in Itgb8 (CD11c-Cre) mice could be due to deletion of the integrin in T-cell subsets, we infected mice lacking integrin αvβ8 on T-cells via expression of CD4-Cre (Itgb8 (CD4-Cre) mice) [11]. In contrast to Itgb8 (CD11c-Cre) mice, Itgb8 (CD4-Cre) mice showed no protection from infection with T.muris (Figure S3A in Text S1) and showed an identical parasite-specific IgG2a/IgG1antibody bias which is associated with development of a chronic infection (Figure S3B in Text S1). Taken together, these data suggest that integrin αvβ8-mediated TGFβ activation by DCs is essential in the promotion of chronic T. muris infection. We next sought to determine the mechanisms responsible for protection from infection in mice lacking the TGFβ-activating integrin αvβ8 on DCs. CD4+ T-cells are key in promoting expulsion of intestinal parasite infection, including T. muris [17], and TGFβ signalling is triggered in these cells early during infection (Figure 1A and B). However, recent evidence has proposed that novel innate lymphoid cells can play crucial roles in the expulsion of several parasite infections [18], [19], [20], [21]. Thus, to determine the function of a CD4+ T-cell response in the expulsion of T. muris observed in Itgb8 (CD11c-Cre) mice, we first bred mice onto a C57BL/6 SCID background lacking all lymphocytes. In the absence of total lymphocytes, protection from infection was completely absent, with Itgb8 (CD11c-Cre) SCID−/− mice showing similar susceptibility to infection as control mice (Figure 3A). To specifically test the role of CD4+ T-cells in protection from infection observed in Itgb8 (CD11c-Cre) mice, we depleted CD4+ T-cells using an anti-CD4 antibody (Figure S4A in Text S1). Absence of CD4+ T-cells restored susceptibility to infection in Itgb8 (CD11c-Cre) mice (Figure 3B). Taken together, these results indicate that protection from infection in the absence of integrin αvβ8 expression on DCs is not via a direct effect of innate lymphoid cells, but driven by a classical CD4+ T-cell response, although a role for innate cells in initial priming cannot be ruled out. CD4+ Foxp3+ regulatory T-cells (Tregs) have been implicated in inhibiting immune responses to helminths [22] including some strains of T. muris [23]. Additionally, we have previously shown that integrin αvβ8-mediated TGFβ activation by specialised intestinal DCs is a crucial mechanism by which Foxp3+ Tregs are induced in the gut [12], and that Itgb8 (CD11c-Cre) mice have reduced Foxp3+ Treg levels in their intestine [11]. Thus, one potential explanation for protection from infection in Itgb8 (CD11c-Cre) mice is that there is a reduced Treg response induced during infection in these mice. To address this possibility, we first directly assessed the role of Foxp3+ Tregs during development of chronic T. muris infection by using the DEREG mouse model on a C57BL/6 background, which allows specific ablation of Foxp3+ Tregs by injection of diphtheria toxin [24]. Despite robust depletion of Foxp3+ Tregs (Figure S4B in Text S1) we did not see any enhanced ability of Foxp3+ Treg-depleted mice to expel worms (Figure 3C). In agreement with a lack of role for Foxp3+ Tregs in the development of chronic T. muris infection, we did not see any enhancement of Foxp3+ Treg levels during the course of infection (Figure 3D). Additionally, to directly assess whether reduced Foxp3+ Treg numbers Itgb8 (CD11c-Cre) mice was responsible for protection from infection, we rescued Treg numbers by adoptively transferred Foxp3+ Tregs from GFP-Foxp3 mice [25] prior to infection. Despite enhancement of Treg numbers in Itgb8 (CD11c-Cre) mice after adoptive transfer of GFP-Foxp3+ Tregs (Figure S5 in Text S1), Itgb8 (CD11c-Cre) mice were still highly protected from development of a chronic infection (Figure 3E). Taken together, these data indicate that the protection from infection observed in Itgb8 (CD11c-Cre) mice is driven by CD4+ T-cells, but independently of Foxp3+ Treg cells. During development of a chronic infection with T. muris, mice develop a Th1-type immune response at the expense of a protective Th2-type response [13]. Thus, an alternative explanation for the expulsion of a normally chronic infection of T. muris by Itgb8 (CD11c-Cre) mice is that, in the absence of early CD4+ T-cell TGFβ signalling, mice produce a Th2-type response instead of the usual non-protective Th1 response. To test this possibility, we analysed the production of Th1 and Th2 cytokines during infection in control and Itgb8 (CD11c-Cre) mice. Strikingly, as early as 3 days post-infection, we observed a significant increase in production of the Th2 cytokine IL-13, which was still elevated at 7 days post-infection (Figure 4A). In contrast, although there was a slight enhancement of the Th1 cytokine IFNγ 3 days post-infection, this was not significantly different between control and Itgb8 (CD11c-Cre) mice (Figure 4B). Control mice developed a marked enhancement in IFNγ production by day 18 post-infection, as expected during development of a chronic infection, and this was not observed in Itgb8 (CD11c-Cre) mice (Figure 4B). Neither control nor Itgb8 (CD11c-Cre) mice produced any detectable IL-4 at any tested timepoint post-infection, a cytokine previously shown to be involved in protection from T. muris infection (Figure S6 in Text S1 and data not shown). We next investigated the cellular source of the early IL-13 production in Itgb8 (CD11c-Cre) mice using flow cytometry. We observed a significant population of IL-13+ CD4+ T-cells within the intestinal lamina propria early during infection in Itgb8 (CD11c-Cre) which was not apparent in control mice (Figure 4C). We also observed a slight increase in IFNγ+ lamina propria CD4+ T-cells in Itgb8 (CD11c-Cre) mice early post-infection; however, these levels were not significantly different from those seen in control mice (Figure 4C). Interestingly, in mice treated with a TGFβ function-blocking antibody which resulted in protection from infection (Figure 1C), we observed a similar increase in CD4+ T-cell IL-13 production, with no difference in IFNγ production observed (Figure 4D). Furthermore, mice treated with TGFβ blocking antibody developed a skewed parasite-specific IgG1 response during infection, indicative of an enhanced type2 immune response (Figure 4E). Taken together, these data indicate that TGFβ activation by DC-expressed integrin αvβ8 is important in controlling IL-13 production in CD4+ T-cells early during development of chronic infection. To test whether the enhanced production of IL-13 early during infection was responsible for expulsion of a chronic T. muris infective dose, we crossed the Itgb8 (CD11c-Cre) mice with C57BL/6 IL-4 knockout mice, which have previously been shown to lack the ability to generate an IL-4/13-mediated Th2 response during T. muris infection [26]. As both control mice and Itgb8 (CD11c-Cre) mice did not display production of IL-4 early during T. muris infection (Figure S6 in Text S1), these mice allowed us to test the role of the enhanced IL-13 response seen early during infection in Itgb8 (CD11c-Cre) mice. As expected, Itgb8 (CD11c-Cre)×IL-4−/− mice no longer demonstrated an early IL-13 production in the intestinal CD4+ T-cells (Figure 5A). Strikingly, Itgb8 (CD11c-Cre)×IL-4−/− mice were completely susceptible to infection, with parasite burdens comparable to those seen in control mice (Figure 5B). Taken together, these data indicate that lack of the TGFβ-activating integrin αvβ8 on DCs results in a heightened CD4+ T-cell Th2 immune response during T. muris infection which is responsible for rapid parasite expulsion. Infection with intestinal helminths can result in either expulsion or development of chronic infection, often depending on the type of CD4+ T-cell response generated. Generally, a chronic infection results when inappropriate Th1 cytokine production occurs, as opposed to an inability of CD4+ T-cells to mount a response. Expulsion of the parasite relies on the production of Th2 cytokines, in particular IL-13 which drives a combination of cytokine-mediated expulsion mechanisms such as increased epithelial cell turnover in the intestine [27], enhanced mucus production [28] and increased production of RELM-β [29]. Our data now demonstrate an essential role for TGFβ and the TGFβ-activating integrin αvβ8 expressed by DCs in promoting chronic intestinal parasite infection, using T. muris, a mouse model of the prevalent human parasite Trichuris trichuria. We observed that TGFβ signalling in CD4+ T-cells is triggered early during T. muris infection, and antibody-mediated blockade of TGFβ function significantly protects mice from infection. Mechanistically, we find that enhanced TGFβ signalling in T-cells during infection occurs via expression of the TGFβ-activating integrin αvβ8 on DCs and that lack of this integrin on DCs completely protects mice from infection due to an enhanced protective Th2 response. We have therefore identified a novel pathway that regulates Th2 immune responses in the gut that could potentially be targeted to upregulate host protective immune responses during gut parasite infection. Recent data suggest that in certain chronic parasite infections, induction of Foxp3+ Tregs is important in suppression of protective immunity and development of chronic infection [30], [31], [32]. Given the fundamental role of TGFβ in induction of Foxp3+ Tregs from naive CD4+ T-cells, and the fact that Itgb8 (CD11c-Cre) mice have previously been shown to have impaired induction of intestinal Foxp3+ Tregs [12], we hypothesised that protection from T. muris infection observed in Itgb8 (CD11c-Cre) mice was due to reduced induction of Foxp3+ Tregs. However, when Foxp3+ Tregs were depleted before and during the course of infection no protection from infection was observed. Indeed, in contrast to Itgb8 (CD11c-Cre) mice, no enhancement of CD4+ T-cell IL-13 production was observed early during infection in Foxp3+ Treg-depleted mice (Figure S7 in Text S1). Additionally, in agreement with previous reports [23] we did not see a significant increase in Foxp3+ Tregs during T. muris infection. Instead, TGFβ activation by DC-expressed integrin αvβ8 appears important in suppression of IL-13 production by CD4+ T-cells early during T. muris infection. This is in agreement with previous data from in vitro studies, suggesting that TGFβ can downregulate expression of GATA-3 in T-cells (a key transcription factor in promoting Th2 cell differentiation) [33], [34]. Indeed, recent data suggest that TGFβ-mediated induction of the transcription factor Sox4 is important in preventing GATA-3 transcription to drive Th2 development [35]. Furthermore, we only observed an early increase in CD4+ T-cell specific pSmad2/3 signalling during a chronic Th1-promoting low dose infection and not during an acute Th2 promoting high dose infection in C57BL/6 mice (Figure S8 in Text S1) Thus, our data suggests that activation of TGFβ by integrin αvβ8 early during T. muris infection is important in suppression of protective Th2 cell development, which leads instead to production of an inappropriate Th1 response and development of chronic infection. Although we did not detect any IL-4 production in Itgb8 (CD11c-Cre) mice during infection, given that we crossed these mice to IL4 KO mice to eliminate enhanced IL-13 production by T-cells, we cannot rule out a potential role for low level production of IL-4 (below our limits of detection) in protection from infection. In addition to effects on T-cells, TGFβ has wide-ranging effects on multiple other immune cell types [36]. Recent reports have highlighted an important role for novel innate lymphoid cells in promoting protective type 2 immunity during certain parasite infections [18], [19], [20], [21]. Hence, it could be postulated that protection from infection seen in the absence of integrin αvβ8 results from an enhanced innate lymphoid cell response. However, protection from chronic T. muris infection observed in Itgb8 (CD11c-Cre) mice did not correlate with enhanced type 2 cytokine production from any cell types apart from CD4+ T-cells (data not shown), and protection from infection was completely dependent on CD4+ T-cells. Thus, although innate lymphoid cell depletion would be required to definitively rule out their role in this enhanced Th2 response, it appears unlikely that lack of integrin-mediated TGFβ activation by DCs is promoting expulsion of the parasite via effects on non-CD4+ T-cells. Given the crucial importance of TGFβ in regulating CD4+ T-cell responses, our current model is that TGFβ activated by DCs acts directly in CD4+ T-cells to regulate type 2 responses during T. muris infection. A recent study by Heitmann et al. (2012) suggests that CD4+ T-cell type 2 responses can be regulated via TGFβ signalling in DCs [36]. Thus, mice expressing a dominant negative construct of the TGFβ receptor II in myeloid cells (hence are refractory to TGFβ signalling) display enhanced Th2 responses during infection with the helminth Nippostrongylus brasiliensis [36]. However, we observed no difference in pSmad2/3 induction in DCs from control versus Itgb8 (CD11c-Cre) mice early during infection (Figure S2 in Text S1). Thus, our data indicate that activation of TGFβ by integrin αvβ8 on DCs does not regulate Th2 cells indirectly via autocrine TGFβ signalling during T. muris infection. Velhoden et al 2008 [37] have demonstrated that mice expressing a dominant negative TGFβ receptor specifically on CD4+ T-cells (CD4-DN-TGFβRII mice, thus T-cells are refractory to TGFβ) are more susceptible to infection with T. muris using an acute model of infection. This finding initially appears to conflict with our data, as we demonstrate that both antibody-mediated blockade of TGFβ and lack of the TGFβ activating integrin αvβ8 on DCs promotes expulsion of the parasite. However, recent data suggests that CD4-DN-TGFβRII mice display high levels of IFNγ level during intestinal helminth infection [38],[39] which, given the known role of IFNγ in promoting chronic T. muris infection [13], may explain the enhanced levels of infection observed in CD4-DN-TGFβRII mice. An important question that remains are which specific subset of intestinal DCs are involved in modulating CD4+ T-cells to suppress Th2 responses? Although a functionally important gut population of CD11c+ T-cells does exist [16], which may be targeted in our CD11c-cre knock-out system, mice lacking integrin αvβ8 on T-cells (Itgb8 (CD4-Cre) mice) were completely susceptible to T. muris infection (Figure S3 in Text S1). These data indicate that it is indeed an αvβ8-expressing DC population (or a related CD11c+ mononuclear phagocyte population) that is important in inhibiting Th2 responses in this infection. An important DC subset likely involved during infection are the migratory CD103+ DC [40], as we have previously demonstrated that this cell subset expresses high levels of integrin αvβ8 [12]. We have observed some integrin αvβ8 expression on the CD11c+ CD103- DC subset in the colon [12], which has been suggested to include both DCs and macrophage-like cell populations [41]. However, although some subsets of CD11c+ CD103- intestinal cells have been shown to migrate to lymph nodes to modulate T-cell responses [42], a large population do not normally migrate. Of note, we did not observe any alteration in the levels of αvβ8 expression on either CD103+ or CD103- LILP subset during the development of a chronic infection (Figure S9 in Text S1). Therefore, the exact DC population involved in downregulation of Th2 responses via integrin αvβ8 remains to be determined. Nevertheless, this key role for DC-expressed integrin αvβ8 in modulating Th2 responses, in addition to its previous essential roles in the induction of Foxp3+ Tregs [12] and Th17 cells [43], places DC-expressed integrin αvβ8 as a key orchestrator of CD4+ T-cell immunity. In summary, we have highlighted an important cellular and molecular pathway by which the TGFβ-activating integrin αvβ8 expressed by DCs represses protective Th2 immunity during intestinal parasite infection with T. muris. Thus, given the poor treatments currently available for chronic parasite infection, further work should focus on the potential for targeting integrin αvβ8 therapeutically to enhance protective immunity during Trichuris infection. Additionally, whether the pathway is involved in the development of other chronic infections and Th2-associated disease is the focus of current studies. C57 BL/6 mice were purchased from Harlan Laboratories. Mice lacking integrin αvβ8 on DCs via expression of a conditional floxed allele of β8 integrin in combination with CD11c-Cre (Itgb8 (CD11c-Cre) mice) [11], DEREG mice [24], GFP-Foxp3 mice [25] and IL-4−/− mice [44], all on a C57BL/6 background, have been previously described. All mice were maintained in specific pathogen-free conditions at the University of Manchester and used at 6 to 8 weeks of age. All animal experiments were performed under the regulations of the Home Office Scientific Procedures Act (1986), specifically under the project licence PPL 40/3633. The project licence was approved by both the Home Office and the local ethics committee of the University of Manchester. The techniques used for maintenance and infection of T. muris were as previously described [45] Mice were orally infected with 20–30 eggs for a low-dose infection and 150 for an acute infection. Worm burdens were assessed by counting the number of worms present in the caecum as described previously [45]. To block TGFβ, mice were injected i.p with 0.5 mg of anti-TGFβ blocking antibody (clone 1d11.16.8) (BioXCell, NH, USA) or control IgG1 every 2 days from 4 days prior to infection. CD4+ cells were depleted via i.p. injection of 2 mg anti-CD4 antibody (YTS 191)47 every 2 days from 6 days prior to infection. Foxp3+ Tregs were depleted in DEREG mice as described [24], via i.p. injection of 200 ng diphtheria toxin (Merck) every 2 days from 2 days prior to infection. Spleens were removed from Foxp3GFP mice, disaggregated and red blood cells lysed. Cells were blocked with anti-FcγR antibody and labelled with anti-CD4 antibody (clone GK1.5; eBioscience) before sorting for CD4+, GFP+ populations using a FACS Aria. Cell purity in all experiments was >99.8%. Mice were injected i.p. with 0.5×106 cells 3 days prior to infection. mLNs were excised from mice and incubated with shaking for 20 min at 37°C in RPMI with 0.08 U/ml liberase blendzyme 3 (Roche) or 1 mg/ml collagenase VIII and 50 U/ml DNAseI, respectively. Cells were blocked with anti-FcγR antibody (clone 24G2; eBioscience) before enrichment using a CD11c enrichment kit and LS MACS column (Miltenyi Biotec). Enriched DCs were labelled with anti-CD11c antibody (clone N418; eBioscience) and sorted using a FACS Aria. In all experiments, subset purity was >95%. DCs were incubated with mink lung epithelial cells transfected with a plasmid containing firefly luciferase cDNA downstream of a TGFβ-sensitive promoter [15] in the presence of 1 µg/ml lipopolysaccharide. Co-cultures were incubated overnight in the presence of 80 µg/ml control mIgG or anti-TGFβ antibody (clone 1d11) and luciferase activity detected via the Luciferase Assay System (Promega) according to manufacturer's protocol. TGFβ activity was determined as the difference in luciferase activity between control mIgG-treated samples and samples treated with anti-TGFβ antibody. Mesenteric lymph nodes (mLNs) were removed from mice and disaggregated through a 100 µm sieve. Caecum and proximal colon were excised and lamina propria lymphocytes were prepared essentially as described [46] with slight modification in the tissue digestion step (digestion medium used was RPMI with 10% Foetal calf serum, 0.1% w/v collagenase type I and Dispase II (both Invitrogen), and tissue was digested for 30 min at 37°C). Cell suspensions were blocked with anti-FcγR antibody (clone 24G2; eBioscience) before labelling with antibodies specific for CD4 (clone GK1.5; eBioscience), Foxp3 (clone FJK-16s; eBioscience), IL-13 (clone eBiol13A; eBioscience), IFNγ (clone XMG1.2; eBioscience) or p-Smad 2/3 (Santa Cruz). For pSmad2/3 staining, an Alexa Fluor 594-labelled donkey anti-goat secondary antibody was used (Invitrogen). All samples were analysed on a FACS LSRII. mLN and LILP cells were prepared as described above before incubating with 50 ug/ml of concavelin A or T. muris excretory/secretory (E/S) antigen for 48 hours. Cell-free supernatants were analysed for cytokine production via cytometric bead array (BD) or paired ELISA antibodies (anti- IFNγ, clone XMG1.2 and biotin anti- IFNγ, clone R4-6A2; anti-IL-13, clone eBio13A and biotin anti-IL-13, clone eBio1316H and anti-IL-4, clone 11B1and biotin anti-IL-4, clone BVD6-2462 (eBioscience). For intracellular cytokine analysis cells were incubated for 12 hours with 50 ug/ml T. muris E/S antigen followed by PMA/ionomycin stimulation for 4 hour with addition of monensin for the final 3 hours. Cells were then stained with antibodies against IL-4, IL-13 and IFNγ using the eBioscience Foxp3 permibilization kit according to the manufacturer's instructions. CD4+ T-cells were isolated from mLN via negative selection using a CD4+ T-cell isolation kit (Miltenyi Biotec) during the course of a chronic T. muris infection and lysed using 1% Triton-X100 in Tris buffer (50 mM Tris-HCl, 150 mM NaCl pH 7.4) plus 5 mM EDTA, 20 µg/ml leupeptin and aprotinin, 0.5 mM AESF and 2 mM NaVO3. Lysates were analysed by Western blot with antibodies to detect p-Smad 2/3 (Millipore) and β-actin (Sigma Aldrich), using the Invitrogen Nupage gel system according to manufacturer's instructions. Total RNA was purified from sorted DC subsets using an RNAeasy minikit according to manufacturer's protocol (Qiagen). RNA was reverse transcribed using Oligo dT primers, and cDNA for specific genes detected using a SYBR green qPCR kit (Finnzymes) Gene expression normalised to HPRT expression. HPRT Forward: GCGTCGTGATTAGCGATGATGAAC, HPRT Reverse: GAGCAAGTCTTTCAGTCCTGTCCA, Integrin β8 Forward: GGGTGTGGAAACGTGACAAGCAAT, Integrin β8 Reverse: TCTGTGGTTCTCACACTGGCAACT. Results are expressed as mean ± SEM. Where statistics are quoted, 2 experimental groups were compared using the Student's t-test for non-parametric data. Three or more groups were compared using the Kruskal–Wallis test, with Dunn's multiple comparison post-test. P≤0.05 was considered statistically significant.

10.1371/journal.pcbi.1004083

Improved Estimation and Interpretation of Correlations in Neural Circuits

Ambitious projects aim to record the activity of ever larger and denser neuronal populations in vivo. Correlations in neural activity measured in such recordings can reveal important aspects of neural circuit organization. However, estimating and interpreting large correlation matrices is statistically challenging. Estimation can be improved by regularization, i.e. by imposing a structure on the estimate. The amount of improvement depends on how closely the assumed structure represents dependencies in the data. Therefore, the selection of the most efficient correlation matrix estimator for a given neural circuit must be determined empirically. Importantly, the identity and structure of the most efficient estimator informs about the types of dominant dependencies governing the system. We sought statistically efficient estimators of neural correlation matrices in recordings from large, dense groups of cortical neurons. Using fast 3D random-access laser scanning microscopy of calcium signals, we recorded the activity of nearly every neuron in volumes 200 μm wide and 100 μm deep (150–350 cells) in mouse visual cortex. We hypothesized that in these densely sampled recordings, the correlation matrix should be best modeled as the combination of a sparse graph of pairwise partial correlations representing local interactions and a low-rank component representing common fluctuations and external inputs. Indeed, in cross-validation tests, the covariance matrix estimator with this structure consistently outperformed other regularized estimators. The sparse component of the estimate defined a graph of interactions. These interactions reflected the physical distances and orientation tuning properties of cells: The density of positive ‘excitatory’ interactions decreased rapidly with geometric distances and with differences in orientation preference whereas negative ‘inhibitory’ interactions were less selective. Because of its superior performance, this ‘sparse+latent’ estimator likely provides a more physiologically relevant representation of the functional connectivity in densely sampled recordings than the sample correlation matrix.

It is now possible to record the spiking activity of hundreds of neurons at the same time. A meaningful statistical description of the collective activity of these neural populations—their ‘functional connectivity’—is a forefront challenge in neuroscience. We addressed this problem by identifying statistically efficient estimators of correlation matrices of the spiking activity of neural populations. Various underlying processes may reflect differently on the structure of the correlation matrix: Correlations due to common network fluctuations or external inputs are well estimated by low-rank representations, whereas correlations arising from linear interactions between pairs of neurons are well approximated by their pairwise partial correlations. In our data obtained from fast 3D two-photon imaging of calcium signals of large and dense groups of neurons in mouse visual cortex, the best estimation performance was attained by decomposing the correlation matrix into a sparse network of partial correlations (‘interactions’) combined with a low-rank component. The inferred interactions were both positive (‘excitatory’) and negative (‘inhibitory’) and reflected the spatial organization and orientation preferences of the interacting cells. We propose that the most efficient among many estimators provides a more informative picture of the functional connectivity than previous analyses of neural correlations.

Functional connectivity is a statistical description of observed multineuronal activity patterns not reducible to the response properties of the individual cells. Functional connectivity reflects local synaptic connections, shared inputs from other regions, and endogenous network activity. Although functional connectivity is a phenomenological description without a strict mechanistic interpretation, it can be used to generate hypotheses about the anatomical architecture of the neural circuit and to test hypotheses about the processing of information at the population level. Pearson correlations between the spiking activity of pairs of neurons are among the most familiar measures of functional connectivity [1–5]. In particular, noise correlations, i.e. the correlations of trial-to-trial response variability between pairs of neurons, have a profound impact on stimulus coding [1, 2, 6–11]. In addition, noise correlations and correlations in spontaneous activity have been hypothesized to reflect aspects of synaptic connectivity [12]. Interest in neural correlations has been sustained by a series of discoveries of their nontrivial relationships to various aspects of circuit organization such as the physical distances between the neurons [13, 14], their synaptic connectivity [15], stimulus response similarity [3–5, 15–22], cell types [23], cortical layer specificity [24, 25], progressive changes in development and in learning [26–28], changes due to sensory stimulation and global brain states [21, 29–33]. Neural correlations do not come with ready or unambiguous mechanistic interpretations. They can arise from monosynaptic or polysynaptic interactions, common or correlated inputs, oscillations, top-down modulation, and background network fluctuations, and other mechanisms [34–39]. But multineuronal recordings do provide more information than an equivalent number of separately recorded pairs of cells. For example, the eigenvalue decomposition of the covariance matrix expresses shared correlated activity components across the population; common fluctuations of population activity may be accurately represented by only a few eigenvectors that affect all correlation coefficients. On the other hand, a correlation matrix can be specified using the partial correlations between pairs of the recorded neurons. The partial correlation coefficient between two neurons reflects their linear association conditioned on the activity of all the other recorded cells [40]. Under some assumptions, partial correlations measure conditional independence between variables and may more directly approximate causal effects between components of complex systems than correlations [40]. For this reason, partial correlations have been used to describe interactions between genes in functional genomics [41, 42] and between brain regions in imaging studies [43, 44]. These opportunities have not yet been explored in neurophysiological studies where most analyses have only considered the distributions of pairwise correlations [2, 4, 5, 13]. However, estimation of correlation matrices from large populations presents a number of numerical challenges. The amount of recorded data grows only linearly with population size whereas the number of estimated coefficients increases quadratically. This mismatch leads to an increase in spurious correlations, overestimation of common activity (i.e. overestimation of the largest eigenvalues) [45], and poorly conditioned partial correlations [41]. The sample correlation matrix is an unbiased estimate of the true correlations but its many free parameters make it sensitive to sampling noise. As a result, on average, the sample correlation matrix is farther from the true correlation matrix than some structured estimates. Estimation can be improved through regularization, the technique of deliberately imposing a structure on an estimate in order to reduce its estimation error [41, 46]. To ‘impose a structure’ on an estimate means to bias (‘shrink’) it toward a reduced representation with fewer free parameters, the target estimate. The optimal target estimate and the optimal amount of shrinkage can be obtained from the data sample either analytically [41, 45, 47] or by cross-validation [48]. An estimator that produces estimates that are, on average, closer to the truth for a given sample size is said to be more efficient than other estimators. Although regularized covariance matrix estimation is commonplace in finance [47], functional genomics [41], and brain imaging [44], surprisingly little work has been done to identify optimal regularization of neural correlation matrices. Improved estimation of the correlation matrix is beneficial in itself. For example, improved estimates can be used to optimize decoding of the population activity [48, 49]. But reduced estimation error is not the only benefit of regularization. Finding the most efficient among many regularized estimators leads to insights about the system itself: the structure of the most efficient estimator is a parsimonious representation of the regularities in the data. The advantages due to regularization increase with the size of the recorded population. With the advent of big neural data [50], the search for optimal regularization schemes will become increasingly relevant in any model of population activity. Since optimal regularization schemes are specific to systems under investigation, the inference of functional connectivity in large-scale neural data will entail the search for optimal regularization schemes. Such schemes may involve combinations of heuristic rules and numerical techniques specially designed for given types of neural circuits. What structures of correlation matrices best describe the multineuronal activity in specific circuits and in specific brain states? More specifically, are correlations in the visual cortex during visual stimulation best explained by common fluctuations or by local interactions within the recorded microcircuit? To address these questions, we evaluated four regularized covariance matrix estimators that imposed different structures on the estimate. The estimators are designated as follows: Csample—the sample covariance matrix, the unbiased estimator. Cdiag—linear shrinkage of covariances toward zero, i.e. toward a diagonal covariance matrix. Cfactor—a low-rank approximation of the sample covariance matrix, representing inputs from unobserved shared factors (latent units). Csparse—sparse partial correlations, i.e. a large fraction of the partial correlations between pairs of neurons are set to zero. Csparse+latent—sparse partial correlations between the recorded neurons and linear interactions with a number of latent units. First, we used simulated data to demonstrate that the selection of the optimal estimator indeed pointed to the true structure of the dependencies in the data. We then performed a cross-validated evaluation to establish which of the four regularized estimators was most efficient for representing the population activity of dense groups of neurons in mouse primary visual cortex recorded with high-speed 3D random-access two-photon imaging of calcium signals. In our data, the sample correlation coefficients were largely positive and low. We found that the most efficient estimator of the correlation matrix in these data was Csparse+latent. This estimator revealed a sparse network of partial correlations (‘interactions’), between the observed neurons; it also inferred a number of latent units interacting with the observed neurons. We analyzed these networks of partial correlations and found the following: Whereas significant noise correlations were predominantly positive, the inferred interactions had a large fraction of negative values possibly reflecting inhibitory circuitry. Moreover, the inferred positive interactions exhibited a substantially stronger relationship to the physical distances and to the differences in preferred orientations than noise correlations. In contrast, the inferred negative interactions were less selective. The covariance matrix is defined as Σ = E [ ( x − μ ) ( x − μ ) T ] , μ = E [ x ] (1) where the p × 1 vector x is a single observation of the firing rates of p neurons in a time bin of some duration, E [ ⋅ ] denotes expectation, and μ is the vector of expected firing rates. Given a set of observations {x(t): t ∈ T} of population activity, where x(t) contains observed firing rates in time bin t, and an independent estimate of the mean firing rates x¯, the sample covariance matrix, C sample = 1 n ∑ t ∈ T ( x ( t ) − x ¯ ) ( x ( t ) − x ¯ ) T , (2) where n is the number of time bins in T, is an unbiased estimate of the true covariance matrix, i.e. E [ C sample ] = Σ. In all cases when the unbiasedness of the sample covariance matrix matters in this paper, the mean activity is estimated independently from a separate sample. Given any covariance matrix estimate C, the corresponding correlation matrix R is calculated by normalizing the rows and columns of C by the square roots of its diagonal elements to produce unit entries on the diagonal: R = ( diag ( C ) ) − 1 2 C ( diag ( C ) ) − 1 2 , (3) where diag(C) denotes the diagonal matrix with the diagonal elements from C. The partial correlation between a pair of variables is the Pearson correlation coefficient of the residuals of the linear least-squares predictor of their activity based on all the other variables, excluding the pair [40, 51]. Partial correlations figure prominently in probabilistic graphical modeling wherein the joint distribution is explained by sets of pairwise interactions [40]. For multivariate Gaussian distributions, zero partial correlations indicate conditional independence of the pair, implying a lack of direct interaction [40, 52]. More generally, partial correlations can serve as a measure of conditional independence under the assumption that dependencies in the system are close to linear effects [40, 53]. As neural recordings become increasingly dense, partial correlations may prove useful as indicators of conditional independence (lack of functional connectivity) between pairs of neurons. Pairwise partial correlations are closely related to the elements of the precision matrix, i.e. the inverse of the covariance matrix [40, 52]. Zero elements in the precision matrix signify zero partial correlations between the corresponding pairs of variables. Given the covariance estimate C, the matrix of partial correlations P is computed by normalizing the rows and columns of the precision matrix C−1 to produce negative unit entries on the diagonal: P = − ( diag ( C − 1 ) ) − 1 2 C − 1 ( diag ( C − 1 ) ) − 1 2 (4) Increasing the number of recorded neurons results in a higher condition number of the sample covariance matrix [45] making the partial correlation estimates more ill-conditioned: small errors in the covariance estimates translate into greater errors in the estimates of the partial correlations. With massively multineuronal recordings, partial correlations cannot be estimated without regularization [41, 45]. We considered four regularized estimators based on distinct families of target estimates: Cdiag, Cfactor, Csparse, and Csparse+latent. In probabilistic models with exclusively linear dependencies, the target estimates of these estimators correspond to distinct families of graphical models (Fig. 1 Row 1). The target estimate of estimator Cdiag is the diagonal matrix D containing estimates of neurons’ variances. Regularization is achieved by linear shrinkage of the sample covariance matrix Csample toward D as controlled by the scalar shrinkage intensity parameter λ ∈ [0, 1]: C diag = ( 1 − λ ) C sample + λ D (5) The structure imposed by Cdiag describes a population with no linear associations between the neurons (Fig. 1 Row 1, A). If sample correlations are largely spurious, Cdiag is expected to be more efficient than other estimators. Estimator Cfactor approximates the covariance matrix by the factor model, C factor = L + D , (6) where L is a p × p symmetric positive semidefinite matrix with low rank and D is a diagonal matrix. This approximation is the basis for factor analysis [51], where matrix L represents covariances arising from latent factors. The rank of L corresponds to the number of latent factors. Matrix D contains the variances of the cells’ independent activity from the latent factors. The estimator is regularized by selecting the rank of L and by shrinking the independent variances in D toward their mean. The structure imposed by Cfactor describes a population whose activity is linearly driven by a number of latent factors that affect many cells while direct interactions between the recorded cells are insignificant (Fig. 1 Row 1, B). Estimator Csparse is produced by approximating the sample covariance matrix by the inverse of a sparse matrix S: C sparse = S − 1 . (7) The estimator is regularized by adjusting the sparsity (fraction of off-diagonal zeros) of S. The problem of finding the optimal set of non-zero elements in S is known as covariance selection [52]. The structure imposed by Csparse describes conditions in which neural correlations arise from direct linear effects (‘interactions’) between some pairs of neurons (Fig. 1 Row 1, C). Estimator Csparse+latent is obtained by approximating the sample covariance matrix by a matrix whose inverse is the difference of a sparse component and a low-rank component: C sparse + latent = ( S − L ) − 1 , (8) where S is a sparse matrix and L is a low-rank matrix. The estimator is regularized by adjusting the sparsity of S and the rank of L. See Methods for more detailed explanations. The structure imposed by Csparse+latent favors conditions in which the activity of neurons is determined by linear effects between some observed pairs of neurons and linear effects from several latent units (Fig. 1 Row 1, D) [54, 55]. We refer to the sparse partial correlations in estimators Csparse and Csparse+latent as ‘interactions’. We next demonstrated how the most efficient among different regularized estimators can reveal the structure of correlations. We constructed four families of 50 × 50 covariance matrices, each with structure that matched one of the four regularized estimators (Fig. 1 Row 2, A–D and Methods). We used these covariance matrices as the ground truth in multivariate Gaussian distributions with zero means and drew samples of various sizes. The sample correlation matrices from finite samples (e.g. n = 500 in Fig. 1 Row 3) were contaminated with sampling noise and their underlying structures were difficult to discern. The evaluation of any covariance matrix estimator, C, is performed with respect to a loss function ℓ(C, Σ) to quantify its discrepancy from the truth, Σ. The loss function is chosen to attain its minimum when C = Σ. Here, in the role of the loss function we adopted the Kullback-Leibler divergence between multivariate normal distributions with equal means, scaled by 2 p to make its values comparable across different population sizes: ℓ ( C , Σ ) = 2 p D K L ( N Σ ‖ N C ) = 1 p [ Tr ( C − 1 Σ ) + lndetC − lndetΣ − p ] (9) Thus ℓ(C, Σ) is expressed in nats/neuron per time bin. When the ground truth is not accessible, the loss cannot be computed directly but may be estimated from data through validation. In a validation procedure, a validation sample covariance matrix C sample ′ is computed from a testing data set that is independent from the data used for computing C. Then the validation loss ℒ ( C , C sample ′ ) measures the discrepancy of C from C sample ′. Here, in the role of validation loss, we adopted the negative multivariate normal log likelihood of C given C sample ′, also scaled by 2 p and omitting the constant term: L ( C , C sample ′ ) = 1 p [ Tr ( C − 1 C sample ′ ) + /lndetC ] (10) Since L( ⋅ , ⋅ ) is additive in its second argument and C sample ′ is an unbiased estimate of Σ, then, for given C and Σ, the validation loss is an unbiased estimate of the true loss, up to a constant: E [ L ( C , C sample ′ ) ] = L ( C , E [ C sample ′ ] ) = L ( C , Σ ) = ℓ ( C , Σ ) + const . (11) Therefore, the validation procedure allows comparing the relative values of the loss attained by different covariance matrix estimators even without access to the ground truth. We drew 30 independent samples with sample sizes n = 250, 500, 1000, 2000, and 4000 from each model and computed the loss ℓ(C, Σ) for each of the five estimators. The hyperparameters of the regularized estimators were optimized by nested cross-validation using only the data in the sample. All the regularized estimators produced better estimates (lower loss) than the sample covariance matrix. However, estimators whose structure matched the true model outperformed the other estimators (Fig. 1 Rows 4 and 5). The validation loss computed by ten-fold cross-validation (see Methods) accurately reproduced the relative values of the true loss as well as the rankings of the estimators even without access to the ground truth (Fig. 1 Row 6). Note that when the ground truth had zero correlations (Column A), Cfactor performed equally well to Cdiag because it correctly inferred zero factors and only estimated the individual variances. Similarly, when the number of latent units was zero (Column C), Csparse+latent performed nearly equally well to Csparse because it correctly inferred zero latent units. With increasing sample sizes, all estimators converged to the ground truth (zero loss) but the estimators with correct structure outperformed the others even for large samples. In Gaussian models, the pairwise partial correlations perfectly characterize the conditional dependencies between the variables. To demonstrate that estimator rankings were robust to deviations from Gaussian models, we repeated the same cross-validated evaluation using pairwise Ising models to generate the data. Ising models have been used to infer functional connectivity from neuronal spike trains [56]. Conveniently, the Ising model has equivalent mathematical form to the Gaussian distribution, x ∼ 1 Z ( J , h ) exp ( 1 2 x T J x + h T x ) (12) but the Ising model is defined on the multivariate binary domain rather than the continuous domain. Both models are maximum-entropy models constrained to match the mean firing rates and the covariance matrix [57]. The partition function Z(J, h) normalizes the distributions on the models’ respective domains. In the Gaussian model, the matrix −J−1 is the covariance matrix; and the mean values are μ = J−1 h. For the Ising model, J is the matrix of pairwise interactions and h is the vector of the cells’ individual activity drives, although they do not have a simple relationship to the means and the covariance matrix. Both distributions have the same structure of pairwise conditional dependencies: zeros in the matrix J indicate conditional independence between the corresponding pair of neurons. Indeed, despite their considerable departure from strictly linear conditional dependencies, Ising models yielded the same relationships between the performances of the covariance estimators as the Gaussian models in cross-validation (Fig. 2). Identical interaction matrices J of the joint distributions over the observable and latent variables were used for both the Gaussian and the Ising models. This simulation study demonstrated that cross-validated evaluation of regularized estimators of the covariance matrices of population activity can discriminate between structures of dependencies in the population. The selection of the most efficient covariance estimators for particular neural circuits is therefore an empirical finding characteristic of the nature of circuit interactions. We recorded the calcium activity of densely sampled populations of neurons in layers 2/3 and upper layer 4 in primary visual cortex of sedated mice using fast random-access 3D scanning two-photon microscopy during visual stimulation (Fig. 3 A–B) [58–60]. This technique allowed fast sampling (100–150 Hz) from large numbers (150–350) of cells in 200 × 200 × 100 μm3 volumes of cortical tissue (Fig. 3 C and D). The instantaneous firing rates were inferred using sparse nonnegative deconvolution [61] (Fig. 3 C). Only cells that produced detectable calcium activity were included in the analysis (see Methods). First, 30 repetitions of full-field drifting gratings of 16 directions were presented in random order. Each grating was played for 500 ms, without intervening blanks. This stimulus was used to compute the orientation tuning of the recorded cells (Fig. 3 D). To estimate the noise correlation matrix, we presented only two distinct directions in some experiments or five directions in others with 100–300 repetitions of each condition. Each grating lasted 1 second and was followed by a 1-second blank. The traces were then binned into 150 ms intervals aligned on the stimulus onset for the estimation of the correlation matrix. The sample correlation coefficients were largely positive and low (Fig. 3 E and F). The average value of the correlation coefficient across sites ranged from 0.0065 to 0.051 with the mean across sites of 0.018. In these densely sampled populations, direct interactions between cells are likely to influence the patterns of population activity. We therefore hypothesized that covariance matrix estimators that explicitly modeled the partial correlations between pairs of neurons (Csparse and Csparse+latent) would have a performance advantage. However, the observed neurons must also be strongly influenced by global activity fluctuations and by unobserved common inputs to the advantage of estimators that explicitly model common fluctuations of the entire population: Cfactor and Csparse+latent. If both types of effects are significant, then Csparse+latent should outperform the other estimators. To test this hypothesis, we computed the validation loss of estimators Csample, Cdiag, Cfactor, Csparse, and Csparse+latent in n = 27 imaged sites in 14 mice. The hyperparameters of each estimator were optimized by nested cross-validation (See S1 Fig. and Methods). Indeed, the sparse+latent estimator outperformed the other estimators (Fig. 4). The respective median differences of the validation loss were 0.039, 0.0016, 0.0029, and 0.0059 nats/cell/bin, significantly greater than zero (p < 0.01 in each comparison, Wilcoxon signed rank test). We examined the composition of the Csparse+latent estimates for each imaged site (Fig. 5 and Fig. 6). Although the regularized estimates were similar to the sample correlation matrix (Fig. 5 A and B), the corresponding partial correlation matrices differed substantially (Fig. 5 C and D). The estimates separated two sources of correlations: a network of linear interactions expressed by the sparse component of the inverse and latent units expressed by the low-rank components of the inverse (Fig. 5 E). The sparse partial correlations revealed a network that differed substantially from the network composed of the greatest coefficients in the sample correlation matrix (Fig. 5 F, G, H, and I). In the example site (Fig. 5), the sparse component had 92.8% sparsity (or conversely, 7.2% connectivity: connectivity = 1−sparsity) with average node degree of 20.9 (Fig. 5 G). The average node degree, i.e. the average number of interactions linking each neuron, is related to connectivity as degree = connectivity⋅(p−1), where p is the number of neurons. The low-rank component had rank 72, denoting 72 inferred latent units. The number of latent units increased with population size (Fig. 6 A) but the connectivity was highly variable (Fig. 6 B): Several sites, despite their large population sizes, were driven by latent units and had few pairwise interactions. This variability may be explained by differences in brain states and recording quality and warrants further investigation. The average partial correlations calculated from these estimates according to Eq. 4 at all 27 sites were about 5 times lower than the average sample correlations (Fig. 6 C). This suggests that correlations between neurons build up from multiple chains of smaller interactions. Furthermore, the average partial correlations were less variable (p = 0.002 Brown-Forsythe test): the coefficient of variation of the average sample correlations across sites was 0.45 whereas that of the average partial correlations was 0.29. While the sample correlations were mostly positive, the sparse component of the partial correlations (‘interactions’) had a high fraction (28.7% in the example site) of negative values (Fig. 5 F). The fraction of negative interactions increased with the inferred connectivity (Fig. 6 D), suggesting that negative interactions can be inferred only after a sufficient density of positive interactions has been uncovered. Thresholded sample correlations have been used in several studies to infer pairwise interactions [26, 62–64]. We therefore compared the interactions in the sparse component of Csparse+latent to those obtained from the sample correlations thresholded to the same level of connectivity. The networks revealed by the two methods differed substantially. In the example site with 7.2% connectivity in Csparse+latent, only 27.7% of the connections coincided with the above-threshold sample correlations (Fig. 5 F, H, and I). In particular, most of the inferred negative interactions corresponded to low sample correlations (Fig. 5 F) where high correlations are expected given the rest of the correlation matrix. We then examined how the structure of the Csparse+latent estimates related to the differences in orientation preference and to the physical distances separating pairs of cells (Fig. 7). Five sites with highest pairwise connectivities were included in the analysis. Partial correlations were computed using Eq. 4 based on the regularized estimate, including both the sparse and the latent component. Connectivity was computed as the fraction of pairs of cells connected by non-zero elements (interactions) in the sparse component of the estimate, segregated into positive and negative connectivities. First, we analyzed how correlations and connectivity depended on the differences in preferred orientations (Δori) of pairs of significantly (α = 0.05) tuned cells. The partial correlations decayed more rapidly with Δori than did sample correlations (Fig. 7 A and D. p < 10−9 in each of the five sites, two-sample t-test of the difference of the linear regression coefficients in normalized data). Positive connectivity decreased with Δori (p < 0.005 in each of the five sites, t-test on the logistic regression coefficient) whereas negative connectivity did not decrease (Fig. 7 G): The slope in the logistic model of connectivity with respect to Δori was significantly higher for positive than for negative interactions (p < 0.04 in each of the five sites, two-sample t-test of the difference of the logistic regression coefficient). Second, we compared how correlations and connectivity depended on the physical distance separating pairs of cells. We distinguished between the lateral distance, Δx, in the plane parallel to the pia, and the vertical distance, Δz, orthogonal to the pia. When considering the dependence on Δx, the analysis was limited to cell pairs located at the same depth with Δz < 30 μm; conversely, when considering the dependence on Δz, only vertically aligned cell pairs with Δx < 30 μm were included. Again, the partial correlations decayed more rapidly both laterally and vertically than sample correlations (Fig. 7 B, C, E, F. p < 10−6 in each of the five sites, for both lateral and vertical distances, two-sample t-test of the difference of the linear regression coefficients in normalized data). Positive connectivity decayed with distance (p < 10−6 in each of the five sites for positive interactions, t-test on the logistic regression coefficient in normalized data) (Fig. 7 E, H, I), so that cells separated laterally by less than 25 μm were 3.2 times more likely to be connected than cells separated laterally by more than 150 μm. Although the positive connectivity appeared to decay faster with vertical than with lateral distance, the differences in slopes of the respective logistic regression models were not significant with available data. The negative connectivity decayed slower with distance (Fig. 7 H and I): The slope in the respective logistic models with respect to the lateral distance was significantly higher for positive than for negative connectivities (p < 0.05 in each of the five sites, two-sample t-test of the difference of the logistic regression coefficients). Functional connectivity is often represented as a graph of pairwise interactions. The goal of many studies of functional connectivity has been to estimate anatomical connectivity from observed multineuronal spiking activity. For example, characteristic peaks and troughs in the pairwise cross-correlograms of recorded spike trains contain statistical signatures of monosynaptic connections and shared synaptic inputs [12, 14, 34, 35, 65]. Such signatures are ambiguous as they can arise from network effects other than direct synaptic connections [66]. With simultaneous recordings from more neurons, ambiguities can be resolved by inferring the conditional dependencies between pairs of neurons. Direct causal interactions between neurons produce statistical dependency between them even after conditioning on the state of the remainder of the network and external input. Therefore, conditional independence shown statistically can signify the absence of a direct causal influence. Conditional dependencies can be inferred by fitting a probabilistic model of the joint population activity. For example, generalized linear models (GLMs) have been constructed to include biophysically plausible synaptic integration, membrane kinetics, and individual neurons’ stimulus drive [67]. Maximum entropy models constrained by observed pairwise correlations are among other models with pairwise coupling between cells [68–72]. Assuming that the population response follows a multivariate normal distribution, the conditional dependencies between pairs of neurons are expressed by the partial correlations between them. Each probabilistic model, fitted to the same data may reveal a completely different network of ‘interactions’, i.e.conditional dependencies between pairs of cells. It is not yet clear which approach provides the best correspondence with anatomical connectivity. Little experimental evidence is available to answer this question. The connectivity graphs inferred by various statistical methods are commonly reported without examining their relation to anatomy. Topological properties of such graphs have been interpreted as principles of circuit organization (e.g. small-world organization) [62–64, 70]. However, the topological properties of functional connectivity graphs can depend on the method of inference [73]. Until a physiological interpretation of functional connectivity is established, the physiological relevance of such analyses remains in question and we did not attempt applying graph-theoretical analyses to our results. Inference of the conditional dependencies also depends on the completeness of the recorded population: To equate conditional dependency to direct interaction between two neurons, we must record from all neurons with which the pair interacts. Unobserved portions of the circuit may manifest as conditional dependencies between observed neurons that do not directly interact. For this reason, statistical models of population activity have been most successfully applied to in vitro preparations of the retina or cell cultures where high-quality recordings from the complete populations were available [67]. In cortical tissue, electrode arrays record from a small fraction of cells in a given volume, limiting the validity of inference of the pairwise conditional dependencies. Perhaps for this reason, partial correlations have not, until now, been used to describe the functional connectivity in cortical populations. Two-photon imaging of population calcium signals presents unique advantages for the estimation of functional connectivity. While the temporal resolution of calcium signals is limited by the calcium dye kinetics, fast imaging techniques combined with spike inference algorithms provide millisecond-scale temporal resolution of single action potentials [74]. However, such high temporal precision comes at the cost of lower accuracy of inferred spike rates. Better accuracy is achieved when calcium signals are analyzed on scales of tens of milliseconds [60, 75]. The major advantage of calcium imaging is its ability to characterize the spatial arrangement and types of recorded cells. Recently, advanced imaging techniques have allowed recording from nearly every cell in a volume of cortical tissue in vivo [59, 60] and even from entire nervous systems [76, 77]. These techniques may provide more incisive measurements of functional connectivity than electrophysiological recordings. The low temporal resolution of calcium signals limits the use of functional connectivity methods that rely on millisecond-scale binning of signals (cross-correlograms, some GLMs, and binary maximum entropy models). Hence, most studies of functional connectivity have relied on instantaneous sample correlations [23, 26, 29, 63]. Although some investigators have interpreted such correlations as indicators of (chemical or electrical) synaptic connectivity, most used them as more general indicators of functional connectivity without relating them to underlying mechanisms. In this study, we sought to infer pairwise functional connectivity networks in cortical microcircuits. We hypothesized that partial correlations correspond more closely to underlying mechanisms than sample correlations when recordings are sufficiently dense. Since neurons form synaptic connections mostly locally and sparsely [78], we a priori favored solutions with sparse partial correlations. Under the assumptions that the recorded population is sufficiently complete and that the model correctly represents the nature of interactions, the network of partial correlations can better represent the functional dependencies in the circuit than correlations. Another approach to describing the functional connectivity of a circuit is to isolate individual patterns of multineuronal coactivations. Depending on the method of their extraction, coactivation patterns may be referred to as assemblies, factor loadings, principal components, independent components, activity modes, eigenvectors, or coactivation maps [79–84]. Coactivation patterns could be interpreted as signatures of Hebbian cell assemblies, i.e. groups of tightly interconnected groups of cells involved in a common computation [79, 82]. Coactivation patterns could also result from shared input from unobserved parts of the circuit, or global network fluctuations modulating the activity of the local circuit [32, 85]. Coactivation patterns and pairwise connectivity are not mutually exclusive since assemblies arise from patterns of synaptic connectivity. However, an analysis of coactivation shifts the focus from detailed interactions to collective behavior. In our study, the functional connectivity solely through modes of coactivations was represented by the factor analysis-based estimator Cfactor. In the effort to account for the joint activity patterns that are poorly explained by pairwise interactions, investigators have augmented models of pairwise interactions with additional factors such as latent variables, higher-order correlations, or global network fluctuations [32, 86–89]. In our study, we combined pairwise interactions with collective coactivations by applying the recently developed numerical techniques for the inference of the partial correlation structure in systems with latent variables [54, 55]. The resulting estimator, Csparse+latent, effectively decomposed the functional connectivity into a sparse network of pairwise interactions and coactivation mode vectors. Inferring the conditional dependencies between variables in a probabilistic model often becomes an ill-posed problem: small variations in the data can produce large errors in the inferred network of dependencies (Fig. 5 C and D). The problem becomes worse as the number of recorded neurons increases until such models lose their statistical validity [90]. As techniques have improved to allow recording from larger neuronal populations, experimental neuroscientists have addressed this problem by extending the recording durations to keep sampling noise in check and verified that existing models are not overfitted [87]. However, ambitious projects already underway, such as the BRAIN initiative [50], aim to record from significantly larger populations. Simply increasing recording duration will be neither practical nor sufficient, and the problem must be addressed by using regularized estimators. Regularization biases the solution toward a small subspace in order to counteract the effects of sampling noise in the empirical data. However, biasing the solution to an inappropriate subspace does not allow significant estimation improvement and hinders interpretation. Several strategies have been developed to limit the model space in order to improve the quality of the estimate. For example, Ganmor et al. [86] developed a heuristic rule to identify the most significant features that must be fitted by a maximum entropy model for improved performance in the retina. As another example of regularization, generalized linear models typically employ L1 penalty terms to constrain the solution space and to effectively reduce the dimensionality of the solution [67]. Our study demonstrates regularization schemes empirically optimized for specific types of neural data. Various model selection criteria have been devised to select between families of models and the optimal subsets of variables in a given model family based on observed data. Despite its high computational demands, cross-validation is among the most popular model selection approaches due to its minimal assumptions about the data-generating process [91]. We evaluated the covariance matrix estimators using a loss function derived from the normal distribution. However, this does not limit the applicability of its conclusions to normal distributions. Other probabilistic models, fitted to the same data, could also serve as estimators of the covariance matrix. If a different model yields better estimation of the covariance matrix than the estimator proposed here, we believe that its structure should deserve consideration as the better representation of the functional connectivity. The results of model selection must be interpreted with caution. As we demonstrated by simulation, even models with incorrect forms of dependencies can substantially improve estimates (Fig. 1). Therefore, showing that a more constrained model has better cross-validated performance than a more complex model does not necessarily support the conclusion that it reveals a better representation of dependencies in the data. This caveat is related to Stein’s Paradox [92]: The biasing of an estimate toward an arbitrary low-dimensional target can consistently outperform a less constrained estimate. We showed that among several models a sparse network of linear interactions with several latent inputs yielded the best estimates of the noise covariance matrix for cortical microcircuits. This finding is valuable in itself: improved estimates of the noise covariance matrix for large datasets are important in order to understand the role of noise correlations in population coding [1, 6, 7, 9, 11] Moreover, this estimation approach provides a graphical representation of the dependencies in the data that can be used to formulate and test hypotheses about the structure of connectivity in the microcircuit. Importantly, the inferred functional interactions were substantially different from the network of the highest sample correlations. For example, the Csparse+latent estimator reveals a large number of negative interactions that were not present in the sample correlation matrix (Fig. 5 F) and may reflect inhibitory circuitry. Distances between cells in physical space and in sensory feature space had a stronger effect on the partial correlations estimated by the Csparse+latent estimator than on sample correlations (Fig. 7 A–F). These differences support the idea that correlations are built up from partial correlations in chains of intermediate cells positioned closer and tuned more similarly to one another, with potentially closer correspondence to anatomical connectivity. These differences may also be at least partially explained by a trivial effect of regularization: the L1 penalty applied by the estimator (Eq. 18) suppresses small partial correlations to a greater extent than large partial correlations, enhancing the apparent effect of distance and tuning. Still, the distinct positive and negative connectivity patterns (Fig. 7 G–I) may reflect geometric and graphical features of local excitatory and inhibitory networks. Indeed, the relationships between patterns of positive and negative connectivities inferred by the estimator resembled the properties of excitatory and inhibitory synaptic connectivities with respect to distance, cortical layers, and feature tuning [23, 78, 93–98]. For example, while excitatory neurons form synapses within highly specific local cliques [78], inhibitory interneurons form synapses with nearly all excitatory cells within local microcircuits [23, 96, 99]. To further investigate the link between synaptic connectivity and inferred functional connectivity, in future experiments, we will use molecular markers for various cell types with follow-up multiple whole-cell in vitro recordings [23, 28] to directly compare the inferred functional connectivity graphs to the underlying anatomical circuitry. Finally, the latent units inferred by the estimator can be analyzed for their physiological functions. For example, these latent units may be modulated under different brain states (e.g. slow-wave sleep, attention) and stimulus conditions (e.g. certain types of stimuli may engage feedback connections) [100, 101]. All procedures were conducted in accordance with the ethical guidelines of the National Institutes of Health and were approved by the Baylor College of Medicine IACUC. The surgical procedures and data acquisition were performed as described in [60]: C57BL/6J mice (aged p40–60) were used. For surgery, animals were initially anesthetized with isoflurane (3%). During the experiments, animals were sedated with a mixture of fentanyl (0.05 mg/kg), midazolam (5 mg/kg), and medetomidine (0.5 mg/kg), with boosts of half the initial dose every 3 hours. A craniotomy was performed over the right primary visual cortex. Membrane-permeant calcium indicator Oregon Green 488 BAPTA-1 AM (OGB-1, Invitrogen) was loaded by bolus injection. The craniotomy was sealed using a glass coverslip secured with dental cement. Calcium imaging began 1 hour after dye injection. All imaging was performed using 3D-RAMP two-photon microscopy [60]. First, a 3D stack was acquired and cells were manually segmented. Then calcium signal were collected by sampling in the center of each cell at rates of 100 Hz or higher, depending on the number of cells. The visual stimulus consisted of full-field drifting gratings with 90% contrast, 10 cd/m2 luminance, 0.08 cycles/degree spatial frequency, and 2 cycles/s temporal frequency. Two types of stimuli were presented for each imaging site: First, directional tuning was mapped using a pseudo-random sequence of drifting gratings at sixteen directions of motion, 500 ms per direction, without blanks, with 12–30 trials for each direction of motion. Second, to measure correlations, the stimulus was modified to include only two directions of motion (in 9 datasets) or five directions (in 22 datasets) and the gratings were presented for 1 second and were separated by 1-second blanks, with 100–300 trials for each direction of motion. All data were processed in MATLAB using the DataJoint data processing chain toolbox (http://datajoint.github.com). The measured fluorescent traces were deconvolved to reconstruct the firing rates for each neuron: First, the first principal component was subtracted from the raw traces in order to reduce common mode noise related to small cardiovascular movements [60]. The resulting traces were high-pass filtered above 0.1 Hz and downsampled to 20 Hz (Fig. 3 C). Then, the firing rates were estimated using by nonnegative deconvolution [61]. Orientation tuning was computed by fitting the mean firing rates for each direction of motion ϕ using two-peaked von Mises tuning functions f ( ϕ ) = a + bexp [ 1 w ( cos ( ϕ − θ ) − 1 ) ] + cexp [ 1 w ( cos ( ϕ − θ + π ) − 1 ) ] where b ≥ c are the amplitudes of the two respective peaks, w is the tuning width, and θ is the preferred direction. The significance of the fit was determined by the permutation test: the labels of the direction were randomly permuted 10,000 times; the p-values of the fits were computed as the fraction of permutations that yielded R2 equal to or higher than that of the original data. Cells were considered tuned with p < 0.05. For covariance estimation, the analysis was limited to the period with two or five stimulus conditions and lasted between 14 and 27 minutes (mean 22 minutes). Cells that did not have substantial spiking activity (those whose variance was less than 1% of the median across the site) or whose activity was unstable (those whose variance in the least active quarter of the recording did not exceed 1% of the variance in the most active quarter) were excluded from the analysis. To compare the performance of the estimators, we used conventional 10-fold cross-validation: Trials were randomly divided into 10 subsets with approximately equal numbers of trials of each condition in each subset. Each subset was then used as the testing sample with the rest of the data used as the training sample for estimating the covariance matrix. The average validation loss over the 10 folds was reported. Since each of the regularized estimators had one or two hyperparameters, we used nested cross-validation: The outer loop evaluated the performance of the estimators with the hyperparameter values optimized by cross-validation within the inner loop. Hyperparameters were optimized by a two-phase search algorithm: random search to find a good starting point for the subsequent pattern search to find the global minimum. The inner cross-validation loop subdivided the training dataset from the outer loop to perform 10-fold cross-validation in order to evaluate each choice of the hyperparameter values. Thus the size of the training dataset within the inner loop comprised 81% of the entire recording. S1 Fig. illustrates the dependence of the validation loss on the hyperparameters of the Csparse+latent estimator for the example site shown in Figs. 3 and 5 and the optimal value found by the pattern search algorithm. When the validation loss was not required, only the inner loop of cross-validation was used on the entire dataset. This approach was used to compute the covariance matrix estimates and their true loss in the simulation study (Fig. 1 Rows 4 and 5) and to analyze the partial correlation structure of the Csparse+latent estimator (Fig. 5–7). Within the inner loop of cross-validation, regularized covariance matrix estimation required only the sample covariance matrix Csample of the training dataset and the hyperparameter values provided by the outer loop. Estimator Cdiag (Eq. 5) used two hyperparameters: the covariance shrinkage intensity λ ∈ [0, 1] and variance shrinkage intensity α ∈ [0, 1]. The variances (the diagonal of Csample) were shrunk linearly toward their mean value 1 p Tr ( C sample ): D = ( 1 − α ) diag ( C sample ) + α 1 p Tr ( C sample ) I (13) The Cdiag estimate was then obtained by shrinking Csample toward D according to Eq. 5. In estimator Cfactor (Eq. 6), the low-rank matrix L and the diagonal matrix D were found by solving the minimization problem ( L , D ) = arg min L ̂ , D ̂ ℒ ( L ̂ + D ̂ , C sample ) , (14) using an expectation-maximization (EM) algorithm for a specified rank of L. After that, the diagonal of D was linearly shrunk toward the its mean diagonal value similar to Eq. 13. In estimator Csparse (Eq. 7), the sparse precision matrix S was found by minimizing the L1-penalized loss with regularization parameter λ: S = arg min S ̂ ≻ 0 L ( S ^ − 1 , C sample ) + λ ‖ S ^ ‖ 1 (15) where S^≻0 denotes the constraint that S^ be a positive-definite matrix and ‖S^‖1 is the element-wise L1 norm of the matrix S^. This problem formulation is known as graphical lasso [102, 103]. To solve this minimization problem, we adapted the alternative-direction method of multipliers (ADMM) [55]. Unlike Cdiag and Cfactor, this estimator does not include linear shrinkage: the selection of the sparsity level provides sufficient flexibility to fine-tune the regularization level. Estimator Csparse+latent (Eq. 8) estimates a larger sparse precision matrix S* of the joint distribution of the p observed neurons and d latent units. S * = ( S S 12 S 12 T S 22 ) , (16) where the p × p partition S corresponds to the visible units. Then the covariance matrix of the observed population is C sparse + latent = ( S − S 12 S 22 − 1 S 12 T ) − 1 (17) The rank of the p×p matrix L = S 12 S 22 − 1 S 12 T matches the number of the latent units in the joint distribution. Rather than finding S12 and S22 separately, L can be estimated as a low-rank positive semidefinite matrix. To simultaneously optimize the sparse component S and the low-rank component L, we adapted the loss function with an L1 penalty on S and another penalty on the trace of L [54, 55]: ( S , L ) = arg min S ̂ , L ̂L ( S ^ − L ^ ) − 1 , C sample + α ‖ S ^ ‖ 1 + β Tr ( L ^ ) (18) The trace of a symmetric semidefinite matrix equals the sum of the absolute values of its eigenvalues, i.e. its nuclear norm; penalty on Tr(L) favors solutions with few non-zero eigenvalues or, equivalently, low-rank solutions while keeping the convexity of the overall optimization problem [104, 105]. This allows using convex optimization algorithm such as ADMM to be applied with great computational efficiency [55]. The partial correlation matrix (Eq. 4) computed from Csparse+latent includes interactions between the visible and latent units and was used in Fig. 5 C and D and Fig. 6 C, and Fig. 7 D–F). The partial correlation matrix computed from S alone expresses strengths of pairwise interactions P sparse = − ( diag ( S ) ) − 1 2 S ( diag ( S ) ) − 1 2 (19) and were used in Fig. 5 F, G, H. The MATLAB code for these computations is available online at http://github.com/atlab/cov-est. Special attention was given to estimating the variances. All evaluations and optimization in this study were defined with respect to the covariance matrices. However, neuroscientists often estimate a common correlation matrix across multiple stimulus conditions when the variances of responses are conditioned on the stimulus [106, 107]. In this study, we too conditioned the variances on the stimulus but estimated a single correlation matrix across all conditions. Here we describe the computation of the validation loss (Eq. 10) when the variances were allowed to vary with the stimulus condition. Let Tc and T c ′ denote the sets of time bin indices for the training and testing samples, respectively, limited to condition c. Similar to Eq. 2, the training and testing sample covariance matrices for condition c are C c , sample = 1 n c ∑ t ∈ T c ( x ( t ) − x ¯ c ) ( x ( t ) − x ¯ c ) T (20) and C c , sample ′ = 1 n c ′ ∑ t ∈ T c ′ ( x ( t ) − x ¯ c ) ( x ( t ) − x ¯ c ) T (21) Here nc and n c ′ denote the sizes of Tc and T c ′, respectively. Note that x ‾ c = 1 n c ∑ t ∈ T c x ( t ) is estimated from the training sample but used in both estimates, making C c , sample ′ an unbiased estimate of the true covariance matrix, Σ. As such, C c , sample ′ can be used for validation. The common correlation matrix Rsample is estimated by averaging the condition-specific correlations: R sample = 1 n ∑ c n c ( V c , sample − 1 2 C c , sample V c , sample − 1 2 ) = 1 n ∑ c ∑ t ∈ T c z ( t ) z ( t ) T , (22) where n = ∑ c n c and Vc, sample = diag(Cc, sample) is the diagonal matrix containing the sample variances. Then Rsample is simply the covariance matrix of the z-score signal z(t)=Vc,sample−12(x(t)−x¯c) of the training sample. For consistency with prior work, we applied regularization to covariance matrices rather than to correlation matrices. The common covariance matrix was estimated by scaling Rsample by the average variances across conditions V sample = 1 n ∑ c n c V c , sample: C sample = V sample 1 2 R sample V sample 1 2 (23) Note that Csample differs from the sample covariance matrix computed without conditioning the variances on c and this computation helps avoid any biases that would be introduced by ignoring changes in variance. The covariance matrix estimators Cdiag, Cfactor, Csparse or Csparse+latent convert Csample into its regularized counterpart denoted here as Creg. To evaluate the estimators, we regularized the conditioned variances by linear shrinkage toward their mean value across all conditions. This was done by scaling Creg by the conditioned variance adjustment matrix Q c = δ I + ( 1 − δ ) V sample − 1 V c , sample to produce the conditioned regularized covariance matrix estimate: C c , reg = Q c 1 2 C reg Q c 1 2 (24) The variance regularization parameter δ ∈ [0, 1] was optimized in the inner loop of cross-validation along with the other hyperparameters. The overall validation loss is obtained by averaging the validation losses across all conditions: 1 ∑ c n c ′ ∑ c n c ′ L ( C c , reg , C c , sample ′ ) (25) With negative normal log-likelihood as the validation loss (Eq. 10) and the unbiased validation covariance matrix Cc, sample, the loss function in Eq. 25 is an unbiased estimate of the true loss. Hence, it was used for evaluations reported in Fig. 4. For simulation, ground truth covariance matrices were produced by taking 150 independent samples from an artificial population of 50 independent, identically normally distributed units. The covariance matrices were then subjected to the respective regularizations to produce the ground truth matrices for the simulation studies (Fig. 1 Row 2). Samples were then drawn from multivariate normal distributions models with the respective true covariance matrices to be estimated by each of the estimators. For Ising models, the negative inverse of the true covariance matrix was used as the matrix of coupling coefficients and the sampling was performed by the Metropolis-Hastings algorithm.

10.1371/journal.pbio.1000473

Brief Bursts Self-Inhibit and Correlate the Pyramidal Network

Inhibitory pathways are an essential component in the function of the neocortical microcircuitry. Despite the relatively small fraction of inhibitory neurons in the neocortex, these neurons are strongly activated due to their high connectivity rate and the intricate manner in which they interconnect with pyramidal cells (PCs). One prominent pathway is the frequency-dependent disynaptic inhibition (FDDI) formed between layer 5 PCs and mediated by Martinotti cells (MCs). Here, we show that simultaneous short bursts in four PCs are sufficient to exert FDDI in all neighboring PCs within the dimensions of a cortical column. This powerful inhibition is mediated by few interneurons, leading to strongly correlated membrane fluctuations and synchronous spiking between PCs simultaneously receiving FDDI. Somatic integration of such inhibition is independent and electrically isolated from monosynaptic excitation formed between the same PCs. FDDI is strongly shaped by I(h) in PC dendrites, which determines the effective integration time window for inhibitory and excitatory inputs. We propose a key disynaptic mechanism by which brief bursts generated by a few PCs can synchronize the activity in the pyramidal network.

The neocortex of the mammalian brain contains many more excitatory neurons than inhibitory neurons, yet inhibitory neurons are essential components of neocortical circuitry. Inhibitory neurons form dense and intricate connections with excitatory neurons, which are mainly pyramidal cells. One prominent pathway formed between pyramidal cells and inhibitory Martinotti cells is frequency-dependent disynaptic inhibition (FDDI), which mediates a strong inhibitory signal in the microcircuitry of the neocortex. Here, we reveal deeper insight into how FDDI is mediated and recruited within the circuit, showing that short simultaneous bursts in four pyramidal cells are sufficient to exert FDDI in all neighboring pyramidal cells within the dimensions of a cortical column. As few as three synchronous action potentials in three pyramidal cells can trigger FDDI. This powerful inhibition is mediated by only a few inhibitory neurons yet correlates membrane potential fluctuations, leading to synchronous spiking between pyramidal cells that simultaneously receive FDDI. The inhibitory signals are independent and electrically isolated from excitation mediated by neighboring PCs via basal dendrites. We propose FDDI as an important pathway that is readily activated by brief bursts of action potentials and correlates neocortical network activity.

The mammalian neocortex consists of neurons that form an intricate network of recurrent circuits [1]–[3]. The synaptic wiring between cells follows a number of stereotypic rules including targeting specific domains of neurons, specific connection probabilities, target neuron preferences, and specific short-term synaptic dynamics [1]–[5]. Revealing these rules is essential to understand the mechanisms that generate the response of a cortical column (or functional unit) to any external input. In particular, it is crucial to identify the synaptic pathways that enable the neocortex to appropriately respond to all possible environmental stimuli. Neocortical neurons receive excitatory and inhibitory inputs over a variety of different network activity states [6] that seem to be proportionally regulated [7]. This balanced excitatory and inhibitory activity is remarkable since the large majority of cells in the neocortex are (excitatory) pyramidal cells (PCs), only around 25% are inhibitory GABAergic interneurons [8],[9], and almost 90% of the neocortical synapses are presumably excitatory [10]. This relatively small population of interneurons is responsible for generating a precisely matched inhibition for a variety of cortical network states. One synaptic principle for dynamically adjusting the level of excitation within a neocortical column is the use of dynamically depressing excitatory synapses [11]–[13], but how inhibitory synaptic pathways ensure dynamic application of balanced inhibition as a function of the moment-to-moment excitation of the neocortical column is not clear. A disynaptic pathway and dynamic circuit mechanism allowing an activity-dependent recruitment of inhibition was recently reported: frequency-dependent disynaptic inhibition (FDDI) between PCs is indeed a common pathway in multiple cortical areas that is dynamically regulated by the firing rate and the number of presynaptic PCs [14]–[17]. In contrast to many other cortical connections, the PC–Martinotti cell (MC) synapse is strongly facilitating. In response to high frequency stimulation of a PC, spiking activity of MCs can be recruited, thus providing a level of inhibition that depends on the previous excitation level in the network. MCs display a characteristic ascending axonal arborization up to layer 1 [18], and they are the only interneurons that target the combination of oblique, apical, and tuft dendrites of their neighboring PCs [3],[14]. FDDI has so far been explored mainly as a pairwise interaction between PCs and MCs, but little is known about how this synaptic pathway could operate to dynamically apply inhibition to the microcircuit as a function of multi-cellular activity. Here, we used multi-neuron whole cell recordings to characterize summation properties of FDDI between layer 5 thick tufted PCs within the dimensions of a neocortical column. FDDI tends to summate linearly with coincident excitatory postsynaptic potentials (EPSPs) from neighboring PCs but may also shunt some input arriving at the apical dendrite. Three to four PCs firing simultaneously are sufficient to generate FDDI in all PCs within the dimensions of a cortical column, and eight to nine PCs can saturate the amount of hyperpolarization recorded from their somata. A brief, high frequency burst in only a few PCs can therefore constitute a gating mechanism for further excitatory input to the apical dendrites of the entire column. This inhibition promotes subthreshold correlations and synchronous spiking in PCs. In order to study the network properties of FDDI, we obtained simultaneous whole-cell recordings from neighboring thick tufted layer 5 PCs and in some cases also layer 5 MCs. In total, 1,185 PCs and 14 MCs in 283 clusters from 133 animals were recorded for this study. Figure 1 illustrates the basic components (A,B) that mediate FDDI. A presynaptic PC (red) projecting onto an MC (blue) excites the MC using a strongly facilitating synapse, which in turn gives rise to a delayed inhibition in another postsynaptic PC (FDDI, black). Monosynaptic excitatory connections between PCs occurred in 14% of all tested cases (probability of occurrence was 0.14; 463 out of 3,342 tested connections), while PC-MC connections occurred far more frequently (0.43; 26/61) and MC-PC connections had a probability of occurrence of 0.31 (18/58). The entire FDDI loop occurs with a probability of 0.283 (859/3,041), which is more than double the monosynaptic connectivity between two PCs. Silberberg and Markram (2007) previously showed a strong modulation of FDDI by Ih currents [14]. Blocking Ih currents with extracellular application of zd7288 leads to larger amplitudes (average 75% increase, n = 23, μctrl = 0.99±0.5 mV, μzd7288 = 1.73±0.99 mV, p = 0.0002, paired t test) and longer decay time constants (250% increase, μctrl = 0.051±0.01 s, μzd7288 = 0.182±0.071 s, p = 7.64e-9) of FDDI (Figure 1C). In some cases (3 out of 26) FDDI disappeared after Ih block. Since zd7288 blocks Ih irreversibly [19], we do not know whether the disappearance is due to a drug action or a general rundown. On the other hand, Ih block never leads to FDDI appearance de novo (n = 19). In order to understand whether the effects can be attributed to Ih on the intermediate interneuron or on the postsynaptic PC, we recorded from the entire disynaptic pathway while Ih was blocked. Facilitating EPSPs from PCs to MCs were only slightly changed in the presence of zd7288 (average 8% decrease of maximal depolarization; n = 5, μctrl = 2.275±1.961 mV, μzd7288 = 2.431±1.825 mV, p = 0.384, paired t test), whereas MC input onto PCs displayed increased synaptic summation (Figure 1D). Thus, the strong effect of zd7288 on FDDI is likely to be mediated by Ih in PCs. PCs receiving both disynaptic inhibition and monosynaptic excitation from their neighboring PCs displayed the tendency of a frequency-dependent transition from a net depolarization to hyperpolarization (Figure 1E, n = 4). Blockage of Ih resulted in increased frequency dependence, enhancing both low-frequency depolarization and high-frequency hyperpolarization. Together with the observed shortening of synaptic events, this suggests that Ih in PCs acts to localize synaptic inputs, both spatially and temporally. Monosynaptic excitation between PCs mainly targets their basal dendrites [20] while FDDI mainly targets their apical and tuft dendrites [14]. It is not clear to what extent these two inputs interact. We therefore activated both pathways simultaneously and quantified the linearity of summation. Clusters of three PCs, with a PC receiving FDDI from a neighboring PC and a direct excitatory connection from another PC, were stimulated in a way that FDDI and a direct EPSP coincided (Figure 2A). We observed supra-, sub-, and linear amplitude summation in the soma (Figure 2B) in different experiments, and on average there was no significant difference in EPSP amplitude between control and coinciding FDDI (Figure 2C, n = 21, μctrl = 1.885±1.334 mV, μFDDI = 1.808±1.154 mV, p = 0.295, paired t test). Inhibition in the distal dendrites may not shunt the peak amplitude of fast AMPA-mediated EPSPs from the basal dendrites but could reduce the total charge. We did not, however, observe any significant change in the integral of the EPSPs (μctrl = 0.08±0.057 mV*ms, μFDDI = 0.074±0.048 mV*ms, p = 0.15, paired t test). Next, we used the same protocol to investigate the summation of FDDI with excitatory input to the apical dendrite (Figure 2A). Instead of stimulating a neighboring PC, we synchronously injected a brief current (aEPSC) into the trunk of the apical dendrite (50–350 µm away from the soma) that mimicked EPSP kinetics (τrise = 0.5 ms, τdecay = 2 ms) and peak amplitude (200–500 pA, tuned to match a somatic voltage depolarization of 1–4 mV). The somatic amplitude (Figure 2D) and integral of dendritic aEPSPs was slightly reduced by FDDI input in a distance dependent manner and as a function of the number of presynaptic PCs. We used fast AMPA kinetics for the aEPSPs, which might underestimate the shunting effect by FDDI on events with slower kinetics, namely NMDA components and EPSPs filtered by dendritic attenuation. A further technical limitation of artificial EPSPs via dendritic recording besides the focalization is the fact that excitatory synapses rather target spines, not the trunk like the patch electrode. Nevertheless, these data suggest that FDDI is more effective in shunting synaptic input from the apical and tuft dendrites than input from the basal dendrites, revealing a dual and separable action between layer 5 PCs: direct excitation mostly onto basal dendrites, and indirect inhibition mostly onto the apical and tuft dendrites. This finding is supported by the anatomical separation of the inputs (Figure 1A, see also [3],[14],[20]). We performed a set of experiments to estimate the number of MCs that meditate FDDI between two PCs. We stimulated a presynaptic PC that synapses onto an MC, which in turn projects to another PC (Figure 3A). Every other iteration, the MC was prevented from discharge by a hyperpolarizing step current, thereby isolating the effect of this one MC on the FDDI recorded in the postsynaptic PC. FDDI amplitude was reduced to 47.5%±38.1% (integral to 45.3%±35%) when the single MC was prevented from participating (Figure 3B, n = 7, amplitude: μwMC = 0.692±0.417 mV, μw/oMC = 0.460±0.446 mV, p = 0.0011, integral: μwMC = 0.08±0.082 mV*ms, μw/oMC = 0.053±0.049 mV*ms, p = 0.1148, paired t tests). These results show that although on average multiple MCs participate in FDDI, a single MC can make a significant contribution to the overall FDDI produced in a target PC. The exact number of intermediate MCs is not straightforward to extrapolate. Assuming linear amplitude summation of the MCs' inhibitory postsynaptic potentials (IPSPs), three MCs (μwMC/(μwMC−μw/oMC)) participate on average in FDDI upon stimulation of one layer 5 PC (range 1–28 MCs). We might have indirectly prevented further neighboring MCs from spiking through electrical coupling by hyperpolarizing the recorded MC, which might have resulted in an underestimate of participating MCs. Figure 3C shows an example of two MCs coupled via electrical synapses. Their coupling coefficient was 0.11 for hyperpolarizing step currents, which is within the range that has been found in previous studies [21],[22]. Due to low pass filtering, miniature EPSPs in one MC do not pass to the other MC (arrows in Figure 3C). For the same reason, the coupling coefficient was only 0.02 for action potentials. Thus, electrical synapses can only play a role in the communication in the FDDI network if synaptic inputs summate with a sufficiently slow time constant so that the signal is not eliminated by low-pass filtering. The same two MCs were targeted by two PCs that were recorded at the same time (Figure 3D) providing direct evidence for PC-MC divergent and PC-MC convergent connectivity. We also found multiple cases of MC-PC divergent connectivity (data not shown), indicating that neighboring PCs might share a common pool of MCs for feed-forward and feed-back inhibition. The high degree of interconnectivity between PCs and MCs results in subthreshold correlations between PCs (Figure 4A,B), showing a high correlation coefficient for simultaneous FDDI in different PCs (n = 28, μFDDI-FDDI = 0.892±0.125) and significantly lower ones for control conditions (n = 28, μCTRL-CTRL = 0.085±0.364, n = 26, μFDDI-CTRL = −0.070±0.331, p<0.00001, ANOVA with Scheffe correction). This correlation was calculated with average traces and is therefore based on mean responses. In order to estimate the similarity of FDDI in different PCs arising from stimulating a single PC, we performed a trial-to-trial analysis of divergent FDDI responses. In principle, divergent FDDI connectivity may be mediated by a high degree of divergence from PCs onto many different MCs and/or a high degree of divergence from MC to PCs (see Figure 4C for illustration). To quantify the amount of common FDDI input, we defined a “Dissimilarity Index” (DI), which is the root mean squared of mean subtracted traces (see Methods). DI was calculated pairwise between single trial traces, either between simultaneous traces of different cells, or, as a control, between traces of the same (or different, data not shown) cells but from different trials. If each postsynaptic PC received FDDI from a different set of interneurons (as illustrated in the left part of Figure 4C), the inhibitory response in the different postsynaptic PCs would not co-vary from trial to trial, resulting in a strong dissimilarity (high DI, as control). In contrast, if each postsynaptic PC received common input from the same set of interneurons (right part of Figure 4C), single-trial FDDI responses between different PCs should be more similar (low DI, smaller than control). If single-trial responses in PCs were identical, DI would be zero. In all tested cases except one, we found a lower DI of simultaneously acquired traces than that of non-simultaneously acquired traces, indicating a high degree of common MC input to neighboring PCs. The data of the illustrated example as well as 43 more cases suggest high MC to PC divergence (Figure 4D, n = 44, μac = 0.0148±0.0033 mV, μar = 0.0178±0.0028 mV, p = 2e-12, paired t test). Direct connections diverging from a PC to two or more postsynaptic PCs did not have a significantly different DI (n = 11, μac = 0.0162±0.0054 mV, μar = 0.0168±0.0058 mV, p = 0.1063, paired t test). These results show that divergent FDDI from a single PC onto multiple neighboring PCs is not because of a large set of MCs but can be accounted for by a highly divergent MC-PC connectivity. Combined with these findings on the contribution of a single MC on FDDI (Figure 3A–C), we conclude that the high prevalence of FDDI is supported by both PC-MC divergence as well as a high degree of MC-PC divergent connectivity. This MC-PC divergence causes the inhibitory inputs onto neighboring PC to be precisely timed and, together with the mean-based correlations (Figure 4B), enables FDDI to facilitate synchronization of PC activity. Figure 5 shows this synchronization of multiple PCs in the suprathreshold regime. A single presynaptic PC (Figure 5A, red) was stimulated with high frequency (15 spikes at 70 Hz) and elicited FDDI in multiple postsynaptic PCs (black, left column). Postsynaptic PCs were stimulated with a suprathreshold step current (resulting in low frequency spiking of 2–8 Hz) in the presence of FDDI input (right column), and as a control, without FDDI input (middle column). Without FDDI input, firing of PCs already displayed some variability from trial to trial, probably due to spontaneous membrane potential fluctuations and drifts over the long duration of the stimulus paradigm. As can be seen in the peristimulus time histogram (Figure 5B), however, the probability of spiking is reduced during the beginning of FDDI (blue color), followed by a period of “rebound spiking” at the end and briefly after FDDI (red color). We quantified this effect by counting spikes during this first (left part of Figure 5C, n = 11, μCTRL = 10.7±3.92, μFDDI = 6.3±5.1, p = 0.0048, paired t test) and second 100 ms time window (right part of Figure 5C, μCTRL = 12.6±7.7, μFDDI = 18±5.1, p = 0.0015, paired t test) of 22 repetitions in control and FDDI condition. This effect is also quantified by a correlation-based spike timing reliability measure (Figure 5D; standard deviation of the Gaussian used for convolution with the spike trains was 10 ms; for details on the method, see [23]). Spike timing reliability between single repetitions of pairs of postsynaptic PCs increases during FDDI (left part of Figure 5D, n = 18, μCTRL = 0.197±0.109, μFDDI = 0.241±0.1034, p = 0.018, paired t test), which also holds true if a time window before FDDI onset is chosen as a control (right part of Figure 5D, see methods, μBEFORE = 0.162±0.064, μDURING = μFDDI, p = 0.003, paired t test). Thus, FDDI can lead to synchronous pauses followed by subsequent synchronous spiking in neighboring PCs. Next, we investigated the spatial and temporal integration properties of FDDI in a single PC when multiple presynaptic PCs are stimulated at the same time. Figure 6A shows that an increased number of stimulated PCs leads to a reduced delay of MC firing. The MC was recorded in cell-attached mode so that the intracellular medium remained undisturbed. Not only does the discharge onset take place earlier by tens of milliseconds (μ3pre = 0.114±0.021 s, μ2pre = 0.202±0.036 s, p = 0.000021, two-sample t test), but also the number of APs fired by the MC increases (Figure 6B). Figure 6C shows the same type of experiment with the MC recorded in whole-cell mode. Stimulation of two PCs simultaneously can lead to earlier and more numerous spikes (Rep. 1, blue traces), a PC-MC convergence configuration, which would lead to earlier and larger FDDI in a PC postsynaptic to the MC (supralinear summation). On the other hand, simultaneous stimulation may also lead to earlier MC spiking only (Rep. 2, light blue traces), which should lead to reduced FDDI amplitude in postsynaptic PCs (sublinear summation). In view of the latency shortening and increased discharge in MCs, we analyzed both amplitudes and onset latencies of FDDI mediated by several presynaptic PCs onto a single postsynaptic one (Figure 6D–F, Figure 7). Similarly to the previous summation experiments we compared FDDI in response to synchronous stimulation of two PCs (Figure 6D, black traces in the right column) to the off-line calculated sum of the separate stimulations (left and middle column, gray dashed traces, and gray traces in right column). As expected, a variety of different responses were found (Figure 6E), ranging from linear summation (left), reduced amplitude with reduced onset delay (left middle), increased amplitude with reduced onset delay (right middle), and increased amplitude with same delay (right). Possible underlying connectivity schemes are depicted in Figure 6F. In cases where the onset delay was shortened, it is very likely that the FDDI is mediated by MCs receiving convergent common excitation from both PCs. The common input decreases the discharge onset of the MC(s) and results in earlier onset of inhibition (see Figure 6A–C). Networks that did not exhibit a latency decrease following co-stimulation may also involve MCs receiving common input but not exclusively (Figure 6F, right). The origin of amplitude summation is more complicated, since both supra- and sublinear summation can be explained by convergent PC-MC inputs: if an intermediate MC can be reliably activated only by convergent input, this will result in an average supralinear increase in amplitude. However, if an MC discharges reliably following inputs from both PCs individually, such that co-activation does not significantly increase the number of APs, the result is sublinear amplitude summation. Our results show that on average the latency was shortened by 33.7±35.8 ms (n = 103, p<0.0001, two-tailed t test), and the amplitude increase was supralinear (μsync = 1.232±0.723 ms, μsummed = 1.111±0.877 mV, n = 103, p = 0.00096, paired t test). These results indicate that co-stimulation of presynaptic PC pairs increases FDDI in a supralinear manner due to the high degree of PC-MC convergence. How does FDDI summate when more than two neighboring PCs are active? We stimulated an increasingly larger number of PCs and recorded FDDI in another PC (Figure 7A, gray shades of the traces according to the number of stimulated cells). FDDI monotonically increased in amplitude and voltage integral, which saturated when eight to nine PCs were simultaneously stimulated (Figure 7B). In order to compare the pooled data of many recorded clusters, we used a nonlinearity index for amplitude and integral summation [15]. FDDI summated on average supralinearly for the amplitude as well as for the integral following stimulation of two presynaptic PCs (Figure S1, n = 103, pamp = 1.977e-6, pint = 2.710e-13, one-sample t test). Stimulation of three or more (Figure 7C and S1) presynaptic PCs increased the supralinearity of integral and amplitude of FDDI, and also decreased the onset delay. Saturation levels of amplitude and integral difference were reached at around 60% and 70% when six to seven PCs were stimulated simultaneously (Figure 7C). A remarkable feature of FDDI was its abundance in the layer 5 network. Upon stimulation of four PCs simultaneously, all recorded neighboring PCs were inhibited (Figure 7D). Our typical stimulation protocol used to elicit and reliably identify FDDI contained multiple APs (15) and high frequencies (70 Hz), a condition that is presumably unlikely to be experienced by PCs in the intact brain. However, the onset of FDDI after this long-train stimulation is variable between cells (Figure 8A), and in several cases, less APs would have been sufficient to trigger FDDI by a single PC, since the hyperpolarization can start off briefly after stimulus onset (Figure 8B, n = 439, mean = 0.110 s, eight presynaptic APs). We also know that synchronous activation of multiple PCs can significantly decrease the FDDI onset (Figures 6, 7 and S1). In order to examine whether FDDI can be triggered with few APs only, we stimulated three presynaptic PCs with only three APs at 70 Hz, mimicking the spiking output evoked by dendritic calcium spikes [24]. As can be seen in Figure 8C, even this condition is sufficient to elicit FDDI reliably, with a probability of occurrence of 0.23 (23 out of 99 tested different quadruplet combinations of cells), a mean onset delay of 0.068±0.012 s, and amplitudes of up to several millivolts (μ = 1.22±0.77 mV; range 0.25–3.7 mV). This illustrates that brief synchronous bursts (∼three APs at 70 Hz) of only three PCs are able to trigger FDDI in neighboring PCs, a condition that is likely to be relevant in the in vivo situation. This study reveals the key properties of one of the physiologically and anatomically most distinguished disynaptic inhibitory pathways in the neocortex, FDDI: the number of PCs required, divergent and convergent properties to and from MCs, spatio-temporal principles that govern the integration of the inhibition applied through this pathway, the dependency of this form of inhibition on Ih currents, and its potential influence on the functioning of the network of thick tufted PCs in the somatosensory neocortex of juvenile rats. Previously, the summation properties for convergent FDDI have been investigated for two [15] or three [14] stimulated presynaptic PCs, and the activity-dependent recruitment of MCs was extrapolated for the case of multiple active PCs [15]. Two of our most important findings are that (a) every neighboring (<150 µm intersomatic distance) thick tufted layer 5 PC is affected by FDDI when four or more PCs burst simultaneously and that (b) the FDDI amplitude saturates at the somatic recording site at resting condition when eight to nine PCs are stimulated simultaneously. We find the low number of PCs necessary to trigger FDDI in all neighboring PCs especially remarkable—it shows that in the high frequency, high correlation range the major signaling between PCs is (after an initial brief excitatory response) inhibitory. The observed FDDI saturation may be caused by several reasons: limited recruitment of MCs (due to limited connectivity or limited number of MCs), reduction in the driving-force of the inhibitory signal in the apical dendrite when it reaches the GABAA reversal potential, saturating firing rates in MCs, and frequency-dependent synaptic depression of the MC-PC connection. It is likely that all these factors contribute to this early saturation. Summation properties as shown in Figures 3, 4, and 6 indicate that only a few MCs are actually recruited by a cluster of PCs. We cannot, however, state that MC recruitment is saturated by stimulation of eight to nine PCs since multiple MCs could mutually shunt their inhibitory signals in a postsynaptic PC and therefore mask the contribution of additional MCs to FDDI. Further, it should be considered that the saturation might not hold in different cortical activity states. For example, if reduced driving force was a major reason for saturation at rest, FDDI might saturate at a later stage at more active cortical states. A high level of excitatory synaptic input to the apical dendrite would require larger activation of the FDDI pathway in order to reach saturation. Kapfer and colleagues [15] extrapolated PC-PC and PC-MC connectivity data to predict a saturation curve for MC recruitment (see Figure 6 therein). Although the studies are not directly comparable and were performed in different cortical layers, our data suggest a smaller dynamic range for recruitment of inhibition and earlier saturation than previously reported. What is not exactly known and not addressed in the current study is the degree of synchrony that brief bursts of neighboring need to have in order to trigger FDDI. We only tested simultaneous stimulations of PCs; a jitter in the PC firing is likely to alter the efficiency of FDDI recruitment and amplitude. Neighboring neocortical cells can show highly correlated activity patterns both in vitro [25] and in vivo [26],[27]. Recently, it has been shown that the synchrony of subthreshold membrane potential fluctuations depends on the behavioral state of the animal [28]. FDDI acts as a synchronizer of subthreshold membrane potential between PCs in two ways. Multiple PCs, targeted by the same MC, receive FDDI simultaneously, resulting in a high correlation coefficient (Figure 5B). Moreover, due to the reliability of the MC-PC synapse, and its high divergence, the inter-trial variability is mainly due to the summation of the facilitating PC-MC synaptic response. A previous study has also shown that the synaptic dynamics from interneurons are virtually identical across postsynaptic neurons of the same class, which may also underlie the high subthreshold correlations mediated by MCs [29]. Simultaneous responses in different postsynaptic PCs are therefore more similar to each other than the responses of the same PC for different iterations. The high correlation in FDDI across PCs suggests that inhibitory inputs from MCs to PCs may contribute to subthreshold correlations observed between neighboring PCs under in vivo conditions [26],[27]. Photostimulation studies have suggested that interneurons with adapting firing pattern (like MCs) are less specific or selective concerning the targeting of their synaptic input and output [30], a finding which is in agreement with the high degree of FDDI divergence we report (Figure 5A). Two dynamically different disynaptic inhibitory pathways have been identified in the neocortex [14] and their equivalents in the hippocampus [31]. The pathways differ in their dynamical as well as morphological properties, with the delayed, frequency-dependent pathway activated by MCs (belonging to the low threshold spiking (LTS) class of interneurons), triggered by facilitating connections from PCs and target PC dendrites. The other inhibitory pathway conversely is “immediate” and time-locked to PC single APs. It is mediated by depressing connections onto fast-spiking cells, typically PV-expressing basket cells, which in turn target PC perisomatic regions. These interneurons also mediate strong feed-forward inhibition activated by the thalamocortical pathway that has received attention in recent studies, showing that FS interneurons respond to thalamic input by discharge that precedes that of their excitatory neighbors [32]–[34]. LTS cells, on the other hand, receive only weak thalamic input [34],[35] (but see [36]), suggesting that their activation is primarily intracortical, optimally driven by high-frequency burst discharge of PCs. One implication of the dendritic locus of MC-PC connections, reaching up to the distal dendritic tuft [14], suggests that FDDI has a role in regulating dendritic excitation, including intrinsic excitability in the form of calcium [24],[37] and NMDA spikes [38]. Indeed, in a recent study, Murayama and colleagues [39] demonstrated direct blocking of dendritic calcium spikes by FDDI in older animals (24–40 d old), showing that FDDI is preserved in development and can regulate dendritic excitability in layer 5 PCs. The authors also showed that GABAergic inhibition to PC dendrites originated from layer 5 interneurons and was crucial for enabling a wide dynamic range of calcium responses in vivo, correlated to the intensity of sensory stimulus. Therefore, FDDI might be a precisely matching antagonist of active excitatory conductances like calcium spikes, both being triggered by high frequency bursts. Our study was performed in younger animals, suggesting that development of FDDI onto PC dendrites precedes the maturation of their excitability, which occurs after the third postnatal week [40]. Ih is a prominent current with increasing channel density along the dendrites of layer 5 PCs [41],[42]. It renders the apical (and presumably also the basal) dendrites disconnected from the soma [43] by counteracting any polarization deviating from the resting potential. The decay times of de- and hyperpolarizing inputs are substantially shortened, allowing for a higher temporal precision in the processing of information. Due to its increasing density along the dendrites it also renders EPSP shape and time course site independent [44]. Here we showed that Ih can change the gain between excitation and inhibition for train stimulations, thus increasing the dynamic frequency range. A modulation of this channel conductance might be an approach to profoundly alter this inhibitory pathway [42],[43]. Studies showed Ih presence in MCs as well, but there seem to be exceptions to this finding, with not all MCs expressing the Ih mediated sag in response to hyperpolarizing step currents [18]. We also did not find a prominent sag in MCs that participated in FDDI (n = 3, see also Figure 3C). The relative contribution of the various MC populations to FDDI remains to be elucidated [45]. MCs can be modulated by various means. Acetylcholine receptor agonists lead to increased firing in MCs [46], which might influence the plasticity rules at the apical dendrite of PCs [47]. LTS cells, have been shown to synchronize and oscillate in response to a G-protein coupled glutamate receptor antagonist [48]. This synchronization, mediated by electrical synapses, should enhance and broaden the effect of FDDI in the PC population. Compared to other cell types, MCs seem to be particularly susceptible to changes in the general cortical activity state [49]. Spiking activity (in certain frequency ranges) in MCs can trigger intracellular endocannabinoid signaling that eventually leads to hyperpolarization, and thus reduced excitability [50],[51]. It remains to be elucidated which modulations play strong roles under physiological conditions and to what extent FDDI properties reported in the present study are altered. Several aspects of FDDI still remain to be elucidated. So far, layer 5 thick tufted PCs and layer 3 PCs have been shown to display FDDI [14],[15]. Cortical-callosal layer 5 PCs with a slender apical dendrite lacking tuft dendrites do not seem to feature this type of inhibition [52]. Also, PCs in layer 6 do not show any measurable FDDI (Berger and Markram, unpublished data). It is, however, not clear whether these potential pathways require a larger number of active neurons to become observable. It remains to be shown whether other PC classes are inhibited in a similar manner and whether this inhibition is mediated via the same MCs. Apart from the FDDI mediated within the same layer, it is possible that presynaptic activity in one layer will inhibit PCs in a different layer. Kapfer and colleagues showed that MCs in layer 5 mediated FDDI between layer 3 PCs [15], which is in agreement with the neurons' axonal terminal distribution [18],[45]. It is not known, however, if these MCs are the ones that also mediate FDDI onto layer 5 PCs and whether they are also recruited by layer 5 PCs. It also remains to be elucidated whether supragranular MCs also participate in FDDI between layer 5 PCs. Layer 5 PCs do innervate layer 2/3, and these MCs target preferentially supragranular layers and possibly also the apical trunk or tuft of layer 5 PCs. Subpopulations of SOM expressing interneurons are now GFP labeled in various mouse strains [45], facilitating future investigations of FDDI in different layers. A recent study described differences in monosynaptic excitatory connectivity between different types of layer 5 PCs [53], according to their long-rage projections. It would be of great importance to determine the properties of disynaptic inhibition between these populations as well. Fourteen- to 18-d-old Wistar rats (mean age 15.0 d, range 14–18 d) were quickly decapitated according to the Swiss national and institutional guidelines. The brain was carefully removed and placed in iced artificial cerebrospinal fluid (ACSF). Three hundred µm thick parasagittal slices of the primary somatosensory cortex (hindlimb area) were cut on a HR2 vibratome (Sigmann Elektronik, Heidelberg, Germany). Slices were incubated at 37°C for 30–60 min and then left at room temperature until recording. Cells were visualized by infrared differential interference contrast videomicroscopy utilizing either a C2400-03 camera (Hamamatsu, Hamamatsu City, Japan) mounted on an upright Axioscope FS microscope (Zeiss, Oberkochen, Germany) or a VX55 camera (Till Photonics, Gräfeling, Germany) mounted on an upright BX51WI microscope (Olympus, Tokyo, Japan). Thick tufted layer 5 PCs were selected according to their large soma size (15–25 µm) and their apparent large trunk of the apical dendrite. Care was taken to use only “parallel” slices, i.e. slices that had a cutting plane parallel to the course of the apical dendrites and the primary axonal trunk. This ensured sufficient preservation of both the PCs' and MCs' axonal and dendritic arborizations. Some experiments included recording of MCs. They were targeted by their soma, which is oval and bitufted, and often oriented sideways. Slices were continuously superfused with ACSF containing (in mM) 125 NaCl, 25 NaHCO3, 2.5 KCl, 1.25 NaH2PO4, 2 CaCl2, 1 MgCl2, and 25 D-glucose, bubbled with 95% O2–5% CO2. The intracellular pipette solution (ICS) contained (in mM) 110 K-gluconate, 10 KCl, 4 ATP-Mg, 10 phosphocreatine, 0.3 GTP, 10 N-2-hydroxyethylpiperazine-N9-2-ethanesulfonic acid (HEPES), and 13 biocytin, adjusted to a pH 7.3–7.4 with 5 M KOH. Osmolarity was adjusted to 290–300 mosm with D-mannitol (25–35 mM). The membrane potential values given were not corrected for the liquid junction potential, which was approximately −14 mV. 4-(N-ethyl-N-phenylamino)-1,2-dimethyl-6-(methylamino) pyridinium chloride (zd7288) was bought from Biotrend (Zurich, Switzerland), and all other drugs and chemicals were from Sigma-Aldrich (St. Louis, MO) or Merck (Darmstadt, Germany). Multiple somatic whole cell recordings (2–12 cells simultaneously) were performed with Axopatch 200B or Multiclamp 700B amplifiers (Molecular Devices, Union City, CA) in the current clamp mode. We selected PCs that were located close to each other, preferentially in clusters of near to adjacent cells. When 12 cells were recorded at the same time, the pairwise intersomatic distance increased due to limited accessibility with multiple patch electrodes in the tissue but did not exceed 150 µm. In some experiments, MCs were first recorded in voltage clamp in the cell-attached configuration, leaving the intracellular medium unperturbed, and then in whole-cell mode, thus perfused with the ICS and Biocytin contained in the pipette, allowing a subsequent staining and cell type identification. In experiments including dendritic recordings, dendrites were patched before the somata. Alexafluor594 (Invitrogen, Eugene, OR) was sometimes included in the dendritic patch electrode, revealing the corresponding soma unambiguously. The temperature was 34°C±1°C during recording. Data acquisition was performed via an ITC-18 or ITC-1600 board (Instrutech Co, Port Washington, NY), connected to a PC or Macintosh running a custom written routine under IgorPro (Wavemetrics, Portland, OR). Sampling rates were 5–10 kHz, and the voltage signal was filtered with a 2 kHz Bessel filter. Patch pipettes were pulled with a Flamming/Brown micropipette puller P-97 (Sutter Instruments Co., Novato, CA) and had an initial resistance of 3–8 MΩ (10–15 MΩ for dendritic patches). 3D morphological reconstruction of biocytin-labeled cells was done under an Olympus BX 51 W microscope fitted with a water-immersion 60× (numerical aperture (NA) 0.9) or an oil-immersion 100× (NA 1.35) objective using Neurolucida software (MicroBrightField, Magdeburg, Germany). Monosynaptic, direct connections were usually identified by stimulation of a presynaptic cell with a 20 Hz train of eight strong and brief current pulses (1–3.5 nA, 2–4 ms), followed by a so-called recovery test response (RTR) 0.5 s after the end of the train, all precisely and reliably eliciting APs. Disynaptic connections were characterized by the same protocol but at a higher frequency (usually 70 Hz) and with longer trains (usually 15 APs). Postsynaptic PCs were slightly depolarized from a potential of ∼−62 mV to −57 to −60 mV to increase the driving force for inhibitory connections. This was usually not necessary to detect FDDI but gave larger amplitudes occasionally. Due to the dendritic location and the resulting space clamp effect in layer 5 PCs, especially MC-PC synapses have a very hyperpolarized apparent somatic reversal potential that deviates strongly from the calculated one [14]. We did not find any depolarizing FDDI responses, possibly because we used rats older than 13 d [54]. Connectivity ratios were calculated as the ratio between observed versus tested connections between a pair of cells. A pair of cells could therefore maximally have two connections (both directions), a triplet could have six connections, and a cluster of n neurons could potentially have n * (n−1) connections. “Autaptic” connections—that is, FDDI elicited and received by the same PC—were not taken into consideration. The balance between de- and hyperpolarization due to FDDI and direct EPSPs as a function of stimulation frequency (Figure 1E) was calculated as the net polarization deviating from baseline in the time window starting from stimulation onset and ending just before the RTR, i.e. 0.5 s after the stimulation train ended. Bath application of zd7288 resulted in a strong hyperpolarization of PCs (∼10–12 mV; [14]), which was counteracted by a positive holding current to reestablish resting membrane potential of around −60 mV. The waiting time between stimulations was 10–20 s. Especially for FDDI summation experiments (Figures 6–8, S1) long waiting times were crucial as FDDI amplitudes would decrease otherwise (much more dramatic than, e.g., EPSP amplitudes). For these figures, we only included “pure” FDDI responses (without monosynaptic EPSP contamination) in the analysis. Stimulations were given in an alternating manner (ABAB… instead of AABB…). For summation experiments as shown in Figure 7 and S1, linearity of amplitude (and likewise integral) was calculated as a normalized difference according to L = (Ainput(1,2,…,n)−(Ainput1+Ainput2+…+Ainputn))/Ainput(1,2,…,n), where Ainput(1,2,…,n) is the amplitude of the simultaneous stimulation and Ainput1+Ainput2+…+Ainputn is the offline calculated sum of the separately stimulated presynaptic cells. For Figures 7C,D and S1, data were included if the FDDI evoked by synchronous stimulation exceeded 0.5 mV, as the signal-to-noise ratio was too high for the difference measures otherwise. All statistical analysis (paired and unpaired student's t test, ANOVA) was done with MATLAB (The Mathworks, Natick, MA, USA). The DI was defined aswhere xi and yi are single repetitions of baseline-subtracted (mean of the first 100 ms before stimulation was taken as a reference) traces of different or identical cells and different or identical repetitions, and x′ and y′ are the baseline-subtracted mean responses. DI is the point-wise squared difference between mean- and baseline-subtracted traces, calculated for every possible pair of traces, i.e. “across cells, same repetition,” “same cell, across repetitions,” and “across cells, across repetitions.” It quantifies the deviation from the average response and shows whether noise coming along the FDDI signal co-varies between two cells or not. Given one stimulated presynaptic PC, two postsynaptic PCs receiving FDDI, and n repetitions of stimulation, one obtains n “across cells, same repetition” conditions, n * (n+1)/2 “same cell, across repetitions” conditions, and n * (n−1) “across cells, across repetitions” conditions. For the latter two conditions the DI measure was nearly identical, therefore the “across cells, across repetitions” condition is not displayed in Figure 4E. DI was taken for the interval from 0 to 0.5 s after stimulation onset. Note that DI is intended to compare traces that have been stimulated in the same way. It is therefore not meaningful to compare DI values of FDDI (stimulated with 15 APs at 70 Hz, disynaptic) with EPSPs (8 APs at 20 Hz, monosynaptic). Cross-correlation and Pearson's correlation coefficient of two mean FDDI responses (Figure 4A and 4B) were calculated with Igor Pro. The effect of FDDI on spiking postsynaptic PCs (Figure 5) was quantified by counting spikes at specific 100 ms time windows of the peristimulus time histogram, namely during the second half of (first window) and immediately after (second window) presynaptic train stimulation. Spiking responses of postsynaptic PCs without coincident FDDI input served as control condition. Peristimulus time histograms contained spike counts of around 22 repetitions. Correlation-based spike timing precision was calculated according to [23] and on 0.5 s long time windows.

10.1371/journal.pgen.1005249

Minor Type IV Collagen α5 Chain Promotes Cancer Progression through Discoidin Domain Receptor-1

Type IV collagens (Col IV), components of basement membrane, are essential in the maintenance of tissue integrity and proper function. Alteration of Col IV is related to developmental defects and diseases, including cancer. Col IV α chains form α1α1α2, α3α4α5 and α5α5α6 protomers that further form collagen networks. Despite knowledge on the functions of major Col IV (α1α1α2), little is known whether minor Col IV (α3α4α5 and α5α5α6) plays a role in cancer. It also remains to be elucidated whether major and minor Col IV are functionally redundant. We show that minor Col IV α5 chain is indispensable in cancer development by using α5(IV)-deficient mouse model. Ablation of α5(IV) significantly impeded the development of KrasG12D-driven lung cancer without affecting major Col IV expression. Epithelial α5(IV) supports cancer cell proliferation, while endothelial α5(IV) is essential for efficient tumor angiogenesis. α5(IV), but not α1(IV), ablation impaired expression of non-integrin collagen receptor discoidin domain receptor-1 (DDR1) and downstream ERK activation in lung cancer cells and endothelial cells. Knockdown of DDR1 in lung cancer cells and endothelial cells phenocopied the cells deficient of α5(IV). Constitutively active DDR1 or MEK1 rescued the defects of α5(IV)-ablated cells. Thus, minor Col IV α5(IV) chain supports lung cancer progression via DDR1-mediated cancer cell autonomous and non-autonomous mechanisms. Minor Col IV can not be functionally compensated by abundant major Col IV.

Collagens, the major extracellular matrix components in most vertebrate tissues, provide cells with structural and functional support. Collagens are trimers of collagen α chains. Multiple trimers are formed by highly homologous α chains for certain types of collagens (e.g. α1α1α2, α3α4α5 and α5α5α6 heterotrimers for type IV collagen). Type IV collagens are named as major type (α1α1α2) or minor type (α3α4α5 and α5α5α6), mainly reflecting the abundance and tissue distribution, but not the importance of their biological functions. High similarity in sequence and domain structure of the α chains does not necessarily imply that major and minor type IV collagens share the same cell surface receptors and intracellular signaling pathways. In this study, we generated an α5(IV) chain deficient mouse model lacking minor type IV collagens. We found that the mutant mice have delayed development of KrasG12D-driven lung cancer without affecting major type IV collagen expression. α5(IV), but not α1(IV), ablation impaired non-integrin collagen receptor discoidin domain receptor-1 (DDR1)-ERK signaling, suggesting that major and minor type IV collagens are functionally distinct from each other.

Basement membranes (BMs), specialized extracellular matrices separating epithelial and endothelial cells from underlying mesenchyme, provide cells with structural support, as well as morphogenic and functional cues [1–3]. Type IV collagens (Col IV) are major components of BMs [1,3]. Three triple helical protomers, α1α1α2, α3α4α5 and α5α5α6, are formed by the Col IV α chains that further form collagen networks [4,5]. α1α1α2, the major Col IV, is widely expressed as a component of all BMs. α3α4α5 and α5α5α6, known as minor Col IV, have much restricted tissue distribution [4,5]. Col IV-initiated signals are essential survival and growth cues for liver metastasis in diverse tumor types [6]. BM proteins produced by mouse Engelbrecht Holm-Swarm sarcoma, known as Matrigel, enhanced the tumorigenicity of human cancer cells [7]. BM proteins, including α1(IV), protect small cell lung cancer cells from chemotherapy-induced apoptosis [8]. Angiogenesis, required by tumors to supply nutrients and oxygen, and to evacuate metabolic wastes, is dependent on correct interaction between endothelial cells and the vascular BMs [1,9,10]. Col IV plays crucial roles in supporting endothelial cell proliferation and migration. Blood vessel formation and survival are connected with proper collagen synthesis and deposition in BMs. Col IV, by binding to cell surface receptors, activates intracellular signaling events to promote cell survival, proliferation and tumorigenesis [5]. Loss of integrin α1β1 ameliorates KrasG12D-induced lung cancer [11,12]. β1 integrin and its downstream effecter focal adhesion kinase (FAK) are critical in mediating resistance to anoikis, chemotherapy-induced cell death and metastasis [6,8,11]. Despite Col IV is extensively studied, majority of the works focused on the functions of major Col IV, or unfortunately did not distinguish the roles of major and minor Col IV. It is largely unknown whether minor Col IV plays a role in cancer development. It also remains to be elucidated whether major and minor Col IV signal through the same cell surface receptors and intracellular signaling pathways and whether they can functionally compensate for each other. In the present study, we demonstrate that minor Col IV α5 chain is indispensable in lung cancer development by using α5(IV)-deficient mouse model. α5(IV) supports lung cancer progression via cancer cell autonomous and non-autonomous mechanisms. α5(IV), but not α1(IV), promotes lung cancer cell proliferation and tumor angiogenesis through non-integrin collagen receptor DDR1-mediated ERK activation. The functions of minor Col IV can not be compensated by abundant major Col IV. A LacZ gene trap cassette including En2 splice acceptor/ECMV IRES/LacZ/SV40 polyadenylation site was inserted into intron 35 of mouse Col4a5 gene on chromosome X to generate Col4a5 knockout mice (S1A and S1B Fig) [13]. RT-PCR analyses demonstrated the absence of Col4a5 mRNA in the KO tissues (S1C and S1D Fig). The LacZ reporter reflects endogenous Col4a5 expression. Strong LacZ staining was observed in lung bronchia (S1E Fig). Immunofluorescent staining demonstrated that α5(IV) chain is expressed in lung bronchia at high levels, and in lung alveolar epithelial cells at lower levels in Col4a5+/Y (hereafter refereed as wild-type, WT) mice (S1F Fig). The α5(IV) signal is absent in Col4a5LacZ/Y (hereafter refereed as knockout, KO) lungs (S1F Fig), further demonstrating that the mutant Col4a5 allele is indeed null. Oncogenic KrasG12D drives lung cancer onset and progression. In contrast to large, multifocal tumors formed in KrasG12D; Col4a5+/Y (Kras/α5 WT) mice, significantly less tumors developed in KrasG12D; Col4a5LacZ/Y (Kras/α5 KO) mice (Fig 1A and 1B). Tumors in Kras/α5 KO mice were significantly smaller than those in Kras/α5 WT mice (Fig 1A and 1C). α5(IV) ablation dramatically reduced the number of large tumors (>0.5 mm2), but had no profound effect on the number of small tumors (<0.1 mm2) (Fig 1D), indicating that α5(IV) is mainly involved in regulating tumor progression, but not tumor onset. BM proteins promote cancer cell proliferation and protect cancer cells from apoptosis. Tumors in Kras/α5 KO mice had significantly reduced tumor cell proliferation (Fig 1E and 1F), compared with those in Kras/α5 WT mice. Few apoptotic signal was evident in both groups (Fig 1E). Hemorrhage was evident in α5 KO lungs, but not in WT lungs (Fig 1A). Hemorrhage lesions indicate improper organization of capillaries and blood vessels in α5 KO lungs. As tumor angiogenesis provides tumor cells nutrients and oxygen necessary for sustained tumor growth, this promoted us to examine whether neo-angiogenesis was compromised in Kras/α5 KO tumors. Indeed, tumors in Kras/α5 KO mice were significantly less vascularized (Fig 1E and 1G). Thus, reduction in tumor cell proliferation and tumor angiogenesis account for delayed tumor progression in Kras/α5 KO mice. α5(IV) is expressed in lung bronchia and alveolar epithelial cells (S1 Fig). To study the functions of epithelial α5(IV) in lung cancer development, endogenous α5(IV) was knocked down in A549 lung adenocarcinoma cells (Fig 2A). α5(IV) knockdown significantly reduced A549 cell proliferation, migration and anchorage-independent cell growth (Fig 2B–2D), compared to cells expressing scramble control shRNA. This is not due to the off-target effect of α5(IV) shRNAs, as expression of mouse α5(IV) could rescue the phenotypes of α5(IV)-knockdown A549 cells (S2 Fig). α5(IV) knockdown in CRL-5810 lung cancer cells similarly resulted in impaired cell proliferation and anchorage-independent cell growth (S3 Fig). Therefore, the endogenous α5(IV)-constituted BMs are essential in supporting lung cancer cell proliferation. To determine whether in vitro phenotypes were reflected in vivo, tumorigenic ability of A549 cells was tested by injecting control or α5(IV)-knockdown cells subcutaneously into nude mice. α5(IV) knockdown resulted in slower growing A549 xenograft tumors (Fig 2E). Less proliferating cells were detected in α5(IV)-knockdown xenograft tumors (Fig 2F and 2G). Kras/α5 KO tumors were significantly less vascularized (Fig 1). However, knockdown of α5(IV) in A549 cells only mildly affected neo-angiogenesis in the xenograft tumors, which was not statistically significant (Fig 2F and 2H). This suggests that less angiogenesis observed in Kras/α5 KO tumors may be mainly due to ablation of stromal α5(IV). To examine the roles of stromal α5(IV) in tumor progression, murine Lewis lung cancer (LLC) cells were implanted in Col4a5 WT or KO mice. Tumors grew significantly slower in KO than in WT mice (Fig 3A). Less proliferating cells were detected in the tumors from KO mice, than in that from WT mice (Fig 3B and 3D). Unlike the Kras-driven lung tumors, which were slowly growing and rare apoptosis was evident (Fig 1E), the LLC transplant tumors grew much faster. Apoptosis was evident in the LLC transplant tumors, due to rapid tumor growth (Fig 3C). More apoptotic cells were detected in the tumors from KO mice, than in that from WT mice (Fig 3C and 3D). These data collectively suggest stromal α5(IV) provides necessary survival and proliferation cues to support rapid LLC tumor growth. Tumors trigger profound angiogenesis to support vast nutrient and oxygen demand during rapid LLC transplant tumor growth in WT mice (Fig 3B). Fewer blood vessels formed in the LLC transplant tumors in the KO mice, compared to that in the WT mice (Fig 3B). The impaired tumor angiogenesis in the KO mice was not only reflected by decreased number of CD31-positive endothelial cells (Fig 3E), but also by dramatically decreased number of sinusoid microvessels (Fig 3F) and average vessel diameter (Fig 3G). To further test if stromal α5(IV) plays a role in regulating angiogenesis, VEGF containing Matrigel plugs were implanted subcutaneously in Col4a5 WT or KO mice. Abundant blood vessels, visualized by FITC-dextran, formed in the Matrigel plugs implanted in the WT mice, but not in the KO mice (Fig 3H). CD31 staining on Matrigel plug sections further revealed ~12-fold reduction of capillary numbers in the plugs in KO mice (Fig 3H and 3I). α5(IV) partially colocalized with endothelial cell marker CD31 in the lung (Fig 4A). Knockdown of α5(IV) in human microvascular endothelial cell-1 (HMEC-1) cells (Fig 4B) significantly reduced endothelial cell proliferation (Fig 4C) and migration (Fig 4D). Knockdown of α5(IV) in HMEC-1 cells also significantly impaired the tubule formation capability of endothelial cells (Fig 4E). Thus, endothelial α5(IV) may be responsible for efficient tumor angiogenesis. Major Col IV is known to provide survival and growth cues to cancer cells. α5(IV) may regulate tumor progression through modulating major Col IV expression and basement membrane assembly. Electron microscopy on the lungs from 6-month old KO mice did not reveal overt defect in the basement membranes underneath lung alveolar epithelial cells (S4A Fig). Relatively more abundant α1(IV) expression was detected in KO lungs, compared to WT tissues (S4B Fig). Ablation of α5(IV) had no significant effect on α1(IV) expression in Kras-driven lung tumors (S4C Fig). Knockdown of α5(IV) in A549 (Fig 5A) and HMEC-1 (S7A Fig) cells did not significantly affect major Col IV α1(IV) or α2(IV) chain expression. Despite knockdown of α1(IV) impaired cellular functions of A549 (S5 Fig) and HMEC-1 (S6 Fig) cells, expression of α5(IV) was not affected (Fig 5A and S7A Fig). All these data collectively suggest that altered behavior of α5(IV)-deficient cells and impaired tumor progression in α5(IV)-deficient mice are not the results of concomitant loss of major Col IV expression or disruption of basement membrane structure. The presence of abundant α1(IV) also suggests that major Col IV can not functionally compensate for the loss of α5(IV) in supporting tumor growth. FAK is one of the major effecters transducing signals from Col IV [14]. FAK further phosphorylates and activates downstream signaling molecules, including Src [14]. Knockdown of α5(IV), however, did not affect phosphorylation levels of FAK and Src in A549 and CRL-5810 lung cancer cells (Fig 5A and S3A Fig). Instead, significantly lower phosphorylation levels of ERK and Akt, kinases essential in supporting cell survival, proliferation and transformation [15,16], were detected in α5(IV)-knockdown A549 and CRL-5810 cells (Fig 5A and S3A Fig). Ectopic expression of mouse α5(IV) in α5(IV)-knockdown A549 cells restored phosphorylation of ERK and Akt (S2E Fig). Interestingly, knockdown of α1(IV) resulted in impaired phosphorylation of Akt and Src, but not ERK or FAK in A549 cells (Fig 5A), reinforcing the notion that major and minor Col IV may regulate cancer cell behavior through overlapping, but not identical intracellular signaling pathways. Similar to that in lung cancer cells, knockdown of α5(IV), but not α1(IV), significantly decreased ERK phosphorylation in HMEC-1 cells (S7A Fig). To study if impaired ERK activation is responsible for the defects in cell proliferation and migration resulted from α5(IV) deficiency, constitutively active MEK1 was expressed in α5(IV)-knockdown A549 and HMEC-1 cells. Expression of constitutively active MEK1 successfully restored ERK phosphorylation in A549 and HMEC-1 cells (Fig 5B and S7B Fig). Expression of constitutively active MEK1 in A549 cells rescued the defects of cell proliferation (Fig 5C), migration (Fig 5D) and anchorage-independent cell growth (Fig 5E). Constitutively active MEK1 also restored the capability of cell proliferation (S7C Fig), migration (S7D Fig) and tubule formation (S7E Fig) of α5(IV)-knockdown HMEC-1 cells. Col IV transduces signals through cell surface integrin and non-integrin receptors [5]. Knockdown of α5(IV) had no effect on cell surface integrin expression (S8 Fig). Knockdown of α5(IV) significantly decreased the expression of non-integrin collagen receptor DDR1 in lung cancer cells (Fig 6A and S3A Fig), which can be restored by ectopic mouse α5(IV) expression (S2E Fig). However, DDR1 expression was not altered in α1(IV)-knockdown lung cancer cells (Fig 6A). Similar to that in lung cancer cells, DDR1 expression was decreased in α5(IV)-, but not α1(IV)-, knockdown HMEC-1 cells (S9A Fig). In addition, significantly less DDR1 expression was detected in KO lungs, compared to WT tissues (Fig 6B). Ablation of α5(IV) also significantly decreased DDR1 expression in Kras-driven lung tumors (Fig 6C). Interestingly, α5(IV) knockdown in A549 cells did not affect DDR1 mRNA levels (Fig 6D), suggesting α5(IV) ablation may regulate DDR1 expression via mechanisms other than transcriptional regulation. A much faster decline of DDR1 protein was observed in α5(IV)-knockdown A549 cells subjected to cycloheximide treatment (Fig 6E and 6F), suggesting that α5(IV) regulates DDR1 expression at least partially by stabilizing DDR1 proteins. Lysosome inhibitor NH4Cl had minimal effect on DDR1 protein levels (Fig 6G). Proteasome inhibitor MG132 treatment restored DDR1 protein levels in α5(IV)-knockdown A549 cells (Fig 6G). α5(IV) knockdown significantly increased DDR1 ubiquitination in A549 cells (Fig 6H). Therefore, α5(IV) ablation downregulates DDR1 expression by accelerating DDR1 ubiquitination and proteasome-dependent degradation. Knockdown of DDR1 in A549 cells resulted in decreased phosphorylation of ERK and Akt (Fig 7A), unaffected phosphorylation of FAK and Src (Fig 7A), as well as impaired cell proliferation (Fig 7B), migration (Fig 7C) and anchorage-independent cell growth (Fig 7D), resembling the phenotypes observed in α5(IV)-knockdown A549 cells. Knockdown of DDR1 in HMEC-1 cells similarly resulted in decreased phosphorylation of ERK and Akt (S9B Fig), impaired endothelial cell proliferation (S9C Fig), migration (S9D Fig) and tubule formation (S9E Fig). The similar phenotypes observed in the α5(IV)- and DDR1-knockdown cells indicate that DDR1 may be the receptor transducing signals from α5(IV). DDR1 is a receptor tyrosine kinase that its phosphorylation is indicative of receptor activation and important in transducing downstream signals. Significantly less phosphorylation of DDR1 was detected in α5(IV)-knockdown A549 cells, compared to that in cells expressing scramble shRNA (Fig 8A). DDR1 expression was reduced in α5(IV)-knockdown cells and less amount of DDR1 was immunoprecipitated (Fig 8A). To more accurately examine DDR1 phosphorylation levels in α5(IV)-knockdown cells, DDR1 was expressed back to endogenous levels. Less DDR1 phosphorylation was detected in α5(IV)-knockdown A549 cells expressing exogenous DDR1 than the control cells, despite similar amount of DDR1 was immunoprecipitated (Fig 8A). Overexpression of DDR1 was not able to restore phosphorylation levels of ERK and Akt in α5(IV)-knockdown A549 cells (S10 Fig). These data collectively suggest that α5(IV) not only affects DDR1 stability and expression, but also is required for DDR1 activation. To further study if DDR1 is functionally downstream of α5(IV), a chimeric Div-DDR1 is expressed in α5(IV)-knockdown A549 cells. The chimeric Div-DDR1 is constructed by replacing the extracellular ligand binding discoidin domain of DDR1 with Div, a coil-coiled domain from Bacillus subtilis DivIVA that forms constitutive dimer/oligomer [17,18]. Replacement of DDR1 ligand binding domain with Div provokes spontaneous DDR1 autophosphorylation and activation (Fig 8B) [18,19]. Expression of such a constitutively active Div-DDR1 in α5(IV)-knockdown A549 cells restored ERK and Akt phosphorylation (Fig 8C), cell proliferation (Fig 8D), migration (Fig 8E) and anchorage-independent cell growth (Fig 8F). Oligomerization capability and kinase activity of DDR1 are necessary for DDR1 function. Div-DDR1 with mutations in the Div coil-coiled domain (mDiv-DDR1) that disrupts Div self-assembly ability [17] or in the DDR1 kinase domain (Div-DDR1 K655A) that impairs DDR1 tyrosine kinase activity [20] failed to activate DDR1 (Fig 8B). Expression of such DDR1 mutants also failed to restore ERK and Akt phosphorylation (Fig 8C), cell proliferation (Fig 8D), migration (Fig 8E) and anchorage-independent cell growth (Fig 8F) in α5(IV)-knockdown A549 cells. To study if the DDR1 signaling pathway is involved in transducing signal from α5(IV) in endothelial cells, constitutively active DDR1 was expressed in α5(IV)-ablated HMEC-1 cells. Expression of constitutively active Div-DDR1, but not mDiv-DDR1 or Div-DDR1 K655A, in α5(IV)-knockdown HMEC-1 cells restored ERK and Akt phosphorylation (S11A Fig), and rescued the defects of cell proliferation (S11B Fig), migration (S11C Fig) and tubule formation (S11D Fig). Col IV, the major BM component, is essential in maintenance of tissue integrity and proper function. In addition to broadly expressed and extensively studied major Col IV α1α1α2, minor Col IV α3α4α5 and α5α5α6 are less abundantly expressed with restricted tissue distribution [4]. Physiological and pathological functions of minor Col IV, however, are less well understood. In this report, we present evidences that minor Col IV α5(IV) is essential in supporting lung cancer development via cancer cell autonomous and non-autonomous mechanisms. Minor but not major Col IV signals through non-integrin receptor DDR1. Delayed tumor progression in α5(IV)-deficient mice suggests proper signal from α5(IV) is important in supporting cancer cell survival and proliferation. Col IV transduces signals through cell surface receptors. Cell surface integrin expression is unaffected in α5(IV)-knockdown cells. However, expression of DDR1, the non-integrin collagen receptor functioning independent of integrins [20–22], is decreased in α5(IV)-knockdown cells. DDR1 is highly phosphorylated in non-small cell lung cancer (NSCLC) [23], and DDR1 overexpression is associated with poor prognosis in NSCLC [24]. Inhibition of DDR1 reduces cell survival, homing and colonization in lung cancer metastasis [25]. Consistently, DDR1 expression is elevated in lung tumors with Kras activation, compared to normal lung tissues (compare Fig 6B and 6C). Ablation of α5(IV) results in decreased DDR1 expression in both normal lung tissues and Kras lung tumors. DDR1-knockdown cells phenocopied α5(IV)-knockdown cells. More importantly, expression of constitutively active DDR1 in α5(IV)-knockdown cells can rescue the proliferation and migration defects, suggesting DDR1 is functionally downstream of α5(IV). α5(IV) knockdown impaired DDR1 phosphorylation. Overexpression of exogenous wild-type DDR1 can not restore ERK phosphorylation in α5(IV)-knockdown cells. These data indicate that the function of DDR1 requires the presence of α5(IV) and DDR1 may directly mediate the functions of α5(IV). Despite α5(IV) knockdown does not affect integrin cell surface expression, the possibility exists that integrins are functional receptors for α5(IV). Col IV was reported to bind integrins through sites within the triple-helical cyanogen bromide-derived fragments and noncollagenous domains [5]. Such studies were largely based on purified Col IV or Col IV fragments. It should be noted that proper collagen network assembly and geometry are critical in the biological functions of Col IV. Ablation of endogenous Col IV using gene knockout or silencing will provide more physiologically relevant insights into receptor binding, signaling and biological functions of Col IV. It remains to be elucidated whether integrins have selectivity and specificity towards major and minor Col IV under different physiological and pathological circumstances. DDR1 and integrins may have cooperative or opposing functions in response to collagens [26,27]. The crosstalk between DDR1 and integrins upon α5(IV) binding may provide the cells more robustness. Ablation of α5(IV) does not affect major Col IV expression, or disrupt basement membrane assembly. The inability of abundant major α1α1α2(IV) to support efficient tumor growth and progression in α5(IV)-deficient mice indicates that major Col IV can not functionally compensate for the deficiency of minor Col IV. This is supported by the fact that mutations of Col IV α chains cause distinct heritable diseases. Mutations in COL4A1 cause encephaloclastic porencephaly, characterized by degenerative cavities and cerebral lesions in the brain [28]. Deletion of Col4a1/Col4a2 locus in mice results in growth retardation and embryonic lethality [29]. However, mutations in COL4A5 (Alport’s syndrome) or auto-antibody recognizing α3(IV) (Goodpasture’s syndrome) result in progressive renal failure [4,5]. Mice deficient of α3(IV) [30,31] or α5(IV) [32] are viable, but develop renal phenotypes reminiscent of that in Alport’s syndrome. Knockdown of major Col IV α1(IV) does not affect DDR1 expression. The overlapping, but not identical spectrum of altered signaling events in α5(IV)- and α1(IV)-knockdown cells suggests that major and minor Col IV may exert their biological functions via different cell surface receptors and intracellular signaling pathways. Major and minor Col IV share same domain structure and high sequence similarity. It is yet unclear how highly similar major and minor Col IV recognize different cell surface receptor and activate different intracellular signaling pathways. α3α4α5(IV) is highly cross-linked due to its larger degree intra- and inter-chain disulfide bonds, relative to α1α1α2(IV) [33]. As a result, α3α4α5(IV) has different biochemical properties from α1α1α2(IV) that α3α4α5(IV) is more resistant to proteolytic degradation [33]. Different biomechanical force from major and minor Col IV may be responsible for the receptor specificity. It should be noted that Col IV protomers further form α1α1α2(IV)-α1α1α2(IV), α3α4α5(IV)-α3α4α5(IV) and α1α1α2(IV)-α5α5α6(IV) networks [4,5]. These networks may differentially recognize cell surface receptors and activate intracellular signaling pathways, thus provide signal specificity and redundancy. α5(IV) regulates cancer progression via cancer cell autonomous and non-autonomous mechanisms. The DDR1-ERK signaling cascade is required for the functions of both cancer cells and endothelial cells. Stromal components, including blood vessels, constitute proper microenvironment to support tumor progression. It is reported that stable microvasculature sustains cancer cells at dormancy, whereas sprouting neovasculature rescues cancer cells from cell cycle arrest and promotes cancer cell proliferation [34]. Col IV assembly is critical for vascular BM integrity and structural organization. Small-molecule inhibitors that interfere Col IV biosynthesis were shown to prevent angiogenesis and tumor growth [35]. α5(IV) is expressed in the endothelium. Deficiency of α5(IV) delayed in vitro and in vivo angiogenesis. It warrants further study if the cancer cells in α5(IV) KO mice remain dormant due to impaired neo-angiogenesis. In summary, we provides evidences in this study that α5(IV) deficiency significantly delays tumor progression. α5(IV) signals through non-integrin collagen receptor DDR1 in lung cancer cells and endothelial cells. α5(IV) promotes tumor growth via both cancer cell autonomous and non-autonomous mechanisms. Abundant major Col IV is not able to compensate for α5(IV) deficiency. All mice were housed in specific pathogen-free environment at the Shanghai Institute of Biochemistry and Cell Biology and treated in strict accordance with protocols approved by the Institutional Animal Care and Use Committee of Shanghai Institute of Biochemistry and Cell Biology (Approval number: SIBCB-NAF-15-003-S325-006). The antibodies used are ERK, ERK pT202/pY204, Akt, Akt pS473, Src, Src pY416, cleaved caspase-3 (Cell Signaling), FAK (BD Transduction Laboratories), FAK pY397 (Millipore), Ki-67 (Novocastra Laboratories), α1(IV) (Abgent), α2(IV) and CD31 (Abcam), α5(IV) (rabbit ployclonal antibody from Proteintech (western blot) and rat monoclonal antibody clone b14 (immunostaining) provided by Dr. Yoshikazu Sado, Shigei Medical Research Institute [36]), DDR1, phosho-tyrosine (pY99), ubiquitin (Santa Cruz), MEK1 (Abmart), Actin (Sigma-Aldrich), biotinylated goat anti-rabbit secondary antibody (Zymed), Alexa Fluor 555/488 conjugated anti-mouse, rat or rabbit IgG secondary antibodies (Invitrogen). Expression level of integrins on A549 cell surface was determined by immunofluorescence flow cytometry with anti-β1 (Thermo Scientific Pierce), α1, α2 and α11 (Santa Cruz) integrin antibodies as described [37]. Cycloheximide, MG132 and NH4Cl were purchased from Sigma-Aldrich. The shRNAs were cloned into pLKO.1-puro lenti-viral vector (Addgene). Viral packaging and infection of cells was performed as previously described [38]. After viral infection, cells were selected with puromycin to generate stable cell lines. At least two batches of stable cell lines were generated for each experiment. Experiments were performed in triplicates and repeated at least twice using each batch of cells. The target sequences are: 5’-CAACAAGATGAAGAGCACCAAC-3’ (shScram), 5’-GGGTGATGATGGAATTCCA-3’ (shCOL4A5-1), 5’-GCAGATCAGTGAACAGAAAAG-3’ (shCOL4A5-2), 5’-TCCAGGATGCAATGGCACAAA -3’ (shCOL4A1-1), 5’-TCCAGGTTCCAAGGGAGAAAT -3’ (shCOL4A1-2), 5’-GGTTACTCTTCAGCGAAAT -3’ (shDDR1-1), and 5’-AGATGGAGTTTGAGTTTGACC -3’ (shDDR1-2). To generate cell lines expressing mouse α5(IV), DDR1 or Div-DDR1, coding sequences were cloned into pCDH-Neo lenti-viral vector (Addgene). α5(IV)-knockdown cells were infected with lenti-virus harboring mouse α5(IV), DDR1 or Div-DDR1 sequences and selected with G418. Mouse Col4a5 sequences were amplified from B16-F10 cDNA using primers 5’-gatcTCTAGAatgcaagtgcgtggagtgtgcc-3’ (forward) and 5’-gatcGCGGCCGCttatgtcctcttcatgcatact-3’ (reverse). Amplicon was inserted into pCDH-Neo vector. HA-tagged human DDR1 was cloned from MCF-7 cDNA using primers 5’-GATCGAATTCATGGGACCAGAGGCCCTGT-3’ (forward) and 5’-GATCGCGGCCGCTCAAGCGTAATCTGGAACATCGTATGGGTACACCGTGTTGAGTGCATCCT-3’ (reverse). Amplicon was inserted into pCDH-Neo. K655A substitution was introduced by two step PCR amplification that was restricted with XhoI and NotI and exchanged for the corresponding wild-type fragment in the DDR1 expression construct. The primers used were 5’-CCCGTCCCCCTCGAGGCCC-3’ (fragment 1, forward), 5’-CCGTAAGATCGCGACAGCTACCAGCAAAGG-3’ (fragment 1, reverse), 5’-GTAGCTGTCGCGATCTTACGGCCAGATGCC-3’ (fragment 2, forward), and 5’-GATCGCGGCCGCTCAAGCGTAATCTGGAACATCGTATGGGTACACCGTGTTGAGTGCATCCT-3’ (fragment 2, reverse). To generate the Div-DDR1 chimeric proteins, the coding sequences of DDR1 discoidin domain (aa 29–367) was replaced by a 51bp oligonecleotide sequences compromising BstBI and BamHI restriction sites by two step PCR. The primers used were 5’-GATCgaattcATGGGACCAGAGGCCCTGT-3’ (fragment 1, forward), 5’-GGATCCGTGATAGTTTTTGCTAAGCAACTCTTCAACTTTATCTTCCAACTGTTTCATTTCGAACTTGGCAGGATCAAAATGTC-3’ (fragment 1, reverse), 5’- TTCGAAATGAAACAGTTGGAAGATAAAGTTGAAGAGTTGCTTAGCAAAAACTATCACGGATCCGTGGTGAACAATTCCTCTCCG-3’ (fragment 2, forward), and 5’-GATCgcggccgcTCAAGCGTAATCTGGAACATCGTATGGGTACACCGTGTTGAGTGCATCCT-3’ (fragment 2, reverse). The Div coil-coiled domain [18] (wild-type: MKQLEDKVEELLSKNYHLENEVARLKKLVGERGSSGSGR; mutant: MKQLEDKVEELLSKNYHVENEVARVKKLVGERGSSGSGR), amplified using primers 5’-GATCTTCGAAATGAAACAGTTGGAAGATAAAG-3’ (forward) and 5’-GATCGGATCCGCGGCCGCTTCCAGAGCTTCC-3’ (reverse), was placed between BstBI and BamHI sites. Human CA-MEK1 was prepared by substituting Ser 218 and Ser 222 in MEK1 with glutamic acids and removing residues 31 to 52 as described [39]. Total RNA was prepared and retrotranscribed as described [40]. The RT-PCR primers used are: human/mouse COL4A5: 5’-TGCCTTCGTCGCTTTAGT-3’ (forward) and 5’-TTGACCTGAGCCTTCTGC-3’ (reverse); Mouse Col4a5: 5’-GGATTGGCTATTCCTTCAT-3’ (forward) and 5’-GCATACTTGACATCGGCTA-3’ (reverse); Human/mouse ACTIN: 5’-cctagaagcatttgcggtgg-3’ (forward) and 5’-gagctacgagctgcctgacg-3’ (reverse). A549 and CRL-5810 cells (ATCC) were maintained in RPMI 1640 (Hyclone) supplemented with 5% FBS (Biochrom). 293T cells and Lewis lung cancer (LLC) cells (ATCC) were cultured in DMEM (GIBCO) with 10% FBS. Human microvascular endothelial cell-1 (HMEC-1) (generously provided by Dr. Zhengjun Chen) was cultured in MCDB131 (GIBCO) with 10% FBS, 10ng/mL EGF and 1μg/mL hydrocortisone. To study the functions of endogenous Col IV, the lung cancer cells and endothelial cells were plated directly on tissue culture plates without exogenous substance coating. MTT, bromodeoxyuridine (BrdU) incorporation, migration and anchorage-independent cell growth assays were performed as described [40,41]. In vitro angiogenesis assay was performed as described [42] by seeding HMEC-1 cells in the rat tail type I collagen sandwich gel in the presence of VEGF. Cells were photographed after 24 hours. In vivo Matrigel plug assay was performed as described [43] by subcutaneously injecting growth factor reduced Matrigel containing 50 ng recombinant human vascular endothelial growth factor into 8-week-old WT or Col4a5 LacZ/Y mice in C57/Bl background. On day 14, Dextran-FITC was injected through the tail vein 30 min before the mice were sacrificed. Matrigel plugs were fixed and sectioned for CD31 staining. Histological vascular parameters, including microvascular density (MVD), sinusoid microvessel number, and vascular diameter, were measured [44]. Total cell lysates were harvested in hot SDS sample buffer. For immunoprecipitation, cells were lysed in RIPA buffer. DDR1 was immunoprecipitated with anti-DDR1 (Santa Cruz) antibody. Immunoprecipitated proteins were eluted with SDS sample buffer. Proteins were separated by SDS-PAGE. After electrophoresis, the proteins were transferred to nitrocellulose membrane. The membrane was incubated overnight at 4°C with primary antibodies, washed with TBS-T (TBS with 0.1% Tween-20), and incubated with HRP-conjugated secondary antibodies at room temperature for 1 hour. Immuno-reactive protein was detected using SuperSignal West Pico Chem KIT (Thermo Scientific, USA). Primary antibodies used were against ERK, ERK pT202/pY204, Akt, Akt pS473, Src, Src pY416 (Cell Signaling), FAK (BD Transduction Laboratories), FAK pY397 (Millipore), α1(IV) (Abgent), α2(IV) (Abcam), DDR1, phosho-tyrosine (pY99), ubiquitin (Santa Cruz), α5(IV) (Proteintech), MEK1 (Abmart) and Actin (Sigma-Aldrich). Western blots were scanned and analyzed with Image J. Immunohistochemistry on 5-μm paraffin sections using antibodies against Ki-67 (Novocastra Laboratories), cleaved caspase-3 (Cell Signaling), CD31 (Abcam), α1(IV) (Abgent) or DDR1 (Santa Cruz) was performed as described [40]. For α5(IV) immuno-staining, 8-μm frozen tissue sections were fixed in cold acetone for 10 min. Samples were incubated with α5(IV) antibody (rat monoclonal antibody clone b14) (1:50–1:100) for 16 hours at 4°C, followed by incubation with Alexa Fluor 555/488 conjugated anti-rat IgG antibody. Immunohistochemistry or immunofluorescence sections were viewed under microscope (IX71; OLYMPUS, Inc.) with a UPlan-FLN 4×objective/0.13 PhL, a UPlan-FLN 10×objective/0.30 PhL, or a LUCPlan-FLN 20×objective/0.45 PhL. Images were captured with a digital camera (IX-SPT; OLYMPUS, Inc.) and Digital Acquire software (DPController; OLYMPUS, Inc.). Perfused blood vessels in Matrigel plugs were viewed by UV-illumination under microscope (SZX16; OLYMPUS, Inc.) with a SDF-PLAPO 1×PF. Images were captured with a digital camera (U-LH100HGAPO; OLYMPUS, Inc.) and Digital Acquire software (DPController; OLYMPUS, Inc.). All mice were housed in specific pathogen-free environment at the Shanghai Institute of Biochemistry and Cell Biology and treated in strict accordance with protocols approved by the Institutional Animal Care and Use Committee. Col4a5LacZ/Y mice were generated and maintained in C57/Bl background by the European Conditional Mouse Mutagenesis Program [13]. KrasG12D mice were back crossed to C57/Bl background 3 generations before cross with Col4a5LacZ/Y mice. LLC cells were transplanted at the armpit of lower limb of 8-week old WT or Col4a5 LacZ/Y mice in C57/Bl background. To minimize the possible effects of mouse genetic background on tumor behavior, wild-type littermates were used as control for Col4a5LacZ/Y mice in all experiments. A549 cells were subcutaneously injected into Balb/c nude mice. Lung tissues isolated from 6-month old Col4a5+/Y and Col4a5LacZ/Y mice were fixed in 3% glutaraldehyde in 0.1M PBS (pH7.4) for 4 hours at room temperature and then in 1% osmium tetroxide overnight at 4°C. The fixed lung tissue were dehydrated through an alcohol series and embedded in Epon812 Resin at 60°C for 48 hours. Ultrathin sections (70 nm) were collected on copper grids. The grids were stained in 2% uranyl acetate for 40 minutes and in 0.5% lead citrate for 8 minutes orderly. The samples were examined under FEI Tecnai G2 Spirit TEM. Data were analyzed using the two-sided Student t test, and considered statistically significant when the P value was less than 0.05.

10.1371/journal.pbio.0050134

A Helical Structural Nucleus Is the Primary Elongating Unit of Insulin Amyloid Fibrils

Although amyloid fibrillation is generally believed to be a nucleation-dependent process, the nuclei are largely structurally uncharacterized. This is in part due to the inherent experimental challenge associated with structural descriptions of individual components in a dynamic multi-component equilibrium. There are indications that oligomeric aggregated precursors of fibrillation, and not mature fibrils, are the main cause of cytotoxicity in amyloid disease. This further emphasizes the importance of characterizing early fibrillation events. Here we present a kinetic x-ray solution scattering study of insulin fibrillation, revealing three major components: insulin monomers, mature fibrils, and an oligomeric species. Low-resolution three-dimensional structures are determined for the fibril repeating unit and for the oligomer, the latter being a helical unit composed of five to six insulin monomers. This helical oligomer is likely to be a structural nucleus, which accumulates above the supercritical concentration used in our experiments. The growth rate of the fibrils is proportional to the amount of the helical oligomer present in solution, suggesting that these oligomers elongate the fibrils. Hence, the structural nucleus and elongating unit in insulin amyloid fibrillation may be the same structural component above supercritical concentrations. A novel elongation pathway of insulin amyloid fibrils is proposed, based on the shape and size of the fibrillation precursor. The distinct helical oligomer described in this study defines a conceptually new basis of structure-based drug design against amyloid diseases.

Diseases associated with the presence of amyloid structures, such as Alzheimer and Parkinson disease, are characterized by the presence of protein aggregates in the form of highly ordered fibrils. This amyloid fibril formation is also commonly observed for a number of protein drugs, such as insulin. Detailed information on how and why these fibrils are formed will be very useful to design compounds and drugs that may reverse or even prevent fibril formation, but existing knowledge in this field is still limited. We have studied, in real time, the fibril formation of insulin using a technique based on scattering of x-rays (small-angle x-ray scattering [SAXS]). Using SAXS, we obtained hitherto unprecedented three-dimensional structural information on these fibrils in solution. Most importantly, we were able to describe the three-dimensional structure of a crucial intermediate, which is probably a structural starting point (nucleus) in the fibril formation process. These results suggest that under our experimental conditions this crucial intermediate serves both as the fibrillation nucleus, as well as the elongating species. We propose that the latter intermediate is an interesting target for small molecules in order to prevent or reduce amyloid fibril formation.

Amyloid fibrils are associated with critical diseases such as Alzheimer disease and Type 2 diabetes. In each amyloid disease, a particular protein or polypeptide aggregates and forms insoluble fibrils [1]. Moreover, amyloid fibrils play a critical role in unwanted degradation of a number of protein-based drugs [2,3]. Insulin fibrillation is a commonly used example of amyloid fibrillation because amyloid deposits have been observed in patients after subcutaneous insulin infusion [4] as well as in vitro, where insulin easily fibrillates, causing problems during production, storage, and delivery of insulin-based drugs [2]. The amyloid fibrils share several common structural properties, in which twisted and unbranched amyloid fibrils typically have a diameter of about 100 Å and highly variable lengths up to several microns [5–8]. Furthermore, mature fibrils are suggested to be composed of intertwined protofibrils built from two to three intertwining protofilaments, each with a typical diameter of 15–50 Å [6,7,9]. The most dominant repeating features of presumably all amyloid fibril structures are suggested cross-β-sheets with the β-strands running perpendicular to the fibril axis and an inter-strand spacing of about 4.8 Å [8,10]. Recently, atomic resolution structures of amyloid-like hexameric peptide segments, which form such cross-β-sheets, have been determined [11]. Amyloid fibrillation is proposed to be a nucleation-dependent process [12–16]. For insulin fibrillation, the initial step has been suggested to proceed through a non-native, partly unfolded, monomeric intermediate [12], which proceeds to form an oligomeric nucleus [17,18] prior to elongation of protofilaments. A more thorough knowledge of the structure and mechanism behind the formation of amyloid fibrils in general, and at the early stages of nucleus formation in particular, is essential for the understanding of the processes of amyloidosis. It is equally crucial for the rational design of drugs to inhibit or reverse amyloid formation, and for the general understanding of the mechanisms of protein folding and stability, since the ability to form amyloid fibrils has been proposed to be a generic feature of all polypeptide sequences [1]. Importantly, it has been suggested that oligomeric precursors of amyloid fibrils, and not the mature fibrils themselves, may be cytotoxic [19,20]. Indeed, there are indications that precursors are the main cause of toxicity [21]. This further emphasizes the importance of characterizing such oligomeric precursors, in the search for therapeutic treatment of amyloid diseases. A structural characterization of the fibrillation nucleus is difficult, because at low protein concentrations, this nucleus is the thermodynamically least-favorable species [22]. However, a recent mathematical model of the nucleation-dependent fibrillation process suggests that the nucleus becomes thermodynamically stable and accumulates above the supercritical concentration [16]. The species is thus, per definition, no longer a thermodynamic nucleus, but is defined as a structural nucleus. The three-dimensional shape of the structural nucleus is thus identical to the thermodynamic nucleus present at low concentrations [16]. However, the structural nucleus is still inherently difficult to isolate and characterize. Most methods applied for biophysical or structural characterization of amyloid fibrillation are problematic because they either disturb the equilibrium of the fibrillation process (such as in x-ray crystallography or nuclear magnetic resonance), or the sample suffers from surface-mediated effects (such as in scanning tunneling microscopy, atomic force microscopy [AFM], and cryoelectron microscopy [cryo-EM]). In contrast, small-angle x-ray scattering (SAXS), which yields low-resolution macromolecular structures, allows examination of the fibrillation process directly in solution, so that the individual components present during the evolving equilibrium can be studied. We report a state-of-the-art SAXS analysis of fibrillating insulin and determine the ab initio low-resolution structure of the amyloid fibril repeating unit. Moreover, the structure of a helical oligomer, presumably representing the insulin structural nucleus, is determined, providing surprising details regarding its oligomeric composition and organization, with far-reaching conceptual impact on the current understanding of the early stages of amyloid fibrillation. Above a certain supercritical concentration, at which the rate of the fibril formation reaction becomes almost independent of the protein concentration, Powers and Powers demonstrate that the structural nucleus becomes more stable than the monomer [16]. The concentration dependency was measured using Thioflavin T (ThT) fluorescence. The experimental conditions used in our SAXS analysis, with insulin concentrations of 5 mg/ml, are above the supercritical concentration (Figure 1). A characteristic x-ray fiber diffraction pattern was observed for insulin samples under the experimental conditions employed during the SAXS studies (Figure S3). A typical meridional signal is seen at 4.8 Å, and an equally typical equatorial signal is recorded at 10.5 Å. In addition to the characteristic ThT fluorescence signal, this confirms the existence of fibrils. Time-resolved synchrotron SAXS was applied to study the insulin fibrillation process in solution, starting from monomeric insulin. During the 5-h lag phase, only monomers were detected (Figure 2 and Table S1), with a small increase in the average size of the molecules towards the end of the lag phase, presumably reflecting the expected partial unfolding of a fraction of the monomers. During the 4-h elongation phase, a significant growth of particles is observed leading to drastic changes in the scattering pattern (Figure 2A). Importantly, the scattering curves recorded during the elongation phase could not be described by linear combinations of the scattering from monomers and mature fibrils. Singular value decomposition (SVD) analysis further confirmed that three components contributed significantly to the scattering during the elongation phase (Figure S1). This suggested to us the presence of monomers, fibrils, and a distinct oligomer, apparently representing the hitherto largely uncharacterized structural nucleus. The scattering pattern from the distinct oligomer (Figure S2A) was evaluated as described in Materials and Methods. All the individual scattering patterns collected during fibrillation could be fitted by a linear combination of the three components yielding their volume fractions (Figure 2A and 2B). As expected, the amount of monomeric insulin decreases gradually. Appearance of large amounts of the distinct oligomer accompanies the monomer decrease. There is an additional lag phase of about 1 h between the initial observation of oligomer formation and the onset of fibril growth. Most remarkably, the growth rate of the mature fibrils is proportional to the amount of the oligomer in solution (Figure 2C), suggesting that the oligomers represent the building blocks for the growing fibrils. Thus, above the supercritical concentration, the insulin fibril elongation likely proceeds primarily by addition of oligomers. Also, the amount of monomers in solution is close to zero during the later stages of fibril formation, which further supports the notion that oligomer addition elongates the protofilaments. This mechanism is further supported by the shape of these oligomers, presented below. As early as in 1957, Waugh proposed that a nearly simultaneous interaction of three to four insulin monomers forms a nucleus [18], and later mass spectroscopy also revealed a relatively small nucleus [17]. An elongated association of monomers prior to fibrillation has been observed by AFM on insulin samples [23]. The existence of significant amounts of an oligomeric species of insulin prior to the appearance of mature fibrils and throughout the fibril elongation process was also detected using dynamic light scattering [24] and time-lapse AFM [25]. In the present study, we further characterize the low-resolution shape of the oligomer ab initio that yields a good fit to its scattering pattern (Figure S2A) and reveals a helical structure with a length of 200 Å and a twist of 270°, or three quarters of a full helical turn in the direction of the long axis (Figure 3). The hydrated volume of the oligomer is 6.4 × 104 Å3, corresponding to an approximate molecular weight of 32 kDa, or 5.6 insulin monomers. Visually, the helical oligomer also appears as a bead-on-a-string assembly of six units, each with dimensions comparable to those of insulin monomers (Figure 3). Moreover, using independent rigid-body modeling (Protocol S1) in terms of the crystallographic model of partially unfolded insulin at pH 2 [26], the models consisting of five or six subunits provide the best fit to the scattering pattern from the distinct oligomer (Figure S2A). A typical rigid-body model (Figure S2B) displays six monomers in a helix-like configuration, very similar to that observed in the ab initio model (Figure 3). The SAXS data collected after 9-h incubation were used to evaluate the shape and dimensions of the mature fibril. Given the resolution of the scattering pattern (from about 900 Å to 12 Å), the SAXS data provide information about the organization of the coherently scattering volume inside the fibril, associated with its repeating unit. Using indirect Fourier transformation (IFT), the repeating unit has a length of about 700 Å and a cross-section diameter of about 300 Å (Tables S1 and S2). Ab initio modeling of the fibril was further performed based on these dimensions, showing an elliptically shaped repeating unit with overall dimensions of 690 × 390 × 160 Å3 (for a fit of the final model and data, see the lowest curve in Figure 2A). The length (690 Å) of this unit is within the range of helical crossovers of 350 to 940 Å observed with cryo-EM of insulin fibrils [6]. The fibril consists of three intertwining helical protofibrils (Figures 3 and 4), as also observed by cryo-EM and AFM [6,7]. Each protofibril has an average diameter of about 100 Å, a protofibril diameter that is in accordance with cryo-EM studies [6]. The fibril has a left-handed helical twist as observed for most amyloid fibrils [6,27,28]. The hydrated volume of the repeating helical unit of the fibril (1.44 × 107 Å3), corresponds to an approximate molecular weight of 7,200 kDa or about 1,240 insulin monomers. Given the potential heterogeneity of the mature fibrils and possible residual contribution of the monomers and oligomers even at the latest stages of fibrillation, the models (Figures 3 and 4) represent average shapes of the most abundant fibrillating units. It should be noted, however, that the heterogeneity of the fibrils appears to be low, given that there are only three components in the above model-free SVD analysis. Note also that the potential heterogeneity of the mature fibrils does not challenge our results concerning the presence and volume fractions of the oligomeric species, since these have been determined independently of the model of mature fibrils. Although one cannot completely exclude the possibility that fibrils may grow by addition of individual insulin monomers, the fact that the growth rate is proportional to the amount of oligomers strongly suggests that incorporation of the latter is the predominant mechanism of elongation. Moreover, the helical shape of the oligomer (Figure 3) is ideally suited for entering a protofilament. We have generated a tentative model of the arrangement of intertwining helical oligomers into dense protofilaments, assuming that elongation of protofilaments is directed by oligomer association. Figure 4 displays such a model with a central cavity in which the cross-section contains eight helical oligomers (see Video S1 for the stepwise construction of the protofilament by the addition of helical oligomers). Fourteen repeats of the length of these eight helical oligomers would form a full turn of one protofilament across three fibril repeating units. Two protofilaments presumably intertwine to form the protofibrils, yielding an average protofibril diameter of about 100 Å. Finally, three protofibrils intertwine to form the mature fibril (see Video S2 for a movie on the stepwise construction of the mature fibril from protofilaments). Hence, a total of 224 oligomers would build up one full turn of the protofibril. When following one single protofibril across three repeating fibril units, the protofibril will have traversed all three positions of the three strands that are visible within one fibril repeat. Thus, the volume and estimated molecular weight of the coherently scattering unit must correspond to the calculated number of 224 helical oligomers that define our structural model, which indeed is the case, and the suggested packing is in agreement with the molecular weights of 32 and 7,200 kDa of the oligomer and the helical repeat of the mature fibril, respectively. In the suggested packing of the protofilament, the helical oligomer is preshaped for interactions with other oligomers, and all subsequent oligomer additions elongate the protofilament. Moreover, oligomer additions are possible in both directions. Thus, our model suggests that above the supercritical concentration, the helical oligomer is both the structural nucleus and the elongating building block of insulin amyloid fibrils. The existence of oligomeric species during the amyloid fibrillation process is not unique for insulin fibrillation and has previously been reported, e.g., for β-amyloid peptide [29], β-2-microglobulin [30], phosphoglycerate kinase [31,32], a fragment of Syrian hamster prion protein [32,33], and myoglobin with polyglutamine insertions [34]. The overall sizes of some of these aggregates are compatible with the insulin oligomer described in this study. An insulin fibrillation study using dynamic light scattering also reports significant amounts of oligomers of similar size [24]. Interestingly, the β-amyloid peptide oligomers revealed by scanning tunneling microscopy appeared as a bead-on-a-string–like assembly of individual monomers [29], strikingly similar to the present model of insulin oligomers. Thus, the new structural findings presented here may reveal generic features of amyloid fibrillation. The importance of monomeric insulin for fibrillation has been reported in many studies [12,13] and explained by the need for a partial unfolding of the monomer prior to association into amyloid fibrils. This is corroborated by Fourier transform infrared spectroscopy and circular dichroism studies, showing that a significant conversion of α-helical to β-sheet structure accompanies formation of insulin amyloid fibrils [2,12]. The presence of repeating cross-β-sheets perpendicular to the fibril axis with an inter-strand spacing of about 4.8 Å in the present insulin fibrils was confirmed by x-ray diffraction (Figure S3). In the case of the oligomers of β-amyloid peptide [29], which have a similar appearance to the insulin structural nucleus, β-strands are directly observed [29], and a structural conversion accompanying oligomer formation is also observed for insulin [24]. However, it remains to be shown whether the structural conversion of insulin to a β-structure occurs prior to or during structural nucleus formation, because insulin monomers substantially refolded into β-structure have never been isolated. Under amyloidogenic conditions, a partly modified structure has been described, although a central natively folded core remains [35]. Assuming that a β-sheet with strands separated by 4.8 Å runs parallel to the long axis of the distinct helical oligomer with a length of 200 Å, this corresponds to approximately 40 β-strands with a relative twist of each β-strand in the β-sheet of about 7°. Based on this rationale, each monomer of the oligomer should accommodate six to seven β-strands. Small peptide segments of only three or four residues and several hexa-peptides from insulin have been shown to form amyloid fibrils [36] and crystals with amyloid features [37], respectively. This suggests that insulin is, at least in theory, able to adopt such a structure, as it has 51 amino acids. It is known that the two interchain and one intrachain disulfide bonds in insulin are accommodated in fibrils [6], and it remains to be shown how insulin is able to accommodate this in such a β-sheet–rich structure. Another issue to consider is the recent theoretically based suggestion that α-sheets, and not β-sheets, may form the repeating backbone in the oligomer [38,39]. These conformations cannot be distinguished within the helical insulin oligomer due to the limited resolution of the SAXS data. The nucleation polymerization model suggests that partially unfolded monomers form an oligomeric nucleus, followed by the elongation process proceeding via direct addition of structurally non-native monomers [12,40]. However, in our studies, performed above the supercritical concentration, the oligomeric species is present in high amounts during fibrillation, and the amount is proportional to the rate of fibril elongation. Although Powers and Powers [16] consider elongation only via monomer addition, they also note that it seems physically realistic that the oligomers would associate. This is further supported by the observation of β-sheets in fibrillar oligomers [29]. Although elongation of protofilaments by helical oligomer association appears structurally feasible, we cannot strictly rule out that elongation occurs at least in part through monomer addition. This would, however, imply that the oligomer dissociates into monomers at rates comparable to the elongation rate of the growing fibrils. Thus, the amount of oligomers would decrease prior to or simultaneous with the appearance of fibrils [16]. In contrast, we observe the highest amount of oligomers not at the start of the fibril growth, but rather when the rate of elongation is maximal. The elongation of fibrils also continues after the monomer concentration is no longer detectable, serving as further argument in favor of the “elongation via the structural nucleus” mechanism. An elongation through oligomer assembly was proposed for a peptide derived from the prion protein Sup35 [41]. Fluid oligomeric species was observed and hypothesized to be structurally distinct from the nucleus. The oligomeric species was suggested to change conformation upon interaction with the nucleus, thus speeding up the assembly process. It is unclear whether the experiments by Serio et al. [41] were performed below supercritical concentration or, like ours, above the supercritical concentration, because the supercritical concentration, to our knowledge, has not yet been established for Sup35. The oligomeric species we observe does accumulate above the supercritical concentration (and thus according to Powers and Powers [16] should correspond to the structural nucleus), and it reveals a distinct helical shape strikingly suitable to associate with other helical oligomers (see Video S1). Of course a conformational change may happen during the association of the structural nuclei in the elongating protofilaments, but our modeling shows that such a conversion is not required for the elongation of protofilaments via the assembly of structural nuclei. In summary, we present experimental evidence for the accumulation of a helical oligomeric species, which has previously been suggested to be the structural nucleus. In contrast to the numerical simulation of elongation via monomer addition proposed by Powers and Powers, our experimental results demonstrate that the amount of the structural nucleus in solution determines the rate of elongation. The shape of the structural nucleus is uniquely well suited for forming the protofilaments, obviating the need for conformational changes in the nucleus before associating into the growing protofilaments. A secondary nucleation mechanism after the primary nucleus formation has been proposed for insulin fibrillation [23,42,43]. Such mechanisms may be branching, fragmentation, and nucleation on the fibril surface. These features would not be distinguishable in the current study, since the scattering units present in solution at all times would still be oligomers and mature fibril repeating units, irrespective of the site of secondary nucleation. Our model of a dense protofilament structure has a central cavity (Figure 4A and 4B). Such a cavity has been proposed in some, but not all, models of amyloid protofilaments [44–46], e.g., in a β-helical model for poly-glutamine [47] and in hollow nanotube structures observed for amyloid fibrils, e.g., Aβ(1–40) [48] and Aβ(1–25) [49]. Although no high-resolution structure of a mature amyloid fibril is currently available, it appears that amyloid fibrils possess common structural features, the most dominant being the high content of β-sheets with β-strands perpendicular to the longest and elongating axis of the fibril [8]. This might suggest a common structural principle for the mechanism of fibril formation for several amyloid proteins. The helical oligomer preshaped for fibril formation could represent a generic principle for the building blocks in amyloid fibril formation, which also may represent the structural scaffold followed during monomer addition at subcritical concentrations. This study demonstrates the power of SAXS analysis on data from amyloid systems, since it is possible to separate scattering contributions from different species present during fibrillation. The method allows determination of both the amount and the low-resolution structure of the species in solution. For amyloid systems characterized by nucleated polymerization, the thermodynamically unfavorable nucleus becomes stable above supercritical concentrations, and is defined as a structural nucleus [16]. Here, we report experimental evidence for the accumulation of the structural nucleus in insulin fibrillation and characterize its structure. This oligomer has a distinct helical structure in solution, and the correlation between the amount of helical oligomer present in solution and the growth rate of the mature fibrils suggests that it is the oligomer that primarily elongates the growing protofilaments. Using supercritical concentrations, the potential toxicity of accumulating oligomers of various amyloid proteins could be investigated. The oligomer is conceptually central for the understanding of the early stages of amyloid fibrillation. The low-resolution model and documented presence of large amounts of oligomers can facilitate the search of new strategies for preventing or reversing insulin fibrillation or amyloid fibrillation in general, e.g., by hindrance of structural nucleus formation and/or packing of the helical oligomer into protofilaments. Human zinc insulin was obtained from Novo Nordisk, http://www.novonordisk.com. All other chemicals were of analytical grade. A total of 0.06–12-mg/ml insulin in 20% acetic acid (pH 2.0), with 0.5 M sodium chloride was fibrillated at 45 °C on a Fluostar Optima plate reader from BMG Labtechnologies (http://www.bmg-labtechnologies.com). Insulin (5 mg/ml) in 20% acetic acid (pH 2.0), with 0.5 M sodium chloride was fibrillated at 45 °C on a Fluostar Optima plate reader from BMG Labtechnologies. Fibrils were centrifuged at 18,000 g on a Hettich centrifuge and dried in a desiccator. X-ray synchrotron data were collected at MAXLAB beamline 911–3, Lund, Sweden, using a MarMosaic 225 detector at 20 °C; λ = 1.3 Å, and a 350-mm specimen:detector distance during 30-s exposure time. A pellet of human insulin was dissolved at 5 mg/ml in 20% acetic acid with 0.5 M sodium chloride, at t = 0 (Table S1). The sample was kept and measured at 45 °C approximately every 15 min. Zinc acetate was added to the background buffers corresponding to two Zn2+ ions per insulin hexamer in the protein samples. Synchrotron scattering data were collected on the X33 camera of the European Molecular Biology Laboratory on DORIS III (Deutsches Electronen Synchrotron [DESY], Hamburg, Germany) at a wavelength of 1.5 Å, using a MAR345 Image Plate detector, in the momentum transfer range 0.01 < s < 0.50 Å−1 (s = 4π sinθ/λ, where θ is half the scattering angle) with 3-min exposure times. Selected insulin samples exposed repetitively did not result in detectable radiation damage. Data analysis was performed using the software suite ATSAS 2.1 [50]. The data reduction was done using the program PRIMUS [51], and the molecular masses of the solutes were estimated by comparing the extrapolated forward scattering I(0) with that of a reference solution of bovine serum albumin (Table S1). IFT were evaluated by the program GNOM [52], providing estimates of distance distribution functions, including the radii of gyration Rg and maximal distances (Dmax) within the particles. SVD analysis was performed by the program SVDPLOT [51] using all the measured datasets to detect the number of independent components (Figure S1). The program OLIGOMER [51] was employed to represent the experimental data by linear combinations of the scattering from individual components (monomers, oligomers, and fibrils) and to compute their volume fractions (Figure 2B). The calculated scattering from a monomer of insulin (partially unfolded monomers at acidic pH, pdb-file 1GUJ) and the low-pass–filtered experimental scattering from the coherent unit of the fibril after 9-h incubation were employed. To find the first representation of the scattering pattern from the oligomer, several tentative models were built assuming the oligomer to be from dimers to hexadecamers constructed in long, flat, zigzag, or compact compositions. The representation was selected for which OLIGOMER yielded overall best agreement between the experimental data measured during the elongation phase and the linear combinations of the three components. Then the volume fraction of the oligomer was set to zero, and those of the monomers and fibrils were kept at the values for the three-component fit. The scattering from the oligomer was computed as the average of all the residuals between the experimental data and two-component fits. This decomposition ensures that the scattering curve representing the oligomer is independent of potential heterogeneity of the mature fibrils, because the contribution of the latter is removed. Thus, the determination of the volume fractions during fibrillation and the modeling of the structural nucleus are independent both of the heterogeneity of the mature fibrils and of the modeling of the helical fibril repeat. Ab initio shapes of the insulin fibril repeat and the distinct oligomer were obtained using the program DAMMIN [53]. The program employs a simulated annealing protocol to search for a compact bead model minimizing the discrepancy between the experimental and calculated curves at low resolution (up to s about 0.1–0.2 Å−1). Based on the best approximating three-parameter body as computed by the program BODIES [51], the search volume for the fibril repeat was an ellipsoid with half-axes of 350, 250, and 150 Å using 35,000 spheres. Individual jobs were loaded into a Linux cluster, and 20 independent models were averaged using the program DAMAVER [54]. The resulting model was used as input for subsequent runs of 20 individual models, for a total of five refinements. Likewise, the oligomer shape was calculated inside a sphere with a diameter of 200 Å, and gradually refined over three runs using an average of 20 individual models as input. The final averaged and filtered models also yielded the excluded volume of the hydrated particle.

10.1371/journal.ppat.1006194

Vaccination has minimal impact on the intrahost diversity of H3N2 influenza viruses

While influenza virus diversity and antigenic drift have been well characterized on a global scale, the factors that influence the virus’ rapid evolution within and between human hosts are less clear. Given the modest effectiveness of seasonal vaccination, vaccine-induced antibody responses could serve as a potent selective pressure for novel influenza variants at the individual or community level. We used next generation sequencing of patient-derived viruses from a randomized, placebo-controlled trial of vaccine efficacy to characterize the diversity of influenza A virus and to define the impact of vaccine-induced immunity on within-host populations. Importantly, this study design allowed us to isolate the impact of vaccination while still studying natural infection. We used pre-season hemagglutination inhibition and neuraminidase inhibition titers to quantify vaccine-induced immunity directly and to assess its impact on intrahost populations. We identified 166 cases of H3N2 influenza over 3 seasons and 5119 person-years. We obtained whole genome sequence data for 119 samples and used a stringent and empirically validated analysis pipeline to identify intrahost single nucleotide variants at ≥1% frequency. Phylogenetic analysis of consensus hemagglutinin and neuraminidase sequences showed no stratification by pre-season HAI and NAI titer, respectively. In our study population, we found that the vast majority of intrahost single nucleotide variants were rare and that very few were found in more than one individual. Most samples had fewer than 15 single nucleotide variants across the entire genome, and the level of diversity did not significantly vary with day of sampling, vaccination status, or pre-season antibody titer. Contrary to what has been suggested in experimental systems, our data indicate that seasonal influenza vaccination has little impact on intrahost diversity in natural infection and that vaccine-induced immunity may be only a minor contributor to antigenic drift at local scales.

Influenza is a significant global health problem. Vaccination is the best way to prevent influenza virus infection, and seasonal influenza vaccines are considered for reformulation each year in order to keep up with the virus’ evolution. Despite these efforts, vaccine recipients often develop an immune response that does not protect from infection. Given the current recommendation that all people over 6 months of age get vaccinated, it is important to understand how vaccination itself may impact viral evolution during natural human infection. We studied how vaccination may alter viral evolution within individuals, as each person harbors many highly-related influenza variants that differ in their ability to escape the immune response. We compared groups of people in a vaccine trial to determine the impact that vaccination has on viral diversity and variant selection within individuals. We did not detect significant differences in the number of variants detected or in the prevalence of mutations that could impact antibody binding based on vaccination group or antibody response. Our work suggests that vaccination is not a major factor in driving the emergence of new influenza strains at the level of the individual host.

Despite recommendations for universal influenza vaccination and the ample availability of vaccines in the United States, influenza continues to cause significant morbidity and mortality [1]. This is, in part, a result of the modest effectiveness of current vaccines, so that considerable numbers of vaccine failures occur each year. Within individuals, influenza populations exist as a collection of closely related, and at times antigenically distinct, variants that may exhibit diverse phenotypes [2–6]. Intrahost single nucleotide variants (iSNV) can be transmitted as part of the infecting population [5,7–10] or generated over the course of an infection due to the virus’ low replication fidelity [11,12]. The evolutionary forces that shape the genetic structure of viral populations within hosts and ultimately give rise to novel antigenic variants at the host population level are poorly understood. A clear understanding of the intrahost diversity of influenza virus populations and its impact on influenza virus evolution is central to many questions of direct clinical and public health relevance [13]. Influenza vaccines are considered for reformulation each year to counter the viral antigenic drift that enables escape from the previous year’s vaccine [14]. Annual influenza vaccine effectiveness is 60% on average, and can be much lower during antigenically unmatched years [15,16]. While antigenic drift is monitored annually on a global scale, the source of antigenic variation is ultimately at the level of the individual host. Phylogenetic studies of whole genome sequences from cities and smaller communities have demonstrated that multiple lineages circulate over the course of a single influenza season [2,17], and individual hosts may harbor mixed infections that include antigenically novel variants [3,4]. While human hosts could be preferentially infected with one lineage over another based on pre-infection immune status, the degree to which circulating escape variants contribute to vaccine failure is currently unknown. Host immune selection is a major driver of influenza virus evolution on the global scale. Both phylogenetic analysis and antigenic cartography have demonstrated that antibodies exert positive selective pressure on the viral hemagglutinin (HA) and neuraminidase (NA) proteins [18,19]. Vaccination and natural influenza infection often lead to partial, or non-sterilizing immunity, and post-vaccination antibody titers are only a moderate predictor of subsequent protection [20,21]. Previous work has demonstrated that sub-neutralizing concentrations of immune sera can promote the generation of antigenic variants, and some have argued that vaccination can accelerate the process of antigenic drift [22]. A recent study in vaccinated people suggested that novel antigenic variants could be present at low frequencies [23]. Importantly, humans often differ in their prior exposure to influenza viruses and vaccines, and pre-existing immunity may confound such studies [24]. Therefore, the extent to which partial immunity selects for antigenically relevant variants during natural infection in humans is unclear. By necessity, most of the available data on vaccination and intrahost evolution have come from analyses of HA sequences in large animal models of infection, including horses, pigs, and dogs. These studies have suggested that intrahost populations include a number of somewhat rare single nucleotide variants that increase and decrease in frequency over the course an infection with sporadic fixation events occurring in some animals. The overall impact of vaccination on antigenic diversification was not clear [7,25–27], and there were differences between experimentally and naturally infected animals [7,8,26]. These differences are also likely to be an issue in extrapolating results from human experimental challenge models [28]. Some have suggested that high intrahost diversity reflects increased viral fitness, and mechanisms that alter intrahost diversity may impact evolutionary trajectories [29–31]. If the transmission bottleneck is sufficiently wide, low frequency variants that arise within a host can plausibly be transmitted and spread through host populations [5,10]. Understanding how intrahost diversity is generated and maintained and the extent to which host immune status impacts this diversity may be important for defining influenza virus’ larger evolutionary patterns [23,26,27]. Here we used next generation sequencing to define the impact of vaccine-induced immunity on the intrahost diversity of influenza virus during natural infection. We specifically asked: (i) whether influenza viruses in vaccinated individuals represent escape variants, (ii) whether novel antigenic variants are found in hosts with non-sterilizing immunity, and (iii) the degree to which vaccine-induced immunity impacts the overall diversity of intrahost populations. Because we analyzed influenza populations from individuals enrolled in a randomized, double-blind, placebo-controlled trial of influenza vaccine efficacy [32–34], we were uniquely positioned to define this aspect of immunity to natural infection. We utilized influenza A RT-PCR positive throat swab samples from a randomized, double-blind, placebo-controlled study of vaccine efficacy that took place during the 2004–2008 influenza seasons at six study sites in Michigan [32–34]. This trial measured vaccine efficacy of both the trivalent inactivated (IIV) and live attenuated influenza vaccine (LAIV) compared to placebo and each other. We sequenced patient-derived influenza populations without culturing from three seasons: 2004–2005, 2005–2006, and 2007–2008. The 2006–2007 influenza season did not have enough influenza-positive samples for our study (total n = 16). Influenza A (H3N2) strains dominated the other three seasons, and the circulating 2004–2005 virus was considered at the time to be only a modest mismatch with the vaccine strain. The other seasons were antigenically matched. The numbers of subjects each year were as follows: 2004–2005 season, 522 IIV, 519 LAIV, and 206 placebo [32]; 2005–2006 season, 867 IIV, 853 LAIV, and 338 placebo [33]; 2007–2008 season, 813 IIV, 814 LAIV, and 325 placebo [34]. Over these 5119 person-years of observation, 165 individuals had culture or RT-PCR confirmed influenza A (H3N2) infection and specimens available for analysis. Of these, 80 individuals had received LAIV, 42 had received IIV, and 43 had received placebo. For 2004–2005, flu-positive samples were available for 28 subjects: 7 IIV, 12 LAIV, and 9 placebo, and for 2005–2006, 32 samples were available to study: 13 IIV, 14 LAIV, and 5 placebo. During the 2007–2008 season, 105 flu-positive samples were available: 22 IIV, 54 LAIV, and 29 placebo. We were able to amplify and quantify genomes for 119 of the 165 influenza-positive samples (Table 1). The average age of this sequenced cohort from all years was 24.5, indicating that participants were generally young and likely shared similar influenza pre-exposure histories particularly after randomization. The age, sex and race of the cohort were similar to that of the overall study cohort for each of the 3 seasons. Despite subject randomization, differences in pre-existing immunity due to prior vaccination or influenza infection could provide a dominant immune background that is not significantly altered by vaccination against that season’s strain. In the 2007–2008 season, 50.4% of individuals in the larger cohort and 44.3% in the sequenced cohort reported having ever received a prior influenza vaccination. To obtain a more reliable metric of strain-specific immunity, we measured pre-season antibody titers by hemagglutination inhibition (HAI) and neuraminidase inhibition (NAI) assays for all study participants against that season’s vaccine strain (S1 Fig). Vaccine-induced antibody titers and overall vaccine efficacy are generally stable for a single flu season [35]. Therefore, our pre-season HAI and NAI titers are likely to be similar to titers at the time of infection. Pre-season (post-vaccination) titers for individuals in the IIV group were above the geometric mean for the entire sequenced cohort, LAIV subjects had titers spanning the mean, and those in the placebo group were generally below the mean for all seasons. Since an HAI titer of 40 is typically considered to be associated with 50% protection given exposure [20,21,36], these data demonstrate that in the IIV group, and to a lesser degree, the LAIV group, individuals had strain-specific antibody levels that are sufficient to apply selective pressure against the infecting virus. We have previously shown that viral load influences the sensitivity and specificity of iSNV detection [6]. In order to determine whether viral load was different among the IIV, LAIV, and placebo samples that we sequenced, we measured genome copy number by RT-qPCR for the 2004–2005, 2005–2006, and 2007–2008 seasons. For the 2007–2008 season, which had the most samples, there were no significant differences in copy number by vaccination group (Fig 1A). In agreement with the 2007–2008 data, we did not detect differences in copy number by vaccination group for the 2004–2005 and 2005–2006 seasons (S2 Fig). Since copy number is dependent on time from illness onset [37,38], we analyzed the data based on sample collection day (Fig 1B). Using days 2–4, for which there were at least 5 data points for each treatment group, we did not find any significant differences (p = 0.24–0.57 for days 2–4, non-parametric one way ANOVA). We divided the larger group of 2007–2008 subjects into groups based on pre-season HAI and NAI titers ≥40 or <40 against that season’s strain. This cutoff was identical to the HAI and NAI geometric mean titers for our sequenced cohort (61.9 and 34.5, respectively), given the dilutions used. We did not detect differences in copy number based on HAI or NAI titer (Fig 1C and 1D), even when accounting for day of symptom onset (p = 0.25 for HAI, p = 0.97 for NAI, Mann-Whitney U test). Because we only measured copy number in the subset of virus populations that were amplified and sequenced, these data should not be interpreted in the context of vaccination and overall shedding [39]. We used the Illumina platform to determine the whole genome consensus sequence and to identify intrahost single nucleotide variants for each patient-derived sample. Importantly, we have developed and rigorously benchmarked a variant calling pipeline that maintains high sensitivity for rare iSNV detection while dramatically reducing false positive variant calls [6] (S1 Table). We have found that the number of false positive iSNV calls is much higher in samples with genome copy numbers <103 per μl of transport media; therefore, we only report iSNV from a high quality dataset that includes 64 samples from the 2007–2008 season. High quality iSNV data for the 2004–2005 and 2005–2006 influenza seasons are included in Supplementary Figures and Tables. Our libraries yielded an average coverage above 20,000 reads per base with even coverage across the coding region of all segments for each season (S3 Fig). Given the frequent observation of community-level diversity in circulating influenza viruses [2,17], we asked whether vaccinated individuals were infected with distinct strains relative to the placebo group. If vaccine failures were due to infection with an antigenically distinct variant, we would expect to see evidence of clustering by vaccination- or sero-status in HA and NA phylogenetic trees. We therefore analyzed HA and NA consensus sequences from all 87 individuals in the 2007–2008 season (Fig 2). There was very little diversity in either gene, and we found that sequences from individuals in each treatment group (e.g. IIV, LAIV, or placebo) were dispersed throughout the tree. More importantly, we found no evidence for clustering based on pre-season HAI and NAI titer (by colors in Fig 2, titers indicated at tips). We obtained similar results from the 2004–2005 and 2005–2006 seasons, albeit with fewer sequenced samples (S4 Fig). These data suggest that within-season and within-host antigenic drift due to higher levels of vaccine-induced antibodies (HAI or NAI >40) were not major determinants of vaccine failure in these seasons. We next analyzed the iSNVs present in 64 of our samples from the 2007–2008 season. We identified 360 minority variants across the entire genome, most of which were present at a frequency of <0.1 (Fig 3A). We did not observe many samples with a large number of higher frequency iSNV, which suggests that there were few, if any, mixed lineage infections in the samples from this season. The vast majority of iSNV were only found once (Fig 3B). We also evaluated whether partial immunity impacts viral diversity by comparing the number of iSNVs per sample in the HA and NA genes based on HAI and NAI titers, respectively. We did not observe a difference in iSNV count based on HAI and NAI titers for either HA or NA (Fig 3C and 3D, p = 0.20 for HA, p = 0.26 for NA, Mann Whitney U test; see also S5 Fig for iSNV stratified by titer). The average number of iSNV per sample was similar across the genome regardless of host treatment group (Table 2, p = 0.35, non-parametric one way ANOVA) or HAI and NAI titers (S6 Fig, p = 0.13 for HAI, p = 0.22 for NAI, Mann Whitney U test). Most iSNV from the 2004–2005 and 2005–2006 seasons were also found in only one sample each and at frequencies <0.1 (S7 Fig). The number of iSNV did not differ based on treatment group or HAI and NAI titer (S2 Table). We did not identify any variants specifically associated with vaccination group or HAI and NAI titers for any of the influenza seasons analyzed. Unlike experimental challenge studies, we only had one sample per person and could not evaluate changes in diversity at the level of the individual host. However, we did not identify any significant differences in diversity by day of infection across the cohort by vaccination status (Fig 4, p = 0.16–0.82 for days 2–4, non-parametric one way ANOVA) or antibody titer. We identified only a marginal difference in the number iSNV across the genome based on HAI titer on day 2 (uncorrected p-value 0.02 with >6 comparisons, Mann Whitney U test). These data suggest that our results are unlikely to be confounded by temporal sampling issues. Some have suggested that vaccine-induced immunity will select for novel antigenic variants within hosts [23]. We therefore compared iSNV in the HA gene by vaccination group and serostatus. Of the 17 variants we identified that resulted in nonsynonymous changes within HA, 11 were in HA1 and 6 were in HA2 (Fig 5, S3 Table). Five antigenic sites, comprising 131 amino acid positions, have been described in HA1 for H3N2 viruses [40–42]. Six of the HA1 variants identified in our study were located in antigenic sites C and D (Fig 5B), one of which was found in two samples (S3 Table). Of these potential antigenic variants, three were found in vaccinated individuals (1 IIV, 2 LAIV) and four were found in those in the placebo group. When grouped by HAI titer, four were found in samples from individuals with titers ≥40, while three were found in individuals with titers <40. No variants were specific to vaccinated individuals, and antigenic diversity was similar across all groups. None of the identified mutations were observed in subsequent circulating H3N2 strains. There were no nonsynonymous iSNV in HA from the 2004–2005 samples and just 7 identified in individuals from the 2005–2006 season. Of these 7, N225D was observed in strains dominating prior, but not subsequent seasons. We set out to define the relationship between vaccine-induced immunity and the intrahost diversity of influenza virus. We hypothesized that non-sterilizing immunity could potentially select for novel antigenic variants and contribute to larger scale patterns of influenza evolution. We were able to definitively address these questions using samples derived from a randomized, placebo-controlled vaccine trial in healthy, young adults who likely had similar prior exposure to influenza viruses [32,34]. The availability of post-vaccination, preseason HAI and NAI titers allowed us to examine directly the impact of measured serologic antibody pressure rather than using vaccination alone as a surrogate marker. Because all individuals were infected naturally, our data provide a rare view of within-host influenza virus diversity in humans. We directly sequenced the samples without passage in cell culture, eliminating the possibility of culture-adapted mutations and employed a well-benchmarked variant calling pipeline [6] that dramatically reduces the false positive iSNV calls that often plague next generation sequencing studies. In this exhaustive and well-controlled study, we found no differences in intrahost influenza diversity based on vaccination status or HAI and NAI titers. Our findings in a natural infection system are concordant with an equine influenza virus evolution study in vaccinated horses. Intrahost variation was similar between naïve and vaccinated horses, regardless of whether they were infected naturally or experimentally [27]. However, not all experimental infection studies mirror results seen during natural infection. A study investigating swine influenza virus found discrepancies in intrahost variation based on whether animals encountered natural infection or were experimentally infected [26]. Our data is in contrast with a study of experimentally-infected dogs that uncovered differences in intrahost diversity and evolution in antigenic sites based on vaccination status [25]. Two other studies of equine influenza virus found mixed infections of multiple influenza lineages during natural infection, which would not be seen in an experimental model but may be relevant to the transmission and spread of novel variants [7,8]. Overall, it is challenging to compare our results to those obtained in various animal models, where the hosts are often immunologically naïve and infected with a defined inoculum of a single genotype. We did not detect phylogenetic clustering of HA and NA sequences based on vaccination status, type of administered vaccine, or pre-season HAI and NAI titers. These results are consistent with those of Dinis et al., who found no segregation of HA sequences based on vaccination status in a case test-negative study of vaccine effectiveness [23]. Together, these data suggest that vaccinated and unvaccinated individuals are infected with similar strains and that within season antigenic drift is not a major contributor to reduced vaccine efficacy in the seasons analyzed. Because we also stratified our analysis by pre-season HAI and NAI titer, we can similarly exclude viral escape from a non-sterilizing antibody response. Our data further suggest that pre-existing antibody against circulating strains does not apply sufficiently strong selective pressure to drive the emergence of antigenically distinct strains within a given host. High intrahost diversity may be an important factor in viral evolution, since it increases the number of novel variants on which natural selection can act. Some have proposed that the intrahost diversity of RNA viruses is linked to virulence [29,30], suggesting that processes that act to restrict or enhance intrahost diversity may alter disease phenotypes. We found that within host diversity of influenza virus was quite low. Most iSNV were present at frequencies of less than 0.1, which means they would only plausibly be spread between hosts if the transmission bottleneck were reasonably large [5]. The number of iSNV in a given host was similar between vaccine and placebo groups, both across the genome and on the segments coding for HA and NA. Furthermore, there were no significant differences in diversity when samples were grouped by HAI and NAI titer, and we did not find evidence of treatment-specific iSNV. Together our data suggest that vaccine-induced immunity does not significantly influence intrahost diversity and is a relatively weak selective pressure at the level of the individual host. While the number of iSNV was similar in all groups, we considered it possible that partial immunity may drive the emergence of specific antigenic variants. Vaccination and natural infection can induce a wide range of immune responses depending on host and viral factors. These can range from complete protection to leaky responses that allow infection, but influence disease severity or duration. Non-sterilizing immune responses have the potential to select for escape variants within each host. In the setting of current recommendations for universal vaccination, a highly vaccinated population could potentially select for antigenically evolved viruses more quickly than the spread of natural infection. For this reason, it is important to tease apart the role of vaccination on influenza evolution. We did find a number of variants in antigenic sites, but these were no more frequent than in other regions of the genome and did not vary with vaccination status or HAI and NAI titer. Importantly, our results are in conflict with earlier work in several model systems that demonstrate the rapid selection of antigenic variants in the presence of sub-neutralizing antibody or in experimentally infected animals [26,27,43–45]. This discrepancy may be due to differences in infectious dose, host genetic background, history of prior influenza infection, immune correlates not captured by HAI and NAI titers, or the strains tested. Animal studies are often performed in immunologically naïve or genetically identical animals, whereas humans have complex genetic backgrounds and immunological histories that could play a role in mediating population-wide immunity. While global patterns of influenza transmission and evolution are complex, our study provides important insights regarding influenza evolution on intra- and interhost scales. We were able to define aspects of intrahost evolution within a geographically-constrained and relatively young cohort with similar vaccination histories and previous influenza exposures across groups, limiting potential confounding factors. Still, we acknowledge that by reducing these confounders, we may be missing important determinants of intrahost evolution. For example, intrahost diversity may be different in children, older adults, or those with high-risk conditions. Our study included 5119 person-years of observations to yield a dataset of 165 viruses, 119 of which were sequenced. Large, placebo-controlled, randomized influenza vaccine trials involving thousands of people are unlikely to be conducted in the future due to the recommendation for universal vaccination. Therefore, our sample set likely represents the best chance to directly assess the impact of vaccination on influenza evolution in the context of natural human infection. Despite our well-controlled study, we found little evidence for vaccine-driven evolution in the context of the community that was sampled. None of the low frequency variants identified in the antigenic sites were found in subsequent seasons. While we cannot rule out the possibility that evolutionary patterns would be different in other geographic regions, different influenza seasons, or with other subtypes, we did not uncover differences between vaccinated and unvaccinated populations with respect to genome-wide or antigenic diversity in three H3N2 seasons. We were not able to evaluate the impact of vaccination on H1N1 or subtype B viruses. Furthermore, our sample set cannot discern whether there are differences in evolutionary pressure based on vaccine match/mismatch, as the vast majority of our samples were from antigenically matched seasons. A major limitation of our study was that only one sample was available for each participant, so we could not track changes in diversity or mutation accumulation within each individual over the course of their infection. However, we did not observe significant differences in the number of mutations over the first 6 days of infection, which is consistent with previous work in horses [27] and a deep sequencing study of 7 humans in an experimental challenge model [46]. Together with these works, our data suggest that within host dynamics are dominated by purifying selection with the transient appearance of minority variants and little sustained fixation. Selection and transmission of antigenically and epidemiologically important variants is likely to be a rare event when studied at this scale. Given that our data are derived from 165 incident infections in 5119 season-years of observation, detection of such events in the course of a natural infection would require an unrealistic sample size. We evaluated the potential for vaccine-induced immunity to drive intrahost evolution, as this is an important issue in light of the current recommendation for universal influenza vaccination. We did not find evidence for vaccine-induced pressure on the intrahost or consensus level, despite employing several methods used to study evolutionary processes. Our study is larger than other reports and involves extensive analysis of both placebo and vaccination groups. While randomization, placebo control, and reliance on a young healthy population allowed us to address this question in a rigorous manner, these factors may have lessened person-to-person variation. By better defining the intrahost evolutionary mechanisms and how they impact population-wide influenza evolution, we hope to address questions of import to clinicians and public health workers, improve vaccine design, and develop more efficient epidemiological control measures. This study was approved by the Institutional Review Board of the University of Michigan Medical School, and all human subjects provided informed consent. We characterized host-derived influenza populations archived from a randomized, double-blind, placebo-controlled, clinical trial of influenza vaccine efficacy that ran from the 2004–2005 through the 2007–2008 influenza seasons (ClinicalTrials.gov number, NCT00133523, [32–34]). Each year, healthy adults, ages 18–49, were randomized to receive trivalent inactivated influenza vaccine (IIV), live attenuated influenza vaccine (LAIV), or placebo. Throat swab specimens were collected from individuals with influenza-like illness within 7 days of onset; residual specimen material was stored in veal infusion broth (VIB) at -80°C. Viral RNA was extracted from 140μl of VIB using the QIAamp viral RNA mini kit (Qiagen 52906), eluted in 50μl buffer, and stored at -80°C. Hemagglutination inhibition (HAI) and Neuraminidase agglutination inhibition (NAI) titers for subjects in this study were previously measured and reported in [20,47]. Quantitative reverse transcription polymerase chain reaction (RT-qPCR) was performed on 5μl RNA from each sample using CDC RT-PCR primers InfA Forward, InfA Reverse, and InfA probe, which bind to a portion of the influenza M gene (CDC protocol, 28 April 2009). Each reaction contained 5.4μl nuclease-free water, 0.5μl each primer/probe, 0.5μl SuperScript III RT/Platinum Taq mix (Invitrogen 111732) 12.5μl PCR Master Mix, 0.1μl ROX, 5μl RNA. The PCR master mix was thawed and stored at 4°C, 24 hours before reaction set-up. A standard curve relating copy number to Ct values was generated based on 10-fold dilutions of a control plasmid run in duplicate. We amplified cDNA corresponding to all 8 genomic segments from 3μl of the viral RNA using the SuperScript III One-Step RT-PCR Platinum Taq HiFi Kit (Invitrogen 12574). Reactions consisted of 0.5μl Superscript III Platinum Taq Mix, 12.5μl 2x reaction buffer, 8μl DEPC water, and 0.2μl of 10μM Uni12/Inf1, 0.3μl of 10μM Uni12/Inf3, and 0.5μl of 10μM Uni13/Inf1 universal influenza A primers [48]. The thermocycler protocol was: 42°C for 60 min then 94°C for 2 min then 5 cycles of 94°C for 30 sec, 44°C for 30 sec, 68°C for 3 min, then 28 cycles of 94°C for 30 sec, 57°C for 30 sec, 68°C for 3 min. Amplification of all 8 segments was confirmed by gel electrophoresis, and 750ng of each cDNA mixture were sheared to an average size of 300 to 400bp using a Covaris S220 focused ultrasonicator. Sequencing libraries were prepared using the NEBNext Ultra DNA library prep kit (NEB E7370L), Agencourt AMPure XP beads (Beckman Coulter A63881), and NEBNext multiplex oligonucleotides for Illumina (NEB E7600S). The final concentration of each barcoded library was determined by Quanti PicoGreen dsDNA quantification (ThermoFisher Scientific), and equal nanomolar concentrations were pooled. Residual primer dimers were removed by gel isolation of a 300-500bp band, which was purified using a GeneJet Gel Extraction Kit (ThermoFisher Scientific). Purified library pools were sequenced on an Illumina HiSeq 2500 with 2x125 nucleotide paired end reads. All raw sequence data have been deposited at the NCBI sequence read archive (BioProject submission ID: SUB1907046) Sequencing reads that passed standard Illumina quality control filters were binned by index and aligned to the reference genome using Bowtie [49]. Single nucleotide variants (SNV) were identified and analyzed using DeepSNV [50], which relies on a clonal control to estimate the local error rate within a given sequence context and to identify strand bias in base calling. The clonal control was a library prepared in an identical fashion from 8 plasmids containing the genome for the respective circulating reference strain and sequenced in the same flow cell to control for batch effects. True positive SNV were identified from the raw output tables by applying the following filtering criteria in R: (i) Bonferonni corrected p value <0.01, (ii) average MapQ score on variant reads >30, (iii) average phred score on variant positions >35, (iv) average position of variant call on a read >32 and <94, (v) variant frequency >0.01. We only considered SNV identified in a single RT-PCR reaction and sequencing library for samples with copy number ≥105 genomes/μl transport media or in two separate RT-PCR reactions and sequencing libraries for samples with copy number 103−105 genomes per μl. For variants at a frequency of 50–99%, we called the minority base (i.e. 1–50%) as an iSNV. Our strategy for variant calling is described in [6] and all code can be found at https://github.com/lauringlab/variant_pipeline. Consensus nucleotide sequences for the HA and NA proteins were aligned using MUSCLE [51]. The best-fit models for nucleotide substitution was identified using jModelTest v2.1.10 [52]. Maximum likelihood phylogenetic trees were generated using RAxML v8 [53] with a GTRGAMMA model, Genbank sequences for vaccine strains as outgroups, and 1000 bootstraps. Trees were visualized and annotated using FigTree (v1.4.2). All statistical analyses were performed using Prism 6 and R. Description of the analysis and annotated code are available at https://github.com/lauringlab/fluvacs_paper. HA structural models were generated and visualized with PyMol.

10.1371/journal.pcbi.1003987

How Anacetrapib Inhibits the Activity of the Cholesteryl Ester Transfer Protein? Perspective through Atomistic Simulations

Cholesteryl ester transfer protein (CETP) mediates the reciprocal transfer of neutral lipids (cholesteryl esters, triglycerides) and phospholipids between different lipoprotein fractions in human blood plasma. A novel molecular agent known as anacetrapib has been shown to inhibit CETP activity and thereby raise high density lipoprotein (HDL)-cholesterol and decrease low density lipoprotein (LDL)-cholesterol, thus rendering CETP inhibition an attractive target to prevent and treat the development of various cardiovascular diseases. Our objective in this work is to use atomistic molecular dynamics simulations to shed light on the inhibitory mechanism of anacetrapib and unlock the interactions between the drug and CETP. The results show an evident affinity of anacetrapib towards the concave surface of CETP, and especially towards the region of the N-terminal tunnel opening. The primary binding site of anacetrapib turns out to reside in the tunnel inside CETP, near the residues surrounding the N-terminal opening. Free energy calculations show that when anacetrapib resides in this area, it hinders the ability of cholesteryl ester to diffuse out from CETP. The simulations further bring out the ability of anacetrapib to regulate the structure-function relationships of phospholipids and helix X, the latter representing the structural region of CETP important to the process of neutral lipid exchange with lipoproteins. Altogether, the simulations propose CETP inhibition to be realized when anacetrapib is transferred into the lipid binding pocket. The novel insight gained in this study has potential use in the development of new molecular agents capable of preventing the progression of cardiovascular diseases.

Cardiovascular disease is a leading cause of morbidity and mortality in Western societies. One of the most encouraging treatment methods to prevent the generation and progression of cardiovascular disease is the elevation of high density lipoprotein (HDL) levels in circulation, as high HDL levels have been found to correlate negatively with the risk of cardiovascular disease. HDL elevation is attainable through inhibition of cholesteryl ester transfer protein (CETP). A novel molecular agent, anacetrapib, fulfills the requirements with an acceptable side-effect profile. In this study, our objective is to gain more detailed information regarding the interactions between CETP and anacetrapib in order to unlock the inhibitory mechanism of the drug that has, to date, remained unclear. Our results point out the primary binding site of anacetrapib and highlight the ability of the drug to regulate the structure-function relationship of those structural regions of CETP that are considered important in CETP inhibition. Our findings could be exploited in the development of new and more efficient molecular agents against cardiovascular disease.

Cholesteryl ester transfer protein (CETP) is a 476-residue-long hydrophobic glycoprotein that transports cholesteryl esters (CEs), triglycerides, and phospholipids between high density lipoprotein (HDL) and other lipoprotein fractions in human blood plasma [1]. To be more specific, CETP exchanges CEs of HDL particles to triglycerides of very low density lipoproteins (VLDL) and low density lipoproteins (LDL), thus increasing the amount of triglycerides in HDL, leading to more rapid catabolism of HDL particles. CETP (Figure 1) carries CEs within a 6-nm-long hydrophobic tunnel that traverses the core of the molecule [2]. The tunnel has two distinct openings, and in the crystal structure [2] both of them are plugged by a dioleoylphosphatidylcholine (DOPC) molecule (Figure 1A). The lipid exchange mechanism of CETP is poorly understood. One plausible mechanism is the so-called shuttle mechanism [1], [3], in which CETP binds only one lipoprotein at a time. CETP attaches to the surface of a lipoprotein via its concave surface where also the two tunnel openings reside [2], [4]. The openings are expected to serve as passages to the flow of neutral lipids (CEs and triglycerides) between the particles, and their location supports the view that the concave surface is the only site able to bind lipoproteins, since other surfaces of the protein lack direct access to the tunnel. Further, the inherent curvature of CETP matches well with the curvature of HDL particles that may result from the fact that a major part of CETP has been shown to be associated with HDL due to higher binding affinity compared with plasma LDL or VLDL [1]. However, the molecular details driving the diffusion of lipids into and out from CETP require further elucidation. Previous experimental studies indicate that helix X located at the C-terminal domain of CETP is detrimental for the neutral lipid exchange, but not for the exchange of phospholipids [5], [6]. Helix X has been proposed to act as a lid conducting the exchange of lipids by alternating its open and closed states [2], [4]. In a recent molecular dynamics simulation study it was shown that after the attachment of CETP to lipoprotein surface, helix X is able to fold into the hydrophobic tunnel and interact with the CETP-bound CE [4]. After the lipids have been exchanged, the tunnel openings are plugged by phospholipids followed by the detachment of CETP from the lipoprotein surface. Meanwhile, in addition to the shuttle mechanism, another transportation mechanism has also been suggested. Here, CETP forms a ternary complex with two lipoprotein particles, and lipids somehow diffuse from one lipoprotein to another through the hydrophobic tunnel [7]. The interest towards CETP and its lipid transfer functions came to the forefront after notable associations between decreased CETP activity, decreased LDL-cholesterol level, increased HDL-cholesterol level, and resistance to atherosclerosis [8]. Atherosclerosis is a leading cause of morbidity and mortality in Western societies, and several clinical trials have shown that low HDL levels correlate with the risk of atherosclerosis due to antiatherogenic properties of HDL to remove cholesterol from atherosclerotic plaques back into liver to be recycled or secreted into bile [9], [10]. Pharmacological CETP inhibition has therefore emerged as a prime target to modulate HDL levels, with an objective to become a potential treatment strategy for preventing various cardiovascular diseases. Unfortunately, the road of CETP inhibition has had a turbulent beginning due to the failure of its two first compounds, torcetrapib and dalcetrapib. Torcetrapib increased the level of HDL as was hoped, but additionally it also increased the blood pressure as well as the mortality rate [11]. Due to these reasons, all clinical trials concerning torcetrapib had to be terminated. In the case of dalcetrapib, the trials were halted for futility as only modest increases in HDL levels were reached [12]. However, subsequent studies indicate that CETP remains a valid target since the lethal side effects seen with torcetrapib are unrelated with CETP inhibition and may not be shared by the other members of CETP inhibitors. A real breakthrough was reached at the end of 2010 when a new variant of the drug, anacetrapib (Figure 1B), was found to inhibit CETP with an acceptable side-effect profile [13]. In addition to anacetrapib, a competing inhibitor currently under phase III clinical development, evacetrapib, has already been shown to dose-dependently inhibit CETP and to increase HDL without severe cardiovascular events [14]. Despite the promising start achieved in the anacetrapib-based CETP inhibition, the precise inhibitory mechanism behind the agent is yet to be proven, regardless of the extensive research efforts its clarification has demanded. Classically, drugs inhibit or promote the functions of an enzyme or a receptor by binding e.g. to the active site thus blocking the binding of a ligand. However, in the case of CETP inhibitors, it has been demonstrated that they promote the formation of CETP-HDL complex, indicating that the current inhibitors do not only compete with CEs and triglycerides in CETP-binding but, additionally, they also hamper the detachment of CETP from HDL surface [15], [16]. The reported ability of anacetrapib, torcetrapib and dalcetrapib to increase the binding affinity between CETP and HDL, and hereby to promote the inhibition of lipid exchange between the particles, does not clarify the actual mechanism of action nor the possible structural changes required. However, explanations concerning especially the role of phospholipids and helix X in this regard have been proposed but the details at an atomistic level are largely unknown, and thus the understanding of the molecular basis of inhibition is quite limited. Recently, the X-ray structures of CETP were published with bound torcetrapib and one of its analogs [17]. Surprisingly, the structure showed that torcetrapib is able to bind CETP even if both CEs are bound to CETP. However, the C-terminal phospholipid was not present in the structure suggesting that torcetrapib abolishes the binding of a phospholipid to the C-terminal tunnel opening. This, again, could make the detachment of CETP from the surface of a lipoprotein more unfavorable, thus stabilizing the CETP-HDL complex. It remains to be seen if this holds for anacetrapib, since it has a similar chemical structure compared with torcetrapib [18]. Another mechanism that has been suggested to stabilize the CETP-lipoprotein complex is the ability of anacetrapib to act as glue between the particles [16]. The above findings and suggestions are insightful and encouraging, but call for better understanding of the inhibitory mechanism of anacetrapib, as well as of the lipid transfer functions related with CETP inhibition. In this regard, our objective is to perform atomistic molecular dynamics simulations to obtain detailed information linking the two processes. Here we have studied the interactions between anacetrapib and CETP with different lipid compositions to gather novel information regarding the inhibition of CETP. Previous studies of HDL-like lipid droplets [4], [19], HDL [20], and LDL [21] have shown that atomistic and coarse-grained simulations of lipoproteins and related transfer proteins can provide significant insight into their nanoscale properties and the mechanisms associated with lipid transfer. Understanding the lipid transfer functions of the protein, as well as the mechanisms behind CETP inhibitors, are important to realize in order to develop safe and efficacious treatment methods for the pharmacological raising of HDL cholesterol levels. We find that anacetrapib has a strong affinity for the region of the N-terminal tunnel opening. The primary binding site of anacetrapib turns out to reside in the tunnel inside CETP, near the residues surrounding the N-terminal opening. Free energy calculations show that when anacetrapib resides in this area, it hinders the ability of CE to diffuse out from CETP. The simulations further bring out the ability of anacetrapib to regulate the structure-function relationships of phospholipids and helix X, the latter representing the structural region of CETP important to the process of neutral lipid exchange with lipoproteins. Altogether, the simulations propose CETP inhibition to be realized when anacetrapib is transferred into the lipid binding pocket. The present study serves as a solid foundation for future studies concerning interactions between anacetrapib and CETP-lipoprotein complexes. To find the most probable sites from the crystal structure of CETP where anacetrapib would favor to attach, we used the energy-minimized and flexible structure of anacetrapib together with the crystal structure of CETP. Here, the aim was not to identify the possible binding poses of anacetrapib, but rather to show the most probable binding sites of the drug molecule. The results are illustrated in Figure 1C and show that, according to the docking calculations, the most favorable binding site for anacetrapib resides in the hydrophobic tunnel of the protein. For ligands colored with red, brown, cyan, and green, the respective binding energies are −47.7 kJ mol−1, −46.4 kJ mol−1, −48.5 kJ mol−1, and −46.9 kJ mol−1. In order to further validate the binding site of anacetrapib, the recently published X-ray structure of CETP with bound torcetrapib [17] was matched with the most probable binding site of anacetrapib gained from the docking calculations, see Figure 1D. The binding sites are in good accordance with each other, although CEs were absent from our calculations. In conclusion, docking calculations suggest that anacetrapib could either compete with CETP-bound CEs or phospholipids in binding, or lock them to reside more tightly inside CETP. Below we discuss this topic in more detail based on the free energy calculations that we have carried out to unlock this issue. We carried out ten 20 ns atomistic simulations for fully hydrated systems containing CETP with anacetrapib placed outside but in a close proximity with the protein (S1-helix, S2-1nm, S3-2nm, S4-3nm, S5-4nm, S6-convex, S7-1N, S8-2N, S9-1C, S10-3C; see Table 1 and Materials and Methods) to study the self-assembly process as well as the interactions between anacetrapib and the concave surface of CETP. The initial configurations were constructed to reflect random initial conditions to better correspond to the biological environment of the particles. We also performed five 200 ns simulations involving CETP with different interior lipid and anacetrapib compositions (L1, L2, L3, L4, L5; see Table 1 and Figure 2) to investigate the effects of anacetrapib to the conformational properties of CETP and bound phospholipids. The simulations without anacetrapib (L1, L3) served as control simulations to enable a more elaborate specification of the structural changes induced by the drug. In addition to the molecular dynamics simulations, we conducted eight umbrella sampling (free energy) simulations where both anacetrapib (U1, U2, U3, U4; see Table 1 and Figure 2) and the N-terminal CE (U5, U6, U7, U8; see Table 1 and Figure 2) were pulled out from the hydrophobic tunnel of CETP through the N-terminal opening. The purpose was to determine whether the anacetrapib located inside the tunnel has an influence on the diffusion of CE out from CETP, and to see the effect of helix X in this process. Root mean square deviation (RMSD) profiles of 20 ns simulations indicate that the structures do not deviate considerably from the corresponding X-ray structure (Figure 3). The radii of gyration fluctuated between 3.33 and 3.54 nm in each simulation (Figure 3). The spatial density maps illustrated in Figure S1 (see Supporting Information (SI)) reveal a disordered motion of the drug molecule around CETP. The distance between the particles ranged from 1 to 2 nm in the simulations S1-helix, S2-1nm, S3-2nm, S7-1N, S8-2N, and S9-1C, and from 3 to 4 nm in the simulations S4-3nm, S5-4nm, S6-convex, and S10-3C. A change in the distance from 2 to 3 nm weakens considerably the interactions between the CETP and anacetrapib and, as a consequence, anacetrapib experiences random movement due to thermal fluctuations. These findings are supported by the interaction energies calculated between the particles (Table 2, Figure 4). Table 2 presents the interaction energies averaged over the course of the simulations, whereas Figure 4 depicts the energies as a function of simulation time. It is apparent that the main force to drive the binding of anacetrapib to CETP in S1-helix, S2-1nm, S3-2nm, S7-1N, S8-2N, and S9-1C is the weak van der Waals interactions. In the remaining simulations both the van der Waals electrostatic interaction forces are substantially weaker. It is worth to notice that in S1-helix, S2-1nm, S3-2nm, S7-1N, S8-2N, and S9-1C, the movement of anacetrapib is highly similar, since the drug moves at the N-terminal tunnel opening mainly around residues Arg197, Ser431, Lysh432, Gly433, Ser435, and Hisb462, which is in good agreement with our results based on molecular docking, concerning especially the binding site of the red ligand (Figure 1C, Figure S1). The visualization of the simulation trajectories also revealed that while the drug moved at the N-terminal opening, the trifluoromethyl- and methyl groups of the drug oriented close to each other (Figure 1E). This indicates that the drug aligned itself to a tighter conformation suggesting a possible movement into the tunnel. The observed finding is supported by the interaction energies calculated from the simulations where anacetrapib was transferred into the hydrophobic tunnel of CETP (L2, L4, L5; Table 2, Figure 4). The strength of van der Waals interactions between the particles is two to two-and-a-half times stronger inside the tunnel than at the N-terminal opening, or at the N- and C-terminal ends of the protein. This strongly suggests that CETP inhibition is enabled when the drug enters the protein, and the drug interacts with the structural regions of CETP important to lipid exchange from the hydrophobic cavity. The proposed movement is further supported by the lack of a significant number of hydrogen bonds between the protein and the drug. Anacetrapib was noticed to form separate hydrogen bonds with CETP while moving at the N-terminal tunnel opening, implicating that when a new bond was formed, the previous one was broken. This implies a weak attachment to CETP, and thus an unobstructed movement of the drug. The present findings based on a self-assembly process highlight the importance of thermal diffusion together with electrostatic and van der Waals interactions during the formation of CETP-inhibitor complexes. Thermal motion predominates for distances above 3 nm between anacetrapib and the tunnel opening of CETP, and only for shorter scales the direct interactions between the molecules become strong enough to drive the complex formation process. Furthermore, the affinity of anacetrapib towards the concave surface of CETP is evident when the distance between the particles is taken into consideration. On the basis of the above findings, we propose that the primary binding site of anacetrapib resides inside the hydrophobic tunnel of the protein, near the residues surrounding the N-terminal tunnel opening, including helix X. Computation of free energy profiles is a major computational challenge in general. The case considered here is not an exception. To get the free energy profiles shown in Figure 5 through umbrella sampling simulations required substantial computer resources (see Materials and Methods). Despite this, the obtained profiles are not fully converged. This issue typically arises from inadequate sampling of regions of conformational space that are likely separated by a large barrier. While the data presented here represents largely the state of the art, we yet wish to stress that the best we can extract from the data are suggestive trends. The obtained free energy profiles (Figure 5) indicate strong attachment between anacetrapib and CETP. They further point towards the possibly hindered ability of the CE molecule to diffuse from CETP to the water phase when both the drug molecule and two CE lipids reside in the hydrophobic tunnel (Figure 5). The corresponding free energy barriers range up to 65 kJ mol−1 for anacetrapib, and between 184 and 197 kJ mol−1 for CE (when anacetrapib blocks the path). The free energy barriers found here, as CE is pulled out from CETP, are profoundly high. This results from the transition path that takes CE completely into the water phase. Importantly, Figure 5 also suggests that the increase in free energy is rather modest at short distances when CE is still inside the hydrophobic environment of CETP. Consequently, the free energy barriers are expected to become lower when CETP binds to a lipoprotein surface, hence facilitating the diffusion and exchange of lipids between the particles. It is worth to notice that the presence of helix X seems to have an influence on both the binding strength of anacetrapib and the diffusion ability of CE, since both molecules are pulled out from the hydrophobic tunnel more easily when helix X is removed from the structure of CETP. Additionally, when also the drug molecule is removed, the free energy barrier of CE movement towards the water phase appears to be the lowest, 153 kJ mol−1. The results point towards the important role of helix X in assisting the lipid exchange, and highlight the possible inhibitory mechanism of anacetrapib as the movement of CE outside from CETP could be hindered in the presence of the drug molecule. There are two thermodynamic cycles for binding and unbinding both anacetrapib and CE present in the free energy calculations, namely, the one with the absence of helix X from the structure of CETP (U1/U6/U2/U5) and the other with the presence of helix X in the structure of the protein (U3/U8/U4/U7). With the current data, these cycles result in values of −28±20 kJ mol−1 for the first and −30±20 kJ mol−1 for the latter cycle, where one is tempted to expect for a value of zero. However, the binding site of anacetrapib is not identical in systems where the hydrophobic tunnel is empty (U1, U2) or contains two CEs (U5, U6). To be more specific, the binding site of anacetrapib resides deeper in the hydrophobic tunnel of CETP as the tunnel does not include lipids, while in the presence of these lipids the binding site of the drug molecule resides closer to the N-terminal tunnel opening. Hence, with the present systems, it is not possible to obtain a thermodynamic cycle that would result in a value of zero, and the values of the thermodynamic cycles given above therefore partly stem from this fact, and partly from inadequate sampling. Given the considerable computing resource to generate the data in Figure 5 (see Materials and Methods), we consider it quite unfeasible to reduce the systematic error, but this is not an issue since the main conclusions that can be drawn from Figure 5 are evident based on the data shown. RMSD profiles of atomistic 200 ns simulations indicate that the protein structures do not deviate considerably from the corresponding X-ray structures (Figure 6). The RMSD values of DOPCs are also plotted in Figure 6, and the profiles depict increased conformational alterations of DOPCs when the drug molecule is transferred into the hydrophobic tunnel of CETP: the RMSD fluctuates between 0.27 and 0.47 nm without the drug (L1, L3), and between 0.26 and 0.61 nm with the drug (L2). This finding is further supported by the atomic RMS fluctuations and spatial density maps calculated for DOPCs (Figure 7). Maps confirm that DOPCs, especially their head groups and sn-2 chains, experience considerable wobbling in the presence of anacetrapib (L2) compared with the absence of the drug (L1). However, there is a contradiction when comparing the RMSD and RMSF curves of DOPC in simulations L1 and L3. The RMSD appears more stable in L3 (Figure 6), but the RMSF is more stable in L1 (Figure 7). A reason for this could be that DOPC shifts from its initial position in L1 and is then stabilized there. Nonetheless, the results are in good accordance with the interaction energies calculated between CETP and the associated lipids (DOPCs and CEs), as well as between the lipids and anacetrapib (Table 3). As Table 3 illustrates, the strength of interactions between CETP and DOPCs are the strongest in the absence of anacetrapib (L1). Furthermore, as the interactions between DOPCs and CEs (L3) seem to be stronger than between the lipids and anacetrapib (L2, L5), it is evident that anacetrapib induces high fluctuations to phospholipids. This provides compelling evidence that the drug interacts with phospholipids, and, as a consequence, could hinder the binding of DOPCs to the tunnel openings, which could play a role in the stabilization of the CETP-lipoprotein complex. As the interaction energies between DOPC and CE molecules imply (Table 3), the fluctuations of the same order of magnitude were observed also when two CEs filled the tunnel (Figures 6 and 7). The spatial density maps reveal a high similarity for the trajectories of DOPCs regardless the type of the tunnel-filling particle. The observed movement could be caused by the residence of neutral lipids inside the hydrophobic tunnel of CETP, since in order for CEs to properly accommodate the cavity, a conformational rearrangement of DOPCs would be required. Nonetheless, the wobbling of phospholipids with CETP-bound CE molecules indicates the structure of CETP to be rather unstable during the transportation of neutral lipids. This could highlight the importance of helix X needed to prevent the structure of the protein from collapsing, as was suggested also previously [4]. The RMS fluctuations of CETP backbone were analyzed in order to find the regions that fluctuated the most during the simulations. This method of analysis can give valuable information regarding the functioning of a protein by highlighting the regions of protein backbone with low and high mobility. First, for comparison, CETP has previously been reported to have mobile structures with elevated B-factors near tunnel openings including the hinge region of helix X (Gly458-Pro460), and in the N- and C-terminal ends including the loop regions represented with omegas one and two [2], [4]. As expected, these regions showed high mobility also in our simulations (Figure 8), with the conformational fluctuation of helix X peaking near the residue 462. In addition, four other regions in the backbone of the protein were found to fluctuate highly during each simulation. These regions were Ω3 (residues 380–400), Ω4 (residues 40–50), Ω5 (residues 90–110) and Ω6 (residues 150–170) which were also earlier shown to have high mobility [4]. All these regions are found in the loops and therefore the high fluctuations can be expected. The results imply that the structure of CETP is elastic, facilitating the binding to lipoprotein surfaces with varying curvatures. In addition, the observed flexibility of the hinge region of helix X (Glu 461-Ser472) suggests that helix X could play a crucial role in assisting the lipid exchange process. During the lipid exchange process helix X may partly move into the N-terminal CE-binding pocket of CETP to facilitate the export of CE out from CETP [4]. Another possibility is that helix X moves aside from the N-terminal tunnel opening, thus generating a wider pathway that facilitates the diffusion of CE out from CETP [22]. The above described free-energy calculations point to this direction. Interestingly, it became apparent based on DSSP calculations that the secondary structure of the helix encountered notable fluctuations between turn (unfolding of the helix) and 310-helix (extension of the helix) when the drug molecule interacted with the concave surface of CETP (S1-helix, S2-1nm, S3-2nm, S7-1N, S8-2N, S9-1C), and when either CEs (L3, L5) or anacetrapibs (L2, L4) were present in the hydrophobic tunnel (Figure 9, Table 4). For comparison, helix X maintained its α-helical form during the simulation L1 where the hydrophobic tunnel was empty (Table 4). The visual inspection of the simulation trajectories revealed that in L1, the N-terminal DOPC maintained 1 nm distance from helix X over the course of simulation, while in L2 these two structures oriented themselves close to each other at the time when helix X experienced conformational fluctuations. The results imply that anacetrapib induces conformational alterations to the helix, and hence affects its stability, by interacting with the N-terminal DOPC. This, in turn, could indicate drastic effects on the lipid transfer functions of CETP. On the basis of earlier clinical trials, both anacetrapib and the flawed torcetrapib were shown to increase the binding affinity of CETP towards lipoproteins, especially towards HDL [1], [15], [19]. They induced a tight reversible binding on the lipoprotein surface stabilizing the HDL-CETP complex, and hereby preventing the capability of CETP to transport neutral lipids between different lipoprotein fractions. Despite the appealing start achieved in the anacetrapib-based CETP inhibition, the actual inhibitory mechanism of the drug remained unknown. In the present study, our objective was to reveal the mechanism of action behind anacetrapib, shed light on its ability to inhibit CETP-mediated lipid transfer, and to unravel the dynamics of related processes. The results showed an evident affinity of anacetrapib towards the concave surface of CETP, especially towards the N-terminal tunnel opening where also helix X resides, and highlighted the importance of electrostatic and van der Waals interactions during the process once the drug was able to migrate to a close enough distance from the tunnel opening. However, the distance between the particles should be taken into consideration in the complex formation, since with too large distances (above about 3 nm) the movement of the drug was noticed to be dominated by thermal motion, eventually resulting in disordered motion. Hence, the question is how the affinity between the particles could be ensured in order to secure the interactions and speed up the complex formation process, as otherwise anacetrapib may experience random motion and may not be suitable for CETP inhibition purposes. For comparison, the formation of CETP-lipoprotein complex has been reported to be modulated by pH, surface pressure, and the introduction of positive divalent ions, such as Ca2+ and Mn2+, into the solution [23], [24]. In this spirit it is justified to assume that at least ion mediated interactions could foster the complex formation process, as electrostatics in terms of charged centers is an integral factor in both molecules but in the present case that we simulated in the absence of salt the screening effects were quite strong, resulting in predominance of thermal motion at long distances. In addition to the observed affinity, anacetrapib was noticed to align itself to a tighter conformation while moving near the N-terminal opening. This finding together with the evidence of stronger interactions prevailing between the particles when the drug was transferred inside the hydrophobic cavity, indicate the primary binding site of anacetrapib to reside in the tunnel, particularly near the residues surrounding the N-terminal opening, including helix X. Hence, we propose CETP inhibition to be realized when the drug is transferred into the lipid binding pocket. The regulatory role of helix X has been identified to play an important role in the lipid exchange process, since it has been suggested to act as a lid that conducts the exchange by alternating its open and closed states [4]. The structure of helix X has been proposed to undergo conformational changes during lipoprotein binding by moving aside from the N-terminal tunnel opening through an oblique penetration into the monolayer [22], or by rearranging and becoming buried inside the hydrophobic pocket [4]. Thus helix X is proposed to be locked in the “open” state for the time of lipid exchange. Considering the detachment as well as the transportation of lipids, helix X is needed to shield the corresponding tunnel opening to make CETP more compatible with the aqueous environment [2], hence the nomination “closed” state. The inhibitory mechanism of anacetrapib has been speculated to be in connection specifically with this regulatory property of helix X. One proposition suggests that CETP-bound anacetrapib alters the conformation of helix X to favor the open state, thus stabilizing the HDL-CETP –complex [4]. The above described findings disclose the flexible nature of helix X (the hinge region) that is essential in assisting the exchange of lipids. Our results are in agreement with these observations, as in the present simulations the hinge region of helix X was noticed to have elevated B-factors and, additionally, the secondary structure of the helix was shown to experience fluctuations between turn and 310-helix while anacetrapib moved near the residues that surround the N-terminal tunnel opening. The results provide compelling evidence about the ability of anacetrapib to induce conformational changes to helix X in order to achieve the needed flexibility. How helix X behaves in the presence of the entire HDL-CETP-anacetrapib complex is a question left for additional simulations to be resolved. Another crucial component in the process of neutral lipid exchange are the phospholipids due to their central role both in the binding and detachment of CETP from lipoprotein surfaces. During binding, phospholipids have been proposed to merge into the monolayer followed by a migration away from the tunnel openings to their edges [2], [4]. The molecular simulation data is strongly in favor of this view [4]. This process induces the formation of a hydrophobic pathway under the concave surface of the protein, which permits the access of lipids into the tunnel. Considering the detachment, the tunnel openings will need to be refilled with phospholipids before the dissociation since otherwise the protein would not be able to return to aqueous environment, or at least it would be much more unfavorable [2]. Hence, changes in the structure of phospholipids could possibly hinder their binding to CETP and the dissociation of CETP from lipoprotein surfaces, further resulting in a weaker capability of the protein to transport neutral lipids. Our results pointed to this direction, since phospholipids were noticed to experience increased structural fluctuations, in addition to the declined electrostatic and van der Waals interactions with CETP when anacetrapib was transferred into the hydrophobic tunnel of the protein. The corresponding interactions were stronger between phospholipids and CE molecules, suggesting the capability of the drug to destabilize the binding of phospholipids to CETP. This view is in accordance with the crystal structure of CETP published in complex with torcetrapib [17]. The structure indicates that the binding of torcetrapib to CETP abolishes the binding of phospholipids to the N-terminal tunnel opening. It is possible that torcetrapib together with CE excludes enough volume inside the hydrophobic tunnel of CETP rendering the binding of a phospholipid to the corresponding tunnel opening impossible, as the hydrophobic acyl chains of the phospholipid can no longer be buried inside CETP. The primary binding site of anacetrapib to reside inside the hydrophobic tunnel is further supported by the free energy profiles that reveal strong attachment between the drug molecule and CETP, especially when two CEs fill the length of the tunnel as shown also previously [17]. When attached as described, anacetrapib hinders the ability of CE to diffuse out from the structure of CETP, thus pointing towards the possible inhibitory mechanism of the drug. The presence of helix X has a strong influence during this process, as both anacetrapib and CE were shown to move into the water phase more easily when the helix was removed from the structure of CETP. The results highlight the crucial role of helix X in assisting the lipid exchange during which helix X could possibly move aside from the N-terminal tunnel opening thus generating a wider pathway for CEs to diffuse out from CETP. It is reasonable that the hinge region of helix X enables the movement of the helix aside from the corresponding tunnel opening. In conclusion, our results show an evident affinity of anacetrapib towards the concave surface of CETP, especially towards the region of N-terminal tunnel opening. The primary binding site for the drug turns out to reside inside the hydrophobic tunnel, near the residues surrounding the N-terminal opening. When residing in this area, anacetrapib was shown to hinder the ability of CE to diffuse out from the structure of CETP. Additionally, the results point towards the encouraging capability of anacetrapib to influence the molecular interactions between phospholipids and helix X, both of which represent the structural regions of CETP important for lipid exchange between lipoproteins, thus giving support for the competency of pharmacological CETP inhibition. The view presented in this article paves the way for extending the scope of computational studies to gain a deeper understanding concerning the pharmacological ways to inhibit CETP and to modulate HDL levels. In this regard, simulations concerning the interactions between HDL-CETP-inhibitor –complex are ongoing (work in progress). The novel understanding could be used in the development of new molecular agents in the fight against the progression of cardiovascular diseases. Here we consider systems where anacetrapib interacts with CETP, which is either empty or carries a number of lipids in its transfer pocket. First, anacetrapib (Figure 1B) is an orally active, potent, and selective agent identified by high-throughput screening to belong to the 1–3-oxazolidin-2-one series of CETP inhibitors developed by Merck [15]. The medicinal chemistry behind the discovery of anacetrapib is described in [25]. The coordinate file for CETP in the PDB format was acquired from the RCSB Protein Databank with an accession code 2OBD. In addition to the protein, the file provides information of the atomic positions of the lipids involved in CETP: there are two CEs located inside the hydrophobic tunnel of CETP, and two DOPCs that cover the two openings of the tunnel. A detailed explanation of the protein structure is given elsewhere [2]. For all molecules considered in this study (CETP, DOPC, CE, anacetrapib), we used the official distribution force field OPLS-AA [26]–[33]. In addition, an extension of OPLS was used for the long hydrocarbon tails of DOPC and CE [34]. Concerning the partial charges, for DOPC molecules they were derived in compliance with the OPLS methodology [35], while for anacetrapib they were fitted to the electrostatic potential by applying the RESP software [36]. Here, the Merz-Kollman molecular electrostatic potential (MEP) was computed for the optimized structure of anacetrapib [37]. The MEP calculations were performed by applying the Gaussian 09 package at the Hartree-Fock level by employing the 6-31G* basis set [38]. The charges were fitted automatically by the RESP and ESP charge derived (R.E.D.) software version III. The derived charges can be found from the topology file Dataset S1 (see SI). Water molecules were described with the TIP3P model since it is compatible with the parametrization of OPLS-AA [39]. Prior to molecular dynamics simulations, we used molecular docking calculations to determine the initial configurations for the simulated systems. The purpose of the calculations was not to identify the possible binding poses of anacetrapib, but rather to explore the most probable sites from the crystal structure of CETP where the drug would desire to attach in terms of the lowest binding energy. The constructed box covered the two tunnel openings and the hydrophobic tunnel of CETP. CETP-bound lipids were not present. A flexible anacetrapib molecule with 9 rotatable bonds was used in the calculations. In total, 1000 runs were carried out with default settings, namely, the maximum number of binding modes to generate/export was set to 9, and the maximum energy difference between the best ligand pose and other ligand poses was set to 12.6 kJ mol−1. Four conformations with the highest binding free energy for anacetrapib are shown in Figure 1C. For ligands colored with red, brown, cyan, and green, the respective binding free energies were found to be −47.7 kJ mol−1, −46.4 kJ mol−1, −48.5 kJ mol−1, and −46.9 kJ mol−1. More detailed technical information for the applied program, AutoDock Vina, can be found in [40]. Based on the docking data (see Results for details; Figure 1C), we constructed initial simulation systems as divided into three groups. The first group consisted of 10 systems with lipids removed from CETP, and anacetrapib placed outside the protein but in the vicinity of its lipid binding pocket, to characterize the self-assembly process as well as to elucidate the interactions between the drug and the concave surface of the protein. In the first simulation (S1-helix), anacetrapib was placed 1 nm away from helix X, whereas in the four following systems (S2-1nm, S3-2nm, S4-3nm, S5-4nm) the drug was placed 1, 2, 3, and 4 nm from the tunnel openings, respectively. The sixth simulation (S6-convex) included anacetrapib at a distance of 3 nm from the convex back of the protein. The remaining four simulations (S7-1N, S8-2N, S9-1C, S10-3C) involved the drug at 1 and 2 nm distances from the N-terminal end, as well as 1 and 3 nm distances from the C-terminal end of the protein, respectively. All of the simulations in the first group were simulated for 20 ns each, thus “S” here in the system name stands for “short”. The second group consisted of five systems with anacetrapib placed inside the lipid binding pocket to study the conformational changes of CETP induced by the drug. In the first simulation system (L1), the two CEs and anacetrapib were removed from CETP, thus only the two DOPC molecules remained inside the protein (Figure 2A). In the second simulation (L2, Figure 2B), two anacetrapib molecules were placed inside the empty hydrophobic tunnel based on the binding sites of cyan and green ligands presented in Figure 1C. Compared with L2, the drug molecules were replaced by two CEs in the third simulation (L3), the one being placed in the N-terminal domain and the other in the C-terminal domain of the protein (Figure 2C). Both L2 and L3 included DOPCs to plug the tunnel openings. All the lipids were removed from CETP in the fourth simulation (L4, Figure 2D), with one anacetrapib located in the empty tunnel between the binding sites determined by the cyan and green ligands (Figure 1C). In the fifth simulation (L5), the N-terminal DOPC was removed from CETP, with two CEs and one anacetrapib filling the length of the hydrophobic tunnel (Figure 2E). CEs were placed as in L3, while the location of anacetrapib was determined by the binding site of the cyan ligand (Figure 1C). Each of these systems, “L” standing for “long” ones, was simulated for 200 ns. The third group consisted of eight systems directed to umbrella sampling simulations. In the first four systems, one anacetrapib molecule was pulled out from CETP through the N-terminal tunnel opening. The interior lipid composition and the presence of helix X in the structure of CETP were varied between the systems. In the first two systems, helix X was removed and the tunnel was either empty (U1, Figure 2F) or contained two CEs (U2, Figure 2G). Helix X was returned to CETP in the following two simulations and the tunnel was both empty (U3, Figure 2H) and filled with two CEs (U4, Figure 2I). In all the remaining four systems, two CEs resided inside the hydrophobic tunnel, the N-terminal CE being the molecule pulled out from the structure of CETP with the varying presence of helix X and one anacetrapib. Both of these were absent in the fifth simulation (U5, Figure 2J), while in the sixth simulation (U6) anacetrapib was added inside the tunnel, near the N-terminal tunnel opening (Figure 2K). Helix X was returned to CETP in the last two simulations without (U7) and with (U8) one anacetrapib locating near the N-terminal tunnel opening (Figure 2L and 2M). The N-terminal DOPC molecule was removed in all eight systems, whereas the C-terminal DOPC plugged the corresponding tunnel opening. Each of these systems, “U” standing for umbrella sampling (see below), were first equilibrated for 100 ns, and then simulated for 60 ns. The systems belonging to the first group were solvated with ∼100,000–160,000 water molecules with counter ions, while in the second and third groups about 100,000 and ∼30,000–60,000 water molecules, respectively, were used for solvation. Altogether, the systems included ∼125,000–500,000 atoms. The number of water molecules used in the simulations depends on the size of the simulation box: the greater the distance between CETP and anacetrapib, the greater the size of the simulation box and, hence, the number of water molecules. The size of the simulation box originates from the requirement of periodic boundary conditions. Before any simulation was started, energy minimization was performed for each system using the steepest descent method with 500 steps [41]. Prior to conducting umbrella sampling simulations, eight pulling simulations were performed in order to generate a series of configurations that served as the starting configurations for umbrella sampling. In the pulling simulations, either anacetrapib or the N-terminal CE, depending on the system, was pulled out from the structure of CETP through the N-terminal tunnel opening. Residues Cysh9, Arg10, Ile11, and Thr12 were used as the reference pull group due to their parallel location with the N-terminal tunnel opening. The force constant applied either to the center of mass of anacetrapib or to the last carbon atom in the acyl chain of CE was 2000 kJ mol−1 nm−2 with a pull rate of 0.003 nm ns−1. After pulling simulations, umbrella sampling was conducted with a total of 25 and 47 umbrella windows when the pulled molecule was anacetrapib and CE, respectively. The umbrella windows were selected at 0.1 nm intervals from the original location of the molecules inside the hydrophobic tunnel towards the water phase. Each window was simulated for 160 ns where the first 100 ns were used for equilibration. These parameters were chosen based on a systematic study for increasing equilibration and simulation times. Altogether, the umbrella sampling simulations required about 600 core-years of computing time. In these simulations, both molecules were restrained to the middle of every umbrella window by a harmonic potential with the same force constant as used in the pulling simulations. The restraints were applied only along the reaction coordinate defined by the vector connecting the reference pull group and the pulled molecule, and the molecules were free to move in the other directions. In order to keep the reaction coordinate along the vector and to prevent CETP from rotating and tilting, position restraints were applied to a carbon atom of Ser115 and to a carbon atom Gly413. These atoms reside in the N- and C-terminal ends of the protein, respectively. The simulations were performed under NpT conditions (constant number of particles, pressure, and temperature) with the GROMACS software package using the version 4.5.4 for the first and second simulation groups and version 4.6.1 for the third simulation group [41]. The reference temperature for all simulated systems was 310 K, and each component of the systems was separately coupled to a temperature bath using the Nόse-Hoover coupling method with a time constant of 0.1 ps [42]. The Parrinello-Rahman barostat was applied to couple the pressure, with a coupling constant of 1 ps and a reference pressure of 1 bar [43]. A time step of 2 fs was used in all simulations, with the LINCS algorithm applied to constrain all the bonds in the system [44]. The van der Waals interactions were calculated up to a cutoff radius of 1 nm and the particle mesh Ewald technique was utilized for long-range Coulombic forces, with a real space cutoff of 1 nm [45], [46]. As mentioned above, the ten systems associated with the first simulation group (“S” standing for short) were simulated for 20 ns, the systems in the second group (“L” standing for long) for 200 ns, and the systems in the third group (“U” standing for umbrella sampling) for 160 ns where the first 100 ns were used for equilibration. DSSP (define secondary structure of proteins) [47] was applied to determine the most likely secondary structure of CETP as a function of time. DSSP was calculated by applying the do_dssp tool of GROMACS. The root mean square deviation (RMSD) was used to evaluate the deviation of the structure of the simulated system from the initial starting structure over the course of simulation. RMSD was calculated by applying the g_rms tool of GROMACS. The radius of gyration was used to measure the size and compactness of a molecule. It is defined as the mean square distance of each particle in the structure with respect to its center of mass. The radius of gyration was calculated by applying the g_gyrate tool of GROMACS. The root mean square fluctuation (RMSF) of atomic positions was used to discover and evaluate the most flexible regions of CETP. RMSFs were calculated by fitting the simulated structure to a reference structure followed by the calculation of the average distance deviation from the reference structure. Typically, the residues of a protein that fluctuate the most can be found in loop regions. For purposes of comparison with experimental data, the RMSFs can be converted into B-factor values. RMSF was calculated by applying the g_rmsf tool of GROMACS. Interaction energies between different molecules were calculated by applying the g_energy tool of GROMACS. The tool calculates the contributions of the energies automatically from the simulation trajectory. Weighted histogram analysis method (WHAM) is a standard technique to compute the potential of mean force (PMF) along the reaction coordinate for a molecule [48], [49]. It estimates the statistical uncertainty of the probability distribution obtained from the umbrella histograms and iteratively computes the PMF that corresponds to the smallest uncertainty in the form of global free energy of the molecule [49]. WHAM was calculated by applying the g_wham tool of GROMACS. Statistical uncertainty of the PMF can be estimated by applying the bootstrap analysis. The idea is to generate new hypothetical observations, that is, a bootstrapped trajectory for each umbrella histogram, thus yielding a new set of histograms for the corresponding umbrella window that are subsequently applied in WHAM to calculate a bootstrapped PMF. The process is repeated for each bootstrapped trajectory when a large number of bootstrapped PMFs can be obtained. The applied bootstrapped sample size was 200. Hydrogen bond formation has a remarkable role, e.g., in the stabilization of the secondary structure of a molecule. In this study the formation of hydrogen bonds is analyzed between CETP and anacetrapib in order to see where and how tightly anacetrapib binds. The formation of hydrogen bonds is analyzed between all possible donors D and acceptors A. OH and NH groups are regarded as donors while O is always an acceptor. The determination for the existence of a hydrogen bond is done by using a geometrical criterion (based on the distance between donor-acceptor rDA) and a criterion for the angle α (between acceptor-donor-hydrogen triplet αADH),(1)(2)Hydrogen bonds were calculated by applying the g_hbond tool of GROMACS.

10.1371/journal.pntd.0004671

Genetic Variation in Autophagy-Related Genes Influences the Risk and Phenotype of Buruli Ulcer

Buruli ulcer (BU) is a severe necrotizing human skin disease caused by Mycobacterium ulcerans. Clinically, presentation is a sum of these diverse pathogenic hits subjected to critical immune-regulatory mechanisms. Among them, autophagy has been demonstrated as a cellular process of critical importance. Since microtubules and dynein are affected by mycolactone, the critical pathogenic exotoxin produced by M. ulcerans, cytoskeleton-related changes might potentially impair the autophagic process and impact the risk and progression of infection. Genetic variants in the autophagy-related genes NOD2, PARK2 and ATG16L1 has been associated with susceptibility to mycobacterial diseases. Here, we investigated their association with BU risk, its severe phenotypes and its progression to an ulcerative form. Genetic variants were genotyped using KASPar chemistry in 208 BU patients (70.2% with an ulcerative form and 28% in severe WHO category 3 phenotype) and 300 healthy endemic controls. The rs1333955 SNP in PARK2 was significantly associated with increased susceptibility to BU [odds ratio (OR), 1.43; P = 0.05]. In addition, both the rs9302752 and rs2066842 SNPs in NOD2 gee significantly increased the predisposition of patients to develop category 3 (OR, 2.23; P = 0.02; and OR 12.7; P = 0.03, respectively, whereas the rs2241880 SNP in ATG16L1 was found to significantly protect patients from presenting the ulcer phenotype (OR, 0.35; P = 0.02). Our findings indicate that specific genetic variants in autophagy-related genes influence susceptibility to the development of BU and its progression to severe phenotypes.

Buruli ulcer (BU) is a neglected tropical disease caused by Mycobacterium ulcerans. Because the exact trigger is still under investigation, current treatment options rely mostly on the surgical excision of the affected site. There is therefore a pressing demand for improved risk prediction and tailored treatment as well as for new drug targets. By resorting to the largest case-control study reported to date, we show that genetic variation in the autophagy-related genes NOD2, PARK2 and ATG16L1 influence the risk and course of BU disease. Thus, our results provide crucial insights into the role of autophagy in the pathogenesis of BU.

Buruli ulcer (BU) is a severe necrotizing human skin disease caused by Mycobacterium ulcerans, representing the third most common mycobacteriosis worldwide [1]. At least 33 countries from Africa, South America and Western Pacific, with tropical, subtropical and temperate climates, have reported BU [1]. Moreover, in 2014, 2200 new cases were reported in 12 of those 33 countries [1]. BU initiates as a small, painless, raised skin papule, nodule, plaque or oedema. Later, destruction of the subcutaneous adipose tissue leads to collapse of the epidermis and formation of a characteristic ulcer with undermined edges [1]. Advanced lesions display massive tissue destruction induced by the action of the exotoxin mycolactone, a potent cytotoxic and immunosuppressive polyketide-derived macrolide released by M. ulcerans [2]. Clinically, presentation is a sum of these diverse pathogenic hits subjected to critical, mainly local, immune-regulatory mechanisms [3]. Among the many immunological mechanisms defining susceptibility to infection and its progression, autophagy has been demonstrated as a cellular process of critical importance to immunity to viral, bacterial and protozoan infections [4]. Autophagy is a regulated process contributing to the innate control of intracellular pathogens by triggering the autodigestion of cytoplasmic components and driving pathogen clearance. Autophagy is known to be dependent on microtubule cytoskeleton and dynein-driven transport, with dynein playing a role in the delivery of autophagosome contents to lysosomes during autophagosome-lysosome fusion [4]. Since microtubules and dynein are affected by mycolactone [5], cytoskeleton-related changes might potentially impair the autophagic process and impact the risk and progression of M. ulcerans infection. The function of specific components of the autophagic machinery, namely nucleotide-binding oligomerization domain-containing 2 (NOD2), E3 ubiquitin-protein ligase parkin (PARK2) and autophagy-related protein 16–1 (ATG16L1), has been associated with resistance to several intracellular pathogens, including M. tuberculosis [4]. Based on reports linking variants in these genes with defective activation of autophagy as well as our own data proposing a central role for autophagy in the intracellular control of M. ulcerans infection through mycolactone-induced impairment of cytoskeleton-dependent cellular functions [5], we designed a case-control genetic association study involving 208 prospectively collected cases of BU to dissect the contribution of selected autophagy-related genes to the risk of disease and its distinct phenotypes. The study population comprised 508 individuals from Zé District (Atlantique Department, Benin), with 208 newly diagnosed BU patients recruited at the Centre de Deépistage et de Traitement de l’Ulceère de Buruli d’Allada after 2005, and 300 unrelated, age and gender-matched controls, with similar water contact habits and the same ethnic background (healthy endemic controls) [1] (Table 1). This area presents a high incidence of BU, low consanguinity and uniform ethnicity [6]. All the subjects enrolled were HIV-negative and BCG-vaccinated. Collection of patient-level data included age, gender, clinical form, number and location of lesions and World Health Organization (WHO) clinical classification—as a severity cataloguing. All the patients enrolled were diagnosed after 2005, were positive for at least two of the three WHO recommended diagnostic tests, and received appropriate treatment. The National Ethical Review Board of the Ministry of Health in Benin (IRB0006860) provided approval for this study (clearance Nu 018, 20/Oct/2011), and written informed consent was obtained from all adult participants. Parents or guardians provided informed consent on behalf of all child participants. Genomic DNA from whole blood samples from patients and donors was isolated using the NZY Blood gDNA Isolation kit (NZYTech) according to the manufacturer's instructions. SNPs were selected based on previous published evidence of association with susceptibility to other mycobacterial diseases (S1 Table), with a particular emphasis on genetic variants with well-described functional consequences. Specifically, genetic variants in the multi-step intracellular xenophagy recognition process of mycobacteria through the NOD2-ATG16L1 axis and the complementary parkin-mediated ubiquitination were selected, thereby reinforcing the probability to detect positive associations. Genotyping of PARK2 (rs1333955, rs1040079, and rs1514343), NOD2 (rs13339578, rs2066842, rs4785225, rs9302752, and rs5743278), and ATG16L1 (rs2241880) SNPs was performed using the KASPar genotyping chemistry (LGC Genomics, UK) following the manufacturer’s instructions. The associations between SNPs and BU was performed using Pearson's χ2 test providing a value of odds ratio (OR) with a 95% confidence interval (CI) for different genetic models (co-dominant, dominant and recessive). A P value lower or equal to 0.05 was considered significant. The linkage disequilibrium (LD) and Hardy-Weinberg equilibrium (HWE) tests were assessed by using the Haploview 4.2 software. Genotype frequencies were used to phase the haplotype configurations by resorting to the same software. A total of 208 newly diagnosed cases of BU and 300 unrelated controls were selected according to fulfillment criteria. Demographics and clinical features of cases and age- and gender-matched controls are summarized in Table 1. The median age of cases was 14 years [interquartile range (IQR): 10–25] and similar to that of controls [17 years (IQR: 11–28)]; P = 0.25. The gender distribution of cases and controls was also not significantly different [89 (43%) females in 208 cases; and 146 (49%) females in 300 controls; P = 0.21]. Clinical features were in concordance with general African characteristics of BU [1]. The dominant clinical form reported was the ulcer (70.2%, including 6 cases with osteomyelitis), the mainly affected site were the limbs (87%), and the WHO categories 1 to 3 were displayed in 18.3%, 53.8% and 27.9% of the cases, respectively. The minor allele frequencies and HWE values for all SNPs are shown in S1 Table. To assess the risk and progression of BU according to NOD2, PARK2 and ATG16L1 SNPs, we compared their genotype frequencies between BU patients and age- and gender-matched healthy controls. Whereas no significant variations in the distribution of genotypes among cases and controls were observed in the overall test of association, the rs1333955 SNP in the PARK2 gene was significantly associated with increased susceptibility to BU upon modelling of a dominant mode of inheritance [OR, 1.43 (95% CI, 1.00–2.06); P = 0.05] (Table 2). Of interest, a similar though less significant association was also observed for patients carrying the rs1040079 SNP in the same gene [OR, 1.45 (95% CI, 0.96–2.18); P = 0.07]. Although the rs1333955 SNP was found to be in strong LD with rs1514343 and rs1040079 (Fig 1), none of the four haplotypes determined was significantly associated with the development of BU (S2 Table). No associations with the risk of BU were detected for SNPs in NOD2 or ATG16L1 (S3 Table). In addition, and although the rs13339578 SNP in the NOD2 gene was in strong LD with both rs5743278 and rs4785225 SNPs (Fig 1), no associations were found for the haplotypes formed by this block (S4 Table). Since the clinical presentation of BU varies dramatically and epidemiological data has pointed out that host genetic factors may be involved in these phenotypes [1], we further evaluated the genetic susceptibility to the severe WHO category 3 or the ulcerative form of BU. We found that both the rs9302752 and rs2066842 (P268S) SNPs in the NOD2 gene significantly increased the predisposition of patients to develop category 3 lesions following a dominant genetic model [OR, 2.23 (95% CI, 1.14–4.37); P = 0.02; and OR, 12.7 (95% CI, 0.60–269); P = 0.03), respectively] (Table 3). None of the other SNPs in NOD2, PARK2 or ATG16L1 revealed association with WHO category 3 (S5 Table). In what regards susceptibility to the ulcerative form of BU disease, the rs2241880 (T300A) SNP in the ATG16L1 gene was found to significantly protect patients from presenting the ulcer phenotype when a recessive genetic model was applied [OR, 0.35 (95% CI, 0.13–0.90); P = 0.02] (Table 3). None of the other SNPs in PARK2 or NOD2 genes revealed associations with the degree of ulceration (S5 Table). We compared the prevalence of SNPs in autophagy-related genes in confirmed cases of BU and in randomly selected community controls equally exposed to similar risk factors such as relationship and same behaviors (recreational or not) related to stagnant waters around villages. We found that the rs1333955 SNP in the PARK2 gene was significantly associated with development of BU. The PARK2 protein—known as parkin—is associated with the process of protein ubiquitination, acting as an E3 ligase and targeting proteins for proteasomal degradation [7]. The ubiquitin-mediated pathway is a complementary system for autophagy activation and that contributes to pathogen elimination, including M. tuberculosis, by surrounding bacteria with conjugated ubiquitin chains. Our findings support a role for the PARK2/PACRG gene cluster in susceptibility to M. ulcerans infection, suggesting that mechanisms linked to ubiquitination and proteasome-mediated protein degradation might unveil a common pathway in the intracellular fate of this pathogen. The fact that the same SNP has been associated with a higher risk of leprosy [8] points to a pertinent role for this gene in both infections. In addition, PACRG has been suggested to preferentially bind to hydrophobic molecules, such as lipids [9]. Mycolactone, a lipid mycotoxin, was recently shown to inhibit translocation of newly translated proteins into the endoplasmic reticulum [10], culminating in their degradation by the proteasome. Accordingly, we have recently reported that mycolactone induces an increased amount of ubiquitinated proteins in the cell by affecting cytoskeleton constituents and cytoskeleton-dependent intracellular trafficking [5]. Ultimately, this points to likely critical consequences of the rs1333955 SNP on the proteasomal degradation induced by mycolactone and might explain, at least in part, its association with risk of BU. Because autophagy is a pivotal immunological mechanism mediating protection to infection by intracellular pathogens [4], mycolactone-induced impairment of autophagy might have implications for the progression of BU disease. Previous studies have revealed that the NOD2-ATG16L1 axis is important for maintaining intracellular immune homeostasis [11]. The rs9302752 and rs2066842 SNPs in the NOD2 gene were found to be significantly associated with a severe phenotype of BU disease, reflected by the WHO Category 3, suggesting a crucial role of genetic variability of the NOD2 locus in defining severity of BU disease. The rs9302752 SNP is located in the upstream region of the gene, and therefore it might deregulate promoter activity and influence gene expression and susceptibility to infection. Indeed, silencing NOD2 expression in human macrophages was reported to result in a local spread of M. tuberculosis, with an impairment in NOD2-mediated production of cytokines [12]. In addition, the rs2066842 SNP underlies the P268S amino acid substitution, and has been found to affect host recognition of bacterial muramyl dipeptide. As such, we hypothesize that a failure in the innate immune recognition of M. ulcerans via NOD2 might divert the proper activation of immunological autophagy, therefore permitting progression of infection and development of more severe phenotypes. The non-ulcerative and ulcerative forms of BU can be observed as stable clinical phenotypes, and not all patients progress to the latter [1]. We found the rs2241880 SNP in the ATG16L1 gene to be associated with protection of BU patients from an ulcerative clinical form. ATG16L1 is a master regulator of the core autophagy machinery that was initially identified as a pivotal risk factor for Crohn’s disease [13]. The rs2241880 variant is located in the coding region of ATG16L1 and leads to the Thr300Ala (T300A) amino acid substitution, which has recently been found to enhance its self-degradation by caspase 3, thereby impairing autophagy activation [14]. Of interest, the T300A variant also decreased selective autophagy, resulting in increased interleukin (IL)-1β signaling and decreased antibacterial defense [14]. Increased levels of IL-1β are also associated with a more exuberant local inflammation. Indeed, non-ulcerative forms of BU, such as edema and plaque, are considered more inflammatory than ulcerative lesions [15]. Our study has some limitations. In particular, the study was conducted in a single population, and therefore it requires confirmation in larger groups and independent cohorts, as well as the assessment of the functional consequences of the associated variants and their influence to the immune response dynamics. It is however important to note that our case-control study has a robust sample size and the critical advantage that controls were carefully matched to cases regarding environmental exposure to mycobacteria. Our findings indicate that specific genetic variants in autophagy-related genes influence susceptibility to the development of BU and its progression to severe phenotypes, highlighting the multiple additive effects of single genetic factors and their complex interactions towards the overall weight of the human immune response to M. ulcerans. Ultimately, this study reinforces the applicability of host genomics as an important factor to be considered in the stratification of infection risk in endemic regions and, more importantly, for the definition of patient groups more likely to advance to more severe and debilitating phenotypes of BU disease.

10.1371/journal.pntd.0006555

Genetic diversity in two Plasmodium vivax protein ligands for reticulocyte invasion

The interaction between Plasmodium vivax Duffy binding protein (PvDBP) and Duffy antigen receptor for chemokines (DARC) has been described as critical for the invasion of human reticulocytes, although increasing reports of P. vivax infections in Duffy-negative individuals questions its unique role. To investigate the genetic diversity of the two main protein ligands for reticulocyte invasion, PvDBP and P. vivax Erythrocyte Binding Protein (PvEBP), we analyzed 458 isolates collected in Cambodia and Madagascar from individuals genotyped as Duffy-positive. First, we observed a high proportion of isolates with multiple copies PvEBP from Madagascar (56%) where Duffy negative and positive individuals coexist compared to Cambodia (19%) where Duffy-negative population is virtually absent. Whether the gene amplification observed is responsible for alternate invasion pathways remains to be tested. Second, we found that the PvEBP gene was less diverse than PvDBP gene (12 vs. 33 alleles) but provided evidence for an excess of nonsynonymous mutations with the complete absence of synonymous mutations. This finding reveals that PvEBP is under strong diversifying selection, and confirms the importance of this protein ligand in the invasion process of the human reticulocytes and as a target of acquired immunity. These observations highlight how genomic changes in parasite ligands improve the fitness of P. vivax isolates in the face of immune pressure and receptor polymorphisms.

Until recently, P. vivax was thought to infect only Duffy positive individuals, due to its dependence on binding the Duffy blood group antigen as a receptor for reticulocyte invasion and to be absent from parts of Africa where the Duffy-negative phenotype is highly frequent. However, a number of recent studies from across sub-Saharan Africa have reported P. vivax infections in Duffy-negative individuals. Invasion into Duffy-positive reticulocytes is mediated by the P. vivax Duffy binding protein (PvDBP). The mechanism for invasion into Duffy-negative reticulocytes is not known. A homologue of PvDBP, namely, P. vivax erythrocyte binding protein (PvEBP), has been recently identified but its role in Duffy independent invasion is not clearly defined. Here, we provide unique insights into the roles of these two key ligands by studying the genetic diversity of P. vivax isolates collected from Cambodia, where most of the individuals are Duffy positive (not all), and Madagascar where both Duffy-positive and Duffy-negative individuals coexists. Our data suggest that PvEBP may play an important functional role in invasion into Duffy-negative reticulocytes. PvEBP appears to be a target of naturally acquired antibody responses following natural exposure to P. vivax infection and such as a consequence an important vaccine candidate, together with PvDBP.

Plasmodium vivax is a predominant cause of malaria outside Africa, which causes significant morbidity (estimate of 8.5 million cases in 2016) and places an enormous economic burden on many resource poor countries [1]. Until recently, vivax malaria was considered a benign infection compared to P. falciparum, although clinical episodes and regular recurrent infections cause significant morbidity [2]. Moreover, P. vivax infections can sometimes lead to severe life-threatening pathologies [3]. Previous data from malaria therapy, which was used extensively for over two decades (1920–1940) for treatment of neurosyphilis (e.g. paralysis of the insane), as well as experimental infections of volunteers have demonstrated that individuals of African origin were naturally resistant to P. vivax infection [4–6]. Thereafter, following identification of the Duffy blood group [7], it was shown that the Duffy blood group antigen (Fya or Fyb) was not expressed on red blood cells (RBCs) of individuals of African origin (Duffy-negative) [8, 9]. Seminal works with controlled experimental infections of volunteers through sporozoite challenge and in vitro invasion studies using P. knowlesi as a model subsequently established the paradigm that the Duffy antigen is required for reticulocyte invasion by P. vivax [9–11]. Consequently, vivax malaria was long thought to be absent from parts of Africa where the Duffy-negative phenotype is highly frequent [12]. Host cell invasion by Plasmodium merozoites is a complex, multi-step process that involves multiple interactions between erythrocyte receptors and ligands on merozoites. Unlike P. falciparum merozoites, which can use several erythrocyte receptors for invasion, it was thought, until recently, that invasion of human reticulocytes by P. vivax is completely dependent on the interaction between the P. vivax Duffy Binding Protein (PvDBP) and the erythrocyte Duffy antigen receptor for chemokines (DARC) [13, 14]. However, over the last decade, two reports both from Brazil [15, 16] and a growing body of studies conducted in Africa (review in [12]) have reported PCR-positive vivax malaria cases in Duffy-negative individuals. These frequent observations raise the emerging issue of P. vivax infection in Duffy-negative populations, and the possibility of alternative invasion mechanism(s). At present, we do not know whether these clinical reports are common and were previously undetected or if they are due to the emergence and spread of specific P. vivax strains that use alternative Duffy-independent pathways to invade Duffy-negative reticulocytes. Whatever the reason for the increasing frequency of observations of P. vivax in Duffy-negative populations, it is a cause for concern and demands attention. Recent whole genome sequencing studies of monkey-adapted P. vivax strains and field isolates (all isolates were collected from Duffy-positive patients) have revealed that the PvDBP gene was duplicated in multiple P. vivax isolates, particularly at high prevalence in Madagascar, a setting where Duffy-positive and Duffy-negative individuals coexist [17]. Initially, this epidemiological pattern suggested that the duplication of this gene was likely associated with the capability of the parasite to overcome the barrier of Duffy negativity. Since then, this hypothesis has been challenged by Hostetler et al. [18] who found that PvDBP gene duplications were widespread even in malaria endemic areas in Southeast Asia where Duffy-negativity is not present. Another recent study [19] reported evidence of PvDBP gene amplification (3 and 8 copies) in two Duffy-negative Ethiopian isolates. In addition, sequence data generated from a P. vivax field isolate (C127 isolate from Cambodia), which used reconstruction of long reads without relying on the reference genome (e.g. the monkey-adapted Salvador I strain), identified 792 predicted genes [20]. Among them, two contigs contained predicted protein coding genes similar to known Plasmodium red blood cell invasion proteins. One of these genes harbored all the hallmarks of a Plasmodium erythrocyte binding protein, including conserved Duffy-binding like and C-terminal cysteine-rich domains. Further analysis showed that this gene, which is present in most of studied P. vivax genomes, clustered separately from all known Plasmodium erythrocyte-binding protein genes [20]. Further functional investigations demonstrated that the recombinant PvEBP (derived from the C127 PvEBP allele) bound preferentially to immature (CD71high) and Duffy-positive reticulocytes [21]. A minimal binding was observed with Duffy-negative reticulocytes, and no binding was observed with mature red blood cells or normocytes. PvDBP and PvEBP were clearly shown to be antigenically distinct. These findings were slightly modulated in another study that reported that the region II of PvEBP expressed in COS-7 cells bound both Duffy-positive and Duffy-negative erythrocytes although at low frequency suggesting that PvEBP may be a ligand for invasion of Duffy-negative reticulocytes [19]. To gain insight into the natural genetic diversity and polymorphisms in the two main protein ligands for reticulocyte invasion (PvDBP and PvEBP), we collected and analyzed P. vivax isolates from two distinct settings: in Cambodia, where the vast majority of individuals are Duffy-positive[2], and Madagascar where an admixture of Duffy-positive and Duffy-negative people coexists [22]. Duffy genotyping data were available for 174 samples and all samples were genotyped as Duffy-positive (T-33C substitution: Cambodia, N = 119, 100% T/T and Madagascar, N = 55, 44% T/T and 56%T/C) (Table 1). PvDBP CNV was assessed in 458 P. vivax isolates collected in Cambodia (N = 392) and Madagascar (N = 66) (Table 1). PvDBP CNV ranged from 1 to 6 copies. No significant differences were observed between P. vivax isolates from both countries in the median gene copy number (1.31 and 1.30 for Madagascar and Cambodia, respectively, P = 0.21, Mann Whitney test) or in the proportion of isolates carrying multiple copies PvDBP (45.5% for Madagascar and 37.5% for Cambodia, P = 0.22, Fisher’s exact test). These data confirm that PvDBP amplification is a common event across isolates from Cambodia and Madagascar (Fig 1). Further analysis showed that the proportion of isolates with multiple copies of PvDBP increased over time in Cambodia: from 16.7% (8/48) in 2003–2004 to 39.8% (142/357) in 2011–2017 (P = 0.001, Fisher’s exact test). This trend was significant for samples collected from western provinces (Battambang, Pursat, Pailin, Kampot) (15.2% vs. 36.6%, P = 0.02, Fisher’s exact test). The proportion of isolates with multiple copies of PvDBP and the median gene copy numbers for PvDBP were similar between paired isolates collected before a standard 3-day course of chloroquine (30 mg/kg) (D0) and at the day of recurrence (Dx) occurring during the 2-months follow up in patients relocated in a non-transmission area (D0: 38.5%, 5/13, median PvDBP copy number = 1.4, IQR: 1.1–2.4 vs. Dx: 23.1%, 3/13, median PvDBP copy number = 1.0, IQR: 1.0–1.5, P = 0.67, Fisher’s exact test and P = 0.40, Mann Whitney test, respectively). More surprisingly, we observed in Malagasy samples that the proportion of isolates with PvDBP amplification or the median copy of PvDBP gene was significantly higher in homozygous Duffy-positive individuals (T/T) compared to heterozygous Duffy-positive individuals (T/C) who are supposed to express less DARC antigen on the surface of their reticulocytes: T/T: 59.1%, 13/22, median PvDBP copy number = 2.2, IQR: 1.0–3.0 vs. T/C: 27.6%, 8/29, median PvDBP copy number = 2.0, IQR: 0.9–1.5, P = 0.04 (Fisher’s exact test and P = 0.01, Mann Whitney test, respectively). To determine whether PvDBP gene amplification was restricted to specific alleles, PvDBPII sequences (from codon 184 to codon 468) were determined among 153 Cambodian and 92 Malagasy P. vivax isolates. As already described, PvDBP sequences were found to be highly polymorphic and multiple alleles were observed both in Cambodia (21 alleles) and Madagascar (15 alleles) (Fig 2 and S1 Table). Numerous SNPs already described were observed [23–35]. Four new SNPs (one silent and three non-synonymous) were detected (K260E, P450R, V453V and P475R). Among the 23 SNPs, some were solely observed in Madagascar (I419M, V453V and I464I) and in Cambodia (K260E, F306L, I367T, N375D, R378R, S398T, T404R, P450R, P475A/R and Q486E). SNPs along with the location of binding residues for DARC, identified by previous studies [36–38], were mapped on to the 3D structure of PvDBPII. Interestingly, majority of the SNPs were found to be distal from the DARC binding residues, and located on the opposite surface compared to the binding residues for DARC (Fig 3, Panel A). A total of 169 P. vivax isolates with both SNP and CNV were available for analysis. Among them, 76 (45%) had multiple copies PvDBP. Gene amplification was observed in 9/18 alleles (50%) in Cambodia and in 8/9 alleles (89%) in Madagascar isolates (Fig 4, Panel A). No specific PvDBP allele was found to have in proportion more isolates than expected with PvDBP gene amplification, indicating that PvDBP gene amplification can occur across multiple alleles. PvEBP CNV was determined in 184 Cambodian and 66 Malagasy samples (Table 1). PvEBP CNV ranged from 1 to 2 copies in Cambodian isolates and 1 to 5 copies in Malagasy samples. The proportion of multiple copies PvEBP isolates was significantly more frequent in Madagascar compared to Cambodia: 56% (37/66) vs. 19% (35/184) (P<10−6, Fisher’s exact test). The median copy number of PvEBP gene was also significantly higher in isolates from Madagascar compared to Cambodia: 1.6 (IQR: 1.3–2.1) vs. 1.3 (IQR: 0.95–1.4) (P<10−4, Mann Whitney test) (Fig 1, Panel B). In Cambodia, the proportion of isolates with multiple copies of PvEBP or the median copy number of PvEBP gene was higher in high transmission areas located in eastern provinces (Ratanakiri, Mondulkiri and Kratie): 24% (31/127) vs. 7% (4/57) (P = 0.004, Fisher’s exact test), respectively and 1.3 (IQR: 1.0–1.5) vs. 1.1 (IQR: 0.85–1.3) (P = 0.003, Mann Whitney test). PvEBP CNV (proportion of isolates carrying PvEBP amplification or median copy of PvEBP gene) were similar between paired isolates collected at time of chloroquine treatment (D0) and on day of recurrence (Dx): D0, 60% (6/10), median PvEBP copy number = 1.6 (IQR: 1.4–1.7) vs. Dx, 50% (2/4), median PvEBP copy number = 1.7 (IQR: 1.0–2.0) (P = 1.0, Fisher’s exact test and P = 0.83, Mann Whitney test, respectively). In Malagasy samples, no significant differences in the proportion of isolates with PvEBP amplification or the median copy of PvEBP gene were observed between homozygous (T/T) and heterozygous (T/C) Duffy-positives individuals: T/T, 54% (12/22), median PvEBP copy number = 1.5 (IQR: 1.2–2.2) vs. T/C, 41% (12/29), median PvEBP copy number = 1.4 (IQR: 1.2–1.7) (P = 0.4, Fisher’s exact test and P = 0.3, Mann Whitney test, respectively). PvEBP sequences obtained from 150 Cambodian and 69 Malagasy isolates (Table 1) were compared to the reference genome (C127) [20]. Eleven non-synonymous SNPs were observed leading to twelve different alleles. Most of them (59%) were the C127-like allele (Fig 5 and S2 Table). Four SNPs were common to both countries (N233K, D311N, E322K and I463V), three specific to Madagascar (N200K, C219S and N449K) and four to Cambodia (D268N, E306K, T421I, S423W). More common alleles were observed between the two countries for PvEBP compared to PvDBP: half of PvEBP alleles were shared between both countries compared to 1/11 for PvDBP alleles (P = 0.02). Given that PvEBPII shares homology with DBL domains, the structure of PvEBPII was modeled based on the structure of PvDBPII (Fig 3, Panel B). The ribbon diagrams for the structures of PvDBPII and PvEBPII were superimposed and found to overlap (Fig 3, Panel C). SNPs were mapped to the PvEBPII structure and were found to be distributed across the domain. Since the receptor-binding site of PvEBPII remains to be defined, we were not able to assess whether the binding site was conserved or variable and deduce if this protein ligand can be targeted by strain transcending inhibitory antibodies to block receptor binding and invasion by diverse strains. A total of 143 P. vivax isolates with both SNP and CNV data for PvEBP were available for analysis. Among them, 28 (20%) were found to have PvEBP gene amplification. PvEBP gene amplification was observed in 6/10 alleles (60%) in Cambodia and in 5/7 alleles (71%) in Madagascar isolates (Fig 4, Panel B). No specific PvEBP allele was found to have in proportion more isolates than expected with PvEBP gene amplification, indicating that PvEBP gene amplification can occur across multiple alleles. The work presented here was focused to explore the polymorphism and the copy number variation of two P. vivax protein ligands involved in invasion into reticulocytes [39]. Unsurprisingly, we observed that SNPs detected in the receptor-binding domain of PvDBPII were similar to those observed previously. The binding residues for DARC in PvDBPII were conserved and majority of the SNPs observed were distal to the binding site. In fact, many SNPs were found on the opposite face of PvDBPII compared to the binding site suggesting that this surface is under immune pressure during natural infection. Genome sequencing of different P. vivax isolates collected in various malaria endemic settings had previously revealed that the PvDBP gene is frequently duplicated [17, 18]. Here, we confirm again that amplification of PvDBPII is a common genomic event, which can frequently occur in isolates from areas where Duffy-negative phenotype is virtually absent such as Cambodia. To encompass the two types of duplication previously observed [17, 18], we designed a novel quantitative real-time PCR with a standard curve that enables precise quantification of the number of PvDBP and PvEBP genes up to 6 copies. Our data confirms the hypothesis that PvDBP amplification is unlikely to play a role in invasion of Duffy- negative reticulocytes or at least that human Duffy-negative populations do not specifically select for parasites with multiple copies of gene encoding PvDBP. The critical role of PvDBP protein ligand in invasion pathway into human Duffy-positive reticulocytes is well established while alternative pathways through other parasite protein ligands involved in the invasion of Duffy-negative reticulocytes are unknown. To date, few biological investigations seem to support a possible role of the newly described PvEBP protein ligand in the invasion of Duffy-negative reticulocytes [19, 21, 39]. By exploring its genetic diversity, we discovered two interesting findings. First, a high proportion of isolates from Madagascar (56%) where Duffy negative and positive individuals coexist had multiple copies of PvEBP compared to Cambodia (19%) where Duffy-negative population is virtually absent. Second, there was evidence for an excess of nonsynonymous mutations and the total absence of synonymous mutations in PvEBP. Indeed, among the 219 P. vivax tested samples we observed only eleven non-synonymous point mutations in PvEBP sequences and twelve different alleles (S2 Table). SNPs found were distributed across PvEBPII but we were not able to define whether polymorphism affects the binding site, which remains to be identified. In any case, the absence of synonymous mutations clearly reveals that this gene is under strong diversifying selection, as has been shown previously for P. falciparum EBA-175 and PvDBP [40]. The most likely agent driving the diversification of these antigens is the human acquired immune response. Indeed, we can speculate that novel alleles encoding parasite antigens occuring in the population would potentially be able to avoid the human immune response and as such give the parasite a survival advantage leading to the allele’s selection. However, it remains unclear whether the elevated frequency of C127 allele in the population sample (~62% and ~52% in Cambodia and Madagascar, respectively) indicates a relatively recent selective increase of a new variant whose sequence may reflect its ability to avoid immune detection. Due to a similar type of selection on P. falciparum EBA-175 and P. vivax EBP, our data confirms the importance of the newly decribed protein ligand PvEBP in invasion of human reticulocytes and as a target of acquired immunity. Furthermore, the role of PvEBP as a protein ligand is supported by recent serological analysis41,42. These studies have shown that antibodies against the region II domain of the PvEBP were commonly detected in sera from individuals in malaria endemic settings such as in Cambodia or in Solomon Islands/Papua New Guinea. In Cambodia, humoral immune response to PvEBP was found to be higher and longer lasting compared to PvDBP, making this marker a better candidate to monitor P. vivax infections [41]. Moreover, Franca et al, in Papua New Guinea identified a significant association between reduced risk of clinical vivax malaria and levels of antibodies against PvEBP [42]. As for PvDBP, we also assessed the PvEBP gene copy number in our array of isolates and observed that contrary to PvDBP, isolates from Madagascar, where Duffy negative and positive individuals coexist, carry more frequently a PvEBP gene expansion. In particular, we detected in Malagasy isolates a bimodal distribution of parasites with multiple copies of PvEBP that includes a specific population of parasites with > 3 copies of PvEBP gene that was not found in Cambodian isolates. Unfortunately, we could not test for association between the PvEBP gene expansion and their capacity to invade Duffy-negative reticulocytes, as reported recently by Gunalan et al. for PvDBP [19]. Indeed, among the few P. vivax samples collected from Duffy-negative individuals we had, we were not able to generate reliable PCR signals, probably because of the very small amount of DNA reflecting the usually low parasitemia found in these individuals. One of the limitations in this work is that the majority of patients likely carry multiple clones of P. vivax [43, 44]. By using a qPCR approach as we did, we are aware that the PvDBP and PvEBP estimated gene copy number reflects the major clone contained in each isolate Next generation sequencing such as single cell DNA sequencing or even droplet PCR approach should be able to overcome this issue and assess CNV for each clone. Similarly, the Sanger sequencing technology used in this work to determine the sequences of PvDBP and PvEBP does allow sequencing only the major clone, preventing us to determine whether each strain in isolates with multiple copies of PvDBP and PvEBP had similar or different alleles. In summary, we provide here data regarding the genetic diversity of the PvEBP gene, a recently described and potential protein ligand involved in invasion into human reticulocytes. Evidence for positive diversifying selection on the region II domain of the PvEBP was observed, which is similar to the evidence for diversifying selection on PvDBP region II. These observations clearly confirm the importance of PvEBP in the invasion process of the human reticulocyes and/or as a target of acquired immunity. In addition, the high proportion of P. vivax isolates with multiple copies of PvEBP gene that we found in Madagascar is intriguing and needs additional in-depth investigations. So far, the association between invasion in Duffy-negative individuals and P. vivax specific genomic traits (SNP or CNV) mainly relies on epidemiological associations. Direct evidence through reticulocyte invasion assays in vitro or in humanized mouse models with P. vivax isolates with diverse alleles and copy numbers of PvDBP and PvEBP and human reticulocytes with different Duffy phenotypes will be of great value in deciphering the molecular basis of Duffy-independent invasion pathway(s) used by P. vivax. One hundred twenty nine Malagasy P. vivax samples were collected in 2015–2017 through cross-sectional surveys in the district of Maevatanana, from asymptomatic individuals during active case detection and from symptomatic patients seeking antimalarial treatment in health centers located in three communes (Andriba, Antanimbary and Maevatanana). P. vivax infections were detected using a malaria rapid diagnostic test (CareStart Malaria Pf/pan RDTs, Accesbio) and, capillary blood samples were spotted into filter papers for each positive case. In Cambodia, 453 P. vivax isolates were collected in 2003–2017 from symptomatic patients, seeking antimalarial treatment in health centers located around the country. Malaria diagnosis was also performed by RDT (CareStart Malaria Pf/pan RDTs, Accesbio) and microscopy. Venous blood samples were collected from confirmed vivax malaria cases into 5 ml EDTA tubes. We also included in our analysis, 16 P. vivax isolates collected from recurrences occurring in the 2-months follow up after a standard 3-day course of chloroquine (30 mg/kg) in patients relocated in a non-transmission area [45]. The study protocols were reviewed and approved by the Cambodian National Ethics Committee on Health Research (IRB 038NECHR) or the National Ethics Committee in Madagascar (Ministry of Health, 141/MSANP/CE). All individuals or their parents/guardians provided informed written consent before sample collection. DNA was extracted from blood spots with Instagene Matrix reagent (BioRad, Marnes-la-Coquette, France) or from whole blood samples using the QIAamp DNA Blood Mini Kit (Qiagen, Courtaboeuf, France), according to the manufacturer’s instructions. Molecular detection and identification of Plasmodium parasites were performed by using real-time PCR targeting the cytochrome b gene as described previously [46, 47]. PvDBP and PvEBP sequences were determined by nested PCR targeting the PvDBP region II and Sanger sequencing (Macrogen, Seoul, South Korea) using the following conditions. A first round PCR was conducted in 25 μL reactions using DNA, 0.2 μM of primers, 250 μM each dNTP, 2 mM MgCl2, and 1.25 units Taq Solis DNA Polymerase (Solis BioDyne, Tartu, Estonia) under the following conditions: 94°C for 15 min, followed by 40 cycles of 94°C for 20 s, 56°C for 40 s, 72°C for 90 s, and a final extension at 72°C for 10 min. The nested PCR was carried out in 55 μL reactions using 2 μL of the primary PCR products diluted at 1/10, 0.40 μM of primers, 250 μM each dNTP, 2.5 mM MgCl2, and 2.5 units Taq Solis DNA Polymerase (Solis BioDyne, Tartu, Estonia) under the following conditions: 94°C for 15 min, followed by 40 cycles of 94°C for 20 s, 60°C for 20 s, 72°C for 60 s, and a final extension at 72°C for 10 min (S3 Table). Nucleotides and corresponding amino acids were analyzed using the CEQ 2000 software (Beckman). The sequences generated were compared to PVX_110810 (Pv_Sal1_chr06:976,329–980,090 (+)) for PvDBP and to P. vivax isolate C127 nEBP gene (KC987954.1) for PvEBP. PvDBP and PvEBP genes copy numbers were measured relatively to the single copy β-tubulin gene (housekeeping gene) using a CFX96 real-time PCR thermocycler (Biorad, Singapore). PCR were performed in 20 μL volumes in a 96-well plate containing 1X HOT FIREPol EvaGreen qPCR Mix Plus (Solis BioDyne, Tartu, Estonia), 0.5 μM of each forward and reverse primer and 2 μL of template DNA. Amplifications were performed under the following conditions: 95°C for 15 min, followed by 45 cycles of 95°C for 15 s, 60°C for 20 s, and 72°C for 20 s. PvDBP and PvEBP genes copy numbers were estimated in triplicate relative to a standard curves by using synthetic genes cloned in pEX-A2 vector (Eurofins Genomics, Greece) (β-tubulin, PvDBP, PvEBP) mixed at different ratio from 1:1 up to 1:6 (1 copy of β-tubulin and 1 to 6 copies of PvDBP or PvEBP) (S3 Table). The ΔCT method (where CT is the cycle threshold) was used to determine the number of copies of each sample. In addition, an isolate with one copy of each gene was used as control. All isolates with copy number estimates of less than 0.5 were discarded. Gene copy number values were rounded up as following: 0.5–1.4 to 1 copy, >1.5 to 2 copies, >2.5 to 3 copies, >3.5 to 4 copies and >4.5 to 5 copies (S3 Table). Duffy genotypes were determined by nested PCR and Sanger sequencing (Macrogen, Seoul, South Korea) as previously described [48]. The inner PCR was conducted in 25 μL reactions using 3 μL of template DNA, 0.4 μM of primers, 250 μM each dNTP, 2 mM MgCl2, and 1.25 units Taq Solis DNA Polymerase under the following conditions: 94°C for 15 min, followed by 40 cycles of 94°C for 30 s, 58°C for 30 s, 72°C for 90 s, and a final extension at 72°C for 10 min. Outer PCRs detecting mutations in the GATA box (T or C) and in the coding sequence were carried out in 55 μL reactions using 2 μL of the primary PCR products diluted at 1/10, 0.36 μM of each primer), 250 μM each dNTP, 2.5 mM MgCl2, and 1.25 units Taq Solis DNA Polymerase under the following conditions: 94°C for 15 min, followed by 40 cycles of 94°C for 20 s, 58°C for 20 s, 72°C for 60 s, and a final extension at 72°C for 10 min (S3 Table). Coordinates of single chain of PvDBP (PDB ID 4NUU) were obtained from Protein databank [49, 50] and used for mapping SNPs. SNPs were mapped on the PvDBPII structure using Chimera software [51]. In addition, putative binding residues of PvDBPII predicted previously [36–38] were also mapped on the PvDBPII structure. The selected primary PvEBPII sequence (allele C127) [20] (S1 Fig) was used for 3D modeling prediction (S1 Fig). 3D Model of the PvEBPII domain was determined by homology based structure prediction online tool using Phyre2 under default mode [52]. 3D structure with maximum score was selected and highly flexible N- and C- terminal ends were truncated from the structure. Refinement of predicted 3D model and minimization of local structural distortions was performed using ModRefiner [53]. The overall quality of the predicted 3D model was evaluated with QMEANDisCo (Qualitative Model Energy ANalysis- Distance Constraint) score. QMEANDisCo is a tool for assessing the agreement of pairwise residue-residue distances with ensembles of distance constraints extracted from structures homologous to the assessed model [54]. SNPs were mapped on the PvEBPII model using Chimera software [51]. Data were analyzed with Microsoft Excel and MedCalc version 12 (Mariakerke, Belgium). Quantitative and qualitative data were expressed as median (IQR) or proportion (%), respectively. The Mann-Whitney U test was used for non-parametric comparisons. For categorical variables, proportions were examined by Chi-squared or by Fisher’s exact tests. Two-sided p-values of <0.05 were considered statistically significant. MUSCLE multiple alignment and evolutionary analyses were conducted in MEGA [55] with 1000 bootstrap replicates. All positions containing gaps and missing data were eliminated. Phylogenetic reconstructions were performed using neighbor joining (NJ).

10.1371/journal.ppat.1004702

γδ T Cells Confer Protection against Murine Cytomegalovirus (MCMV)

Cytomegalovirus (CMV) is a leading infectious cause of morbidity in immune-compromised patients. γδ T cells have been involved in the response to CMV but their role in protection has not been firmly established and their dependency on other lymphocytes has not been addressed. Using C57BL/6 αβ and/or γδ T cell-deficient mice, we here show that γδ T cells are as competent as αβ T cells to protect mice from CMV-induced death. γδ T cell-mediated protection involved control of viral load and prevented organ damage. γδ T cell recovery by bone marrow transplant or adoptive transfer experiments rescued CD3ε−/− mice from CMV-induced death confirming the protective antiviral role of γδ T cells. As observed in humans, different γδ T cell subsets were induced upon CMV challenge, which differentiated into effector memory cells. This response was observed in the liver and lungs and implicated both CD27+ and CD27− γδ T cells. NK cells were the largely preponderant producers of IFNγ and cytotoxic granules throughout the infection, suggesting that the protective role of γδ T cells did not principally rely on either of these two functions. Finally, γδ T cells were strikingly sufficient to fully protect Rag−/−γc−/− mice from death, demonstrating that they can act in the absence of B and NK cells. Altogether our results uncover an autonomous protective antiviral function of γδ T cells, and open new perspectives for the characterization of a non classical mode of action which should foster the design of new γδ T cell based therapies, especially useful in αβ T cell compromised patients.

γδ T cells are unconventional T lymphocytes that play a unique role in host protection against pathogens. Human Cytomegalovirus (HCMV) is a widespread virus that can cause severe organ disease such as hepatitis and pneumonitis in immune-compromised patients. Our decade-long study conveys compelling evidence for the implication of human γδ T cells in the immune response against HCMV, but their protective role could not be formally demonstrated in humans. In the present study we use the murine model of CMV infection which allows the spatial and temporal analysis of viral spread and anti-viral immune responses. We show that, in the absence of αβ T cells, γδ T cells control MCMV-induced hepatitis, pneumonitis and death by restricting viral load in the liver, lungs and spleen. γδ T cells expand in these organs and display memory features that could be further incorporated into vaccination strategies. In conclusion, γδ T cells represent an important arm in the immune response against CMV infection that could be particularly important in the context of αβ T cell immune-suppression.

Human CMV (HCMV) is a universally distributed pathogen that infects 50–90% of the world's population. Asymptomatic in healthy people, HCMV infection may lead to increased morbidity and mortality in immunocompromised individuals. Overall survival following transplantation is decreased when either the donor or the recipient is HCMV-seropositive [1,2,3]. Because of drug-related adverse effects and drug resistance there is growing interest for immunotherapy as an adjunct to antiviral therapy. Understanding the mechanisms developed by the immune system to control HCMV is therefore critical to enable the design of new curative or preemptive protocols aimed at enhancing patient immune defense against this virus. Effective immune control of HCMV has been compellingly shown to rely on both conventional lymphocytes and NK cells [4]. However, as we initially reported, HCMV also induces a robust γδ T cell response in organ transplant recipients [5]; and later, γδ T cell response to HCMV was extended to several other situations not always associated to immunosuppression; such as immunodeficiencies, bone marrow transplantation, pregnancy, elderly and also in healthy individuals [6,7,8,9,10,11,12]. HCMV-mediated persistent expansion of γδ T cells in transplant recipients is associated with infection resolution [13], and implies tissue-associated Vδ2-negative γδ T cells which acquire a terminally differentiated phenotype upon HCMV pressure [10,14]. When isolated in vitro, these lymphocytes were shown to kill HCMV-infected cells, limit virus propagation and produce IFNγ through recognition of opsonized viruses [15,16]. Several features of γδ T cells might explain their specific relationship to HCMV: (i) they are not MHC restricted, and thus not affected by HCMV strategies to inhibit HLA molecules, (ii) they recognize self-antigens on the surface of stressed cells such as virus infected cells [17,18] and (iii) they are located at external body surfaces (eg gut and lung) and organs (eg liver) involved in HCMV transmission and replication [19]. Moreover, HCMV-reactive γδ T cells exhibit dual reactivity against tumor cells, due to the recognition of stress-induced self-antigens shared by HCMV-infected and tumor cells [15,18,20]. In agreement with this, HCMV-infection and/or γδ T cell expansion have been associated with reduced cancer risk in kidney transplant recipients [21] and with graft-versus leukemia effect in bone marrow transplant recipients [22,23,24]. All these specificities are consistent with an antiviral protective role of γδ T cells against HCMV and they thus represent valuable candidates for anti-HCMV immunotherapy especially in immunocompromised patients vulnerable to neoplasia. However, their role in protection and specific contribution within the global anti-CMV immune response has not been firmly established, nor their anatomical sites of activation and intervention. The aim of the present study was therefore to take advantage of the murine model of CMV infection to address these questions and to assess the respective ability of αβ and γδ T cells alone to protect mice from CMV infection. Murine CMV (MCMV) has been widely used to model the immune response to HCMV in mice since it reproduces with reasonable accuracy the antiviral response of CD8 T cells and NK cells [25]. Murine γδ T cells have been implicated in MCMV infection only once [26], and their sufficiency for protection has not yet been addressed. We show herein that γδ T cells are as competent as αβ T cells to control MCMV infection and protect mice from death encouraging the development of novel anti-viral immunotherapeutic protocols based on γδ T cell manipulation. In mice, MCMV-specific αβ T cells control viral spread and protect infected mice from death [27] but little is known regarding the implication of γδ T cells. To evaluate the respective contribution of αβ and γδ T cells to the immune response against MCMV, mice deficient for γδ T cells (TCRδ−/−), for αβ T cells (TCRα−/−) or for both T cell subsets (CD3ε−/−) were challenged with 105 plaque forming units (PFU) of salivary gland MCMV. This dose was reported to be sublethal for C57BL/6 mice (as described at http://mutagenetix.utsouthwestern.edu/protocol/protocol_rec.cfm?protocolid=5). Accordingly, 100% of CD3ε+/− control mice survived MCMV infection, whereas CD3ε−/− died about 4 weeks after viral challenge (Fig. 1A), confirming the critical role of T cells in controlling MCMV infection. CD3ε−/− mice were extremely sensitive to MCMV despite the presence of NK cells [28] since they died at doses of MCMV as low as 2.103 PFU (Fig. 1B). Unexpectedly, both TCRδ−/− and TCRα−/− mice survived as long as CD3ε+/− control mice. These results reveal that the presence of either αβ or γδ T cell subset was sufficient to protect mice from MCMV infection, disclosing the potentially critical function of γδ T cells in the immune response against MCMV. To examine whether this protection against CMV by γδ T cells relies on the control of viral loads, the kinetics of MCMV spread in T cell deficient versus T cell competent mice was determined in various organs. Comparison between each mouse line is shown in Fig. 2 and comparison between different time points is shown in S1 Fig. In the absence of T cells, MCMV DNA copy numbers increased substantially from day 3 to 24, with up to 107 copies (/100ng DNA) in the spleen and lungs of CD3ε−/− mice before death. Interestingly, γδ T cells alone (in TCRα−/− mice) were sufficient to prevent an increase of viral load in all organs, except the salivary glands which are known to support prolonged virus replication even in wild-type mice (S1 Fig.). At the end of these experiments, MCMV copies were much lower in T cell bearing mice than in mice without T cells (Fig. 2), underlining the inability of C57BL/6 mice to control MCMV infection in the absence of T cells. It was of particular interest to see that in the lungs γδ T cells were as potent as αβ T cells to control the viral load except at day 14. As a whole, these results suggest independent control of MCMV spread by the αβ and γδ T cell subsets, revealing that γδ T cells are sufficient to control viral load and can substitute for the absence of αβ T cells. Hepatitis and pneumonitis are common features of CMV pathogenesis in both humans and mice. Hepatitis can be assessed in living infected mice through the quantification of transaminase levels in the serum. As shown in Fig. 3A, aspartate aminotransferase (AST) and alanine aminotransferase (ALT) only increased in the absence of all T cells (CD3ε−/− mice), reaching up to 8 fold the basal level before death of CD3ε−/− mice. Accordingly, histological analysis of livers from CD3ε−/− infected mice before death (day 22) showed typical features of active hepatitis, with many large granulomas mainly composed of histiocytic cells associated with multiple apoptotic hepatocytes (Fig. 3B). In contrast, only a few small granulomas were observed in TCRα−/− mice livers at that time point. Furthermore, CD3ε−/− mice presented an active pneumopathy with large granulomas and hemorrhagic foci at day 22, while TCRα−/− lung histology was close to normal with only a slight increase of inflammatory cells in the inter-alveolar septa (Fig. 3B). In conclusion, CD3ε−/− mice showed clear evidences of both liver and lung diseases 3 weeks post MCMV infection, in agreement with the high viral loads found at that time in these organs. In contrast, liver and lung disorders were not observed in TCRα−/− mice, emphasizing the ability of γδ T cells to control MCMV infection and associated organ disease. Whether γδ T cells limit organ disease only as a consequence of viral replication control or also by producing mediators of tissue repair deserves further attention. We next sought to analyze whether the control of MCMV spread was associated with an amplification of γδ T cells in infected organs. S2 Fig. shows the gating strategy used for γδ T cell flow cytometry analysis. After a slight decrease at day 3, γδ T cell numbers increased importantly in the lungs until day 21 (approximately 8 fold), and this rise persisted until the end of the experiment. A significant but more modest and transient increase was also observed in the liver (approximately 2 fold from day 3 to 7). By contrast and to our surprise given their preponderance in gut intraepithelial lymphocytes, no significant variation of γδ T cells was observed in the intestine. In the spleen, γδ T cells levels remained stable until day 21 when they decreased (Fig. 4A). In conclusion, control of MCMV infection by γδ T cells in TCRα−/− mice is associated with a transient γδ T cell increase in the liver, and a delayed but strong and persistent expansion of γδ T cells in the lungs. We next asked whether γδ T cells responding to MCMV differentiate into effector-memory cells as we observed previously in humans [10,14]. After a transient decrease early post MCMV challenge, the proportion of effector memory (EM, CD44+CD62L−) γδ T cells increased in the spleen, liver and lungs concomitantly with a decrease of central memory (CM, CD44+CD62L+) γδ T cells. Effector memory γδ T cells reached more than 80% in the liver and lungs at day 56 (Fig. 4B and 4C). Consistent with the absence of variation in γδ T cell numbers in the intestine, no modification of γδ T cells phenotype could be observed in this organ. These results confirm that MCMV induces a marked response of γδ T cells in the lungs and liver, which is more modestly seen in the spleen and absent from the intestine. The subsets of murine γδ T lymphocytes expressing the Vγ1 or Vγ4 chains of the TCR predominate in the spleen, liver and lungs, whereas intestinal γδ T cells are almost exclusively Vγ7+ (nomenclature of Heilig and Tonegawa [29]). We assessed the quantity, repertoire and memory phenotype of these γδ T lymphocyte subsets in the liver, spleen and lungs. Not surprisingly, low proportions of Vγ1+ γδ T cells were found in the intestine (S2 Fig.). As observed in Fig. 5A, the expansion of γδ T cells in the lungs and liver after day 3 concerned mainly Vγ1+ but also Vγ4+ γδ T cells. Both subsets followed the kinetics of total γδ T cells (Fig. 4A). Analysis of subsets also showed a response of Vγ1+, but not Vγ4+ T cells, in the spleen (Fig. 4A and Fig. 5A). The proportion of EM cells among both Vγ1+ and Vγ4+ γδ T cells increased after day 3 in the lungs, liver and spleen (Fig. 5B). In contrast, Vγ7+ γδ T cell numbers/memory phenotype did not vary significantly upon MCMV infection (Fig. 5A and Fig. 5B), as could be expected from the analysis of the whole γδ T cell population in the intestine (Fig. 4A and Fig. 4C). The complementary-determining-region (CDR3)γ1 and CDR3γ4 length profile of liver, spleen and lung-derived γδ T cells were not different between uninfected and infected mice for 14 days (S3 Fig. and S4 Fig.), indicating that there were no major changes in these CDR3 repertoires after expansion. γδ T cells development in CD3ε−/− mice was reconstituted by bone marrow (BM) transfer experiments using TCRα−/− mice as donors (referred to as TCRα−/− > CD3ε−/− mice). This method allowed the generation of the BM-derived Vγ1+ and Vγ4+ γδ T cell subsets that were increased upon MCMV infection. Control BM transplants were also performed with TCRδ−/− donors (TCRδ−/− > CD3ε−/− mice) and with CD3ε+/− donors (CD3ε+/− > CD3ε−/− mice). γδ and/or αβ T cell reconstitution was allowed to establish for 3 months before MCMV infection of the mice. γδ T cell subset percentages were analyzed in blood from live mice throughout reconstitution (Fig. 6A). Two months after grafting, the percentages of blood γδ and/or αβ T cells (among total lymphocytes) had reached a plateau (Fig. 6A). The proportion of peripheral blood γδ T cells in CD3ε+/− > CD3ε−/− mice was lower than that found in TCRα−/− > CD3ε−/− mice (Fig. 6A, lower panel), in accordance with previous findings which showed that γδ T cells in TCRα−/− outnumbered γδ T cells in C57BL/6 mice [30]. When infected with MCMV at 3 months post-graft, TCRα−/− > CD3ε−/− mice survived MCMV infection as efficiently as CD3ε+/− > CD3ε−/− and TCRδ−/− > CD3ε−/− mice, in marked contrast with CD3ε−/− > CD3ε−/− mice (Fig. 6B). In a second experimental scenario γδ T cells were purified from TCRα−/− splenocytes and injected intravenously (i.v.) into CD3ε−/− hosts one day before MCMV infection. Surprisingly, very low protection was obtained when γδ T cells isolated from control mice were transferred, whereas γδ T cells from MCMV-infected mice conferred good protection (Fig. 6C). All together our results confirm the protective anti-CMV role of BM-derived γδ T cells, and show that priming of splenic γδ T cells with MCMV in donor mice is necessary for protection against MCMV after their adoptive transfer. We next sought to gain insight into the mechanism by which γδ T cells exert their antiviral function. CD27 expression was shown to segregate γδ T cells into two functional subsets in mice: CD27+ γδ T cells being the main producers of the antiviral cytokine IFNγ and CD27− γδ T cells being prone to secrete IL-17A which is not classically considered as important in antiviral responses [31] [32]. To determine which of these subsets respond to CMV, we analyzed their evolution in organs from MCMV-infected mice. As evidenced in S5A Fig., CD27− cells dominated the γδ T cell response in the lungs, while CD27+ and CD27− subtypes were roughly equally implicated in the liver. However, IL-17A transcripts were barely detected in these organs (S5B Fig.). By contrast, IFNγ was expressed in both these organs but noticeably peaked as early as day 3, before the rise of γδ T cell numbers and Cδ transcripts (S5B Fig.). Since the presence of IFNγ transcripts in organs from TCRα−/− infected mice could be attributed to NK cells, we determined IFNγ production at the cellular level by intracellular staining of lymphocytes and using the gating strategy shown in S6 Fig. As shown in Fig. 7A, the proportion of IFNγ-producing NK cells peaked at day 3 in all organs. IFNγ-producing γδ T cells also peaked 3 to 7 days post-infection (Fig. 7A), but represented a minor population when compared to IFNγ-producing NK cells at similar time points (Fig. 7B). Consequently, NK cells were the largely preponderant producers of IFNγ during early acute MCMV infection (Fig. 7B), accounting for 2% of lymphocytes at day 3 in the liver and lungs (i.e. when the relative expression of IFNγ was the highest, S5B Fig.). Similarly, during the course of infection, the proportions of CD107a+ NK cells were higher than that of CD107a+ γδ T lymphocytes (S7 Fig.). These experiments are in accord with the substantial role of NK cells in the control of early MCMV infection through IFNγ production and cytotoxicity [33], and suggest that the antiviral role of γδ T cells might not principally rely on these two functions. Considering the above results we hypothesized that γδ T cells could exert an indirect antiviral effect by promoting NK cells accumulation as has been previously reported [34]. We therefore compared the evolution of NK cell numbers early post-MCMV infection in TCRα−/− and CD3ε−/− mice. For both mouse lines and as depicted in C57BL/6 wt mice, the overall kinetic was organ-specific with an early decrease of NK cells in the spleen in contrast to liver (Fig. 8A) [35][36]. In contrast to our hypothesis and despite the MCMV-induced death of CD3ε−/− mice, NK cell numbers were globally higher in CD3ε−/− mice than in TCRα−/− mice at all early time points tested (Fig. 8A), showing that γδ T cells antiviral activity was not due to an early increase of NK cells. In addition, when transferred into B/NK/T cells immunodeficient Rag−/−γc−/− mice, MCMV-primed γδ T cells were also strikingly sufficient to long term protect these mice from death (Fig. 8B). At day 56, γδ T cells could easily be detected in the liver, spleen and lungs of Rag−/−γc−/− recipient mice in contrast to NK cells, demonstrating that the protective function of γδ T cells could act in the total absence of NK cells (Fig. 8C). Previous work conveys compelling evidence for the implication of human Vδ2neg γδ T cells in the immune response against HCMV infection [5,6,7,9]. However, key questions that cannot easily be addressed in humans remain unanswered, such as the spatial and temporal regulation of the anti-HCMV γδ T cell response and its protective role. Because of its similarity with the human CMV pathogenesis and immune response, the mouse model of MCMV infection has been extensively used and is well characterized. The goal of this study was to take advantage of this model to address these questions concerning the protective role and localization of the γδ T cell response. Herein, we show that γδ T cells are as competent as αβ T cells to protect against CMV challenge, a finding that can be of particular relevance in clinical settings, situations or diseases where αβ T lymphocytes are compromised (hypomorphic Rag1 mutations, individuals treated with immunosuppressive drugs, foetuses or neonates, …) and where γδ T cells have already been shown to expand [6,7,8,9,10,11,12]. This protective function of γδ T cells, under conditions of suboptimal αβ T cell response, has previously been observed earlier in mice in the context of infection by Herpes Simplex Virus type 1 (HSV-1) [37] or by the gut coccidian parasite Eimeria vermiformis [38]. These results also corroborate the conserved level of protection against infection observed in patients lacking TCR αβ T cells due to a mutation in the gene coding the TCR α chain [39]. Since γδ T cells have been shown to play an important role in young mice in other infectious models, it would be interesting to evaluate this role in the context of MCMV infection [40]. In addition to extending our results to more a “natural setting” of suboptimal αβ T cells responses, it would allow analysis of the role of non BM-derived γδ T cell subtypes [41]. Finally this MCMV model could be used to evaluate the importance of γδ versus αβ T cells in the context of immunosuppression as used in transplant recipients. After administration of MCMV via the intraperitoneal route, MCMV targets the liver and spleen as cell free viruses within the first hours before dissemination to the other organs [42]. Accordingly, viral loads were the highest at day 3 in the liver and spleen but peaked at day 7 in the lungs and intestine in all mouse lines tested in the present study. In TCRα−/− mice, viral loads were the lowest at day 14 in the liver and spleen and at day 21 in the lungs (Fig. 2), i.e. after the significant increase of both Vγ1+ and Vγ4+ γδ T cell subsets in the liver and lungs (Fig. 4A), and of Vγ1+ γδ T cells in the spleen (Fig. 5A). Three weeks post-MCMV infection, high viral loads and liver/lung injury were evidenced in CD3ε−/− mice despite normal development and function of NK cells in these mice [28]. In contrast, liver and lung disorders were not observed in TCRα−/− mice at that time. These results are consistent with a role for γδ T cell response/expansion in these organs to control virus multiplication and associated organ damage in the absence of αβ T cells. The protective role of γδ T cells was ascertained by reconstituting γδ T cells in CD3ε−/− mice by bone marrow transplantation, or by adoptive transfer of splenic γδ T cells from TCRα−/− MCMV infected mice. However, when isolated from the spleen of TCRα−/− uninfected mice, γδ T cells were inefficient to induce protection in CD3ε−/− recipients. We can exclude the possibility that lack of protection in CD3ε−/− mice which received naïve γδ T cells was due to an absence of engraftment, because both naïve and MCMV-primed γδ T cells were found in the liver, spleen and lungs of recipient mice (S8 Fig.). The absence of protection by non-primed γδ T cells purified from splenocytes may be due to a delay of reconstitution/differentiation in recipient mice that allow the virus to overwhelm the γδ T cell response. Infection of donor mice by CMV most likely prime γδ splenocytes to readily respond to CMV once transferred in CD3ε−/− mice, compensating this reconstitution limitation. The development of the anti-CMV immune response involves a complex network of cells from the innate and adaptive immunity that act sequentially to favor health over disease. Research in mice has paid a lot of attention to the early control of MCMV by NK cells, which are responsible for the enhanced resistance of the C57BL/6 mouse strain when compared to BALBc mice. In C57BL/6 mice, NK cell antiviral activity relies on both perforin and IFNγ-release that control viral loads in the liver, spleen and lungs [33,43]. Our ex vivo analysis of lymphocytes from C57BL/6 TCRα−/− infected organs show that the early boost (days 3–7) of IFNγ expression and cytotoxic granule exocytosis is mostly due to NK cells, while γδ T cells participate only modestly to these functions (Fig. 7 and S7 Fig.). Thus, although we cannot exclude that this modest contribution might help in controlling MCMV loads, these results rather raise the possibility that γδ T cells operate either by regulating other immune cells or through the production of unknown antiviral mediators. Strikingly, however, our adoptive transfer experiment into Rag−/−γc−/− immunodeficient hosts showed that γδ T cell antiviral protective function can be independent of NK/B/αβ T cells. This emphasizes their efficiency and opens interesting perspectives for their possible manipulation in clinical situations where other immune cells are defective. The kinetics of γδ T cell response was organ specific, with a progressive increase and accumulation of γδ T cells in the lungs, whereas γδ T cells quickly increased and dropped at day 21 in the liver and spleen (Fig. 4A). The persistence within the lungs of memory γδ T cells contrasts with the transient increase of pulmonary γδ T cells that was observed in other murine infectious contexts [44,45,46]. However it reproduces the persistence of γδ T cell expansion in human blood during HCMV-infection which could result from persistent activation of γδ T cells in chronically infected tissues [5,10]. This suggests that the lungs could be an anatomical site for replication of HCMV and chronic activation of γδ T cells, consistent with the fact that HCMV is frequently found in lungs of solid organ transplant patients where it can induce tissue invasive disease [4]. The γδ T cell response to MCMV implicates bone marrow derived Vγ1+ and Vγ4+ T cells. It will be interesting in the future to determine whether these subsets play similar functions in the response to MCMV, since evidence for distinct roles of Vγ1+ and Vγ4+ T cells in the protection and/or pathogenesis during infection of mice has been reported [46,47,48]. The involvement of several subsets in the response to MCMV is in agreement with the implication of diverse Vδ2neg T cell subsets (Vδ1, Vδ3, Vδ5) in the response to HCMV [5]. In contrast to long term HCMV-induced γδ T cells that display a restricted CDR3δ length repertoire [5], the CDR3γ1 and γ4 length repertoire of liver, spleen and lung-derived γδ T cells was equivalent in 14-days MCMV-infected and uninfected TCRα−/− mice (S3 Fig. and S4 Fig.). This could reflect a TCR-independent innate-like response of γδ T cells and/or high frequencies of MCMV-specific γδ T cells already existing in naïve mice. However, we cannot exclude the presence of a shared antigen-recognition motif in CDR3γ of different lengths (as observed for the CDR3δ of T22-specific γδ T cells [49]). The number of CDR3γ1 peaks (4 or 5) confirms previous analysis of CDR3 repertoire in mice [50]. Another interesting question concerns the memory function of γδ T cells during MCMV infection, as recently described for CD44+CD27− γδ T cells in mouse models of bacterial infections [51,52]. Adaptive and innate like γδ T cells could both participate to memory, in light of the emerging role for innate cells in this context [53]. Previous contact with HCMV induced a rapid recall expansion of effector memory Vδ2neg γδ T cells, which coincided with better infection resolution of HCMV reactivation in transplant recipients [10]. CMV infection in mice also induces CD44+CD62L− effector memory γδ T cells that are maintained and outnumber CD44+CD62L+ central memory γδ T cells at day 56 in all organs (Fig. 4B and Fig. 4C). By definition, effector memory cells are prone to exert rapid functions at the aggression site and the results shown here support the hypothesis that peripheral blood effector-memory human Vδ2neg γδ T cells are re-circulating cells that originate from CMV-targeted organs. It remains to be investigated whether murine γδ T cells recognize self-encoded stress-regulated antigens on CMV-infected cells, as demonstrated for human γδ T cells [18]. Acute infections with HCMV can result in serious disease in infected neonates and in the context of immunosuppression linked to transplantation. Inducing or enhancing the antiviral response of γδ T cells in this context is an attractive objective. Our findings open new perspectives for the use of the murine model of MCMV infection to define the precise mechanism of antiviral activity of γδ T cells and to develop new strategies to induce their activation in vivo. Their absence of MHC restriction, their combination of conventional adaptive and innate-like responses, their particular anatomical localization and their dual reactivity against infected and tumor cells, are specific features that place γδ T cells as unique effectors for clinical manipulation. In conjunction with the identification of stress antigens recognized by γδ T cells on infected cells, these results open new avenues for clinical manipulation of γδ T cells against CMV-mediated disease. All experimental procedures involving animals were conducted according to European Union guidelines (European Directive 2010/63/UE) (http://ec.europa.eu/environment/chemical​s/lab_animals/home_en.htm) and approved by the local ethics committee: Comité d'éthique pour l'expérimentation animale de Bordeaux (CE50), [project n° 50120197-A]. We used C57BL/6 mice. CD3ε−/− [54], TCRα−/− [30] and Rag−/−γc−/− mice [55] were from the CDTA (Centre de Distribution, Typage et Archivage Animal, Orléans, France). TCRδ−/− [56] were a gift from Dr Malissen (Centre d’Immunologie de Marseille Luminy, France). Mice were used between 8–12 weeks of age and kept under pathogen-free conditions (Animalerie spécialisée, Université Bordeaux Segalen, France). CD3ε−/− were bred to C57BL/6 mice (C57BL/6J, Charles Rivers laboratory, Larbresle, France) to obtain CD3ε+/− control mice. MCMV-infection was performed in an appropriate animal facility (Animalerie A2, Université Bordeaux Segalen, France). MCMV was acquired from the American Type Culture Collection (Smith strain, ATCC VR-194) and propagated into BALBc mice (BALBcBy/J, Charles Rivers laboratory, Larbresle, France) to generate MCMV salivary gland extracts. Virus titers were defined by standard plaque assay on monolayers of mouse embryonic fibroblasts (MEF). Unless indicated, infections were performed by i.p. administration of 2.103 PFU of the salivary gland viral stock. Mice were bled via the retroorbital sinus after anesthesia (one eye every other week) and the serums collected and frozen. AST and ALT were quantified using standard enzymological methods (laboratoire de Biochimie, CHU Bordeaux, France). Mice were euthanized by cervical dislocation. Liver and lungs were removed, fixed for 24 h in 3.7% neutral-buffered formalin (Sigma-Aldrich), followed by standard histological processing and paraffin embedding. Sections of 4 μm thickness were processed for Hematoxylin/Eosin/Safran (HES) staining (following standard protocols). Genomic DNA was isolated from organs using Nucleospin tissue purification kit (Macherey Nagel). Real time PCR to quantify MCMV was performed in Step one plus thermocycler (Applied biosystem) using GoTaq qPCR Master Mix (Promega) with primers specific for MCMV glycoprotein B (gB) (gi330510, forward primer: AGGCCGGTCGAGTACTTCTT and reverse primer: GCGCGGAGTATCAATAGAGC). Known quantities of plasmid comprising MCMV gB were used for the titration curve. Total RNA from immune cells was prepared with Nucleospin RNAII kit (Macherey Nagel). Goscript reverse transcriptase (Promega) was used to generate cDNA. Real time PCR was performed in CFX 384 (BioRad). The relative expression of transcripts was determined using the GAPDH reference gene. For spectratyping analysis, PCR (40 cycles) was performed with Vγ1 and Cγ4 or with Vγ4 and Cγ1 primers, resulting in amplification of the sequences containing the CDR3γ1 or CDR3γ4, respectively. Then a run-off reaction (one cycle) was performed using a fluorescently labeled Jγ4-FAM primer for CDR3γ1 and with a Jγ1-FAM primer for CDR3γ4 (primers sequences from [50]). The labeled reaction products were run on a capillary sequencer (ABI3730xl analyzer) at ImmuneHealth (Gosselies, Belgium). The fluorescence intensity was analyzed using Peak Scanner 1.0 (Applied Biosystems). List of primer Fw (forward) and Rv (Reverse): GAPDH (Genbank NM_008084): Fw 5’-AATGGGGTGAGGCCGGTGCT-3’ Rv 5’-CACCCTTCAAGTGGGCCCCG-3’ IFNγ (NM_008337.3) Fw: 5’-ACTGGCAAAAGGATGGTGAC-3’ Rv 5’-TGAGCTCATTGAATGCTTGG-3’ IL17-A (NM_010552.3) Fw 5’-TCATCTGTGTCTCTGATGCTGTT-3’ Rv 5’-TTGGACACGCTGAGCTTTGA-3’ Cδ (X12729.1) Fw 5’-CTGTGCACTCGACTGACTTTGAACC-3’ Rv 5’-CCCAGCACCGTGAGGGACATC-3’ CDR3γ1 Fw Vγγ1 5'-CCGGCAAAAAGCAAAAAAGT-3 Rv Cγ4 5’-AAGGAGACAAAGGTAGGTCCCAGC-3’ Jγ4-FAM 5'-TACGAGCTTTGTCCCTTTG-3' CDR3γ4 Fw Vγ4 5’-CTTGCAACCCCTACCCATAT-3’ Rv Cγ1 5’-CCACCACTCGTTTCTTTAGG-3’ Jγ1-FAM 5'-CTTAGTTCCTTCTGCAAATACC-3’ We used cell strainers to mash the spleens and livers in RPMI-1640 with 8% FBS; red blood cells were lysed with NH4Cl. For the liver, immune cells were isolated by centrifugation (2000 rpm, 20 min) over a 40/80% discontinuous Percoll gradient (GE Healthcare). Pulmonary mononuclear cells were isolated as described [57]. Intestinal intraepithelial mononuclear cells were isolated as described elsewhere [58]. Total organ live cells (unstained with Trypan blue) were then counted using a hemocytometer (Malassez chamber). The proportion of γδ T cells (CD3ε+panγδ+) and NK cells (NK1.1+NKp46+) among total organ live cells (7AAD−) was evaluated by FACS using a large FSC/SSC gate that included all cells but debris. This proportion was then multiplied by total organ cell number to obtain the absolute number of γδ T cells and NK cells. The following monoclonal antibodies were from BD Pharmingen: anti-CD3ε (145–2C11), anti-TCRδ (GL3), anti-CD44 (IM7), anti-CD62L (MEL-14), anti-CD27 (LG.3A10), anti-NK1.1 (PK136) and anti-NKp46 (29A1.4). Anti-IFNγ (XMG1.2), anti-CD107a (1D4B) and respective isotype control mAbs: Rat IgG1κ (eBRG1) and Rat IgG2aκ (eBR2a) were purchased from eBioscience. Anti-Vγ1 (2.11), anti-Vγ4 (49.2) and anti-Vγ7 (F2.64) mAbs were a kind gift from P. Pereira (Institut Pasteur, Paris). For flow cytometry analysis, immune cells were first incubated with anti-mouse CD16/32 (eBioscience) and stained with relevant antibodies and 7-AAD (BD Pharmingen). Fixed cells were acquired using a LSRFortessa (BD Biosciences), and analyzed using FlowJo software (Tree Star). For intracellular IFNγ staining, cells were incubated in complete medium for 2h at 37°C; 10μg/ml of Brefeldin A (Sigma-Aldrich) was added during the last hour. Intracellular staining was performed after cell surface staining, using BD Cytofix/Cytoperm Fixation/Permeabilization Kit and according to the manufacturer’s instruction (BD Biosciences). For CD107a staining, cells were incubated in complete medium for 2h at 37°C; 10μg/ml Brefeldin A (Sigma-Aldrich) and anti-CD107a or isotype control mAb were added during the last hour. Cells were then stained with relevant monoclonal antibodies. Mice femora and tibia from CD3ε+/−, TCRα−/−, TCRδ−/− and CD3ε−/− were isolated and the BM was flushed with 1 ml of IMDM with FBS (1%). BM cells from one donor were injected to one CD3ε−/− mice (8–10 per group), intravenously (i.v.) through the retrobulbar sinus in a volume of 0.2 mL IMDM. Mice were conditioned by i.p. injections of Busulfan 22.5 mg/kg (Pierre Fabre laboratory) two days and one day prior to transplantation [59]. 10 TCRα−/− mice were uninfected, or 14 days infected with 2.103 PFU of MCMV. Immune cells were prepared from spleens and pooled before γδ T cell sorting using the TCRγ/δ+ T Cell Isolation kit (Miltenyi Biotec). Purity was verified by flow cytometry and 8.105 to 1.106 γδ T cells i.v. transferred into CD3ε−/− or Rag−/−γc−/− recipients. 24h after γδ T cell transfer, recipient mice were infected i.p. with 2.103 PFU of MCMV and followed daily. 2–3 months after infection, recipient mice were sacrificed to verify the presence of γδ/NK cells in organs. Differences were evaluated by the Mann-Whitney test and represented as follows: * = p<0.05, ** = p<0.01, *** = p<0.001, **** = p<0.0001.

10.1371/journal.ppat.1001021

Plasmodium falciparum Adhesion on Human Brain Microvascular Endothelial Cells Involves Transmigration-Like Cup Formation and Induces Opening of Intercellular Junctions

Cerebral malaria, a major cause of death during malaria infection, is characterised by the sequestration of infected red blood cells (IRBC) in brain microvessels. Most of the molecules implicated in the adhesion of IRBC on endothelial cells (EC) are already described; however, the structure of the IRBC/EC junction and the impact of this adhesion on the EC are poorly understood. We analysed this interaction using human brain microvascular EC monolayers co-cultured with IRBC. Our study demonstrates the transfer of material from the IRBC to the brain EC plasma membrane in a trogocytosis-like process, followed by a TNF-enhanced IRBC engulfing process. Upon IRBC/EC binding, parasite antigens are transferred to early endosomes in the EC, in a cytoskeleton-dependent process. This is associated with the opening of the intercellular junctions. The transfer of IRBC antigens can thus transform EC into a target for the immune response and contribute to the profound EC alterations, including peri-vascular oedema, associated with cerebral malaria.

Cerebral malaria, a major cause of death during malaria infection, is characterised by the sequestration of infected red blood cells (IRBC) in brain microvessels. This study describes the interactions between plasmodium infected red blood cell and human brain endothelial cells. It highlights the activation of a trogocytosis-like mechanism followed by an engulfment of the infected red blood cells by endothelial cells (EC). This transfer concerns up to 20% of the IRBC cocultured with EC. This means that the parasite infected erythrocyte can mimic the leukocytes interaction with endothelial cells. This process is associated with i) a transfer of malaria antigens to the EC which can inappropriately activate the immune system and ii) an opening of the intercellular junctions, which can trigger blood-brain-barrier leakage during cerebral malaria. This transfer of IRBC antigens can thus transform EC into a target for the immune response and contribute to cerebral malaria pathogenesis.

Each year 3.2 billion people worldwide are exposed to the threat of malaria, resulting in around 2 million deaths [1]. Even with the best antiparasitic treatments, patients with cerebral malaria (CM) have no significant improvement in their prognosis, with an average fatality rate of 30 to 50% [1]. This outcome is due to a neurovascular pathology characterised by the accumulation of both infected red blood cells (IRBC) and host cells (leucocytes and platelets) in deep brain microvessels, leading to microcirculation impairment [2]. Leakage of the blood-brain-barrier and local arrest of leucocytes is associated with cytoadhesion of IRBC and microcirculation impairment, increased blood volume due to sequestration, and increased blood flow resulting from seizures and anaemia [2]. Cytokines and parasite toxins have also been shown to cause direct damage to the blood-brain barrier [3], [4]. On the whole, severe increased intracranial pressure [2] and brain oedema [5] are associated with poor outcome for patients with CM. Of note however, the sequestration of IRBC in deep vessels is a normal step in the life cycle of the Plasmodium parasite and does not always trigger severe disease. Most of the molecules implicated in the adhesion of IRBC on EC have previously been described. They involve a series of endothelial molecules and parasite membrane proteins such as PfEMP1 and RIFINS [6]. This cytoadhesion helps the parasite avoid splenic clearance and favours its development in a low oxygen pressure microenvironment. Specific localisation of the parasites in the brain seems to be a complex feature involving both expression of human adhesion molecule isoforms and parasitic var proteins polymorphism [7]. However, the fine mechanisms of microvascular endothelial cell alterations are not yet fully understood. Cellular interactions in microvessels have mostly been described in the context of transmigration of cells through the endothelial layer. Such transmigration occurs through either openings between adjacent endothelial cells or through a single endothelial cell, mainly when leukocytes migrate to tissues or during the diffusion of metastatic cells. This is a stepwise process involving rolling, adhesion, firm adhesion, and finally diapedesis. The formation of a docking structure or “transmigratory cup” was recently described and involves several endothelial adhesion and signalling molecules [8]–[10]. Human lymphocytes use another process and palpate the surface of EC with podosomes before forming transcellular pores through the endothelium [11]. Yet another mechanism of cell-cell interaction, named trogocytosis, was more recently described in mice and human immune cells, but not in EC. Trogocytosis involves the transfer of membrane compounds during short term cell interactions [12]–[13], and is defined as the uptake of membrane fragments and associated molecules from one cell to another. Moreover cell interaction by any of the above mechanisms can activate EC. Therefore, it is essential to understand the binding mechanism and the repercussions of the IRBC adhesion on EC in order to develop new preventive or therapeutic interventions for the treatment of cerebral malaria treatment. The present study demonstrates, that IRBC undergo close association with EC in a manner reminiscent of both trogocytosis and transmigration. In this in vitro system, adhesion and transfer of material involved 10 to 20% of the IRBC in contact with the EC. This process involves engagement of ICAM-1, or other EC adhesion molecules, in the binding of IRBC. It triggers transfer of membrane material and malaria antigens from the IRBC to the brain EC in a trogocytosis-like manner and in a second step the development of a transmigration cup-like structure, which results in EC activation and opening of the intercellular junctions. Upon incubation with human brain EC (HBEC 5i and hCMEC/D3 lines), IRBC bound tightly onto the endothelial surface. To study the transfer of IRBC constituents to EC we fluorescently-labelled both the IRBC membrane, using the membrane-intercalating agents PKH26 or PKH67, and the cytoplasm, using calcein-AM. We then followed IRBC material transfer by confocal microscopy and estimated the amount of fluorescence incorporated into the EC. As seen in Fig. 1, after 30 min, IRBC were seen attached to hCMEC/D3, with a halo-like diffusion from the IRBC PKH-labelled elements present on the HBEC surface around the area of attachment of IRBC (Fig. 1A–C). This process was enhanced after one hour of co-incubation, with a pattern of dense punctuate patches visible on the HBEC surface after 90 min (Fig. 1D)., After 3 h the fluorescent dye had migrated into the HBEC away from the original IRBC attachment point (Fig. 1E). The labelled elements transferred from IRBC were detected on the HBEC surface up to 24 h after their co-incubation, even if the IRBC had been removed from the culture at 3 h post-incubation. On the other hand, calcein remained localised within the IRBC for up to 1 h after incubation with HBEC (Fig. 1D). Calcein-labelled IRBC could still be detected as globular shapes on the HBEC, after 90 min of co-culture. However, after 3 h of incubation, the green dye was detectable in the HBEC far from the IRBC docking area (Fig. 1E). Time-dependent transfer of PKH67 dye to HBEC was quantified by the amount of fluorescence present in HBEC after extensive washing of the endothelial monolayer to remove any unbound IRBC (Fig. 2A). This transfer was significantly enhanced by TNF stimulation of the HBEC prior to incubation with IRBC (Fig. 2 A–B). No dye transfer was detected when HBEC were incubated either with fluorescent non-infected RBC (Fig. 2 A) or when the labelled IRBC were mildly trypsinised prior to incubation with HBEC (Data not shown, DNS). We assessed cytoskeletal involvement in this transfer process by incubating the HBEC with optimal concentrations of cytoskeleton mobilisation inhibitors prior to co-culture with IRBC, followed by quantitation of the transferred fluorescence. Nocodazole (1 or 10 µM), a microtubule stabiliser, and cytochalasin D (20 or 200 µM), an actin reorganisation inhibitor, were the most effective in blocking the transfer of IRBC membrane elements, reaching a 6-fold inhibition (Fig. 3). Amiloride (5 or 50 µM), a micropinocytosis inhibitor, was less effective at blocking the transfer of IRBC membrane elements. We attempted to identify the cellular compartment involved in the IRBC material transfer to the HBEC by selectively labelling early endosomes (EE), lysosomes or clathrin-coated pits. After less than 90 min of incubation, only the surface of HBEC appeared labelled with PKH26, in a trogocytosis-like process (Fig. 4 A). At longer incubation times (90 min to 3 h) PKH26 was partially detected in EE (Fig. 4 B) but not in either lysosomes or clathrin-coated vesicles (DNS). We evaluated the transfer of malaria antigens upon adhesion of IRBC to HBECs using a pool of human adult immune serum (HIS) originating from a dispensary in South Senegal, an area where malaria transmission is mesoendemic and where adults develop premunition against malaria. A pool of serum from age-matched adults from non malaria endemic areas was used as control. Western blot analyses indicated that the HIS pool recognised parasite antigens as well as parasite-encoded erythrocyte membrane proteins. In contrast, HIS could not detect any malaria antigens in proteins extracted from HBEC containing fluorescent compounds transferred from IRBC. This could be due either to sensitivity, with insufficient amounts of parasite proteins transferred to the HBEC to allow detection by Western blot or to degradation of the parasite antigens during the cell-to-cell transfer. However, using immunofluorescence deconvolution microscopy, HIS readily detected parasite proteins transferred to HBEC. It strongly detected the IRBC, but not the non-infected RBC on the HBEC surface (Fig. 5). This labelling can be detected even without any permeabilization of the cell membrane with triton. After 30 to 90 min of co-culture, HIS revealed a malarial antigen pattern on HBEC closely related to that obtained with PKH26-labelled IRBC (Fig. 5 A–C). However, after a 2 h co-culture, PKH-labelled elements and malarial antigens could clearly be located in different areas on the surface of HBEC (Fig. 5 A–C and Fig. 5F). This disparity may result from the different metabolic pathways to which the lipids and proteins transferred from the IRBC are subjected. At 24 h post-incubation, malarial antigens were still detected inside HBEC (Fig. 5 D–E). However there was no longer any detectable labelling of the HBEC plasma membrane as it is not detectable without permeabilization of the cell membrane with triton data not shown). We also used HIS to evaluate the role of IRBC-surface antigens in IRBC/HBEC adhesion with a view to determine the putative role of the antibodies present in immune sera in protection against CM. Pre-incubation of IRBC with HIS, but not with non-immune serum, abrogated IRBC adhesion on HBEC as well as the transfer of PKH26-labelled material, as observed by both microscopy (data not shown) and by fluorescence quantification (Fig. 2 B–C). As expected this pre-incubation had no effect on the adhesion of non-infected RBC used as control (Fig. 2C). This observation strongly supports a role for humoral immunity in the protection against the IRBC/HBEC binding process. At present, the fine IRBC/EC contact structure had not been totally described. We conducted a detailed study of the IRBC/EC adhesion area using two HBEC cells lines, named 5i and hCMEC/D3 which presented marked differences in their surface morphology in standard culture conditions. HBEC-5i have defined surface structures, such as microvilli, podocytes and cups (Fig. 6), close to those described in vivo [14]–[15]. In the opposite way, as described by Weksler (2005), cultured hCMEC/D3 present a very smooth surface (Fig. 7A). Structures found on the 5i surface in resting condition could be related to a partial activation of these cells, as suggested by their high basal level of ICAM-1 expression. Overnight incubation of the HBEC 5i cells with 10 to 100 ng/ml of TNF induced dramatic changes in these structures, with an enlargement of the microvilli into leaf-like structures (at 10 ng/ml) (Fig. 6C–D) and to finger-like structures (at 100 ng/ml) (Fig. 6E). The highest TNF concentration induced large areas of bubbling microparticles, as described earlier [16]–[17] (Fig. 6F). We observed differences in the first contact between IRBC and HBEC according to the type of HBEC lines used and the type of surface involved. In the case of HBEC-5i, IRBC adhesion occurred on the microvilli, in a capture-like process (Fig. 7B–C). The subsequent engulfing structure appeared to be closely related to that described for leukocytes (Fig. 7D–I). The microvilli were progressively emitted from the surface at the same time as a transmigration cup-like structure was formed (Fig. 7D–E). This cup-like structure progressively covered and engulfed the IRBC (Fig. 7F–I). Fig. 8C–D shows a confocal image of IRBCs in these cup-like structures. FRC 3Ci is a parasite strain selected for its ability to bind to CD36 and ICAM-1, whereas the CS2 strain mainly binds to CSA. hCMEC/D3 are known to poorly express CD36, a fact we confirmed by flow cytometry (DNS). However both hCMEC/D3 and -5i are known to express ICAM-1 on their plasma membrane, especially following stimulation by TNF [17]. We analysed whether actin, ICAM-1 and/or VCAM-1 play an active role in this engulfing process by imaging their distribution during the co-culture. Actin displayed a sub-membranous labelling pattern (Fig. 8E), originally described as stress fibres [18]. However, a crown of actin was also detected, preferentially located around the parasite in the digitations of the cup (Fig. 8C–E). ICAM-1 displayed a pattern of membrane folds on HBEC or at the border of cells when there are not fully confluent (Fig. 8A–B). At higher magnification, was also detected at the bottom of the cup-like structures, under the IRBC and forming part of the cup digitations themselves (Fig. 8A–D). While VCAM-1 was weakly expressed on the HBEC surface, a definite labelling was found at the bottom but not on the borders of the of cups (Fig. 8F). In combination, these results suggest that IRBC engulfment is a process closely related to leukocyte transmigration, which involves adhesion to ICAM-1. On the IRBC side, membrane molecules such as PfEMP1 are likely to be involved in the binding to ICAM-1, and their removal by trypsin would explain the abolishment of the binding to HBEC. Consistent with this, hCMEC/D3 incubated with anti-ICAM-1, but not with anti-VCAM-1, antibody prior to co-culture with IRBC displayed reduced binding of 3Ci, but not of CS2 IRBC (Fig. 2C). However, other adhesion molecules could be involved in this EC/IRBC binding which could explain this partial inhibition. A major event driving brain pathophysiology during malaria infection is the opening of the blood-brain barrier (BBB). We used hCMEC/D3, capable of forming a monolayer and establishing tight intercellular junctions involving VE-cadherin and ZO-1 [19], to analyze the signals responsible for opening of intercellular junctions. Electric cell–substrate impedance sensing (ECIS) of the HBEC monolayer was used to assess trans-endothelial electrical resistance (TEER), which reflects opening of the intercellular junctions. Experiments were done more than ten times, but only illustrations are showed in figures as summarizing ECIS data is not accurate. Four days after seeding, impedance of the monolayer was stable, with only minor fluctuations (see “control” Fig. 9A–D). Overnight pre-activation of the confluent monolayer with 10 ng TNF/ml did not cause any modification in the TEER (DNS). Histamine (100 µM) was used as a positive control and induced a two-phase junction opening process, with a rapid decrease in TEER, within 30 min of addition, followed by a slow and steady TEER decrease, lasting for several hours (Fig. 9B). We found that 3Ci-IRBC, but not non-infected RBC, induced a opening of the junctions after 2 h of incubation (Fig. 9A). This opening depended on parasiteamia and lasted for over 24 h, even if the RBCs were carefully removed after 4 h of incubation (Fig. 9A). In contrast, CS2 IRBC were only capable of inducing a minor decrease in the monolayer's TEER (Fig. 9B). When the RBCs were not removed after just 4 h but co-cultured with the HBEC monolayer overnight, they induced a slight decrease in TEER which was apparent even with non-infected RBC (DNS). This effect may be due to changes in the culture medium composition due to RBC lysis. We then proceeded to use the ECIS to test compounds that would inhibit the junction-opening effect of the 3Ci IRBC. Pre-incubation of the HBEC monolayer with 10 µM nocodazole abrogated the IRBC-induced junction opening (Fig. 9C). In contrast, 10 µM rolipram, a compound known to strengthen junctions by maintaining/stimulating cAMP signalling, had no effect on 3Ci IRBC-induced opening of the junctions (Fig. 9D). Pre-incubation of HBEC with anti-ICAM-1 antibodies (10 µg/ml) had no effect on monolayer TEER (DNS), but partially inhibited the opening of the junctions induced by 3Ci IRBC (Fig. 9E). This effect is consistent with the partial inhibition that the anti-ICAM-1 antibodies caused on the IRBC binding to HBEC. Here we demonstrate that the first step during the IRBC/HBEC interaction is a diffusion of IRBC membrane elements on the surface of HBEC, with features similar to those of trogocytosis. This process originally described during amoeba infection [20], is also used by all hemopoietic cells [21] and plays a major regulatory role in immunity. Both T and B cells acquire their antigens by trogocytosis, in the same way that Natural Killer (NK) cells modulate IL-4-polarised monocytes [22], regulatory T cells acquire their allo-antigens to kill syngeneic CD8 T cells [23], and CTLs capture membrane fragments from their targets [24]. Additionally, T cells, NK, gamma-delta T cells and monocytes use trogocytosis to interact with cancer cells [12], [25]. However, it was not until this year that this mechanism was encountered in non-immune cells when Waschbish et al found that the capture of myoblast membrane patches by T-cells occurred by trogocytosis [26]. Here we describe, for the first time, this trogocytosis-like interaction process between two non-immune cells, i.e. endothelial cells and red blood cells. Using human malaria-immune serum, we were able to demonstrate diffusion of malaria antigens from IRBC to HBEC during the early stage of trogocytosis. This process could result, during malaria infection, in the transfer of malaria antigens to HBEC during short interactions, such as the rolling of infected cells on the endothelium. It has also been shown that in T cells [27] trogocytosis requires actin polymerization and involves kinase signaling pathways. Our results, showing that both nocodazole and cytochalasin-D were capable of inhibiting this process, as well as the actin redistribution observed by microscopy, strongly support actin involvement in the HBEC/IRBC interaction. After 1 to 3 h of co-culture, the contact between IRBC and HBEC became tighter and involved the formation of an engulfing cup-like structure. The formation of the cup was morphologically related to the structure formed during leukocyte transmigration. We found reorganisation of actin in the protrusions of the cup as well as a localized concentration of ICAM-1, and, to a lesser extent, of VCAM-1, in the bottom of the cup. This engulfing process has been analyzed in depth for leukocyte transmigration. During this process, a cup is formed with projections surrounding the leukocyte [9]–[10]. These projections were enriched in actin, but not microtubules, and required both intracellular calcium mobilisation and intact microfilament and microtubule cytoskeletons [28]. Disruption of these projections with cytochalasin D or colchicine had no affect on the adhesion of leukocytes but affected the cup formation itself [29]. Importantly a similar engulfing process has already been described for pathogens such as bacteria [30]–[31] and yeasts [32]. Bacteria were previously described to aggregate, before engulfing, in a specific structure called “invasome” which is highly enriched in actin, ICAM-1 and phosphotyrosine [30]. Formation of the invasome was found to be inhibited by cytochalasin D but not by nocodazole. Here we describe a potentially similar cup-like formation and engulfing process during IRBC/HBEC interaction. However, the fact that both cytochalasin-D and nocodazole inhibited the transfer of material from IRBC to EC suggested that, in our case, there are most likely different steps involved in the interaction. The type of adhesion molecules involved in this interaction process depends on the type of cells interacting with the EC. For leukocytes, in response to LFA-1 engagement, the endothelium forms an ICAM-1-enriched cup-like structure that surrounds adherent leukocytes. Polymorphonuclear neutrophils use ICAM-1, but not VCAM-1, to move across activated EC monolayers [33]. On the other hand, the transmigration of THP-1 cells was reduced only when VCAM-1 or both ICAM-1 and VCAM-1 were blocked [34]. Similarly, a large panel of molecules is involved in the adhesion of IRBC on EC, including PECAM-1, CD36, chondroitin-sulphate A, ICAM-1, thrombospondin, αvβ3 E-selectin, P-selectin, and VCAM-1 [35]–[36]. The data we obtained for the two different parasite strains; 3Ci (which binds ICAM-1 and CD36) and CS2 (which mainly binds chondroitin-sulphate A), illustrates this diversity of these interactions. However, ICAM1 is a major adhesion molecule in HBEC largely increased during TNF stimulation, thus facilitating leukocytes migration through HBEC-D3 [33]. Additionally, upregulation of ICAM-1 has been described on brain microvessels in patients who died with CM [37]–[38] and correlates with adhesion of IRBC and the severity of attack in patients [39]. Taken together, our results provide new explanations for this increase in malaria pathology. After adhesion, IRBC were progressively engulfed in the EC monolayer and subsequently altered. As revealed by the use of human immune sera, our results highlight the diffusion of malarial antigens into HBEC. These antigens were first transferred onto the endothelial surface, in a trogocytosis-like process. Later on, the membrane dye, the cytosolic material of the red blood cells and malaria antigens were recycled in the endosomal compartment of the HBEC. Of note they were readily detected in these cells up to 24 h after incubation. As HBEC are known to be antigen-presenting cells [40], we attempted to detect the malaria antigens on their surface after a 24 h co-incubation. After this time, the anti-malaria serum pool very slightly detects malaria antigens on the HBEC surface, however this could be due to degradation of the IRBC proteins (and epitopes) by EC. The observed transfer of antigens to the EC involves dramatic implications for the interaction of EC with the immune system, as it could transform the EC into a new target for the immune response, especially during the rolling of immune cells on EC, and trigger major pathophysiological changes during CM. More experiments are required using immune cells from donors from endemic areas, to study the interaction between the immune system and the “IRBC-loaded” HBEC. As occurs with sepsis and viral infections, a major element in the pathophysiology of CM is the opening of the intercellular junctions. We report here that the opening of these junctions occurred shortly after the beginning of the IRBC/EC co-culture and was closely related to adhesion of the IRBC on the HBEC, as the non-infected RBC or non-binding IRBC, such as CS2, did not induce any significant junction opening. This also implies that neither metabolic modifications of the culture medium nor proteins secreted by the parasite were, in our culture conditions, sufficient to induce junction opening. A decrease in trans-endothelial electrical resistance (TEER) in the endothelial monolayer is well known during leukocyte transendothelial migration [41] and has also been suggested during in vivo infection of mice with P. berghei [42]–[44]. Three primary signaling pathways are activated by leukocyte adhesion: Rho GTPases, reactive oxygen species, and tyrosine phosphorylation of junctional proteins [45]. The pathways activated during both cup formation and opening of the EC junctions, need to be further explored for IRBC attachment and transmigration. ICAM-1 engagement by leukocytes has been shown to enhance trans-endothelial permeability by tyrosine phosphorylation of VE-cadherin [46]. However, in our results, treatment with anti-ICAM-1 antibodies had no effect on the increase of permeability induced by leukocytes on HBEC-D3, although it decreased leukocyte migration [33]. Interaction of ICAM-1 and VCAM-1 with adhesive molecules regulates the different steps of diapedesis by modulating either i) the GTPase pathway (Rho and Rac) [29], [34] ii) and the MAP-kinase pathway (Ca++, CaMKK, and AMPK) [47]. In our study, the opening of the EC junctions was independent from cAMP activation and suggested a MAP kinase pathway activation, as previously reported for P. falciparum [48]. However, the direct pathogenic effect of IRBC adhesion on the HBEC TEER must also be taken into account, as it up-regulates several TNF-superfamily genes and apoptosis-related genes such as Bad, Bax, caspase-3, SARP2, DFF45/ICAD, IFN-receptor2, Bcl-w, Bik, and iNOS [49]. This could account for the increase of permeability of the HBEC monolayer we observed after 6 or 8 h of IRBC/HBEC co-culture. Adhesion of IRBC on EC is a key step in the life cycle of the plasmodium parasite and this study highlights new implications for this adhesion. We showed that a first step of rapid transfer of material from IRBC to HBEC presented features of a trogocytosis-like mechanism. This was shortly followed by a tighter adhesion, which appears to divert the natural transmigration pathway of leukocyte-EC and involves a cup-like engulfing process. Malarial antigens then entered the HBEC endosomal pathways and were detected inside the HBEC up to 24 h later. This could be followed by their presentation to the immune system. IRBC transfer was closely followed by a rapid opening of the EC intercellular junctions, an event that may contribute to cerebral oedema. All the mechanisms hereby described can have dramatic implications in the pathophysiology of CM. The relevance of these in vitro observations during CM is first supported by the experimental conditions used. Activation of the ECs and detection of TNF secreting monocytes in brains vessels in the same time as IRBC were reported by Pongponratn and Porta et al [37], [50] who showed images of IRBC sequestered in vessels in the same time as leukocytes. Sequestration of IRBC and CM seem closely related to TNF overproduction, as reported by Grau et al and Kwiatkowski et al [51], [52]. The range of TNF we used to activate ECs and inducing adhesion of IRBC is the same as already detected in vivo (100 pg/ml with high range near 500 pg/ml [53], [54]). This elevated level was also detected in Cerebrospinal Fluid [55], [56] especially in patients who died from CM. Upregulation of TNF-receptor 2 (TNFR2) seems also related to CM [57]. Along the same line, focal induction of ICAM-1 expression in infected brain vessels was also reported years ago. Brown et al [58] described activation of EC and macrophages and disruption of endothelial intercellular junctions in vessels containing sequestered parasitized erythrocytes. These findings suggest that BBB breakdown occurs in areas of parasite sequestration during CM in African children. Porta et al [37], showed CD68 leukocytes coexisting with infected erythrocytes in capillaries, whereas in venules the monocyte population outnumbered the erythrocytes. They also showed expression of ICAM-1 on EC surface in vessels with sequestered cells but not in unaffected vessels. Similarly, Esslinger et al reported that in vivo stimulation of human vascular EC with P. falciparum-infected erythrocytes resulted in the non-transient up-regulation of ICAM-1 expression on endothelial surfaces [59]. The soluble form of ICAM-1was also found significantly higher during acute malaria in children and correlated with levels of TNF, IL-1 alpha and interferon gamma [60]. Clark et al [61] showed inducible nitric oxide synthase staining of ECs which suggested intense inflammatory mediator activity. These alterations can be detected in other organs than brain especially in children who died during severe malaria without true CM. Deposits of malaria antigens in vessels were also reported earlier during autopsies of patients dying from CM. Pongponratn et al first reported images of IRBC sequestered in vessels some associated with or beneath ECs [50]. Boonpucknavig et al demonstrated intense deposition of P. falciparum antigens, IgG and fibrin in cerebral vessels associated with hemorrhages [62]. Immunofluorescent studies also demonstrated the extravascular deposits of P. falciparum granular antigens associated with acute inflammatory lesions in cerebral tissue. IgE was also reported in these depositions especially in the white matter [63]–[65]. They were also found beneath ECs suggesting transfer of material. In the same line pLDH or pAldolase were detected in a variety of organs during CM but were most abundant in the blood vessels of brain, heart, and lung tissues, also detectable in ECs [66]. All these data strongly suggest transfer of malaria antigen in or beneath the EC wall in the brain. Our in vitro observations are thus relevant in regard to these in vivo findings and can explain a part of the pathophysiology of CM. Presentation of malaria antigens by ECs to immune cells and activation of cytotoxic mechanisms could be another step in the explanation of this pathology. Analysis of this mechanism will require malaria immune cells from tropical area countries and is currently in process. The following antibodies and dyes used were: PKH-26 or -67 (MINI26-1KT or 67-IKT) and Phalloidin-FITC (P5282) from Sigma; anti-ICAM1 (0544) and anti-VCAM1 (1244) from Immunotech; ER tracker red BodiPyTR (E34250), Calcein-AM (C3099) and Lysotracker Red DND-99 (L7528) from Invitrogen; anti-EEA1 (610457) and anti-clathrin (07339) from BD Bioscience; Hoechst 33258; anti-human-IgG FITC (733175) and anti-glycophorin A (PN IM2210) from Beckman Coulter. Other compounds used were: albuMAX II (10021-037 Gibco), HMDS hexamethyldisilazane (H4875, Sigma); collagen (BD Biosciences); TNF (PreproTech, 300-01A). Immortalised human brain endothelial cells 5i (CDC Atlanta) were grown in DMEM/F12, whereas hCMEC/D3 [19] were grown in EBM-2 medium (Lonza CC-3156), in 24 well plates or on glass cover-slips coated with collagen. TNF activation of HBEC was carried out by treating the cells with 10ng/ml TNF for 18h. Plasmodium falciparum strains FCR 3Ci and CS2 (kindly donated by S. Rogerson) were grown in RPMI+0.5% Albumax, as previously described [67] and periodically selected for knob expression [68] Late stage IRCB were selected and concentrated using Automacs (MiltenyR), according to manufacturer instructions, to an average of 80–90% parasitaemia. The IRBC were then co-incubated with HBEC [69] at a 20 RBC: 1 HBEC ratio (i.e. 2×106 IRBC per well in a 24-well plate). For overnight incubations, non-adherent RBCs were gently removed from the HBEC, after 4 hours of incubation, by washing the cells three times with pre-warmed medium. For trogocytosis studies, IRBC/RBC were labelled, according to manufacturers' instructions, with either PKH-26 or -67 [70], calcein-AM or Hoechst. Labelled RBCs were incubated, at 37°C for 30min, in parasite media prior to the last wash. Immunofluorescence detection was carried out on cells fixed in 2% paraformaldehyde (PAF) for 10min. Cells were treated with 25mM NH4Cl, permeabilized in 0.1% Triton X-100, and incubated for 30min in 3% bovine serum albumin (BSA) prior to antibody reactions (see list below) for 45min (in PBS containing 0.3% BSA). Detection of P. falciparum (Pf) antigens was performed using a pool of 15 human plasma samples selected out of 100 samples from Senegalese Pf-immune adults (used at 1∶500). The samples were selected for their high titer (>1024) in plasmodium antibodies [71] and for their low background on HBEC (checked individually by immunofluorescence). A pool of plasma from non-immune adults was used as a negative control. For microscopy examination cover-slips were mounted in moviol and examined either under an Olympus FV1000 confocal microscope (magnification 600 or 1000) or on an Olympus-IX71 fluorescent microscope equipped with a F-View CCD camera (Soft Imaging Sys.). Laser 405nm was used with Differential Interferential Contrast (DIC) to generate bright field images. Quantitation of fluorescence was carried out in 24-wells plates on a Fluostar Optimax Spectrophotometre (BMG LabTech) at the relevant wavelengths (100 spots/well were read). For scanning electron-microscopy, HBEC were grown on 10 mm glass coverslides in 24-well plates for 5 days. They are co-cultured with IRBC as previously described. were fixed first in PAF 2% for 10 min and glutaraldehyde 2% in cacodylate buffer for 30min, followed by potassium ferrocyanate–osmium (1% each) post-fixation. Dehydration was performed in grading alcohols, with a final step in HMDS for 3 min. Transendothelial impedance was measured every 10 minutes over the course of experiments using electric cell–substrate impedance sensing (ECIS) system. hCMEC/D3 were seeded at 20,000 cells/wells in 8-well slides and allowed to grow for 3 to 4 days, until confluent, in complete HBEC medium. Confluent hCMEC/D3 monolayers (estimated 100,000 cells/well) were activated with TNF (10 ng/ml) for 18 hours. Drugs were then added directly to wells 40 min prior to addition of IRBC (3Ci- or CS2) or NRBC at a ratio of 20 IRBC/cell. Red blood cells were removed from the HBECs by gentle washing, 4 hours after beginning of incubation. Histamine was used as a positive control for TEER modification. For each well, impedance at time t was normalized according to the impedance of the well at the beginning of the co-culture (t0) and plotted according to time as: (I(t)−I(t0))/I(t0). I(0) was evaluated as the mean impedance over 50 min just before beginning of experiments.

10.1371/journal.pcbi.1000449

Small RNAs Originated from Pseudogenes: cis- or trans-Acting?

Pseudogenes are significant components of eukaryotic genomes, and some have acquired novel regulatory roles. To date, no study has characterized rice pseudogenes systematically or addressed their impact on the structure and function of the rice genome. In this genome-wide study, we have identified 11,956 non-transposon-related rice pseudogenes, most of which are from gene duplications. About 12% of the rice protein-coding genes, half of which are in singleton families, have a pseudogene paralog. Interestingly, we found that 145 of these pseudogenes potentially gave rise to antisense small RNAs after examining ∼1.5 million small RNAs from developing rice grains. The majority (>50%) of these antisense RNAs are 24-nucleotides long, a feature often seen in plant repeat-associated small interfering RNAs (siRNAs) produced by RNA-dependent RNA polymerase (RDR2) and Dicer-like protein 3 (DCL3), suggesting that some pseudogene-derived siRNAs may be implicated in repressing pseudogene transcription (i.e., cis-acting). Multiple lines of evidence, however, indicate that small RNAs from rice pseudogenes might also function as natural antisense siRNAs either by interacting with the complementary sense RNAs from functional parental genes (38 cases) or by forming double-strand RNAs with transcripts of adjacent paralogous pseudogenes (2 cases) (i.e., trans-acting). Further examinations of five additional small RNA libraries revealed that pseudogene-derived antisense siRNAs could be produced in specific rice developmental stages or physiological growth conditions, suggesting their potentially important roles in normal rice development. In summary, our results show that pseudogenes derived from protein-coding genes are prevalent in the rice genome, and a subset of them are strong candidates for producing small RNAs with novel regulatory roles. Our findings suggest that pseudogenes of exapted functions may be a phenomenon ubiquitous in eukaryotic organisms.

Pseudogenes are “defunct” copies of protein-coding genes that have been accumulated in a genome. They have conventionally been considered the junk byproducts of genome evolution, as they cannot code for proteins due to sequence degeneration. Recent important studies, however, have discovered that a subset of them have unforeseen roles as sources of non-coding RNA transcripts that can regulate the expression of functional coding genes. In this work, we have explored this perspective by studying to what extent rice pseudogenes can encode small RNAs, especially the antisense type. After cross-examining thousands of rice pseudogenes and several millions of small RNA sequences, we found that pseudogene-derived small RNAs are abundant in rice, and furthermore, many of them may function as natural antisense RNAs by interacting complementarily (based on sequence similarity) with sense RNAs from either parental genes or other pseudogenes. Some of these pseudogene RNAs may also be involved in repressing pseudogene transcription. Our findings shed new light on the multifaceted biological roles of pseudogene-derived siRNAs that may or may not affect their functional counterparts directly, indicating that more studies are required to explore in molecular detail the diverse functions of siRNAs originating from pseudogenes in both animals and plants.

Pseudogenes are genomic sequences derived from functional genes, but are often considered non-functional due to the accumulation of various deleterious mutations over their evolutionary history [1]–[6]. Compared with its parental gene (more precisely, the direct descendent of the ancestral gene that gave rise to the pseudogene), a pseudogene generally contains sequence features such as premature stop codon or frameshift mutations, due to relaxation of or entirely lack of functional constraints. Two major classes of eukaryotic pseudogenes have been described: processed and duplicated. Processed pseudogenes arose from retrotransposition events, i.e., the insertions of DNA materials into a genome via RNA intermediates. Duplicated pseudogenes, on the other hand, originated from DNA duplications. As a result, duplicated pseudogenes often retain the exon-intron structures of their parent genes, a characteristic absent in processed pseudogenes [1],[2],[4],[6]. Genome-wide pseudogene annotations have been carried out for several mammals and prokaryotes but to date not for rice or any other plants [3], [5], [7]–[11]. While previous studies have demonstrated that retrotransposition is the major mechanism for generating mammalian pseudogenes [3],[4],[6], it remains to be established whether duplication or retrotransposition is the predominant mechanism in plants such as rice. Rice (Oryza sativa) is a very important crop species that supports about one half of the human population. The genomes of two sub-species (indica and japonica) were sequenced completely in 2005 [12],[13]. Based on current annotation (TIGR V5), the rice (japonica) genome contains 41,046 protein coding genes (excluding transposable elements, TEs), 763 tRNA genes, and 2,859 novel genes seemingly unique to rice and other cereal. To this date, no systematic and comprehensive annotation of pseudogenes has been done to address their impact on the structure and the function of the rice genome. However, a number of rice duplicated pseudogenes have been reported lately, including several arising from MADS-box genes, which encode a large family of transcription factors [7], and 99 pseudogenes in Cyt P450 family [8]. Moreover, current annotation has identified 15,232 TE-related genes and retrotranspositions have been documented to play a significant role in shaping the rice genome [14], suggesting that pseudogenes is a significant component of the rice genome. While pseudogenes are usually considered non-functional, many transcriptionally active pseudogenes and several pseudogenes with exapted functions had been identified experimentally or suggested [2], [6], [9]–[11]. Generally, pseudogenes cannot be transcribed due to the lack of a functional promoter and auxiliary regulatory elements [1],[15]. However, a previous study revealed that at least a fifth of human pseudogenes could be transcribed to different degrees based on a variety of empirical transcription evidence, such as 5′ RACE (Rapid Amplification of cDNA Ends), tiling microarray analysis and high throughput sequencing data [5]. Other studies have also independently estimated that 5–20% of human pseudogenes exhibit evidence of transcription [16]–[19]. Likewise, a significant percentage of mouse pseudogenes were also found to produce stable RNA transcripts following the analysis of 100,000 mouse full-length cDNA [20]. The biological and functional implications of such pseudogene transcripts are largely unexplored, but a few of them have been indicated to play important biological roles mostly in gene regulation [1],[2],[6],[21],[22]. Moreover, direct evidence has been established for a functional NOS (nitric oxide synthase) pseudogene that is transcribed in specific neurons in the central nervous system of L. stagnalis, where its transcript forms a RNA duplex with the mRNA of its parental gene and curtains the production of NOS proteins [22]–[24]. The human XIST non-coding gene, the key initiator of X chromosome inactivation, also arose from the relic of a pseudogene [25]. Recent studies from deep sequencing of small RNA libraries have provided further strong evidence that a significant number of pseudogenes may serve as a genomic reservoir for functional innovation, e.g., as the source of small regulatory RNAs [9]–[11],[26]. Three major classes of small RNAs have been described in plants and animals: microRNAs (miRNAs), small interfering RNAs (siRNAs), and Piwi-interacting RNAs (piRNAs) [10],[27],[28]. MiRNAs are usually 20 to 24 nt long and they interact with targeted mRNAs to modulate their translations [27]. siRNAs, usually 21-nt long, are generated from double-stranded RNA precursors such as those from viruses or endogenous transposons [28]. In plants, five major groups of siRNAs have been reported: transacting siRNAs (ta-siRNAs), natural antisense transcript-derived siRNAs (nat-siRNAs), repeat-associated siRNAs (ra-siRNAs), heterochromatic siRNAs, and long siRNAs (lsi-RNA) [29]–[36]. In animals, piRNAs derived from repetitive elements via a Dicer-independent pathway have also been shown to repress the activity of mobile genetic elements [9], [37]–[39]. Meanwhile, siRNAs derived from long hairpin RNAs (hp-RNAs) can also repress endogenous target transcripts in Drosophila [40]. Most recently, it was shown that siRNAs originated from pseudogenes can regulate gene expression in mouse oocytes [9],[26]. More importantly, it was found that some of these pseudogene siRNAs were processed through Dicer and the loss of Dicer significantly reduced the number of pseudogene siRNAs, which in turn led to the up-regulation of their targeted genes [9],[26]. In plants, however, a plant-specific siRNA biogenesis pathway has been suggested to be responsible for the production and function of pseudogene-derived siRNAs. It was shown that the production of siRNAs from Arabidopsis pseudogenes depended on RNA-dependent RNA polymerase (RDR2) and Dicer-like protein 3 (DCL3) [41]. With a characteristic feature of 24-nt length, these pseudogene-derived siRNAs were thought to repress local transcription in a similar manner as the cis-acting siRNAs originated from transposons or retroelements [41], which can mediate RNA-directed DNA methylation (RdDM) and heterochromatin formation [42]–[47]. While this hypothesis is wholly consistent with the significant findings from the Arabidopsis work [41] and warrants more detailed investigations, it is not immediately clear how pseudogenes, especially those arising from gene duplications, can sustain DRD2 and DCL3 activities that seem to work most efficiently on tandem repeating sequences [42]. Motivated by these recent advances, we have carried out a genome-wide annotation of candidate pseudogenes in rice (japonica genome) and then carefully interrogated these pseudogenes for their capability of generating siRNAs. We annotated a total of 11,956 pseudogenes using a recently developed and validated pseudogene annotation pipeline, PseudoPipe [48]. Characterization of these pseudogenes with a library of ∼1.5 million small RNAs from developing rice grains [35] identified 145 pseudogenes as strong candidate loci for producing antisense small RNAs, with many of them having the potential to regulate the expression of their parental genes. Further survey of additional small RNA libraries indicated that the production of specific pseudogene-derived siRNAs could be restricted to particular developmental stages or growth conditions. We have assigned a total of 11,956 pseudogenes, most of which contained premature stop codons and frameshift mutations (Table 1), in the rice genome using the PseudoPipe software [48]. Of the 41,046 non-TE-related protein-coding genes used as our query sequences, 4,946 (12.1%) had at least one pseudogene. Respectively, 3,392 and 2,350 of the rice pseudogenes were classified as processed pseudogenes and duplicated pseudogenes by PseudoPipe (Table 1). The rest (6,214) were designated as pseudogene fragments, since they did not contain sequence features that were considered by PseudoPipe to distinguish between retrotranspositions and DNA duplications. Such fragments are usually derived from duplications, so together with duplicated pseudogenes they are referred as non-processed pseudogenes. The average nucleotide sequence identity between pseudogenes and their parents was 70.5%, 76.6%, and 74.0% for processed pseudogenes, duplicated pseudogenes, and pseudogene fragments, respectively. The mean alignment coverage on the parental genes was 87.9%, 59.6%, and 35.2% for processed, duplicated, and fragments, respectively, suggesting that most past retrotranspositions generated rice processed pseudogenes of full length (in relation to the CDS of their parental genes). Detailed information for individual pseudogenes can be found in the annotation file available on the web (http://www.pseudogene.org/rice09/). Overall, the processed pseudogenes are randomly distributed in the rice genome, a pattern apparently different from the genome-wide distribution of non-processed pseudogenes (Figure 1 and Supplementary Figure S1). Moreover, non-processed pseudogenes are prevalently located in duplicated regions as expected from well documented extensive past duplication events in the rice genome. For example, a comparison of our pseudogene annotation with the whole genome duplication (WGD) data, described in [49], showed that 6,071 (70.9%) of rice non-processed pseudogenes are located in WGD regions, verse 2,085 (61.5%) of processed pseudogenes, in consistent with their distinct generation mechanisms (p = 1.88e−8). It has been reported that most of the rice genes can not be clustered into gene families [12],[13],[50], i.e., they do not have a functional rice paralog and thus are considered singleton genes. This is somehow counterintuitive with respect to the past active DNA duplications in rice. To investigate how this has impacted pseudogene biogenesis, we transferred the family annotation of each parental gene to its pseudogene(s), using gene family annotation from TIGR v5. Interestingly, this analysis showed that more than 65% of pseudogenes (vs ∼50% of coding genes) were from singleton gene families (Figure 2), and conversely 13.3% of singleton genes (vs 12.1% for all rice genes) had at least one pseudogene. These numbers imply that rice pseudogenes preferentially come from singleton families. Upon more careful investigation, however, we found that this bias was introduced largely by the specific method employed in annotating rice gene family, which clustered genes based on domain architectures of their protein products [50]. As a consequence, two genes sharing all but one domain would have been assigned to distinct families. We therefore utilized an alternative method that is better suited for our purpose in designating singleton genes: a gene is considered singleton if it does not share significant sequence similarity (BLAST e-value <1e-14) with other rice genes. By this criterion we found that only 7.6% (1,262) of singleton genes had a corresponding pseudogene, indicating that coding genes in singleton families are less likely to have a detectable pseudogene. Nevertheless, our results indicate that many singleton genes actually have a pseudogene relative even though they do not have a functional paralog. In conjunction with the TIGR rice gene annotation, our analysis suggests that the overabundance of singletons in rice genome is the outcome of sequence loss rather than sequence degeneration (i.e., pseudogenization). Overall, we found that family size was negatively correlated with the number of pseudogenes in a family (R = −0.93, p<1e-15; Spearman's correlation computed for family size of 2 to 15), suggesting that large gene families do not necessarily have more “dead” (pseudogene) members (Figure 2), an interesting observation to be further studied. The top ten genes with the most pseudogenes are LOC_Os01g10030 (550 pseudogenes), LOC_Os09g23670.1 (226), LOC_Os10g41910.1 (125), LOC_Os03g32980.1 (113), LOC_Os11g14500.1 (90), LOC_Os06g11360.1 (90), LOC_Os11g10210.1 (89), LOC_Os11g36210.1 (79), LOC_Os08g39680.1 (73), and LOC_Os01g59540.1 (67). The functions for eight of them have not been annotated but two appear to be housekeeping genes. This is strikingly different from what has been reported in mammalian genomes, where a large fraction of pseudogenes are derived from known gene families such as ribosomal protein genes and olfactory receptor genes [3],[5]. More specifically, the ribosomal protein genes have generated about 2,000 pseudogenes in both humans and mice [51] but only 50 in rice based on our current study. As the significant enrichment of processed pseudogenes from housekeeping genes in mammals is considered to be relevant to their high expression levels, current finding suggests that distinct evolutionary events are probably responsible for the generation and subsequent retainment of pseudogene populations in plants and animals. Recent studies have revealed that pseudogenes are an important source of non-coding RNAs [6],[9],[26]; furthermore, genome-wide small RNA analyses have indicated that pseudogene-originated siRNAs could either regulate the expression of their parental genes in mice [9],[26] or be implicated in silencing pseudogenes themselves in Arabidopsis [41]. To explore this, we have searched a library of small RNAs from developing rice grains [35] for antisense RNAs from pseudogene loci. Such RNAs would be complementary to the mRNAs from parental genes and thus may function as nat-siRNAs. We found that 2,867 and 2,582 pseudogenes had at least one small RNA mapped to their sense and antisense strands, respectively. By a threshold of >4 small RNAs/100 nucleotides (corresponding to ∼1× coverage, see Materials and Methods for details), we considered 145 pseudogenes as good candidates for producing antisense small RNAs (Figure 3, group A). Three-quarters (75.9%) of these candidates were from singleton families, in consistent with their abundant representation in the rice pseudogene population (Figure 2B). Additionally, 24 (16.6%), 39 (26.9%) and 82 (56.5%) of these pseudogenes were processed, duplicated and fragments, respectively, indicating a slight bias (p = 0.2) of non-processed pseudogenes in generating antisense small RNAs. For the convenience of description, we will use small RNAs and siRNAs inter-changeably below even thought the functions of the small RNAs in our dataset have not been established experimentally. The sizes of antisense siRNAs, however, suggest that most of the rice pseudogene-derived siRNAs may not serve as nat-siRNAs, as the majority (53.4%) of them are 24-nt long (Figure 4), which is a common signature feature of small RNAs derived from plant RDR2 pathway. Our finding is rather consistent with what has recently been shown for pseudogene-derived siRNAs in Arabidopsis that were predominantly 24-nt long and depended on RDR2 and DCL3 for their accumulation [41]. The potential involvement of RDR2/DCL3 pathway is further supported by the observations that the majority of the 145 pseudogenes also produced some sense small RNAs and that the size distributions of sense and antisense small RNAs were identical (Figure 4). All together, these data indicate that ∼50% of pseudogene siRNAs in rice grains could have been produced by the Pol IV/RDR2/Pol V pathway. As plant 24-nt siRNAs and the Pol IV/RDR2/Pol V pathway are specifically implicated in RNA-directed DNA methylation (RdDM) and heterochromatin formation, two processes important for silencing transposons and other retroelements in plants [42]–[47], these results suggest that some pseudogene siRNAs may be important for cis-repression (more precisely, local-repression) of pseudogene transcription. More careful analysis of small RNA sizes suggested that pseudogene-derived siRNAs might have other regulatory roles besides cis-acting repression in rice. It is known that trans-acting siRNAs (tasiRNAs) or nat-siRNAs in plants exhibit a size range quite different from that of repeat-associated siRNAs. For example, a recent analysis of rice nat-siRNAs found that their lengths varied from 17- to 31-nt with a detectable peak in 21-nt [36]. Interestingly, small RNAs from rice miRNA loci also displayed two peaks, at 21 nt and at 24 nt (Figure 4). In contrast, small RNAs from repeats (LINE, SINE, LTR and DNA transposons) exhibited a single strong peak at 24 nt (Figure 4). Therefore, we decided to study 24-nt and non 24-nt small RNAs separately, with the assumption that these two groups are generated from different pathways and have largely distinct functions. A comparison of the genome-wide distributions of these two groups of small RNAs, to our surprise, did not find an enrichment of 24-nt siRNAs in the centromeric and pericentromeric regions where repetitive elements are concentrated (Figure 1). These data accumulatively suggest that the repeat-relevant RDR2 pathway is not necessarily the sole pathway contributing 24-nt siRNAs to our small RNA library, and that pseudogene-derived siRNAs may have functions other than inducing local heterochromatin formation. It is worth to mention here that the first example of trans-acting nat-siRNAs, formed between SRO5 and P5CDH transcripts through the action of RDR6/DCL2 in Arabidopsis, is indeed 24-nt long [31]. In the studies of siRNAs in mice [9],[26], it was considered that simultaneous accumulation of sense-strand small RNAs from parental genes and antisense-strand siRNAs from pseudogenes in the complementary region(s) is good evidence for the production and as well as regulatory potential of pseudogene-derived siRNAs. Following this idea, we have searched for parental gene-pseudogene pairs in which both the parental genes and pseudogenes can produce significant numbers of sense and antisense small RNAs, respectively (see Materials and Methods for details). The complementary interaction of those RNAs can potentially affect the expression of the parental genes (or pseudogenes). Application of this strategy has yielded 38 gene-pseudogene pairs (Figure 3, group B) with the capability of producing trans-nat-siRNAs in developing rice grains. The over representation of non-processed pseudogenes (32 out of 38) lends a good support to our hypothesis, since small RNAs of trans-regulatory function arising from duplicated pseudogenes have previously been documented [6],[41],[52]. Again, one caveat here is that many of these 38 group B pseudogenes produced both sense and antisense siRNAs with a mixture of 24 and non-24 nt (Table 2). Based on sign-test statistical analysis, we found that 8 of the 38 group B pseudogenes (vs 16 of the 145 group A) had significantly more non-24 than 24 antisense siRNAs (at the false discovery rate of 5%), indicating that a RDR2 independent pathway is very likely involved. As an extra line of evidence for the existence of regulatory pseudogene siRNAs in rice grains, we found that none of the parental genes of the 38 group B pseudogenes had a matching EST from either rice seeds or seedlings, whereas seven and zero of the 145 group A pseudogenes had ESTs from seeds and seedlings, respectively. Intriguingly, the pseudogene and its parental gene in 21 of these 38 pairs in group B (55.3%) are very close to each other on chromosomes (<10 kb) (Table 2). By comparison, we found that only 437 of our 11,956 pseudogenes (3.7%) were within 10 kb of their parental genes. These data indicate that siRNAs with a potential of forming (or from) nat-RNA duplex are significantly more likely to be generated from gene-pseudogene pairs of short distances (χ2 = 171.0, p<1e-15). It should be mentioned here that none of these 21 pairs are located in the previously identified siRNA generation hotspots [53]. As an example, one pair defined by the gene LOC_Os10g04950 and its pseudogene 2,594-bp downstream contained large numbers of unique small RNAs exclusively mapped (based on polymorphism) to the sense strand of gene or the antisense strand of pseudogenes (Figure 5, A and B). In this case, we also noticed some antisense RNAs from the parent gene and sense RNAs from the pseudogene. The importance and implication of their presence need to be explored in the future. The discovery of nat-siRNAs from gene-pseudogene pairs prompted us to search for other types of mechanisms that could also be important for the production of functional pseudogene siRNAs. It has previously been shown that siRNAs can be derived from hairpin RNAs [9],[26],[40]. In the search of such small RNAs (Figure 3), we identified two pairs of pseudogenes that potentially could generate transcripts forming hairpin structures (Table 3). One example is a pseudogene located on chromosome 6 (from 12,656,179 to 12,656,385-bp) and its downstream paralogous pseudogene (12,654,917 to 12,654,733-bp; the parent of these two pseudogenes is LOC_Os02g46460). The RNA transcript(s) of these two pseudogenes was predicted to form a 185-nt duplex linked by a 1,283-nt loop (Figure 5). A large number of small RNAs were uniquely mapped to the inverted repeat duplex predominantly to the minus strand, indicating that the transcripts were produced from minus strand of the two pseudogene loci (Figure 5, C and D). We did not detect any small RNAs from the parental gene of these two pseudogenes. In summary of our analysis of small RNAs from developing rice grains, a total of 145 pseudogenes were identified as good candidates for generating pseudogene-derived antisense endo-siRNAs, with the parental genes for 38 of them harboring significant number of sense siRNAs, suggesting that a small subset of rice pseudogenes might have evolved exapted functions to produce regulatory antisense small RNAs. Our finding extends the previous observations from mouse oocytes [9],[26] and Arabidopsis [41], and suggests that siRNA-modulated regulatory function of pseudogenes may be conserved from animals to plants. We have further explored the diversity of pseudogene-derived siRNAs using additional five small RNA datasets acquired in different rice developmental stages and physiological growth conditions (details in Figure 6 and Supplementary Table S1). Obtained with different high-throughput sequencing techniques, these five datasets contained (a) 285,873 [53], (b) 108,472 [53], (c) 299,454 [54], (d) 182,792[36], and (e) 11,809 [55] small RNAs, approximately half of which in each dataset could be mapped to the rice genome uniquely (Figure 6). These new datasets are considerably smaller than our primary dataset described above, which had ∼1.5 million mapped RNAs, therefore we did not apply the threshold of 4 RNAs/100 nt to them. The numbers of pseudogenes with 1∼5 antisense small RNAs in these libraries and their comparison with our primary RNA data can be found in the Supplementary Table S1, indicating that 2,350 and 82 rice pseudogenes had ≥1 or ≥5 antisense small RNA in at least one of the five datasets, respectively. Moreover, 43.9% (1,134/2582) of the pseudogenes with ≥1 antisense siRNA(s) in our primary grain library were found to have siRNA(s) in the pool of these five new datasets. Among the 145 group A pseudogenes (Figure 3), 119 (82.1%) could be detected with ≥1 antisense small RNA, and more strictly, 58 (40%) with ≥5 antisense RNAs in at least one of the five datasets (Figure 6A, library all), strongly supporting the existence of endo-siRNAs from those pseudogenes. When considering the source of rice materials used for individual libraries, however, we found that most pseudogenes with antisense siRNAs from rice grains (≤10 days-after-fertilization) did not have matched siRNAs in other libraries, including those from more developed rice seedling. For example, 91 of the 145 (62.8%) group A pseudogenes were detected with antisense siRNAs from the dehulled grain library obtained by Heisel et al. (Figure 6A, library a), but only 13 of them (9.0%) had siRNAs in the library of 23 days old seedling (Figure 6A, library b). The difference here (p = 1.2e-6) is not simply the consequence of sequencing depth, as library a is no more than three times bigger than library b. Similarly, antisense siRNAs were also detected for most (31) of our 38 group B pseudogenes (Figure 3) in the Heisel's rice grain dataset (a) but not (only 5) in non grain-related datasets (b, d, e) (Figure 6B). Such skewness is even more pronounced when we examined pseudogenes with more than one siRNA in the new libraries (Figure 6). These comparisons of small RNA libraries from a variety of sources showed that many pseudogenes seemed to produce antisense siRNAs only under specific developmental stages or physiological conditions. Therefore, the pseudogene-derived siRNAs detected only in rice grains may play important roles in early rice grain development, a topic worthy of future investigation. Our analysis of Gene Ontology (GO) showed that the majority of the parent genes with pseudogenes producing siRNAs in developing rice grains were implicated in gene regulation. For all rice pseudogenes, the most representative function was hydrolase activity (14.8%) (Figure 7A). This bias seems to be specific to plants, but consistent with previous observations for mammalian genomes housekeeping functions such as protein binding (12.4%) and nucleotide binding (9.7%) were also relatively abundant [3],[5]. However, it has to be cautioned that these statistics might have been complicated by that fact that 32% of 11,956 pseudogenes have not been assigned a GO functional category. In comparison to all rice pseudogenes, we found that the 145 group A pseudogenes with many antisense siRNAs in developing rice grains were substantially enriched in GO terms implicated in DNA binding and transcription regulation (p<0.05) (Figure 7B). This result indicates that pseudogene-derived antisense siRNAs could affect more downstream targets in a broader content by regulating the expression of the parental genes. Pseudogenes are conventionally viewed as the by-products of genome evolution. About one-fifth of the rice protein coding genes are transposable elements [14], and the continuous activity of TEs has generated numerous TE-related pseudogenes (∼26,000 identified in this study), a feature common to many plant genomes. It is also known that gene duplications have occurred quite frequently in the rice genome [56]–[58]. For example, it was estimated that 15% to 62% of the rice genome underwent a whole genome duplication (WGD) ∼70 million years ago [56],[57]. A more recent segmental duplication event has also left ∼3 Mb synteny between chromosome 11 and 12 [58],[59]. Our observation of a high percentage of non-processed pseudogenes in the rice genome, and particularly their relative enrichment in the WGD region over processed pseudogenes, is consistent with these previous reports. On the other hand, at least 35% of the rice genome is considered the product of retrotransposition events mediated by retrotransposons such as copia and gypsy elements [13],[60]. That estimation is in line with the percentage (28.4%) of processed pseudogenes annotated in current analysis. However, in comparison to the human or rodent genomes, in which about half of their pseudogenes are derived from retrotranspositions, rice has much fewer processed pseudogenes. Moreover, the rice genome has a lower pseudogene to gene ratio (12000 vs 41,000) than the mammalian genomes (approximately 20,000 vs 22,000) [51]. It will be interesting to study in the future how much this is related to the rice genome's compact size and rapid sequence loss after retrotranspositions or gene duplications [57]. Another avenue to explore is whether many of the predicted rice genes without experimental evidence are actually pseudogenes, an important concern not addressed in current work. Although our pseudogene annotation is largely consistent with the known evolutionary history of the rice genome, the exact ratio of rice processed verse duplicated pseudogenes could deviate from the number reported here, since PseudoPipe [48],[61] was originally developed for mammalian genomes. It is very intriguing that so many rice genes (∼50%) are presented themselves as singletons (TIGR v5) [50] even though duplications have been highly active during the evolution of the rice genome. In current work, we found that 13.3% of rice singleton genes have pseudogene relatives despite that the majority (1,729, 64.2%) of them have only one pseudogene. We had expected more singleton genes to have pseudogene relatives, based on high frequency of DNA duplications in rice and the reasonable assumption that most of the duplicated sequences associated with singleton genes should be detected as pseudogenes. The small percentage of singleton genes with pseudogenes (actually 7.6% based on a more strict definition of singleton family, see Results), in conjunction with the large number of singleton coding genes, indicated that most of the rice sequences resulted from past duplications have been either deleted or altered too substantially to be recognized by simple sequence comparison. This is in full accordance with the general view in the field that deletion and degeneration are the predominant outcome for duplicated sequences [62]–[64]. Moreover, an analysis of the synonymous substitution ratio (Ks) between pseudogenes and their parental genes showed that the non-processed pseudogenes from singletons (Ks = 0.26±0.39) appear younger than those from multi-gene families (0.46±0.61) (p<0.001), implying that the failure of detecting “old” pseudogenes from singletons might indeed be a reason. As a comparison, we found that 36.5% of Arabidopsis pseudogenes were generated from singleton genes and conversely 11.5% of Arabidopsis singleton genes had a pseudogene relative. These results suggest that domestication process could have introduced significant interference to the nature selection in rice. Nonetheless, there are rice singleton genes that have generated many pseudogenes. For example, the top four singleton genes (LOC_Os01g10030, LOC_Os08g13800, LOC_Os09g02850 and LOC_Os11g25600) with the most pseudogenes have generated a total of 717 pseudogenes, accounting for 6.0% of the rice pseudogenes. This is remarkable especially considering that so many (singleton) genes do not have a protein coding paralog but their pseudogene derivatives may produce antisense siRNAs. Our finding might also be relevant to the differences in functions and alternative splicing between singletons and multi-gene families [50]. While pseudogenes derived from functional genes have lost their protein coding potential, they may have gained novel biological functions. This topic has not received adequate attentions, but it is supported by the discovery of a functional NOS pseudogene in snails [22]–[24], and furthermore by recent studies that employed deep sequencing to show that a subset of pseudogenes in mouse oocytes can produce regulatory siRNAs [9],[26]. Our discovery of a large number of antisense siRNAs from rice pseudogenes and especially the finding of 38 parental gene-pseudogene pairs with many complementary small RNAs suggest that some rice pseudogenes could have evolved novel functions by encoding nat-siRNAs. To certain extent, our result is complementary to a recent finding that a large number of rice coding genes can produce cis-nat and trans-nat siRNAs, the majority of which also seem to be associated with specific growth conditions or developmental stages [36]. Alternatively, the pseudogene-derived siRNAs might not regulate functional genes but play a cis-acting role in recruiting RNAi machinery to suppress local transcription through the plant-specific Pol IV/RDR2/Pol V pathway [42]–[47]. The latter scenario, however, cannot explain completely the existence of sense siRNAs in parental genes. These two possibilities are not necessarily mutually exclusive. Firstly, siRNAs from processed pseudogenes can be involved in cis-acting as they are generated from retrotranspositions, while siRNAs from non-processed pseudogenes can function in trans similar to how some known miRNAs evolving from pseudogenes regulate their targets [6],[41],[52],[65]. Secondly, the same pseudogene might produce both cis-acting (presumably 24-nt) siRNAs and trans-acting (non 24-nt) siRNAs depending on how the pseudogene transcript is processed and utilized by RNAi machinery. Even in the case of cis-acting 24-nt siRNAs, their potential roles in regulating local gene expression should not be ignored. For example, RdDM and formation of repressive chromatins mediated by pseudogene siRNAs can affect the expression of a parental gene if it is sufficiently close to its pseudogene in the chromosomal space due to repression spreading. In our study, we also observed small RNAs derived from adjacent paralogous pseudogenes on opposite strands. Although these pseudogene siRNAs might be taken by the cellular siRNA machinery to modulate the expression of functional genes, they can also be important for suppressing pseudogene expression. It is conceivable that double stranded RNAs can form in vivo between pseudogene antisense siRNAs and their complementary sequences, but only carefully designed molecular and cellular experiments will resolve the different functional scenarios discussed here, for example, by identifying the specific Argonaute proteins and RNA-induced silencing complex (RISC) interacting with pseudogene siRNAs. As most of the studies on 24-nt repeat-associated siRNAs, RDR2/DCL3 and RdDM have been conducted in the model plant organism Arabidopsis, we hope our analysis and similar recent work [35],[36] will draw the interests of rice molecular biologists to study endo-siRNAs as there seems to be a difference in the association of 24-nt siRNAs with repetitive regions especially in the centromeric regions between rice and Arabidopsis genomes (Figure 1 here vs Figure 2 in ref 41). In our current study, we have focused on pseudogenes that produce a significant number of antisense small RNAs. This does not mean that small RNAs from other pseudogene loci are spurious and less likely to be biologically relevant. Rather, the lack of support by multiple RNAs is largely technical, namely the insufficient depth of high-throughput sequencing. For example, 82–92% of the small RNAs from rice grains were only detected once in deep-sequencing using Illumina or 454 sequencing technology [35]. Therefore, low expression and inadequate sampling are probably the reasons for not finding more pseudogene siRNAs. These surely have added a caveat to our results in the comparison of different small RNA datasets with a variety of sequencing depths. On the other hand, pseudogene-derived small RNAs can be much more complicated than what is described here. For example, we found that two pseudogenes with high sequence similarity in very narrow complementary regions could also generate many siRNAs (data not shown). As suggested by previous investigators, such small RNAs may actually be an important source of novel miRNAs [6],[11]. Finally, we have also explored the potential regulators for the transcription of these pseudogenes derived endo-siRNAs. In particular, we have examined the genomic distance of pseudogenes to various transposable elements (included DNA transposon, LTR element, LINE and SINE). A total of 1,062 and 1,108 rice pseudogenes were found to be near (<1 kb) active transposable elements (defined as >90% of the full length TEs) in their 5′ sense and 3′ antisense directions, respectively (Table 4). DNA transposons were the most predominant TEs both in the 5′ sense (788 or 73.3%) and the 3′ antisense directions (811, 73.2%). The enrichments of these TEs in the flanking regions of pseudogenes are significant based on randomization simulations. However, only seven of the 145 pseudogenes with abundant antisense small RNAs have a TE oriented in the antisense direction in their 3′ flanking regions, suggesting that TE promoters are unlikely the primary driver for the production of endo-siRNAs from rice pseudogenes. The initiation of the transcription of these pseudogenes therefore remains perplexing regardless whether the primary transcripts can be subsequently used to produce cis or trans-acting siRNAs. Rice genomic sequences and their annotations (release 5) were downloaded from the Rice Genome Annotation of TIGR (The Institute of Genomic Research, ftp://ftp.tigr.org). About 5.5 million small RNA sequences, generated by CSIRO (Commonwealth Scientific and Industrial Research Organization) for 1–5 days-after-fertilization (DAF) and 6–10 DAF rice grains using high throughput sequencing [35], were downloaded from the Gene Expression Omnibus (GEO) (GSE11014, http://www.ncbi.nlm.nih.gov/geo/). From this dataset a total of 1,483,951 small RNA sequences matching to rice genome uniquely (i.e, with a single best aligned location in up to one mismatch) were obtained as our primary RNA dataset. Five additional small RNA datasets were also obtained, including (a) 285,873 unique small RNAs from the tissue of rice grains assayed with three replicates of independent libraries and (b) 108,472 small RNAs from rice seedling downloaded from the GEO (GSE13152) [53], (c) 299,454 17-bp MPSS small RNA sequences from University of Delaware (http://mpss.udel.edu/rice/) [54], (d) 182,774 unique small RNAs matching perfectly to the rice genome collected from the library of control (58,781), drought (43,003) and salt (80,990) seedlings [36], and (e) 11,809 small RNA sequences downloaded from CRSDB (Cereal small RNA database, http://sundarlab.ucdavis.edu/smrnas/) [55]. Small RNAs from these five additional datasets were mapped to rice genome by the program BLAST [66]. The results were parsed by in-house scripts to extract the chromosomal coordinates of RNAs matching to rice genome uniquely with at most one mismatch in order to ensure comparability between data from our primary library and these additional new libraries. All subsequent analyses utilized genomic coordinates unless specified otherwise. For pseudogene assignment, we applied the PseudoPipe program [48],[61] to the rice genome with its repetitive sequences masked by the RepeatMasker (http://www.repeatmasker.org). Briefly, PseudoPipe scanned the rice genome for DNA sequences similar to a library of annotated rice protein sequences in the TIGR release V5. After those overlapping with annotated genes were discarded, the rest of matching DNA fragments was assembled into pseudogene candidates based on their structural similarities to the query proteins. Details of the pseudogene discovery procedure have been described previously [48],[61]. At the end, a total of 11, 956 pseudogenes derived from non-TE protein coding genes were identified in the rice genome. These pseudogenes were classified into duplicated pseudogenes, processed pseudogenes, and pseudogene fragments based on the potential mechanisms of their generations [3]–[5]. Thus, each of these pseudogenes was defined by its sequence and structural similarity to a functional gene (or protein), commonly referred to as the parental genes. Our pseudogene annotation is available publicly (http://www.pseudogene.org/rice09/). The sequence identity and coverage (the percentage of parental DNA material present in the alignment) between a pseudogene and its parent are part of the information in our annotations. The high sequence similarity between a pseudogene and its parent gene suggests that a natural trans RNA duplex can be formed between the antisense transcript from the pseudogene and the sense transcript from the parent gene. The cellular siRNA/RICS machinery can often use such double strand RNAs to produce mature and functional siRNAs. Here we focused our study on antisense small RNAs from pseudogenes by cross-referencing our pseudogene annotation with the library of small RNAs from ≤10 DAF rice grains, which has the largest collection of RNAs among all the publicly available datasets. As described above, we excluded small RNAs that were aligned to more than one location in the rice genome. For each pair of pseudogene and parent gene, their putative coding DNA sequences were first re-aligned using the global sequence alignment program NEEDLE in the EMBOSS package [67]. We then counted the numbers (N) of small RNAs mapped to the alignment region(s) to generate two numbers for genes, one for the sense and the other for the antisense strand, and likewise two corresponding numbers for pseudogenes. These numbers were then divided by the length of the alignment (L) to obtain the densities (N/L) of small RNAs for each pair of gene and pseudogene. We chose the density of 0.04 as a cutoff in order to select candidate pseudogenes with high potential to produce RNAs. This threshold is equivalent to ∼1× sequence coverage of small RNAs as the length of small RNAs is 18∼25 nt [35]. Furthermore, for a pair of parental gene and pseudogene to be considered as a strong candidate with the potential to produce nat-siRNAs, the sense RNAs of the gene and the antisense RNAs of the pseudogene must both meet the density cutoff. The overall protocol and the resulting different groups of pseudogenes are shown in Figure 3. We also searched for cases in which the transcripts of two adjacent pseudogenes can form RNA hairpin structures. To do so, we used BLASTN program to search paralogs of each pseudogene with antisense RNAs. Paralogous pairs within 2 kb of each other but on opposite strands were evaluated for the potential of producing hairpin RNAs. We used two criteria to define pairs that could generate siRNAs from the RNA hairpin: 1) the stem of the hairpin >100-bp and the loop length <2 kb, and 2) the density of small RNAs (from the same strand) mapped exclusively to the inverted hairpin >0.04 (Figure 3). All statistical analyses were carried out in the R language. When characterizing all pseudogenes, we used the GO function categories of their parents and simply reported GO terms that were highly represented among rice pseudogenes without further inferring statistical significance. To evaluate the functional significance of the 145 pseudogenes enriched with antisense siRNAs (density >0.04 siRNA per nt), we considered the GO distribution of all pseudogenes as the genome-wide background and then employed a re-sampling approach to infer the enrichment of a specific GO term. We selected 145 pseudogenes randomly from all pseudogenes and then recorded the number of pseudogenes in each GO. After 10000 iterations, we obtained an empirical p-value for each GO term measuring the number of iterations that more than X pseudogenes was observed in a GO, where X was the number of pseudogenes in this GO from the 145 pseudogenes. To identify transposable elements in close proximity of pseudogenes, we considered LTR, SINE, LINE and DNA transposons defined by RepeatMasker annotation with these two criteria: (1) distance to pseudogene is <1 kb and (2) annotated TE covers >90% of its consensus full length (thus considered as “active” TE). So, data discussed here are for pseudogenes located downstream (<1 kb) of an active TE. A re-sampling protocol was used to test the significance of a TE enrichment, which was done by choosing a random region with the same length of a pseudogene from the same chromosome that this pseudogene was located. A total of 11,956 such regions were selected and then the number of TEs within flanking regions of these randomly selected regions was counted. By repeating this process 1,000 times, we derived an empirical p-value for TE enrichments. We carried out this analysis for TEs oriented in the 5′ sense and 3′ antisense directions of pseudogenes separately.

10.1371/journal.pgen.1005578

The GTP- and Phospholipid-Binding Protein TTD14 Regulates Trafficking of the TRPL Ion Channel in Drosophila Photoreceptor Cells

Recycling of signaling proteins is a common phenomenon in diverse signaling pathways. In photoreceptors of Drosophila, light absorption by rhodopsin triggers a phospholipase Cβ-mediated opening of the ion channels transient receptor potential (TRP) and TRP-like (TRPL) and generates the visual response. The signaling proteins are located in a plasma membrane compartment called rhabdomere. The major rhodopsin (Rh1) and TRP are predominantly localized in the rhabdomere in light and darkness. In contrast, TRPL translocates between the rhabdomeral plasma membrane in the dark and a storage compartment in the cell body in the light, from where it can be recycled to the plasma membrane upon subsequent dark adaptation. Here, we identified the gene mutated in trpl translocation defective 14 (ttd14), which is required for both TRPL internalization from the rhabdomere in the light and recycling of TRPL back to the rhabdomere in the dark. TTD14 is highly conserved in invertebrates and binds GTP in vitro. The ttd14 mutation alters a conserved proline residue (P75L) in the GTP-binding domain and abolishes binding to GTP. This indicates that GTP binding is essential for TTD14 function. TTD14 is a cytosolic protein and binds to PtdIns(3)P, a lipid enriched in early endosome membranes, and to phosphatidic acid. In contrast to TRPL, rhabdomeral localization of the membrane proteins Rh1 and TRP is not affected in the ttd14P75L mutant. The ttd14P75L mutation results in Rh1-independent photoreceptor degeneration and larval lethality suggesting that other processes are also affected by the ttd14P75L mutation. In conclusion, TTD14 is a novel regulator of TRPL trafficking, involved in internalization and subsequent sorting of TRPL into the recycling pathway that enables this ion channel to return to the plasma membrane.

Protein trafficking in neurons occurs throughout the lifetime of a cell and includes the internalization and redistribution of plasma membrane proteins. Regulated protein trafficking controls the equipment of the plasma membrane with receptors and ion channels and thereby attenuates or enhances neuronal function. Defects in recycling of plasma membrane proteins can cause detrimental neurodegenerative diseases such as Alzheimer’s disease, Parkinson’s disease and Down´s syndrome. In Drosophila photoreceptors, the TRPL ion channel, together with the TRP channel, mediates vision and light-dependently shuttles between an endomembrane storage compartment and the apical plasma membrane. Here, we report the identification of a mutation in the ttd14 gene that inhibits TRPL-trafficking in both directions and also results in photoreceptor degeneration. The TTD14 protein contains a region with weak homology to a PX-domain, which is also found in proteins that sort cargo in the endosome and enable protein recycling. We characterize TTD14 as a new regulator of photoreceptor maintenance and ion channel trafficking that binds to GTP and PtdIns(3)P, a phospholipid enriched in early endosomes.

Photoreceptor membrane proteins undergo a carefully regulated turnover that helps to adjust the sensitivity of the receptors and to renew old and possibly worn out proteins. Throughout the lifetime of a photoreceptor cell, new proteins are synthesized and transported to the photoreceptive membrane while other proteins are removed from this membrane and are either recycled or degraded in the lysosome. Major integral membrane proteins of the Drosophila photoreceptive membrane comprise the G protein-coupled receptor rhodopsin and two ion channels, transient receptor potential (TRP) and TRP-like (TRPL). Defects in rhodopsin turnover can result in degeneration of photoreceptors in humans and flies [1–3]. The signaling cascade operating in fly photoreceptors is a G protein-coupled, phospholipase Cβ-mediated signaling pathway that is initiated by the absorption of a photon by rhodopsin and results in the opening of TRP and TRPL channels and subsequent influx of sodium and calcium ions. TRP and TRPL are the founding members of the large family of transient receptor potential channels that comprises 28 members in mammals [4–7]. TRP channels function in sensory systems as well as in calcium regulation in non-neuronal cells, for example in kidney or heart cells. In Drosophila, newly synthesized rhodopsin of R1-6 photoreceptor cells (Rh1) and the two light-activated ion channels TRP and TRPL are transported via the secretory pathway from the endoplasmic reticulum (ER) to the apical plasma membrane that forms a light-sensitive microvillar compartment, termed rhabdomere. Precise folding and successful transport of Rh1 and TRP channels to the rhabdomere are crucial for photoreceptor function. A number of reports have identified proteins required for the anterograde transport of Rh1 and for its endocytosis [8–19]. These include chaperones, Rab GTPases (Rab1, Rab6, Rab11), Rab-interacting proteins, a COPII-interacting phosphatidic acid phospholipase A1, and myosin V. Among these proteins, the chaperone XPORT, Rab11, and the Rab11-interacting guanine nucleotide exchange factor Crag were shown to also be required for TRP trafficking, but not for TRPL trafficking [9, 15, 18]. Although Rh1 undergoes a turnover that is enhanced in the light most of the Rh1 is detected in the rhabdomeres in both light- and dark-adapted flies. A significant portion of Rh1 that is removed from the rhabdomere in the light ends up in the lysosome and becomes degraded [20–22]. However, the rhabdomeral Rh1 content does not change significantly under physiological light conditions, indicating that degraded Rh1 is replenished by newly synthesized protein. It has recently been reported that a fraction of internalized Rh1 is not degraded but enters a recycling pathway that requires components of the retromer complex [23]. The retromer is a hetreromultimeric protein complex composed of Vps26, Vps29, Vps35, and sorting nexins [24–27]. It is a principle component of the retrograde transport from endosomes to the trans-Golgi network or for recycling of proteins to the plasma membrane for many recycling membrane proteins, including Wntless [28], the β2 adrenoreceptor [29], the Drosophila adherens junction protein Crumbs [30], and vertebrate AMPA receptor subunits [31–33]. Wang et al. [23] showed that Vps26 and Vps35 are required for Rh1 recycling and that mutations in these retromer proteins cause retinal degeneration. The Drosophila ion channel TRPL is a recycling photoreceptor membrane protein that undergoes light-dependent translocation between the rhabdomere, where it is located in dark-adapted flies, and a storage compartment in the cell body, to which it is transported upon illumination within several hours [34]. The removal of TRPL from the rhabdomere depends on activation of the phototransduction cascade and the resulting Ca2+ influx through TRP channels [35]. It has been described as a two-step process in which TRPL first (within 5–10 minutes) is transported to the base of the microvilli and adjacent stalk membrane via lateral membrane transport and then (within several hours) becomes internalized by a vesicular transport mechanism [36–38]. The majority of internalized TRPL does not enter the lysosomal pathway but is stored in the cell body and recycled back to the rhabdomere when the flies are transferred from light to darkness [34, 37]. In order to identify components required for TRPL transport, we previously performed a genetic screen for TRPL translocation-defective mutants [39]. This screen used a TRPL-eGFP reporter gene to monitor TRPL localization in intact flies and was based on FRT/FLP-driven mitotic recombination enabling the generation of homozygous mutant eye clones of otherwise lethal genes. In the present study, we mapped the mutant ttd14 and identified the mutated gene. The encoded protein TTD14 is a cytosolic GTP-binding protein that interacts with phospholipids and functions in the internalization and recycling of TRPL during light-triggered TRPL translocation between cell body and rhabdomere. In addition, the ttd14 mutant results in photoreceptor degeneration and larval lethality indicating a vital role of ttd14 in other contexts. The recessive mutant ttd14 is homozygous lethal during the larval stage and was obtained from a genetic screen of chromosome arm 2R that used ethyl methanesulfonate as a mutagene and was based on FRT/FLP-driven mitotic recombination of the chromosome arm to obtain homozygous mutant eye clones [39]. In the genetic screen, trafficking of the TRPL-eGFP reporter protein was visualized in mosaic eyes by the fluorescence of the deep pseudopupil and the ttd14 mutant was identified by its defective TRPL-eGFP internalization to the cell body after 16 hours orange light illumination. A detailed description of the screen, including crossing schemes, can be found in Ref. [39]. A more detailed analysis using age-matched flies revealed a complex phenotype including defective light-induced TRPL internalization to the storage compartment in young flies, TRPL depletion in the rhabdomere after long-term dark adaptation, and defective redistribution of TRPL to the rhabdomere upon light adaption and subsequent dark adaptation (Fig 1A). To identify the gene affected in the ttd14 mutant, we made use of the lethal phenotype of ttd14, assuming that lethality in larvae and impaired TRPL trafficking in mosaic eye clones is caused by the same mutation. We performed a mapping approach of the mutated chromosome arm 2R with deficiency strains. A lethal mutation could be mapped between 55C8–55C9, a region of 30 kilobases containing six candidate genes (Fig 1B). Among these candidate genes, a lethal P-Element mutation in the orphan gene CG30118 (CG30118KG03769, see Fig 1C) failed to complement the lethality of the ttd14 mutant demonstrating that ttd14 is a mutant allele of the CG30118 gene. As shown in Fig 1C, the ttd14 gene encodes three predicted transcripts (ttd14-A, ttd14-B, and ttd14-C). Sequencing of the coding sequence of the ttd14 gene derived from cDNA of wild type fly heads confirmed expression of ttd14-A and -B but not ttd14-C in Drosophila heads. Sequencing of genomic DNA obtained from a heterozygous ttd14 mutant revealed that the mutation present in the ttd14 allele is a C → T transition in the second exon altering the codon for proline75 to a codon for leucine (P75L) (Fig 2C). This mutation affects all ttd14 transcripts as P75 is encoded by a common exon. In order to demonstrate that the lethal phenotype and the TRPL transport defect are caused by the same mutation in ttd14, we generated myc-tagged and untagged rescue constructs driving the expression of ttd14-A or ttd14-B in R1-6 photoreceptor cells under the control of the Rh1 promoter. Upon expression of the constructs in ttd14P75L homozygous mutant eye clones, both ttd14 isoforms rescued the TRPL trafficking defects (S1A Fig). In addition, water immersion microscopy analysis of homozygous mutant eye clones harboring the lethal P-element insertion ttd14KG03769 in the ttd14 gene revealed the same TRPL trafficking defect as was observed in the ttd14P75L mutant (S1B Fig). These results demonstrate that the lethal ttd14P75L mutation causes the TRPL trafficking defect in homozygous mutant eye clones. The protein encoded by the ttd14 gene will hereafter be referred to as TTD14. No functional data about TTD14 are available so far. Protein domain prediction revealed a P-loop-containing nucleoside triphosphate hydrolase domain (P-loop) and a CYTH-like domain (Fig 2A). The predicted P-loop of TTD14 contains a bona fide GxxxxGKT Walker A motif (amino acids 73–80) and a possible hhhhDxG Walker B motif at position 152–158 (canonical Walker B motif: hhhhDxxG, where h is a hydrophobic and x is any amino acid; [40–42]). The mutated amino acid P75 is located in the Walker A motif and is highly conserved in the TTD14 homologs of other invertebrate species. CYTH domains were identified in bacterial adenylyl cyclases and mammalian thiamine triphosphatases [43]. Although highly conserved orthologs of TTD14 are present in other invertebrates such as bees or C. elegans (Fig 2B), BLAST sequence similarity searches did not reveal full-length homologs in vertebrates with significant sequence similarity to TTD14. Hydrophobicity analysis of TTD14 suggested that the protein is a soluble or a peripheral membrane protein since no sufficiently hydrophobic regions for putative transmembrane domains were predicted (Fig 3A). In order to determine the subcellular localization of TTD14 experimentally, we performed biochemical fractionation experiments and immunocytochemistry. We generated a polyclonal antibody against a recombinantly-expressed full length TTD14-A protein. To test the specificity of the generated antibody, an immunoblot experiment was carried out using protein extracts of dissected Drosophila eyes of wild type flies, of flies overexpressing TTD14-A in photoreceptor cells under the control of the Rh1 promoter, and of flies having homozygous mutant ttd14 eye clones (comprising 80–90% of the eye) (Fig 3B). The anti-TTD14 antibody detected two protein bands with an apparent molecular weight of 55 kDa and 130 kDa, respectively, in wild type eyes. These bands were reduced or undetectable in ttd14P75L and ttd14KG03769 mosaic eyes, respectively, and greatly enhanced in eyes overexpressing TTD14-A. Although it cannot be excluded that residual TTD14 protein detected in eyes with ttd14P75L mutant eye clones results from the eye parts that are heterozygous for ttd14P75L, the loss of TTD14 in these eyes seems to be less severe than in ttd14KG03769 mosaic eyes. Therefore, the P75L point mutation may less severely affect the stability of the TTD14 protein but rather impair protein function. Since the calculated molecular weight is 54,3 kDa for TTD14-A and 53,5 kDa for TTD14-B, we assume that the 55 kDa and 130 kDa bands detected by the antibody represent monomers and dimers of TTD14-A and -B isoforms. For fractionation experiments, Drosophila heads were homogenized in Tris/NaCl-buffer and soluble and membrane proteins were separated by ultracentrifugation (Fig 3C). Using immunoblot analysis, the TTD14 protein was detected exclusively in the soluble fraction. Although the antibody specifically detected TTD14 on immunoblots, it did not yield specific signals when used in immunocytochemistry. Therefore, we used flies expressing a myc-tagged TTD14 (TTD14-myc) in photoreceptor cells under the control of the Rh1 promoter and anti-myc antibodies for immunocytochemistry on cross sections through eyes of 24 hours dark- or 16 hours orange light-adapted flies. These experiments revealed a rather uniform signal in the cell body but no signal in the rhabdomeres (Fig 3D). No obvious differences in localization of TTD14-myc between photoreceptor cells of light- or dark-adapted flies were observed. Co-localization of TRPL and TTD14-myc on cross sections through eyes showed that there is indeed a considerable overlap in TRPL and TTD14-myc staining in eyes of light-adapted flies but not in eyes of dark-adapted flies where TRPL is localized in the rhabdomeres (Fig 3D). It has to be noted, however, that in eyes of light-adapted flies, TTD14-myc staining does not completely overlap with TRPL staining, suggesting that the TTD14 protein is not specifically enriched in membrane compartments occupied by TRPL. In order to test the predicted nucleotide binding activity of the TTD14 protein, we performed an in vitro nucleotide binding assay. Purified recombinant full-length TTD14-A protein was incubated with agarose beads coupled to ATP, GTP, or no nucleotide (control). After removal of unbound protein, TTD14 was eluted from the beads with either 25 mM ATP or 25 mM GTP. TTD14 protein was recovered from GTP beads but not from ATP or control beads indicating that TTD14 is a GTP-binding protein (Fig 3E). In order to test an effect of the P75L mutation present in the ttd14P75L allele on GTP-binding, we generated a plasmid for recombinant expression of TTD14 protein harboring the P75L amino acid substitution. This amino acid substitution did not affect protein stability as the mutated protein could be purified in an amount comparable to the wild type TTD14 protein (Fig 3E). However, the P75L mutation abolished GTP binding of TTD14 in the in vitro nucleotide binding assay (Fig 3E) indicating that the predicted Walker A motif (amino acids 73–80) is likely to be involved in GTP binding. In addition, the mutant phenotype of the ttd14P75L allele might be attributed to the loss of GTP binding activity, suggesting that GTP binding is essential for the biological function of the TTD14 protein. As differentially-distributed phosphoinositides play a crucial role in membrane trafficking by recruiting trafficking proteins to the membrane [44], we tested a possible interaction of TTD14 with membrane lipids. In a lipid binding assay we employed nitrocellulose strips spotted with 100 pmol of various phospholipids (PIP-Strips). The PIP-Strips were incubated with purified recombinant full length TTD14-A protein and bound protein was detected with the anti-TTD14 antibody. As a result, among the 15 lipids tested, TTD14 bound to 3-phosphoinositide (PtdIns(3)P) and to phosphatidic acid (PA) (Fig 3F). PtdIns(3)P is predominantly localized at the cytosolic side of membranes of early endosomes [44], suggesting that TTD14 interacts with early endosome membranes. Collectively, our results reveal that TTD14 is a soluble, most likely cytosolic protein expressed in the Drosophila eye that binds GTP and interacts with PtdIns(3)P and PA. For a detailed analysis of the TRPL trafficking defect in ttd14P75L, flies expressing TRPL-eGFP in R1-6 photoreceptor cells were subjected to water immersion microscopy of intact eyes and the TRPL-eGFP fluorescence in rhabdomeres was determined as in [35] (Fig 4A). As the generated Rh1>ttd14-A and Rh1>ttd14-B rescue constructs are associated with a white+-marker and result in an orange eye color that interferes with the quantification of water immersion images, a yellow+-marked Rh1>ttd14-A-myc construct bearing a C-terminal myc-tag was expressed in white-eyed flies and used for this quantitative analysis. Qualitatively, rescue of the ttd14P75L phenotype by untagged ttd14 constructs is shown in S1A Fig. In addition, the phenotype in homozygous ttd14KG03769 mutant eye clones, that have red eye color, is depicted in S1B Fig and compared to red-eyed wild type flies. In wild type flies (Fig 4B green bars), long-term dark incubation for up to 7 days only slightly reduced the rhabdomeral TRPL-eGFP fluorescence. In contrast, after 16 hours illumination with orange light, rhabdomeral fluorescence was reduced by 78% in 1 day old flies, indicating that most of TRPL-eGFP had translocated to the cell body. Light-triggered TRPL translocation was less efficient in flies that were previously kept in darkness for 3 days or 7 days but still resulted in a reduction of the rhabdomeral TRPL-eGFP fluorescence by more than 55%. Subsequent dark adaptation for 24 hours fully restored the original rhabdomeral TRPL-eGFP fluorescence irrespective of the duration of the initial dark incubation. In ttd14ttd14 mutant eye clones (Fig 4B, red bars) of 1 day old flies, TRPL-eGFP was properly located in the rhabdomere in the dark, but hardly translocated to the cell body upon illumination with orange light for 16 hours. With increasing time of dark incubation, TRPL-eGFP fluorescence progressively disappeared from the rhabdomere resulting in a rhabdomeral TRPL-eGFP fluorescence of only 40% in 7 day old flies. Orange light illumination in these flies did not further reduce the rhabdomeral TRPL-eGFP content. Of primary significance here, subsequent dark adaptation for 24 hours did not affect TRPL-eGFP distribution in 3 day or 7 day old flies, indicating that recycling of TRPL-eGFP to the rhabdomere was severely impaired. This latter phenotype is less prominent in young flies, as the initial rhabdomeral TRPL content before dark adaptation is already elevated. In conclusion, mutation of the ttd14 gene did not affect the initial localization of TRPL-eGFP in the rhabdomere but resulted in reduced transport of TRPL-eGFP from the rhabdomere to the cell body in young flies. Upon prolonged incubation of flies in the dark, net transport of TRPL-eGFP from the cell body to the rhabdomere was reduced as most of TRPL-eGFP was located in the cell body and failed to recycle to the rhabdomere irrespective of the light condition. All aspects of the mutant phenotype described above were rescued by expression of ttd14 wild type constructs (Fig 4B, blue bars and S1A Fig) in R1-6 photoreceptor cells under control of the Rh1 promoter showing that TTD14 function is required in photoreceptor cells. In order to reveal the subcellular localization of TRPL with higher resolution, we performed immunocytochemistry of 7 day old wild type, ttd14P75L, and rescue flies (Fig 4C). Like in the water immersion experiments, the flies were first kept in constant darkness, then illuminated with orange light for 16 hours and then again kept in darkness for 24 hours. As observed previously [34, 37], wild type flies revealed TRPL labeling in the rhabdomeres when kept in darkness and a relatively uniform labeling of the cell bodies (except for nuclei) when kept in light. The same labeling pattern was observed in ttd14 mutant eye clones of flies expressing the rescue construct Rh1>ttd14-myc. Without the rescue construct, ttd14 mutant eye clones revealed a TRPL labeling pattern in the cell body in all light and dark conditions tested that was indistinguishable from that of wild type flies kept in the light. We previously reported a localization defect of a mutated TRPL, in which eight C-terminal phosphorylation sites were abolished [45]. However, the mislocalization of phosphorylation-deficient TRPL resulted in labeling of distinct spots in the cell body quite different from the uniform TRPL labeling pattern observed in the ttd14 mutant (S2 Fig and see Ref. 45). In addition to the localization defect, removal of TRPL phosphorylation sites affected TRPL stability in the dark and resulted in progressive TRPL degradation [45]. To assay TRPL stability in the ttd14P75L mutant, we carried out immunoblot analyses with protein extracts obtained from 1 day and 7 day old wild type flies and from the ttd14P75L mutant subjected to different light conditions (Fig 4D and 4E). A lower amount of TRPL in 1 day old flies, which were assayed immediately (that is before further incubation in light or darkness, Fig 4D, lanes 1 and 7), was observed in both wild type and mutant, suggesting that the TRPL content in freshly eclosed flies is lower than in older flies. No indication of TRPL degradation in the ttd14P75L mutant flies was observed. Indeed, except for flies freshly eclosed in the dark (Fig 4E, first and 7th column), ttd14P75L mutant flies exhibited significantly (p < 0.05) higher TRPL protein levels as compared to the corresponding condition in wild type flies. Thus, TRPL in the ttd14P75L mutant is not targeted to the lysosomal degradation pathway. It rather seems that the ttd14 mutation affects the recycling of TRPL which accumulates in the internal storage compartment. While TRPL becomes internalized from the rhabdomere in the light and is recycled back to the rhabdomere in the dark, two other important Drosophila photoreceptor membrane proteins, Rh1 and TRP, do not seem to undergo such a regulated change in their subcellular localization. Rh1 is constantly renewed in illuminated photoreceptor cells as the internalized Rh1 becomes partially degraded in the lysosome and rhabdomeral Rh1 is replenished by newly synthesized protein [20]. In addition, Rh1 is also partially recycled via the retromer complex [23]. However, both mechanisms do not result in a major change of the subcellular localization, as most of the Rh1 protein remains localized in the rhabdomere. Albeit less well studied, there is no evidence for a light-dependent translocation of the TRP protein. Indeed, a number of mutations in chaperones and transport proteins have been described that affect the anterograde transport of both Rh1 and TRP, but not TRPL suggesting a common transport pathway for these two photoreceptor proteins and a different pathway for TRPL (see introduction). Assuming that loss of functional TTD14 does not induce a general cytological defect but a specific defect in the internalization and recycling of membrane proteins such as TRPL, one might expect that the ttd14 mutation does not affect trafficking of Rh1 and TRP. To test this assumption, we carried out water immersion microscopy with flies expressing eGFP-tagged Rh1 and TRP in R1-6 photoreceptor cells and also performed immunocytochemical experiments (Fig 5). While we had observed that TRPL-eGFP disappears from the rhabdomeres of ttd14 mutants upon dark incubation for 7 days (see Fig 4), no age-related changes in the rhabdomeral content of Rh1-eGFP or TRP-eGFP were observed in water immersion microscopy for up to 28 days in darkness. The fluorescence pattern of Rh1-eGFP or TRP-eGFP expressed in ttd14P75L mutant eye clones was the same as in wild type (Fig 5A). Likewise, immunocytochemistry of cross sections through ommatidia of 7 day old flies kept in the dark, revealed no signs of mislocalization of Rh1-eGFP or native TRP that were both confined to the rhabdomeres in wild type flies as well as in ttd14P75L mutant eye clones (Fig 5B). In addition, we assayed the Rh1 and TRP protein content of 1 day and 7 day old ttd14P75L mutant flies, which were kept in darkness and then subjected to orange light illumination for 16 hours, by immunoblot analysis and observed the same protein content as in wild type flies (S3 Fig, Fig 4). Taken together, these results strongly suggest that the transport of Rh1 or TRP to the rhabdomere is not affected in the ttd14P75L mutant. Internalization of Rh1 can be assessed by immunoblot analysis when flies are illuminated with bright white light for at least 16 hours. In wild type flies, this illumination results in a reduced Rh1 level due to lysosomal degradation and the original Rh1 level is recovered after a subsequent incubation in darkness for 6 hours [23]. In mutants with defective Rh1 internalization, reduced lysosomal degradation would be expected. In contrast, in mutants affecting components required for Rh1 recycling like the retromer complex, enhanced Rh1 degradation results in a stronger reduction of the Rh1 level as compared to the wild type [23]. In illuminated ttd14P75L mutant flies, the Rh1 level was reduced to a similar degree as in wild type flies (Fig 5C and 5D) arguing against a major role of the TTD14 protein in Rh1 trafficking. Finally, in order to show that the ttd14 mutant does not affect the function of Rh1 and TRP, we performed electroretinogram (ERG) recordings of wild type and ttd14P75L mutant flies kept in darkness for 7 d. At that time, TRPL was present in the rhabdomeres of wild type flies but not in the rhabdomeres of ttd14P75L mutant flies (Fig 4A–4C). A loss of rhabdomeral TRPL is not expected to have a major impact on the ERG, as the loss of TRPL in the trpl302 null mutant, has an effect on light adaptation, but no impact on the shape of ERG recordings [46]. In contrast, impaired Rh1 or TRP function causes characteristic changes in ERG recordings (Fig 6A). A severe reduction in the amount of Rh1, which can be achieved by feeding flies with a vitamin A-deprived diet (see below), resulted in a loss of the prolonged depolarization afterpotential (PDA) after bright blue light illumination while a complete loss of Rh1, like in the ninaE17 mutant, was readily indicated by a dramatic reduction of the ERG amplitude (Fig 6A). Loss of TRP function, like in the trpP343 mutant, can be detected by the characteristic transient response to a light stimulus (Fig 6A). Using a stimulus protocol containing both orange and blue light illumination, we did not observe any obvious differences in ERG recordings from ttd14P75L mutant eye clones as compared to wild type (Fig 6A). We conclude that besides TRPL, neither Rh1 nor TRP nor any other major component of the phototransduction cascade is severely affected in ttd14P75L mutant flies. In order to asses long-term effects of the ttd14P75L mutation on photoreceptor function, we carried out ERG recordings of wild type flies, ttd14P75L mutant flies, and flies expressing the rescue construct Rh1>ttd14-A-myc in a ttd14P75L mutant background (Rescue) for up to 21 d. In a 12 hours light / 12 hours dark cycle, ttd14P75L mutant flies exhibited a decline in ERG amplitude after 7 days and an almost abolished photoresponse after 14 days (Fig 6B and 6C). After 21 days in constant darkness, ttd14P75L mutant flies also displayed a reduction in the ERG amplitude, albeit to a much lower extent than in a 12 hours light / 12 hours dark cycle (Fig 6B and 6C). In mutants of the retromer complex, photoreceptor degeneration that results in a loss of photoreceptor function is based on the mislocalization of Rh1 and can be attenuated by a reduction of the Rh1 level [23]. The Rh1 level can effectively be reduced by raising the flies on a vitamin A-deprived diet as the opsin protein becomes degraded in the absence of its chromophore (Fig 6D). In contrast to mutants of the retromer complex, diet-induced reduction of the Rh1 level had no significant effect on the ERG amplitude in response to an orange light stimulus in wild type or ttd14P75L mutant flies and did not rescue the declined ERG observed in ttd14P75L mutant flies kept in a 12 hours light / 12 hours dark cycle for 21 days (Fig 6E). In line with our results showing that the ttd14P75L mutation does not affect Rh1 localization (see Fig 5), this finding suggests that aberrant Rh1 localization is not causal for the loss of photoreceptor performance in this mutant. In order to assess if the decline of the ERG amplitude was associated with morphological alterations in the rhabdomeres, we assessed rhabdomeral structure in wild type and ttd14P75L mutant eye clones for up to 21 days in a 12 hours light / 12 hours dark cycle by transmission electron microscopy and by monitoring TRP-eGFP fluorescence (Fig 7). While no obvious changes could be detected in wild type eyes, electron microscopy revealed severe degeneration of photoreceptor cells in ttd14P75L mutant flies kept in a 12 hours light / 12 hours dark cycle for 21 days. Most rhabdomeres of photoreceptor cells R1-6 and also the rhabdomere of R7 were absent in these flies. Degeneration was much less pronounced when the ttd14P75L mutant was kept in the dark for 21 days. In these flies the R7 cell was affected frequently while almost all rhabdomeres of R1-6 cells remained intact. These findings show that degeneration of photoreceptor cells in the ttd14P75L mutant is enhanced by light and affects inner and outer photoreceptor cells. Of note, no signs of degeneration of inner or outer photoreceptor cells were detected in ttd14P75L mutant flies kept in the dark for 7 days followed by 16 h orange light, a condition in which TRPL failed to recycle back to the rhabdomere. This finding suggests that the TRPL trafficking defect in the ttd14P75L mutant is not a result of photoreceptor cell degeneration. As observed before Vitamin A deprivation resulted in diminished rhabdomere size [23]. However, degeneration was also observed in vitamin A-deprived ttd14P75L mutant flies showing that a reduction of the rhodopsin content cannot rescue the degeneration phenotype. Using water immersion microscopy with ttd14P75L mutant eye clones that express TRP-eGFP as a rhabdomeral marker we analyzed the time course for the loss of rhabdomeres both under regular and vitamin A-deprived conditions (Fig 7B and S4 Fig). In flies kept in a light-dark cycle first signs of degeneration were detected after seven days while almost all rhabdomeres were lost after 21 days (Fig 7B). Vitamin A-deprivation had a small but not statistically significant effect on the degeneration time course and slightly slowed down the speed of degeneration in the ttd14P75L mutant. Taken together, loss of functional TTD14 in the ttd14P75L mutant resulted in late onset light-dependent, but Rh1-independent retinal degeneration. Recycling of plasma membrane proteins plays a pivotal role in the function and maintenance of neurons. Defects in protein recycling caused, for example, by mutations in components of the retromer complex that regulates recycling and retrograde transport of proteins in early endosomes have been implicated in detrimental neurodegenerative diseases such as Alzheimer’s disease, Parkinson’s disease and Down´s syndrome [32, 47–49]. The TRPL ion channel of Drosophila photoreceptor cells is a useful model for studying recycling of neuronal plasma membrane proteins as its internalization from and redistribution to the plasma membrane can be triggered simply by exposing the flies to light or darkness, respectively [34–36, 45]. Due to its genetic versatility, Drosophila offers the possibility to identify novel components of protein recycling pathways in neurons by unbiased genetic screens. We had previously performed a genetic screen for mutants defective in TRPL ion channel transport [39] and in this study, we identified the orphan gene ttd14 as a major regulator of TRPL trafficking in Drosophila photoreceptor cells. The mutant ttd14P75L was initially identified by the internalization defect of TRPL-eGFP after illumination. However, TRPL internalization is not completely abolished in ttd14P75L mutant flies suggesting that the TTD14 protein is not crucial for TRPL internalization but rather acts as a modulator that facilitates TRPL internalization. Alternatively, TTD14 might play an indirect role in TRPL internalization and promote TRPL internalization by enabling recycling of a rhabdomeral protein that is required for TRPL internalization. The most prominent phenotype of the ttd14P75L mutant is the failure of TRPL trafficking from the storage compartment back to the rhabdomere during dark adaptation. This phenotype is not evident in young flies, in which TRPL is localized in the rhabdomeres, suggesting that delivery of newly synthesized TRPL to the rhabdomere via the secretory pathway is not affected by the ttd14P75L mutation. However, when TRPL is redistributed from the rhabdomere to the storage compartment in illuminated 3 day or 7 day old ttd14P75L mutant flies, subsequent dark adaptation for 24 hours does not result in redistribution of TRPL to the rhabdomere (see Fig 4). This finding indicates that internalized TRPL fails to become recycled to the plasma membrane in ttd14P75L. Continuous dark adaptation of the ttd14P75L mutant for 7 days also results in localization of TRPL in the cell body. This finding indicates that there is a basal level of TRPL internalization in the dark. In wild type flies, TRPL that becomes internalized in the dark is readily recycled back to the rhabdomere resulting in little if any TRPL in the cell body in the dark. In ttd14P75L, where TRPL recycling is blocked, prolonged dark adaptation depletes TRPL from the rhabdomere and results in TRPL accumulation in the cell body. We had previously observed that mutation of C-terminal phosphorylation sites of TRPL also results in depletion of mutated TRPL from the rhabdomere when the flies are kept in darkness for five days [45]. This finding again argues for a basal rate of TRPL internalization in the dark. The phosphorylation-deficient TRPL, however, does not accumulate in the cell body but becomes degraded. Therefore, this mutation might shift the balance between recycling and lysosomal degradation of TRPL towards degradation rather than hindering TRPL entry into the recycling pathway. The proposed cellular trafficking pathways of TRPL in light and darkness and the steps that are presumably disturbed in the ttd14 mutant are illustrated in Fig 8. Our biochemical analysis revealed that TTD14 is a soluble protein that binds GTP and the phospholipids PtdIns(3)P and phosphatidic acid (see Fig 3). Protein domain searches revealed a P-loop nucleotide binding domain comprising amino acids 65–167 of TTD14. This domain contains a bona fide Walker A and a possible Walker B motif and thus represents a likely site for binding of GTP or ATP. When compared with vertebrate proteins the domain exhibited 37% amino acid identity with the respective domain of mitochondrial GTPase Era (S1 Table). Binding assays with recombinant TTD14 revealed binding to GTP but not to ATP. The P75L mutation of ttd14P75L affects a conserved proline in the Walker A motif and abolished binding of TTD14 to GTP. This finding indicates that the P-loop domain indeed is the GTP binding site. It also suggests that GTP binding is important for TTD14 function as mutation of the site results in the observed ttd14 phenotype. A role of GTP binding proteins in the context of protein trafficking is well established. Important examples are the families of small GTPases, including Rho proteins that regulate actin organization, Rab proteins that regulate docking and fusion of vesicles at different organelles of the endocytic and secretory pathways, and Arf proteins that are involved in protein transport from Golgi to endoplasmic reticulum or at the trans-Golgi network [50]. Other examples for GTP binding proteins in this context are the microtubule forming subunits α and β tubulin and dynamin, which is required to pinch off clathrin-coated vesicles from membranes [51, 52]. Except for dynamin, GTP hydrolysis by these proteins is not used to generate force but rather to induce conformational changes in the GTP binding proteins that regulate protein function. Besides to GTP, TTD14 also bound to the phospholipids PtdIns(3)P and phosphatidic acid (PA). Phosphoinositides that differ in the phosphorylation state of the inositol ring are distributed differentially in cellular compartments [44]. Accordingly, trafficking proteins that bind specific phosphoinositides are recruited to distinct membrane compartments [44]. PtdIns(3)P is located mainly in the early endosome [44]. Therefore, binding to PtdIns(3)P may recruit TTD14 to the early endosome in order to perform its function in TRPL recycling. PA is an intermediate in the biosynthesis of phosphoinositides and other phospholipids. In Drosophila photoreceptor cells, PA is generated during phototransduction that generates diacylglycerol (DAG) by phospholipase Cβ-mediated hydrolysis of phosphatidylinositol 4,5-bisphosphate. DAG in turn is translocated to the submicrovillar cisternae located at the base of the rhabdomere where it is phosphorylated to PA by the rdgA-encoded DAG kinase as a first step in the regeneration of rhabdomeral phosphatidylinositol 4,5-bisphosphate [53, 54]. Therefore, upon illumination, TTD14 could be recruited to the base of the rhabdomere via binding to PA, where it could assist light-triggered internalization of TRPL. However, an obvious light-dependent redistribution of myc-tagged TTD14 was not observed in immunocytochemical experiments (see Fig 3D). It has also been suggested that PA levels are critical for apical membrane transport events required for rhabdomere biogenesis in Drosophila photoreceptors as elevated PA levels in Drosophila mutants caused defects in rhabdomere biogenesis [55]. However, defective biogenesis of rhabdomeres during development was not observed in ttd14 mutants and TRPL was properly transported to the rhabdomeres in young flies, arguing against a role of TTD14 in apical membrane trafficking events required for rhabdomere biogenesis. As protein domain prediction did not reveal any lipid binding domain in the TTD14 protein, we performed a BLAST search with low stringency which revealed regions within TTD14 that showed amino acid sequence identities to vertebrate proteins between 30% and 37%. Homologous regions in vertebrate proteins that appeared in more than one species in the BLAST screen and contained specific protein domains within the homologous region included the Phox homologous- (PX) domain within vertebrate sorting nexin-8 homologs. This domain displayed sequence similarity with amino acids 254–320 of TTD14 (S1 Table). PX-domains were first identified in NADPH oxidase subunits, sorting nexins, and PtdIns(3)P-kinases [56]. These domains bind to phosphoinositides and sorting nexins are typically recruited to the early endosome membrane by binding to PtdIns(3)P via their PX-domain [57, 58]. Besides in the PX-domain there is little amino acid sequence conservation among the various sorting nexins and the PX-domain is located at varying positions. Sorting nexins are involved in the recycling of internalized proteins. They can be part of retromer complexes that sort internalized proteins in early or late endosomes away from the lysosomal degradation pathway and enable their trafficking back to the trans-Golgi network or to the plasma membrane [24, 59]. This function would be in accordance with a function of TTD14 in TRPL recycling. Yet, TTD14 is not an ortholog of a known sorting nexin as there is no significant sequence similarity over the whole protein length to any of the described 33 vertebrate sorting nexins. Given the weak sequence conservation between different classes of sorting nexins and the functional analogy of TTD14 with sorting nexins, it is tempting to speculate that TTD14 represents a new class of sorting nexin, composed of a GTP-binding domain and a putative PX-domain that is present in Drosophila and other invertebrates but not in vertebrates. The ttd14P75L mutant displays a light-dependent late onset photoreceptor degeneration as ERG amplitudes diminished and rhabdomere structure became distorted when flies were kept in a 12 hours light / 12 hours dark cycle for more than 7 days (see Figs 6B, 6C and 7 and S4 Fig). While defects in the phototransduction cascade that result in constitutively open TRP channels, for example in rdgA or trpP365 mutants, lead to fast photoreceptor degeneration within days after eclosion [54, 60, 61] light-dependent late onset photoreceptor degeneration of Drosophila photoreceptor cells can result from defective Rh1 transport [2, 62]. Accordingly, mutations in proteins required to transport Rh1 to the rhabdomere like XPORT, Rab11 or Crag cause photoreceptor degeneration [9, 18, 63]. Importantly, mutations in Vps26 or Vps35, components of the retromer complex, also cause light-dependent late onset photoreceptor degeneration [23]. This has been attributed to a failure in Rh1 recycling, which then accumulates in late endosomes or lysosomes and exerts a toxic effect on photoreceptor cells possibly due to an overload of the endolysosomal system [23]. Since ttd14 mutants display a defect in TRPL recycling we wondered if they also have a defect in Rh1 recycling that would then cause photoreceptor degeneration. In contrast to TRPL, in flies dark-adapted for 28 d, Rh1 (and also TRP) were properly located in the rhabdomere as revealed by water immersion microscopy (Fig 5A). In addition, in the ttd14P75L mutant, over night illumination of flies with white light under conditions, which markedly reduced Rh1 levels in the Vps261 mutant [23], resulted in the same Rh1 levels as in wild type. Finally, while a reduction of the Rh1 content by genetic means or by vitamin A deprivation rescued photoreceptor cells from degeneration in the Vps261 mutant [23], Vitamin A deprivation did not rescue degeneration in the ttd14P75L mutant kept under a 12 hours light /12 hours dark cycle (see Figs 6E and 7 and S4 Fig). Thus, the defective ttd14 allele does not seem to affect Rh1 recycling or trafficking of Rh1 to the rhabdomere, and photoreceptor degeneration caused by ttd14P75L cannot be attributed to defective Rh1 trafficking. These findings indicate that TTD14 is specifically required for TRPL recycling but not Rh1 recycling, either because TRPL utilizes a different recycling pathway than Rh1 or because TTD14 specifically recruits TRPL to a common recycling pathway. What then might be the reason for photoreceptor degeneration in ttd14P75L flies? It can be excluded that the lack of TRPL in the rhabdomeres, observed in older ttd14P75L flies irrespective of the light condition, underlies photoreceptor degeneration as photoreceptors of the trpl302 null mutant do not degenerate [46]. Rather than by Rh1 accumulation, degeneration of photoreceptor cells in ttd14P75L could be caused by accumulation of TRPL in the endocytic pathway. However, the amount of TRPL in photoreceptor cells is much smaller than the amount of Rh1, and dark-kept ttd14P75L flies, in which TRPL was found to be accumulated in the cell body, showed only little photoreceptor degeneration. Since a homozygous ttd14P75L mutant is lethal during the larval stage while all known trpl mutants are viable, it is likely that other membrane proteins require TTD14 for their recycling. Therefore, other proteins requiring TTD14 for recycling might be involved in photoreceptor degeneration. These proteins remain to be identified. The following strains of Drosophila were used: yw, w1118 (here referred to as wild type), Oregon R (wild type with red eyes), ninaE17 [64], trp343 [65], yw;;P[Rh1>TRPL-eGFP,y+] [35], yw;P[Rh1>Rh1-eGFP,w+] [66], yw;;P[Rh1>TRP-eGFP,y+] [67], yw,ey>flp; FRT42D,ttd14P75L/CyO, yw; FRT42D,w+,2R11.5(lth)/CyO [39], yw; FRT42D,w+,2R11.5(lth)/CyO; P[Rh1>TRPL-eGFP,y+], yw,P[Rh1>ttd14-A-myc,y+], yw;;att86Fb[Rh1>ttd14-A,w+], yw;;att86Fb[Rh1>ttd14-B,w+], yw, ey>flp; FRT42D,ttd14KG03769[w+]/CyO, yw,ey>flp; P[Rh1>Rh1-eGFP,w+], FRT42D,ttd14P75L/CyO, yw; FRT42D,ttd14P75L/CyO, P[ActGFP,w+]JMR1, yw; FRT42D,ttd14KG03769[w+]/CyO, P[ActGFP,w+]JMR1. Flies were raised on standard cornmeal food at 25°C unless indicated otherwise. Flies were either kept in the dark or in a 12 hours light / 12 hours dark cycle using white light illumination (1300 lux). For vitamin A deprivation, flies were raised and kept on food containing 10% dry yeast, 10% sucrose, 2% agar, and 0,02% cholesterol. For the analysis of TRPL ion channel translocation, flies were kept in the dark for the indicated period and were then illuminated with orange light (acrylic glass cut off filter transmitting light >560 nm, ~200 lux) for 16 hours. Light-raised flies were dissected under white light whereas dark-raised flies were dissected under dim red light (Schott RG 630, cold light source KL1500, Schott). For the analysis of the amount of Rh1 protein flies were illuminated with white light (1800 lux). To map the mutation in ttd14, yw,ey>flp, P[Rh1>TRPL-eGFP,y+]; FRT42D,ttd14P75L/CyO mutant flies were crossed to the Bloomington 2R Deficiency Kit and the offspring was screened for lethality and TRPL translocation. In addition, the deficiency stocks Df(2R)ED3610 (54F1-55C8), Df(2R)BSC483 (55A1-55B7), Df(2R)334 (55B2-55C4), Df(2R)3636 (55B8-55E3), Df(2R)Excel7153 (55B9-55C1), Df(2R)BSC337 (55B11-55C9), Df(2R)ED3683 (55C2-56C4), Df(2R)BSC335 (55C6-55F1), Df(2R)BSC399 (55D1-55E10), Df(2R)BSC339 (55E2-55F6), and Df(2R)Excel7158 (55E9-55F6) were used to refine the region of the lethal ttd14P75L mutation. For the analysis of TRPL-eGFP localization in the eye of ttd14P75L mutant Drosophila, mosaic eyes were generated by crossing yw,ey>flp; FRT42D,ttd14P75L/CyO females with yw; FRT42D,w+,2R11.5(lth)/CyO; P[Rh1>TRPL-eGFP,y+] males. Female offspring with mosaic eyes (yw/yw,ey>flp; FRT42D,ttd14P75L/FRT42D,w+,2R11.5(lth); P[Rh1>TRPL-eGFP,y+]/+) was analyzed for the localization of TRPL-eGFP using the deep pseudopupil and water immersion microscopy. For wild type flies, yw females were crossed with yw;;P[Rh1>TRPL-eGFP,y+] males. Female offspring of the genotype yw;;P[Rh1>TRPL-eGFP,y+]/+ was analyzed. To analyze the rescue construct (Rh1>ttd14-A-myc) in ttd14P75L mutant mosaic clones, we crossed females expressing the rescue construct on the X-chromosome (yw,P[Rh1>ttd14-A-myc,y+]) with yw; FRT42D,w+,2R11.5(lth)/CyO; P[Rh1>TRPL-eGFP,y+] males. F-1 males without the CyO balancer (yw,P[Rh1>ttd14-A-myc,y+]; FRT42D,w+,2R11.5(lth)/+; P[Rh1>TRPL-eGFP,y+]/+) were then crossed with yw,ey>flp; FRT42D,ttd14P75L/CyO females. In the F-2 generation, females carrying mosaic eyes were selected for the presence of the TRPL-eGFP reporter by fluorescence microscopy using a Leica MZ16 F stereomicroscope prior to water immersion microscopy. Analyzed flies had the following genotype: yw,P[Rh1>ttd14-A-myc,y+]/yw,ey>flp; FRT42D,ttd14P75L/FRT42D,w+, 2R11.5(lth); P[Rh1>TRPL-eGFP,y+]/+. For the analysis of the rescue by the non-tagged Rh1>ttd14-A and Rh1>ttd14-B construct, the transgenic males yw;;att86Fb[Rh1>ttd14-A,w+] and yw;;att86Fb[Rh1>ttd14-B,w+] were crossed to yw,ey>flp; FRT42D,ttd14P75L/CyO females. The heterozygous rescue construct can be traced by the orange eye color. F-1 males with orange colored eyes and without the CyO balancer (yw,ey>flp; FRT42D,ttd14P75L/+; att86Fb[Rh1>ttd14-A or B,w+]/+) were then crossed with yw; FRT42D,w+,2R11.5(lth)/CyO; P[Rh1>TRPL-eGFP,y+] females. In the F-2 generation, females carrying mosaic eyes composed of orange and red clones (yw/yw,ey>flp; FRT42D,ttd14P75L/FRT42D,w+,2R11.5(lth); P[Rh1>TRPL-eGFP,y+]/att86Fb [Rh1>ttd14-A or B,w+]) were analyzed. For the analysis of the TRPL localization in the ttd14KG03769[w+] allele, this mutant allele was recombined with the FRT42D locus. Next, yw, ey>flp; FRT42D,ttd14KG03769[w+]/CyO males were crossed with yw; FRT42D,w+,2R11.5(lth)/CyO; P[Rh1>TRPL-eGFP,y+] females. In the F-1 generation, females lacking the CyO balancer (yw/yw,ey>flp; FRT42D,ttd14KG03769[w+]/FRT42D,w+,2R11.5(lth); P[Rh1>TRPL-eGFP,y+]/+) were analyzed. For the wild type control yw;;P[Rh1>TRPL-eGFP,y+] flies were crossed with Oregon R flies and F-1 females (yw/+;; P[Rh1>TRPL-eGFP,y+]/+) were analyzed. For the analysis of TRP-eGFP localization, flies carrying the TRP-eGFP reporter (yw;;P[Rh1>TRP-eGFP,y+]) were crossed with yw,ey>flp; FRT42D,ttd14P75L/CyO females. F-1 males lacking the CyO balancer (yw,ey>flp; FRT42D,ttd14P75L/+; P[Rh1>TRP-eGFP,y+]/+) were then crossed to yw; FRT42D,w+,2R11.5(lth)/CyO females. In the F-2 generation female flies with mosaic eyes were selected for the presence of the TRP-eGFP reporter using a Leica MZ16 F stereomicroscope prior to water immersion microscopy. Analyzed flies had the following genotype: yw/yw,ey>flp; FRT42D,ttd14P75L/FRT42D,w+,2R11.5(lth); P[Rh1>TRP-eGFP,y+]/+. For the wild type control, yw females were crossed to yw;;P[Rh1>TRP-eGFP,y+] males and F-1 females (yw;;P[Rh1>TRP-eGFP,y+]/+) were subjected to water immersion microscopy. For the analysis of Rh1-eGFP localization in water immersion microscopy and immunocytochemistry, the P[Rh1>Rh1-eGFP,w+] construct was recombined to the FRT42D,ttd14P75L chromosome resulting in the fly stock yw,ey>flp; P[Rh1>Rh1-eGFP,w+], FRT42D,ttd14P75L/CyO. Flies from this stock were crossed with yw; FRT42D,w+,2R11.5(lth)/CyO flies and F-1 females with mosaic eyes were analyzed (yw/yw,ey>flp; P[Rh1>Rh1-eGFP,w+], FRT42D,ttd14P75L/FRT42D,w+,2R11.5(lth)). The w+ marker of the heterozygous Rh1>Rh1-eGFP construct results in a faint orange eye color while flies homozygous for the Rh1>Rh1-eGFP construct exhibit dark orange colored eyes. As the F-1 females with mosaic eye clones display a faint orange eye color, we assume that the Rh1>Rh1-eGFP construct is present with one copy in this genetic background. We conclude that the Rh1>Rh1-eGFP construct is localized on the left arm of the second chromosome which is not affected by FRT-mediated mitotic recombination. Therefore, for the wild type control, yw females were crossed to yw;P[Rh1>Rh1-eGFP,w+] males and F-1 females (yw;P[Rh1>Rh1-eGFP,w+]/+) were subjected to water immersion microscopy. For immunoblot, immunocytochemistry and ERG experiments, yw,ey>flp; FRT42D,ttd14P75L/CyO flies were crossed with yw; FRT42D,w+,2R11.5(lth)/CyO flies and female F-1 flies with mosaic eyes (yw/yw,ey>flp; FRT42D,ttd14P75L/FRT42D,w+,2R11.5(lth)) were used for analysis. w1118 was used as wild type control. To analyze the rescue construct (Rh1>ttd14-A-myc) in ttd14P75Lmutant mosaic clones, we crossed females expressing the rescue construct on the X-chromosome (yw,P[Rh1>ttd14-A-myc,y+]) with yw; FRT42D,w+,2R11.5(lth)/CyO males. F-1 males without the CyO balancer (yw,P[Rh1>ttd14-A-myc,y+]; FRT42D,w+,2R11.5(lth)/+) were then crossed with yw,ey>flp; FRT42D,ttd14P75L/CyO females. In the F-2 generation, females carrying mosaic eyes (yw,P[Rh1>ttd14-A-myc,y+]/yw,ey>flp; FRT42D,ttd14P75L/FRT42D,w+,2R11.5(lth)) were analyzed. For the analysis of the subcellular localization of the TTD14 protein by immunocytochemistry yw,P[Rh1>ttd14-A-myc,y+] flies were crossed with yw;;P[Rh1>TRPL-eGFP,y+] flies and female F-1 flies, carrying both reporter constructs (yw,P[Rh1>ttd14-A-myc,y+]/+;;P[Rh1>TRPL-eGFP,y+]/+) were analyzed. To address the larval lethality of mutant alleles of the ttd14 gene, the mutant alleles were crossed with a GFP-labeled CyO-Balancer (yw; FRT42D,ttd14P75L/CyO, P[ActGFP,w+]JMR and yw; FRT42D,ttd14KG03769[w+]/CyO, P[ActGFP,w+]JMR1 respectively) and development of non-fluorescent homozygous mutant F-1 larvae was traced. For the generation of the Rh1>ttd14-A and Rh1>ttd14-B rescue constructs, the entire coding sequence, but not the 5´ and 3´ untranslated region of ttd14-A and ttd14-B was PCR amplified from cDNA derived from fly heads. Restriction sites were introduced in the primer pair 5´-TTCCCGAATTCGAAGACATG-3 (EcoRI) and 5´-GAGTCAACATAATCGATAGCCA-3´ (ClaI) for the amplification of ttd14-A and 5´-TTCCCGAATTCGAAGACATG-3´ (EcoRI) and 5´-GTCCCTGAATCGATTTTGCACAC-3´ (ClaI) for the amplification of ttd14-B. PCR fragments were cloned into a modified pBluescript II SK vector (Stratagene) between the Rh1 minimal promoter (base pairs -833 to +67) and the last 0.6 kb of the Rh1 3´ untranslated region using the EcoRI and ClaI restriction sites. After a KpnI site in the Rh1 promoter was eliminated with a site directed mutagenesis Kit (Agilent) using the mutagenesis primer 5´-CAGAATCCAGGAACCCTGAGTACCGGATCC-3´, the Rh1>ttd14-A(B) construct was excised from the pBluescript vector using a NotI/KpnI digest and cloned into the pattB vector [68]. The final pattB Rh1>ttd14-A and -B clones were verified by DNA sequencing (Qiagen). Transgenic flies were generated using site specific recombination by injecting the pattB Rh1>ttd14-A/B vector (200 ng/μl) into y1 M{vas-int.Dm}ZH-2A w*; M{3xP3-RFP.attP}ZH-86Fb embryos. For the generation of the Rh1>ttd14-A-myc rescue construct, a PCR amplified ttd14-A construct encoding amino acids 1 to 472 of TTD14-A and a C-terminal myc-tag obtained from the vector pSF-CMV-COOH-EKT-CMyc2 (Oxford Genetics) was cloned in the pENTR 1A vector (Invitrogen). Using the Gateway system (Invitrogen), the ttd14-A-myc sequence was recombined with a modified pYC4 vector containing a DEST cassette between the Rh1 minimal promoter (base pairs -833 to +67) and the last 0.6 kb of the 3´untranslated region of Rh1 [67]. The final pYC4 Rh1>ttd14-A-myc clone was verified by DNA sequencing (GATC Biotech). The construct was injected into yw embryos. Full length ttd14-A was PCR amplified from cDNA derived from fly heads and cloned in the pQE30 vector (Qiagen), which encodes a N-terminal His-tag, using BamHI and SalI sites. For generating a construct encoding the mutated TTD14[P75L] protein, the respective codon was mutated using the Quick Change Lightning Site-Directed Mutagenesis Kit (Agilent) as described in the instruction manual. The resulting pQE30 ttd14-A and pQE30 ttd14-A[P75L] clones were verified by DNA sequencing (Qiagen) and transformed in E. coli M15 cells (Qiagen). For the purification of native TTD14 or TTD14[P75L] protein, expression of the recombinant protein was induced by 1 mM IPTG for 2 hours at 30°C. The pellet of 1 liter bacteria culture was lysed in 20 ml lysis buffer (50 mM Tris, pH 8.0, 300 mM NaCl, 10 mM Imidazol, 1% Triton X-100, 2 mM DTT, 50 μM APMSF) using a French press (OneShot; Constant Systems) and incubated with 500 μl Ni-NTA agarose (Qiagen) for 60 min at 4°C. After three wash steps in 20 ml wash buffer each (50 mM Tris/HCl, pH 8.0, 300 mM NaCl, 50 mM Imidazol, 50 μM APMSF), the TTD14 protein was eluted from the Ni-NTA agarose with 400 μl elution buffer (50 mM Tris, pH 8.0, 300 mM NaCl, 250 mM Imidazol, 50 μM APMSF). For the purification of TTD14 protein under denaturating conditions, expression of the recombinant protein was induced by 1 mM IPTG for 2 hours at 37°C. The pellet of a 300 ml bacteria culture was homogenized in 10 ml lysis buffer (8 M Urea, 100 mM Na2HPO4,/NaH2PO4, pH 8.0, 10 mM Tris/HCl pH 8.0, 20 mM β-mercaptoethanol, 10 mM Imidazol) and incubated with 2 ml of Ni-NTA agarose (Qiagen) for 30 min at room temperature. After three wash steps in 20 ml wash buffer each (8 M Urea, 100 mM Na2HPO4,/NaH2PO4, pH 6.3, 10 mM Tris/HCl pH 6.3, 10 mM Imidazol), the TTD14 protein was eluted with 1 ml elution buffer (50 mM Tris/HCl, pH 7.5, 300 mM NaCl, 250 mM Imidazol). 1 μg of TTD14 protein purified under denaturating conditions was used to immunize two rabbits for 61 days (Pineda Germany). IgG were purified from the serum using a HiTrap Protein A HP-column (GE Healthcare) according to the manufacturer’s instructions. PIP-Strips (Echelon Research Laboratories) which contain various phospholipids at distinct spots (100 pmol) were blocked in TBS-T (10 mM Tris/HCl pH 8.0; 150 mM NaCl, 0.1% Tween-20) containing 3% bovine serum albumin (BSA). The membrane was incubated with recombinant TTD14 protein purified under native conditions (0.4 μg/ml) in TBS-T plus 3% BSA for 90 min at 4°C. After three wash steps for 5 min each in TBS-T, the membrane was incubated with the α-TTD14 antibody in TBS-T plus 3% BSA for 90 min at 4°C, washed again three times for 5 min each in TBS-T and finally incubated with horseradish peroxidase-coupled anti-rabbit IgG (1:10,000 Sigma) in TBS-T plus 3% BSA for 90 min at 4°C. Lipid-bound protein was detected by enhanced chemiluminescence (0.091 M Tris-HCl pH 8.6; 0.0227% (w/v) luminol; 0.01% (w/v) para-hydroxycoumarin acid 0.01% H2O2) using a ChemiDocXRS+ Imaging system (Bio-Rad). 20 μg of recombinant wild type TTD14 or TTD14[P75L] protein was diluted to 225 μl with binding buffer (20 mM Tris-HCl pH 7.0; 150 mM NaCl; 5 mM MgCl2; 0.1% (v/v) Triton X-100) and incubated with agitation (1000 rpm) at 4°C for 1 hour with 25 μl of either control agarose (Pierce), ATP-agarose (Sigma Aldrich) or GTP-agarose (Sigma Aldrich) pre-equilibrated in binding buffer. Beads were centrifuged at 4°C at 13,000 g for 2 min and washed three times with 1 ml ice-chilled binding buffer. Proteins were eluted by adding 50 μl of binding buffer containing either 25 mM ATP or 25 mM GTP for 5 min on ice. After centrifugation (2 min at 13,000g at 4°C) 15 μl of the resulting supernatant was subjected to immunoblot analysis. For immunoblot analyses of TRPL, TRP, Rh1, and Tubulin in wild type and ttd14P75L mutant flies (Figs 4D, 5C and 6D and S3 Fig), fly heads were homogenized in SDS extraction buffer (4% SDS, 1 mM EDTA, 75 mM Tris/HCl, pH 6.8) using 4 μl of extraction buffer per head and incubated for 1 hour at room temperature. After 10 min centrifugation at 16,000 g to remove debris the supernatant was subjected to SDS-PAGE. For immunoblot analysis of TTD14 protein in wild type and ttd14 mutant flies (Fig 3B), eye cups were dissected and homogenized in Tris-buffer (50 mM Tris/HCl pH 8.0; 150 mM NaCl; 50 μM APMSF; 1 μl per eye) and extraction was carried out for 30 min on ice. The supernatant obtained after 10 min centrifugation at 16,000 g was subjected to SDS-PAGE. For immunoblot analysis of membrane and soluble proteins (Fig 3C), fly heads were homogenized in Tris-buffer (50 mM Tris/HCl pH 8,0; 150 mM NaCl; 50 μM APMSF; 2 μl per head) and extraction was carried out for 30 min at room temperature. A 3 min centrifugation step at 2,500 g was applied to remove cuticular particles of the fly head. The supernatant was subjected to ultracentrifugation (10 min at 100,000 g, 4°C) and the resulting supernatant was loaded as soluble fraction on a SDS-gel. The membrane pellet was washed three times in Tris-buffer and solubilized in SDS extraction buffer (4% SDS, 1 mM EDTA, 75 mM Tris/HCl, pH 6.8; 1 μl per head) for 20 min at room temperature and loaded as membrane fraction on a SDS-gel. SDS-PAGE was performed according to Laemmli [69] using 10% or 12% polyacrylamide gels. Separated proteins were electrophoretically transferred to polyvinylidene difluoride membranes (Bio-Rad). The membrane was then blocked for 20 min in TBS-T with 5% skim milk (10 mM Tris/HCl, pH 7.5, 150 mM NaCl, 0.1% Tween 20, 5% skim milk). α-TRPL [34], α-TRP (Mab83F6; Developmental Studies Hybridoma Bank, University of Iowa), α-Rh1 (4C5; Developmental Studies Hybridoma Bank), α-tubulin (E7; Developmental Studies Hybridoma Bank, University of Iowa) and α-TTD14 antibodies were used for immunological detection in TBS-T with 5% skim milk over night at 4°C. Signals were detected by enhanced chemiluminescence ((0.091 M Tris-HCl pH 8.6; 0.0227% (w/v) luminol; 0.01% (w/v) para-hydroxycoumarin acid; 0.01% H2O2) using the ChemiDocXRS+ Imaging system (Bio-Rad). Quantification of immunoblot signals was performed with Image Lab 4.0 (Bio-Rad) by determining the integrated density of each protein band. The Rh1 signals were normalized by the TRP signals of the same sample. Fluorescence in the deep pseudopupil was observed in CO2-anaesthetized flies using a Leica MZ16 F stereomicroscope with 63x magnification and the GFP3 filter set. Images were captured with a Leica DFC420 C camera. Water immersion microscopy of eGFP-tagged proteins in intact eyes was performed as previously described [35]. Living flies were anaesthetized on ice, spiked on an insect needle, mounted with plasticine on an object slide and covered with ice-chilled distilled water. The eGFP fluorescence was observed with an AxioImager.Z1m microscope (Zeiss, Germany; objective: Achroplan 20X/0.5 water immersion). Images were captured with the AxioCamMrM (Zeiss) camera and the Axio-Vision 4.6/4.8 or Zen 2012 (Zeiss) software. For images of flies with white and orange colored eyes, exposure time was determined individually for every single eye to be closely below overexposure. For the analysis of the red colored wild type and ttd14KG03769 eyes, an exposure time resulting in images closely below overexposure in 1 day old dark-adapted wild type flies was determined and applied for all other experimental conditions. For quantitative analyses of the relative TRPL-eGFP fluorescence in the rhabdomeres, fluorescence images obtained with the water immersion technique were analyzed with ImageJ 1.42j software (National Institute of Health, USA). The relative amount of TRPL-eGFP present in the rhabdomeres (R) was calculated using the formula R = (Ir–Ib)/[(Ir–Ib)+(Ic–Ib)], where Ir, Ib, and Ic are the fluorescence intensities in the rhabdomeres, in the background, and in the cell body, respectively. Background intensities were determined in the center of the ommatidium where the rhabdomere of the R7/R8 cells is located. For each eye, three ommatidia were analyzed and five individual flies were analyzed per data point. The data were normalized to the value obtained for 1 day old dark raised wild type flies that was set to 100%. For determining the time course of rhabdomere degeneration flies expressing TRP-eGFP in R1-6 photoreceptors were inspected for the presence of R1-6 rhabdomeres using water immersion microscopy. Clearly visible rhabdomeres were scored 2, weakly visible rhabdomeres were scored 1 and absent rhabdomeres were scored 0. Three ommatidia per eye were scored, resulting in a degeneration score of 36 for fully intact eyes or in a score of 0 for fully degenerated eyes. The maximal degeneration score of 36 was set to 100%. Five individual flies were analyzed per data point. For immunocytochemical analyses, Drosophila eyes were fixed in 2% paraformaldehyde (PFA) in PBS (175 mM NaCl, 8 mM Na2HPO4, and 1.8 mM NaH2PO4, pH 7.2) for 1 hour at room temperature, and then washed twice in 0.1 M phosphate buffer (0.1 M Na2HPO4 and 0.1 M NaH2PO4, pH 7.2). Subsequently, three wash steps in 10% sucrose and two wash steps in 25% sucrose were performed for 15 min each. Eyes were finally infiltrated with 50% sucrose overnight at 4°C, cryofixed in melting pentane, and sectioned at 10 μm thickness in a Leica CM3050S cryostat (Leica, Germany) at −25°C. Cryosections were first fixed in 2% PFA in PBS for 10 min and then washed twice in PBS. After blocking of sections in PBS-T (1% BSA, 0.3% Triton X-100 in PBS) for 2 hours at room temperature, sections were incubated with α-TRPL [35], α-TRP (Mab83F6; Developmental Studies Hybridoma Bank of the University of Iowa) and α-myc (Sigma) in PBS-T overnight at 4°C. The sections were subsequently washed three times in PBS and were then incubated either with α-mouse-Alexa Fluor 488 and α-rabbit-Alexa Fluor 680 or with α-mouse-Alexa Fluor 680 and α-rabbit-Alexa Fluor 488 (Life Technololgies) in 0.5% fish gelatine and 0.1% ovalbumin in PBS for at least 4 hours at room temperature. Phalloidin-Alexa Fluor 546 (Life Technologies) was added to stain F-actin in rhabdomeres. The eGFP-tagged Rh1 protein was visualized by its own fluorescence. The sections were finally mounted in Mowiol 4.88 (Polyscience) and examined with an AxioImager.Z1m microscope (objective: EC Plan-Neofluar 40×/1.3 Oil, Zeiss, Germany) using the ApoTome module (Zeiss, Germany) at room temperature. Images were captured with the AxioCamMrM (Zeiss) camera and the Axio-Vision 4.6/4.8 or Zen 2012 (Zeiss) software. For electroretinogram recordings, flies were immobilized in a pipette tip and mounted with a mixture of colophonium and bee’s wax (1:3). Electroretinogram recordings were performed at room temperature after 3 minutes of dark adaptation prior to the first orange light-stimulus. Light-stimuli of 5 s duration were delivered by an orange light-emitting diode (Roithner, Austria) and a blue light-emitting diode (Roithner, Austria) in a setup of two collimating lenses (Linos, Germany) within the light path. The light intensity at the position of the fly eye was 2.15 mW/cm2 for orange light and 1.3 mW/cm2 for blue light. A DPA-2FS amplifier (NPI electronic, Germany) with a low pass filter (700 Hz) was used for signal amplification. Analog-to-digital conversion was accomplished with a BNC-2090A rack-mounted terminal block (National Instruments, Germany) and a PCI-6221 PC card (National Instruments, Germany). Data recording was achieved by the Whole Cell Analysis Program software WinWCP 4.7.6. (University of Strathclyde). The recording electrode glass capillary was filled with Davenport solution (100 mM NaCl, 2 mM KCl, 1 mM CaCl2, 1.8 mM NaHCO3, pH 7.2). Fly heads were separated from the body, dissected into two halves and incubated in fixative solution (4% paraformaldehyde, 2.5% glutaraldehyde in 1 x PBS, pH 7.4) for 1 h at room temperature. The semi-heads were washed 3 times with 0.1 M sodium cacodylate buffer (pH 7.4) for 10 min and postfixed in 2% OsO4 in 0.1 M cacodylate buffer (pH 7.4) for 1 h. After 3 washes in 0.1 M sodium cacodylate buffer for 10 min the semi-heads were dehydrated through a graded series of ethanol from 30% to 100%. The dehydrated semi-heads were incubated in 100% propylene oxide twice for 10 min each, and then transferred to 50% propylene oxide: 50% Renlam M-1 resin (Serva Electrophoresis, Heidelberg, Germany) and incubated overnight. After that the semi-heads were incubated in 100% Renlam® M-1 resin overnight, embedded in 100% Renlam® M-1 resin and polymerized at 60°C for two days. Ultrathin sections (60–70 nm) were obtained using a Reichert Ultracut E microtome (Leica). Sections were counterstained with heavy metal staining (2% uranyl acetate in 50% ethanol; aq. 2% lead citrate). Ultrathin sections were analyzed in a Tecnai 12 BioTwin transmission electron microscope (FEI, Eindhoven, The Netherlands. Images were obtained with a charge-coupled device SIS Megaview3 SCCD camera (Surface Imaging Systems, Herzogenrath, Germany). Image contrast was adjusted with Adobe Photoshop CS using different tools.

10.1371/journal.pbio.1000307

mRNA Secondary Structures Fold Sequentially But Exchange Rapidly In Vivo

RNAs adopt defined structures to perform biological activities, and conformational transitions among alternative structures are critical to virtually all RNA-mediated processes ranging from metabolite-activation of bacterial riboswitches to pre-mRNA splicing and viral replication in eukaryotes. Mechanistic analysis of an RNA folding reaction in a biological context is challenging because many steps usually intervene between assembly of a functional RNA structure and execution of a biological function. We developed a system to probe mechanisms of secondary structure folding and exchange directly in vivo using self-cleavage to monitor competition between mutually exclusive structures that promote or inhibit ribozyme assembly. In previous work, upstream structures were more effective than downstream structures in blocking ribozyme assembly during transcription in vitro, consistent with a sequential folding mechanism. However, upstream and downstream structures blocked ribozyme assembly equally well in vivo, suggesting that intracellular folding outcomes reflect thermodynamic equilibration or that annealing of contiguous sequences is favored kinetically. We have extended these studies to learn when, if ever, thermodynamic stability becomes an impediment to rapid equilibration among alternative RNA structures in vivo. We find that a narrow thermodynamic threshold determines whether kinetics or thermodynamics govern RNA folding outcomes in vivo. mRNA secondary structures fold sequentially in vivo, but exchange between adjacent secondary structures is much faster in vivo than it is in vitro. Previous work showed that simple base-paired RNA helices dissociate at similar rates in vivo and in vitro so exchange between adjacent structures must occur through a different mechanism, one that likely involves facilitation of branch migration by proteins associated with nascent transcripts.

Properly folded RNAs are critical for virtually all RNA-mediated processes ranging from feedback regulation of gene expression to RNA maturation. The ability of RNAs to adopt specific structures in living cells is remarkable given their propensity to become trapped in a mixture of stable, misfolded structures in vitro. Using mRNA with an inserted ribozyme and self-cleavage to monitor competition between mutually exclusive structures, we previously showed that upstream structures dominated folding outcomes during RNA synthesis in vitro, suggesting that folding occurs sequentially. However, when studied in vivo upstream and downstream structures blocked ribozyme assembly equally well in yeast, providing evidence that intracellular folding outcomes reflect the relative stability of alternative structures. We find that very stable upstream structures can block assembly of downstream structures in vivo even when the downstream structures are more stable, and that a narrow threshold of stability determines whether folding and unfolding rates or thermodynamic stability govern folding outcomes. Thus, mRNAs fold sequentially in vitro and in vivo but exchange between adjacent structures is faster in vivo than in vitro. Simple RNA structures unfold at similar rates in vivo and in vitro, so exchange between adjacent structures in vivo probably occurs through a distinct, step-wise mechanism that could be facilitated by proteins associated with nascent RNAs.

RNAs adopt specific secondary structures to carry out their biological functions, and exchange among alternative secondary structures plays essential roles in virtually all RNA-mediated processes ranging from RNA silencing and metabolite-activation of bacterial riboswitches to pre-mRNA splicing and viral RNA replication (Figure 1A) [1]–[6]. The ability of RNAs to assemble into precise structures and undergo transitions from one defined structure to another on a biological time scale is remarkable since RNAs tend to adopt a mix of misfolded structures with slow exchange kinetics in vitro [7]–[9]. Thus, detailed understanding of the mechanisms of RNA assembly and exchange as it occurs in vivo is critical for understanding RNA function. Two kinds of mechanisms have been proposed to explain how RNA secondary structures form and assemble into tertiary structures precisely and efficiently in vivo. First, RNA chaperones might facilitate thermodynamic equilibration by lowering the free energy barrier for RNA unfolding and refolding [10]–[14]. Many proteins, particularly basic unstructured proteins, exhibit general RNA chaperone activity in vitro [15]–[17]. The DEAD-box family of putative RNA helicases has been implicated in virtually every aspect of RNA metabolism including ribosome biogenesis, pre-mRNA splicing, RNA interference, translation, mRNA transport, and decay [18]–[20]. Although they exhibit little substrate specificity in vitro, most DEAD-box proteins function as part of a large macromolecular complex, such as a spliceosome or degradosome, that is devoted to a particular process. Although certain DEAD-box proteins have been shown to facilitate self-splicing and translation of a variety of RNAs in vivo and in vitro [21],[22], it is not yet clear whether nonspecific chaperones act generally to promote assembly of RNAs into functional structures or accelerate exchange among alternative structures in vivo. Second, the sequence and timing with which regions of a nascent RNA become available to fold during transcription also might channel it into a productive folding pathway. RNA secondary structure folding occurs on a microsecond time scale [23],[24], a rate that is much faster than elongation by RNA polymerase II, which transcribes at a rate of about 100 nucleotides per second in vivo [25]. Therefore, folding of a nascent transcript as it emerges from the polymerase could favor local secondary structures and limit long-range interactions. Evidence that elongation kinetics, transcriptional pausing, and circular permutations of RNA sequences influence folding patterns support the idea that RNA secondary structures fold sequentially in vivo [26]–[30]. Probing RNA folding mechanisms in a biological context is challenging because many components interact in complex pathways and several steps usually intervene between assembly of an RNA structure and execution of a biological function. We developed a system to investigate intracellular RNA folding that relies on hairpin ribozyme (HP) cleavage kinetics to report directly and quantitatively on partitioning between two mutually exclusive RNA secondary structures, helix 1 (H1) and alternative helix 1 (AltH1), in chimeric mRNAs (Figure 1B) [31]–[33]. The 3′ untranslated region (UTR) of a chimeric mRNA contains a self-cleaving ribozyme sequence and a complementary sequence, located either upstream or downstream of the ribozyme, that has the potential to anneal with part of the ribozyme sequence and block formation of the H1 helix needed for assembly of a functional ribozyme. Thus, part of the ribozyme sequence can participate in one of two mutually exclusive base-paired structures, similar to the kinds of RNA conformational switches that have been implicated in biological regulation of RNA silencing, pre-mRNA splicing, mRNA turnover, viral genome replication, translation initiation, transcription attenuation, and in metabolite-triggering of bacterial riboswitches, as examples [1]–[6]. The alternative secondary structures formed by these chimeric RNAs were designed to have well-defined structures and thermodynamic stabilities that facilitate quantitative analyses, but yeast do not normally have RNAs like HPs and are not likely to contain any ribozyme-specific ligands. Therefore, the behavior of these chimeric mRNAs should reflect general features of RNA folding in an intracellular environment. Differences in the folding behavior of chimeric RNAs with inserts located upstream or downstream of the ribozyme reflect the influence of 5′ to 3′ transcriptional polarity, and the behavior of RNAs with different H1 and AltH1 structures reflects the influence of folding and unfolding kinetics and thermodynamic stability on folding outcomes. Self-cleaving RNAs are expressed in yeast as chimeric mRNAs under the control of a glucose-repressible promoter that enables quantification of intracellular RNA turnover rates (Figure 1C). Chimeric mRNAs that assemble into functional ribozyme structures decay through self-cleavage and through endogenous mRNA degradation pathways while chimeric mRNAs with mutationally inactivated ribozymes decay only through endogenous degradation pathways. Therefore, the difference between intracellular decay rates for mutant and self-cleaving mRNAs reflects partitioning between the H1 helix of a functional ribozyme and nonfunctional AltH1 structures. We previously examined chimeric RNAs with the potential to form an H1 helix with eight base pairs in competition with AltH1 helices with 10 base pairs that have greater thermodynamic stability (Figure 1B) [32]. Complementary inserts located upstream of the ribozyme inhibited ribozyme assembly more than downstream inserts during transcription in vitro, consistent with a sequential folding mechanism in which a stable structure that forms first dominates the folding outcome. These H1 and AltH1 structures with eight or 10 base pairs have sufficiently high thermodynamic stability that they are not expected to dissociate for months, or even years, under standard conditions in vitro [31],[34]–[37]. Therefore, it was not surprising that a stable upstream 5′ AltH1 could prevent H1 folding from a downstream sequence that was not transcribed until after the 5′ AltH1 had formed. When the same variants were expressed as chimeric mRNAs in yeast, however, upstream and downstream inserts blocked ribozyme assembly equally well. The ability of a downstream 3′ AltH1 structure to interfere with assembly of an upstream ribozyme that can fold first suggested that structures that are kinetically stable in vitro undergo rapid equilibration in vivo and allow intracellular folding to reach thermodynamic equilibrium or that AltH1 folding from contiguous sequences had a kinetic advantage over H1 folding from separate ends of the ribozyme. We have extended these studies to learn when, if ever, thermodynamic stability becomes an impediment to exchange between alternative RNA secondary structures in vivo. We found that stable upstream structures can block folding of downstream structures in vivo even when downstream structures have greater thermodynamic stability, consistent with a sequential folding mechanism. However, the thermodynamic stability needed to inhibit exchange was much greater in vivo than in vitro. In contrast, the simple helix dissociation reactions required for cleavage product release occur at virtually the same rates in vivo and in vitro [36],[38]. Differences between slow rates of simple helix dissociation and rapid exchange between adjacent secondary structures with moderate stability might be explained by the ability of proteins associated with nascent transcripts to facilitate branch migration. In order to determine whether thermodynamic stability ever becomes an impediment to secondary structure exchange in vivo, we systematically increased the thermodynamic stabilities of competing H1 and AltH1 structures relative to the structures that exchanged freely in our previous study. We began by adding two base pairs to H1, increasing its length from eight to 10 base pairs to create HP210 (Figure 2A). Addition of two base pairs enhances H1 stability by about 2 kcal/mol [35],[37], a change that is expected to slow dissociation of this H1 helix by more than 20-fold relative to the H1 helix with eight base pairs that we examined previously (Table 1). Complementary inserts located upstream of the ribozyme can anneal with the 5′ strand of H1 to form 5′ AltH1 stem loops with 10 base pairs, in HP210-510, or 12 base pairs, in HP210-512. Likewise, complementary inserts located downstream of the ribozyme can anneal with the 3′ strand of H1 to form 3′ AltH1 stem loops with 10 base pairs, in HP210-310, or 12 base pairs, in HP210-312. The H1 helices with 10 base pairs still have lower thermodynamic stability than AltH1 structures with 10 or 12 base pairs by 3 or 6 kcal/mol, respectively [35],[37]. Ribozyme variants with upstream or downstream inserts displayed very different self-cleavage activity during co-transcriptional assembly in vitro (Figure 2B). Upstream inserts with the potential to form 5′ AltH1 structures with 10 or 12 base pairs inhibited ribozyme assembly and self-cleavage almost completely. In contrast, a downstream insert with the potential to form a 3′ AltH1 structure with 10 base pairs reduced self-cleavage rates only 2-fold. Chimeric RNAs with the potential to form a downstream 3′ AltH1 structure with 12 base pairs partitioned almost equally between fully functional H1 and inactive AltH1 structures. Thus, transcription polarity influenced folding outcomes during co-transcriptional folding in vitro, consistent with the sequential mechanism of secondary structure assembly that we inferred from previous results [32]. In yeast, the two-base-pair extension of H1 did not rescue ribozyme assembly in chimeric mRNAs with an upstream insert capable of forming a 5′ AltH1 structure with 10 base pairs. That is, HP210-510 still displayed no acceleration of intracellular decay kinetics relative to its mutationally inactivated counterpart, and no cleavage product RNAs were detected in RNase protection assays (Figure 3B, Table 1). However, competition between H1 with 10 base pairs and a downstream 3′ AltH1 structure with 10 base pairs in HP210-310 mRNA had a very different outcome in vivo than we observed previously when H1 contained only eight base pairs [32]. HP210-310 mRNA decayed faster than its mutationally inactivated counterpart, indicating that intracellular cleavage combined with normal mRNA degradation to accelerate intracellular decay kinetics, and RNase protection assays revealed intracellular cleavage products (Figure 3C, Table 1). The nonfunctional AltH1 structures in HP210-510 and HP210-310 RNAs are expected to dominate the folding outcome by about 150-fold relative to functional H1 structures in a rapid equilibration mechanism in which secondary structure folding reaches thermodynamic equilibrium. However, the intracellular cleavage rate of 0.049 min−1 calculated for HP210-310 mRNA was only 2-fold lower than the rate of 0.082 min−1 measured for the HP210 RNA that lacked any complementary insert. The ability of an upstream H1 helix with 10 base pairs to dominate the folding outcome, even when the alternative downstream structure has greater thermodynamic stability, suggests that secondary structures formed sequentially both in vitro and in vivo. It was important to confirm that this change in intracellular secondary structure partitioning resulted from the increased kinetic stability of H1 and not from inaccuracy in the free energy calculations that indicated that the H1 helix was less stable than the 3′ AltH1 structure in HP213-310 RNA. HP210-512 and HP210-312 RNAs have the same H1 sequence as HP210-510 and HP210-310 RNAs, but they have two additional base pairs in the AltH1 stem loops that are expected to lower the AltH1 free energy by 3 kcal/mol (Figure 2A). In these variants, the H1 helices were calculated to be less stable than the AltH1 helices by 6 kcal/mol. With a thermodynamic advantage of 6 kcal/mol, nonfunctional AltH1 structures with 12 base pairs would dominate the folding outcome by more than 104-fold relative to functional H1 structures with 10 base pairs if folding reaches thermodynamic equilibrium. An upstream insert able to form a 5′ AltH1 structure with 12 base pairs inhibited assembly of a functional ribozyme much more than a downstream insert during co-transcriptional assembly of HP210-512 RNA in vitro, as we previously observed for HP210-510 RNA with 10-base-pair 5′ AltH1 structures (Figure 2B) [32]. A chimeric mRNA with 12 base pairs in an upstream 5′ AltH1 stem loop also exhibited no detectable self-cleavage activity in yeast, evidence that folding of a stable, upstream 5′ AltH1 dominated the folding outcome as expected (Figure 3D, Table 1). However, chimeric mRNA with a downstream insert capable of forming a 3′ AltH1 with 12 base pairs displayed an intracellular cleavage rate that was reduced only 4-fold relative to chimeric HP210 mRNA with no insert (Figure 3E, Table 1). The resistance of H1 sequences in each of these HP210 variants to chemical modification by dimethyl sulfate (DMS) in vivo suggests that chimeric mRNAs with the potential to form 10 base pairs in H1 adopt functional ribozyme structures in vivo, consistent with the activity observed in functional assays (Figures S1 and S2). The ability of HP210-310 and HP210-312 mRNAs with 10-base-pair H1 helices to resist competition from a downstream 3′ AltH1 structure that has 10 or 12 base pairs supports the conclusion that the two additional base pairs added to H1 rescued intracellular self-cleavage activity by slowing exchange between functional and nonfunctional structures. In the third series of variants, both H1 and AltH1 helices are much more stable than the secondary structures in the chimeric RNAs examined previously (Table 1). HP214-512 and HP214-312 RNAs have the potential to form 14 base pairs in H1 and 12 base pairs in AltH1 (Figure 4A). Free energy calculations indicate that H1 is more stable than 5′ AltH1 and 3′ AltH1 by 3.2 kcal/mol so the functional ribozyme structure is expected to dominate the folding outcome by more than 200-fold if secondary structure assembly reaches thermodynamic equilibrium. HP214-512, in which 5′ AltH1 forms from upstream sequences, displayed very little cleavage activity during co-transcriptional folding in vitro while a large fraction of HP214-312, the variant with the potential to form a downstream 3′ AltH1, assembled into a functional ribozyme (Figure 4B). This pattern is consistent with a sequential mechanism of secondary structure folding, as observed for other chimeric RNAs with stable AltH1 structures during co-transcriptional folding in vitro (Table 1). Chimeric HP214-512 mRNA, with an upstream insert capable of forming a 5′ AltH1 structure with 12 base pairs, exhibited no detectable intracellular cleavage activity (Figure 4C). HP214-512 mRNA decayed at the same rate as its mutationally inactivated counterpart, and no products of intracellular cleavage were detected in RNase protection assays. Thus, HP214-512 mRNA appeared to fold exclusively into a nonfunctional 5′ AltH1 structure despite the potential to form a downstream H1 helix with greater thermodynamic stability, an interpretation supported by the susceptibility of a 5′ AltH1 loop nucleotide to DMS modification (Figure S3). In contrast, a downstream insert that had the potential to form a 3′ AltH1 structure with 12 base pairs had virtually no inhibitory effect on the ability of chimeric mRNA to form a functional ribozyme structure in vivo. Chimeric HP214-312 mRNA exhibited virtually the same intracellular decay kinetics as HP214 mRNA that lacks a complementary insert (Figure 4C). Furthermore, the 5′ strand of H1 in HP214-312 mRNA is relatively resistant to chemical modification, consistent with the conclusion that HP214-312 mRNA adopts a functional ribozyme structure in vivo (Figure S3). These results are consistent with a sequential folding mechanism in which the H1 helix folds first and does not exchange with a downstream 3′ AltH1 structure that has lower thermodynamic stability. Secondary structure folding from contiguous strands to form AltH1 stem loops is expected to be faster than H1 folding from noncontiguous regions of the RNA because folding rates decrease with increasing loop size [39]. To probe how topology and folding kinetics affect partitioning between alternative structures, we examined circularly permuted ribozymes in which H1 stem loops fold from adjacent strands and AltH1 folding requires interaction between noncontiguous regions of the RNA (Figure 5A). In HPC28-510 and HPC28-310 RNAs, H1 helices, with eight base pairs, and AltH1 helices, with 10 base pairs, have similar sequences and calculated thermodynamic stabilities as the variants examined previously in which folding of both 5′ and 3′ AltH1 structures completely inhibited intracellular ribozyme assembly (Table 1) [32]. Circularly permuted variants were equally functional during co-transcriptional assembly in vitro (Figure 5B). Chimeric HPC28-510 and HPC28-310 mRNAs containing circularly permuted ribozyme sequences with inserts located upstream or downstream of the ribozyme were less abundant in yeast relative to their mutationally inactivated control mRNAs and displayed the accelerated decay kinetics indicative of efficient intracellular self-cleavage (Figure 5C). H1 sequences in HPC28-310 mRNAs also resisted DMS modification in vivo, consistent with the assembly of functional ribozyme structures (Figure S4). The high self-cleavage activity of circular permutants suggests that H1 helices that fold from contiguous sequences, and are expected to fold rapidly, are better able to resist competition from noncontiguous AltH1 stem loops, even when AltH1 stem loops have greater thermodynamic stability. An intracellular environment contains high concentrations of macromolecules, described as “molecular crowding” [41], that might influence the stability of RNA structures [42]–[45]. We investigated the effect of molecular crowding on RNA secondary structure exchange by combining the products of co-transcriptional folding reactions with PEG or Ficoll, two crowding agents that are believed to mimic molecular crowding in vitro (Figure 6). If crowding agents lower the activation barrier to exchange between otherwise stable RNA secondary structures, HP214-512 RNAs that are kinetically trapped in a 5′ AltH1 structure that has lower thermodynamic stability than the downstream H1 helix would be expected to exchange rapidly into the thermodynamically favored ribozyme structure and self-cleave. However, we observed no change in cleavage extents for any of the HP214 variants after dilution of co-transcriptional folding reactions into high concentrations of PEG or Ficoll. We have examined the folding behavior of chimeric mRNAs with the potential to adopt defined alternative secondary structures during co-transcriptional folding reactions in vitro and in living cells. We previously found that folding patterns were consistent with a sequential mechanism in which stable upstream structures dominate the folding outcome during co-transcriptional folding in vitro but the most thermodynamically stable structures dominated folding outcomes during assembly of the same chimeric mRNAs in yeast [32]. The current experiments were designed to probe contributions of folding kinetics and dissociation kinetics to intracellular RNA assembly and to determine whether RNA secondary structures can ever be sufficiently stable to resist thermodynamic equilibration in vivo. The folding behavior of the RNAs with extremely stable secondary structures examined here revealed that there is a threshold where intracellular secondary structure folding does occur sequentially in vivo and does not reach thermodynamic equilibrium. Furthermore, the threshold for exchange is higher and the rate of exchange between alternative secondary structures is much faster in vivo than it is in vitro. The H1 helix of HP210-510 and HP210-310 has greater thermodynamic stability than the H1 helix with eight base pairs in the chimeric mRNAs that we examined previously by about 2 kcal/mol and is expected to dissociate more than 20-fold more slowly [32]. Functional ribozymes with 10 base pairs in H1 were able to resist competition from the downstream 3′ AltH1 in HP210-310, which has a thermodynamic advantage of −3 kcal/mol. Likewise, HP214-512 mRNAs exhibited no detectable intracellular cleavage activity even though an H1 helix with 14 base pairs has a thermodynamic advantage of −3.2 kcal/mol relative to an upstream 5′ AltH1 helix with 12 base pairs. These results suggest that these upstream 5′ AltH1 and downstream H1 structures formed sequentially during transcription and remain folded despite the potential to form alternative structures with greater thermodynamic stability by interacting with downstream sequences. HP210-310 and HP210-312 mRNAs chimeric mRNAs share the same H1 structure with 10 base pairs but the 10- and 12-base-pair 3′ AltH1 structures differ in thermodynamic stability by 5 kcal/mol. If partitioning between H1 and 3′ AltH1 structures reflected their relative thermodynamic stabilities, the functional form of HP210-312 mRNA would have been more than 50-fold more abundant than the functional form of HP210-310 mRNA. The observation that both chimeric mRNAs exhibit similar intracellular self-cleavage kinetics suggests that folding outcomes are determined by slow H1 dissociation kinetics and not by thermodynamic equilibration. Stem loop folding rates decrease linearly with increasing loop size in vitro [39], so folding of nonfunctional AltH1 structures from complementary strands separated by four nucleotides could have a kinetic advantage relative to H1 helices that fold from noncontiguous strands at opposite ends of the ribozyme sequence that are separated by 63 nucleotides. In the first set of ribozyme variants we examined, AltH1 stem loops folded from contiguous sequences while the H1 stem loops folded from sequences at opposite ends of the ribozyme [32]. If AltH1 structures fold first and dissociate slowly, the ability of the downstream 3′ AltH1 stem loops with moderate thermodynamic stability to inhibit ribozyme assembly could have reflected the importance of folding kinetics in folding outcomes. Indeed, chimeric mRNAs with circularly permuted ribozymes in which eight-base-pair H1 stem loops folded from contiguous strands were able to resist competition from noncontiguous AltH1 stem loops, even when upstream and downstream AltH1 stem loops had greater thermodynamic stability. Thus, contiguity might influence folding outcomes by conferring a kinetic advantage on local secondary structures. Our previous studies revealed that kinetic and equilibrium parameters for intermolecular and intramolecular ribozyme reactions in yeast agree remarkably well with the same parameters measured in vitro provided that in vitro reactions approximate an intracellular ionic environment [34],[36],[38],[46]–[48]. The 5′ and 3′ products of HP self-cleavage associate through intermolecular base pairs in H1 so product dissociation kinetics reflect H1 dissociation. Cleavage products that associate through an intermolecular H1 helix with six base pairs exhibited no detectable product dissociation in vivo, and the dissociation rate constant of about 3 min−1 measured for a complex with four base pairs in H1 in vivo agreed remarkably well with the rate constant expected for dissociation of the same cleavage products in vitro [38]. Thus, intracellular product dissociation kinetics provided no evidence that any component of the intracellular environment significantly altered H1 stability in ribozymes without the potential to form AltH1 structures. The slow dissociation rate of an H1 helix with four base pairs is difficult to reconcile with a rapid conformational exchange model in which nonspecific RNA chaperones act generally to destabilize all RNA helices in vivo. Free energy calculations suggest that a helix with eight base pairs should dissociate about 160-fold more slowly than a helix with four base pairs but an eight-base-pair H1 helix seemed to exchange rapidly with an adjacent AltH1 helix in vivo [32]. This contrast between slow kinetics of simple helix dissociation and rapid exchange between adjacent secondary structures suggests that the intracellular mechanisms of exchange between adjacent secondary structures and simple helix dissociation are qualitatively different. Folding of a 10 base pair AltH1 during stepwise dissociation of an eight base pair H1 might facilitate exchange between neighboring structures without incurring the large energy cost required for complete dissociation of a long, stable helix. Exchange between adjacent structures might occur much faster than simple helix dissociation if RNA secondary structures exchange through a branch migration mechanism (Figure 7). A branch migration mechanism previously was proposed to explain the lower-than-expected activation barrier observed for exchange between alternate secondary structures of a spliced leader RNA in vitro [49]. However, stable upstream 5′ AltH1 structures consistently and effectively inhibited assembly of downstream ribozymes during co-transcriptional assembly in vitro, so rapid exchange between adjacent helices was not a spontaneous process under our in vitro transcription conditions. An intracellular environment contains high concentrations of macromolecules, described as “molecular crowding” [41],[44], which is thought to influence the stability of nucleic acid structures through effects on the activity of water that modulate hydration states [42],[43],[45],[50]–[52]. However, we found no evidence of secondary structure rearrangements even when PEG or Ficoll were added to mimic crowding effects in vitro. Rapid exchange between adjacent secondary structures during co-transcriptional assembly in vivo might be explained by the ability of proteins associated with nascent transcripts to facilitate branch migration. Further work will be needed to identify which protein(s) might modulate exchange kinetics in vivo and learn whether transcripts produced by different RNA polymerases, by different forms of RNA polymerase II, or in different physiological states, exhibit different exchange kinetics. Many RNA processing and assembly events occur co-transcriptionally in vivo [53]–[58]. The H1 and AltH1 structures examined here, with free energies ranging from −15 to −25 kcal/mol, are similar in thermodynamic stability to secondary structure elements found in internal ribosome entry sites, iron response elements, selenocysteine insertion sites, histone stem loop structures, and structures implicated in alternative mRNA splicing that are found in eukaryotic mRNAs, and the secondary structures that regulate transcription and translation in the 5′ UTRs of bacterial mRNAs (Figure 1A) [3],[5],[59]–[65]. Therefore, these quantitative relationships between the thermodynamic stability of RNA secondary structures and intracellular secondary structure folding and exchange kinetics during transcription are likely to have important implications for understanding the mechanisms of RNA and RNP assembly and RNA-mediated processes in biological systems. Some of the most detailed studies of RNA secondary structure folding and exchange mechanisms have been carried out with an adenine-responsive riboswitch found in the 5′ UTR of an mRNA in Bacillus subtilis that encodes a purine efflux pump [3],[5],[66]–[68]. This riboswitch consists of an upstream aptamer domain that binds adenine and a downstream expression platform domain that has the potential to terminate transcription before RNA polymerase reaches the coding region. Adenine binding to the aptamer domain affects partitioning between aptamer and terminator structures to provide feedback regulation of an adenine biosynthetic gene in response to intracellular adenine concentrations. Part of the riboswitch sequence has the potential to participate in mutually exclusive aptamer or transcription termination structures (Figure 1A), similar to the chimeric mRNAs examined here, depending on whether adenine is bound. In a rapid exchange model of the switch mechanism, bound and unbound conformations are in rapid exchange and ligand binding drives folding of the ligand-bound structure by increasing its thermodynamic stability. With the free energies of the aptamer and terminator structures on the order of −12 and −32 kcal/mol, respectively [67], exchange between alternative structures could be too slow for adenine binding to drive conversion from the terminator to the aptamer structure on a biologically relevant time scale. Indeed, results of careful bulk and single molecule analyses of folding, ligand binding and transcription elongation kinetics of an adenine-responsive riboswitch in vitro argue against a thermodynamic equilibration model of riboswitch activation and point to a kinetically controlled process in which ligand binding to the nascent transcript stabilizes the bound aptamer conformation before transcription and assembly of the downstream transcription termination sequence is complete. Kinetic control of partitioning between alternative secondary structures has also been proposed in regulation of translation initiation and viral replication, for example [28],[69],[70]. It is not clear yet how predictions based on riboswitch folding behavior in vitro relate to metabolite-regulation of gene expression in vivo. Chimeric mRNAs that have competing secondary structures with nine or fewer base pairs and free energies above −15 kcal/mol appeared to exchange rapidly in vivo (Table 1) [32]. However, chimeric mRNAs became kinetically trapped in upstream secondary structures with free energies ranging from −17 to −25 kcal/mol even when downstream sequences had the potential to form alternative structures that were more stable by 3 to 6 kcal/mol (Table 1). These results delineate a very narrow threshold of thermodynamic stability that determines whether thermodynamics or folding and unfolding rates govern the folding outcome for a particular mRNA. This narrow range of free energy over which folding outcomes reflect thermodynamic equilibria or the kinetics of folding and unfolding suggests an elegant mechanism for regulating a switch through ligand binding. In the case of the adenine riboswitch, for example, adenine binding was found to stabilize an adenine aptamer structure by about 4 kcal/mol [68]. In our yeast mRNA system, a decrease in free energy from −12 to −16 kcal/mol would be sufficient to shift from a rapid to a slow exchange folding mechanism. It is important to note that our studies address secondary structure exchange in eukaryotic mRNAs transcribed by RNA Pol II in yeast, but similar RNA switches also are likely to participate in eukaryotic gene regulation. Further studies will be needed to establish whether the free energy threshold that distinguishes between rapid and slow exchange regimes varies among different biological systems. Nonetheless, it is intriguing to speculate that RNA binding proteins or small molecules like adenine with equilibrium dissociation constants in the millimolar range could provide more than enough stabilizing energy to drive an RNA secondary structure across this threshold and kinetically trap a specific ligand-bound secondary structure in vivo. Plasmid templates for in vitro transcription were derived from pTLR28, a pUC18 variant in which ribozyme-coding sequences are fused to a T7 RNA polymerase promoter [46]. Sequence changes were introduced using QuikChange™ mutagenesis (Stratagene) and the primers shown in Table S1. To construct plasmids for T7 RNA polymerase transcription of circularly permuted ribozymes in vitro, a DNA fragment that encodes a circularly permuted ribozyme fused to a T7 RNA polymerase promoter was obtained by overlapping polymerase chain reaction (PCR) using the four oligonucleotides listed in Table S1, and the fragment was digested with Kpn I and EcoR I and inserted into the pUC18 polylinker. Further changes to helix and loop sequences were performed using QuikChange™ mutagenesis and the primers shown in Table S1. For expression of chimeric mRNAs in yeast, sequences encoding ribozyme variants were inserted into the 3′ UTR of the yeast PGK1 gene in pGAL28, a pRS316 derivative in which the PGK1 gene is fused to the GAL1 upstream activation sequence [36],[38],[71]. The unique Cla I site in the 3′ UTR of PGK1 was replaced with Mlu I and Afl II sites, the same sites were introduced at the opposite end of the ribozyme sequences in pUC18 derivatives using primers shown in Table S1, and the Mlu I Afl II fragments were ligated to produce pGAL28 derivatives. Ribozyme names reflect the nature of the interdomain junction (two-way), the number of base pairs in H1, the location of a complementary insertion (5′ or 3′), and the number of base pairs in AltH1. For example, HP210-510 is a HP with a two-way helical junction that has 10 base pairs in H1 and a complementary insert located on the 5′ side of the ribozyme with the potential to form a 5′ AltH1 stem loop that has 10 base pairs. “m” indicates the presence of an inactivating G+1A mutation. Plasmids were propagated in Escherichia coli strain DH5α [72] or XL-Blue (Stratagene) or in S. cerevisiae strain HFY114 (MATa ade2-1 his3-11,15 leu2-3 112 trp1-1 ura3-1 can1-100) [73]. Co-transcriptional self-cleavage kinetics were measured in vitro as described [32],[74]. Briefly, linearized plasmid template DNA, at a concentration of 29 nM, was pre-incubated at 30°C for 10 min in 30 mM HEPES (pH 7.8 at 30°C), 1 mM DTT, 0.25 mM EDTA, 96 mM sodium glutamate and either 16 mM magnesium acetate, along with 4 mM of each NTP, 1µL of RNAsin (40U/µL, Promega), and 10–40 µCi [α-32P] ATP (3,000 Ci/mmole, NEN) in a volume of 57 µL, as described [75], except for the change in magnesium concentration. To start the reaction, 3 µL of T7 RNA polymerase, freshly prepared at a concentration of 0.2 mg/mL in transcription buffer with 1% Tween 20 (Sigma), was added for a final reaction volume of 60 µL with 0.01 mg/mL T7 RNA polymerase and 0.05% Tween 20. Aliquots were removed at intervals over 2 h, quenched by the addition of gel loading buffer (90% formamide, 25 mM EDTA, 0.002% xylene cyanole, and 0.002% bromophenol blue) and fractionated by denaturing gel electrophoresis. In every case, care was taken to ensure that transcription rates remained linear throughout each time course so that self-cleavage rates could be computed accurately from the fit to single exponential or double exponential rate equations [74]. The plots shown in the figures represent the results of a single representative experiment. Reported values represent the mean and standard deviation obtained from two or more experiments. In experiments designed to measure the effects of crowding agents, transcription reactions were diluted 2-fold into transcription buffer containing either 40% PEG 8000 or 40% Ficoll 400 and incubated at 30°C for up to 1 h before aliquots were combined with gel loading buffer. For intracellular self-cleavage assays, RNA was extracted from log phase yeast cultures grown at 30°C in minimal medium after the addition of glucose to inhibit transcription and quantified using RNase protection assays, as described [32],[33]. The 32P-labeled RNAs used as hybridization probes were transcribed from linearized pGEM-4Z derivatives (Promega), as described [38]. When 32P-labeled self-cleaving RNA was combined with yeast pellets and subjected to extraction and analysis procedures, in control experiments that were carried out in parallel with every assay, less than 10% of uncleaved ribozyme RNAs underwent cleavage, confirming that conditions used for RNase protection assays do not support ribozyme activity. Intracellular chimeric mRNA decay rates were calculated by fitting to a single exponential rate equation. Intracellular self-cleavage rates were calculated from the difference between decay rates of uncut self-cleaving mRNAs and chimeric mRNAs with an inactivating G+1A mutation as described [33]. Uncut HP mRNA abundance was normalized by comparison with ACT1 mRNA. The plots shown in the figures represent the results of a single representative decay time course experiment. Reported values represent the mean and standard deviation obtained from two or more experiments. All mutationally inactivated chimeric mRNAs displayed the same degradation rate of 0.043±0.003 m−1. At steady state, self-cleaving mRNAs are present at lower levels than mutationally activated chimeric mRNAs because both kinds of chimeric mRNAs are synthesized at the same rate, but self-cleaving RNAs decay both through self-cleavage and through normal mRNA degradation pathways while mutationally inactivated RNAs only decay through intrinsic degradation pathways. Therefore, intracellular cleavage rates also can be calculated from the relative abundance of self-cleaving and mutationally inactivated chimeric mRNAs at steady state when the intrinsic degradation rate is known [33]. Intracellular cleavage rates determined using both methods typically agree within 30% and never vary more than 2-fold. Chemical structure mapping was used to confirm that nonfunctional structures contained AltH1 and not H1 structures, as expected (Figures S1, S2, S3, S4). DMS modification of intracellular yeast RNA was performed essentially as described [76] using chimeric mRNAs that contained the inactivating G+1A mutation [33]. Yeast at mid-log phase were pelleted and resuspended in 1/50 vol minimal medium, combined with 4 µL DMS and allowed to react at 30°C for 2 min with frequent mixing. The modification reaction was quenched with 25 µL ß-mercaptoethanol, then yeast were washed by vigorous mixing with 0.25 mL of ice-cold 0.7 M ßME, pelleted, and then washed with 1 mL ice-cold water. DMS modified adenosine and cytosine residues were identified as blocks to reverse transcription [77]. For primer extension reactions, 20 µg of yeast RNA and 0.2 pmole of [5′-32P] PX4 primer were annealed in 3 µl 50 mM Tris Cl (pH 8.3 at 42°C), 0.1 mM EDTA by heating to 95°C and cooling to 50°C over 45 min, then adjusted to 50 mM Tris Cl pH 8.3, 50 mM KCl, 10 mM MgCl2, 10 mM DTT, 0.5 mM spermidine, 0.5 mM each dNTP, and 0.17 units/µl AMV reverse transcriptase in 6 µl, and incubated at 50°C for 45 min. Parallel sequencing reactions also contained 0.5 mM ddATP, ddCTP, ddGTP, or ddTTP. Reaction products were fractionated by gel electrophoresis and quantified through radioanalytic imaging. Profiles represent the relative amounts of primer extension products after normalization to the intensity of the band corresponding to unmodified uridine at position U+5 of the ribozyme unless otherwise indicated.

10.1371/journal.pntd.0000901

Targeting the Midgut Secreted PpChit1 Reduces Leishmania major Development in Its Natural Vector, the Sand Fly Phlebotomus papatasi

During its developmental cycle within the sand fly vector, Leishmania must survive an early proteolytic attack, escape the peritrophic matrix, and then adhere to the midgut epithelia in order to prevent excretion with remnants of the blood meal. These three steps are critical for the establishment of an infection within the vector and are linked to interactions controlling species-specific vector competence. PpChit1 is a midgut-specific chitinase from Phlebotomus papatasi presumably involved in maturation and degradation of the peritrophic matrix. Sand fly midgut chitinases, such as PpChit1, whether acting independently or in a synergistic manner with Leishmania-secreted chitinase, possibly play a role in the Leishmania escape from the endoperitrophic space. Thus, we predicted that silencing of sand fly chitinase will lead to reduction or elimination of Leishmania within the gut of the sand fly vector. We used injection of dsRNA to induce knock down of PpChit1 transcripts (dsPpChit1) and assessed the effect on protein levels post blood meal (PBM) and on Leishmania major development within P. papatasi. Injection of dsPpChit1 led to a significant reduction of PpChit1 transcripts from 24 hours to 96 hours PBM. More importantly, dsPpChit1 led to a significant reduction in protein levels and in the number of Le. major present in the midgut of infected P. papatasi following a infective blood meal. Our data supports targeting PpChit1 as a potential transmission blocking vaccine candidate against leishmaniasis.

For a successful development within the midgut of the sand fly vector, Leishmania must overcome several barriers which are imposed by the vector. The ability to overcome these barriers has been associated with species specificity, and interference with the sand fly vector-parasite balance can change the outcome of the infection in the vector. Recently, our group has carried out a transcriptome assessment of the sand fly Phlebotomus papatasi midgut, uncovering many transcripts possibly associated with the barrier to Leishmania development. In order to validate the role of such genes, we have developed a dedicated RNA interference (RNAi) platform to assess whether RNAi targeting such genes can reduce Leishmania major development. PpChit1 is a midgut-specific chitinase presumably involved in the maturation/degradation of the peritrophic matrix in the gut of the sand fly after a blood meal. Our results show that knockdown of PpChit1 via RNAi led to a significant reduction of Le. major within the gut, supporting the potential use of PpChit1 as a target for transmission blocking strategies against sand fly-transmitted leishmaniasis.

Emerging and reemerging vector-borne diseases pose significant threats to human and animal health [1]. The emergence of insecticide resistance as well as the lack of other efficient insecticidal tools to control disease vectors imply that new methodologies need to be developed in order to reduce vector-borne disease transmission [1]. For this, the study of vector-pathogen interaction pinpointing factors underlying vector competence can reveal new molecular targets to be disrupted, preventing pathogen transmission [2], [3]. In sand flies, midgut molecules are known or believed to be involved in defining a species ability to transmit Leishmania in nature. For a successful development within the midgut of the sand fly vector, Leishmania must overcome several barriers that include an early proteolytic attack [4], [5], [6], [7], [8], the need to escape the peritrophic matrix (PM) [8], [9], [10], [11], [12], and attachment to the midgut epithelia to prevent excretion with the remnants of the blood meal [13], [14], [15], [16]. Attachment to midgut epithelia has long been associated with the type of lipophosphoglycan (LPG) present on the surface of Leishmania, and is associated with defining sand fly-Leishmania pairs in nature [15], [16], [17]. For Leishmania major V1 strain, with LPG displaying highly decorated side chains with prominent galactose residues, we demonstrated that PpGalec, a P. papatasi galactose-binding protein, is the docking site for Le. major on the midgut epithelium of Phlebotomus papatasi [14]. Recently, LPG-independent midgut binding has been associated with the degree of glycosylation detected on proteins expressed by midgut epithelial cells [18]. For events leading up to the midgut binding, such as early parasite survival during the proteolytic attack and escape of the endoperitrophic space, some investigators suggested that midgut proteases, such as trypsins and chymotrypsins, also are responsible for defining vector-Leishmania specificity [4], [5], [6], [7]. Such proteases were shown to be specially harmful to transitional stages amastigotes [8]. A role of the PM on sand fly vector competence was suggested through comparisons of Leishmania development in different sand fly species displaying different PM degradation rates [9], [10], [11]. Studies later revealed a dual role for the sand fly PM in parasite development; protecting Leishmania from digestive enzymes in the beginning of blood digestion, yet becoming a barrier to parasite escape when mature [8]. Recent data also indicate that an anterior PM plug located at the junction between the anterior and posterior midgut acts as a barrier to Leishmania migration towards the stomodeal valve [12]. Regarding Leishmania escape from the PM, it was firstly proposed to be solely accomplished by a parasite chitinase [19]. Further work demonstrated that a Le. mexicana chitinase-overexpressing strain had an accelerated escape from the PM in Lutzomyia longipalpis [20]. However, since the characterization of a blood induced chitinolytic system in the sand fly midgut [21], it became apparent that the parasite must take advantage of the sand fly peak chitinolytic activity within midgut, approximately 40–48 hours after a blood meal, for their escape [12], [21]. PpChit1 is presumably involved in PM maturation/degradation in P. papatasi [8]. Based on the fact that Leishmania must escape the PM, and that this escape may be aided by the vector's own chitinase, we predicted that PpChit1 knock down (via RNAi) would interfere with Le. major development. Our data indicates that dsRNA-mediated silencing of PpChit1 transcripts leads to a reduction in the parasite load within the midgut of P. papatasi, pointing to the role of this molecule in P. papatasi vector competence and its potential for the development of a transmission-blocking vaccine. The use of animals during this study was reviewed and approved by the Kansas State University Institutional Animal Care and Use Committee (KSU-IACUC). P. papatasi (Israeli strain -PPIS) was reared in the Department of Entomology, Kansas State University, according to conditions described [21]. For all experiments, three-to-five day old female sand flies were used. Blood feeding was performed through a chicken skin membrane attached to a feeding device. Prior to sand fly feeding, fresh mouse blood was heat inactivated for 30min at 56°C and supplemented with 50 µl/ml of Pen/Strep solution (MP Biomedicals, Solon, OH, USA) as well as 1 mM ATP (MP Biomedicals). Sixteen to twenty four hours after blood feeding, fully engorged females were separated from partially engorged and non-blood fed by anesthetizing flies with CO2 and examining the midgut distension under a stereoscope microscope. Only fully fed individuals were maintained for further analyses. Fully engorged sand fly midguts were individually dissected on RNAse free (cleaned with ELIMINase, Fisher Scientific, Pittsburgh, PA, USA) glass slides, transferred to 50 µl of 1× PBS buffer (RNase free, pH 7.4; Fisher Scientific), and thoroughly homogenized using a hand held tissue homogenizer and RNAse-free pestle. Half the homogenate volume (25 µl) was transferred to 350 µl of RLT buffer (supplemented with 1% β-mercaptoethanol) provided by the RNA extraction kit (RNAeasy mini kit, Qiagen, Valencia, CA, USA) and stored at −80°C for quantitative real-time PCR assays. The remaining 25 µl of midgut homogenate was used in Western blot assays, as described below. Infections of sand flies with Le. major amastigotes V1 strain were done by addition of 5×106 parasites/ml into the blood meal. Le. major amastigotes were harvested from BALB/c mouse footpads lesions formed roughly 30 days after inoculation with 5×105 parasites from late phase culture according to [22]. dsRNA for PpChit1 were synthesized using the primers PpChit1/T7i_2–F (5′–TAATACGACTCACTATAGGGAGAATGAAGATATCATTGTGTGC-3′) and PpChit1/T7i_2–R (5′– TAATACGACTCACTTAGGGAGATCAGCATTGGACCAGGAAGG-3′), which contain the complete T7 promoter and amplify the full length sequence encoding the mature PpChit1. PCR reactions were performed with 0.5pmoles of each primer along with 1 µl of cDNA (synthesized from midgut dissected at 72 h post-blood meal, PBM), and 10 µl of GoTaq PCR master mix (Promega, Madison, WI, USA). The 20 µl PCR reactions were carried according to the conditions: 10 cycles of 95°C for 1 min, 55°C for 1 min, and 72°C for 1 min and 15 sec, followed by 35 cycles 95°C for 1 min, 65°C for 1 min, and 72°C for 1 min and 15 sec. The reaction products were purified and concentrated using the YM-100 filters (Millipore, Billerica, MA, USA), and 1 µg DNA was used for dsRNA synthesis using the Megascript RNAi kit (Ambion, Austin, TX, USA). dsRNA synthesis reactions were performed for four hours at 37°C, and the products were further purified following manufacturer's recommendations. Thereafter, dsRNAs were suspended in ultra-pure water and further purified and concentrated to approximately 3.5 mg/ml or 4.5 mg/ml using the YM-100 filters (Millipore). The positive control provided by the Megascript RNAi kit (Ambion; used in Real-Time PCR and Western blot assays) or a dsRNA specific to a green fluorescence protein gene (dsGFP [23]; for parasite counting assays) was used as controls for dsRNA injection assays. For dsRNA injections, individual females were anesthetized with CO2, kept on a cold dish, and injected intra-thoraxically with either 23 nl (3.5 mg/ml, 80.5 ng) or 32 nl (4.5 mg/ml, 144 ng) of dsRNA using Nanoject II microinjector (Broomall, PA, USA). Immediately following injection, flies were transferred to a 500 ml plastic container, provided with 30% sugar embedded cotton, and maintained inside a high humidity chamber (85–95% humidity at 25°C). Flies were allowed to recover for 48 hours and blood fed on an uninfected blood meal through a chicken membrane, as described above. Total RNA was isolated from individual midguts dissected as described above. RNA extraction was carried out using the RNAeasy mini kit (Qiagen) following manufacturer's instructions. Following extraction, the Turbo DNA-free kit (Ambion) was used to eliminate DNA contamination. After quantification, 25 ng total RNA was used for cDNA synthesis using 200 units of SuperScript III Reverse Transcriptase (200 u/µl), 2.5 µM Oligo (dT)20 primer, and 0.5 µM dNTPs (10 mM). These reagents were incubated at 65°C for 5 minutes (min) and kept in ice for at least 1 min. This step was followed by addition of a mix containing 4 µl 5× SuperScript III Reverse Transcriptase First-Strand Buffer, 5 mM DTT (0.1 M), 20 Units of RNaseOUT to the reaction. The mixture was incubated for one hour at 50°C and stored at −20°C. All the reagents for cDNA synthesis were purchase from Invitrogen (Carlsbad, CA, USA). Real-Time PCR reactions were performed using BioRad SYBR green and BioRad iCycler (BioRad, Hercules, CA, USA). The reactions were carried out in duplicate using 0.5 µl cDNA, 6pmoles of each primer (10 µM), 10 µl of 2× SYBR green, and 8.3 µl of Ultra Pure DNase/RNase-Free Water (Invitrogen). The primers used for chitinase amplification were PpChit_137F (5′ - ATGATCTGCATGGTTCTTGG - 3′) and PpChit_137R (5′ - GGAGCTCCATTTCGAATCC - 3′) while the S3 primers (Pp40S_S3_136F: 5′ - GGACAGAAATCATCATCATG – 3′ and Pp40S_S3_136R: 5′ – CCTTTTCAGCGTACAGCTC – 3′) were used for amplifications of the housekeeping control gene (encoding the protein S3 of ribosomal subunit 40S). The reaction cycle of 94°C for 1 min, 57°C for 1 min, and 72°C for 30 sec was repeated 40 times, and the amplification profiles were assessed using the BioRad iCycler software (BioRad). Polyclonal anti-PpChit1 sera were obtained by injecting three month old female BALB/c mice subcutaneously into the ears. Mice were injected three times in two weeks intervals with approximately 10 µg of purified VR2001 plasmid [24] encoding the mature chitinase protein [21] per mouse ear. Blood was collected from the submandibular vein (“cheek bleed”) of injected animals and antibody levels accessed via Easy-Titer IgG Assay Kit (Pierce, Rockford, IL, USA). Sera were maintained at −20°C until used. For Western blots, seven midgut extracts from flies injected with dsPpChit1 and dsControl were pooled together in RNasefree microcentrifuge tubes containing 1 µl of complete protease inhibitor (Thermo Scientific, Rockford, IL, USA) and concentrated using the YM-10 filters (Millipore). Total protein concentration in midgut extracts was quantified using BCA Protein Assay Kit (Thermo Scientific). Similar proteins amounts (5 µg per lane) from midguts of dsPpChit1 and dsControl injected sand flies were fractionated on 10% Bis-Tris NuPAGE gels (Invitrogen). Proteins were transferred to a nitrocellulose filter (Whatman, Dassel, Germany), incubated with PpChit1 antisera (1:100 dilution) overnight at 4°C, washed three times in TBS-T (1× TBS buffer with 0.1% tween-20) for 15 minutes each time. Blot was incubated with anti-mouse conjugated to alkaline phosphatase (1:10,000 in TBS-T) antibodies (Promega) for one hour at room temperature and washed in TBS-T as indicated above. The protein bands (56 kDa, [21]) were visualized using the Western Blue substrate for Alkaline Phosphatase (Promega). Alternatively, Western blot was incubated with anti-mouse-Horseradish Peroxidase secondary antibody (1:10,000) and detected with SuperSignal West Pico Chemiluminescence Substrate (Thermo Scientific) in chemiluminescence assays. Densitometry analysis was performed using the TotalLab TL100 software (Nonlinear Dynamics, Durham, NC, USA). In order to assess the PpChit1 knockdown effects on Le. major development, 80.5 ng of dsRNA was injected intra-thoraxically into P. papatasi, and flies were fed on an infected blood meal as described above. Midguts from fully engorged-only flies were dissected at 48 h and 120 h after the infective blood meal and homogenized in 30 µl of PBS buffer (pH 7.4). Parasites were counted with a hemocytometer. Two independent experiments were carried out for each time point. Mann-Whitney U test was performed to compare expression profiles as well as parasite numbers between sand flies injected with either dsRNA targeting PpChit1 transcripts (dsPpChit1) or the dsRNA control (dsControl) injected flies. D'Agostino & Pearson omnibus normality test was performed to assess whether parasite numbers followed a normal distribution. The Chi-square test (or Fisher's exact test) was performed in order to assess whether dsPpChit1-injected flies exhibit altered Le. major load compared to the dsControl-injected flies. Parasite infection load in flies dissected at 48 h post infection was scored according to parasite numbers in the sand fly midgut as no parasite, light infection (1–1,000 parasites), moderate infection (1,001–10,000), or heavy infection (>10,000), in accordance to [25]. For flies dissected at 120 h PBM parasite loads were categorized in two groups: zero or light infections (0–1,000 parasites) was arranged as one group, and moderate infection (>1,000 parasites) as another. Differences were considered statistically significant at p<0.05, and tests were carried out using GraphPad Prism v. 5.01 software (GraphPad Software, Inc). Injection of 80.5 ng of dsRNA into the sand fly thorax targeting the midgut expressed PpChit1 gene led to a significant decrease in PpChit1 mRNA levels in comparison with the control dsRNA-injected flies (Figure 1). Reduction of PpChit1 expression after a blood meal varied over time. Twenty four hours after blood meal (and 72 h after injection of dsPpChit1), a 27% reduction of PpChit1 transcripts was detected (Figure 1A). At 48 h PBM (previously shown to be the maximum activity for PpChit1 [21]) and at 72 h PBM, reductions of 58% and 53% on average of the PpChit1 expression were observed (Figure 1A). Finally, at 96 h PBM (120 h after dsRNA injection), when no chitinolytic activity was detected [21], the reduction in PpChit1 expression was 72%. On the other hand, injection of 144 ng of dsPpChit1 into P. papatasi thorax displayed a weaker reduction in PpChit1 expression levels than injection of 80.5 ng (Figure 1B). Although similar expression reduction at 24 h PBM was exhibited (26%, Figure 1B), expression differences between dsPpChit1 and dsControl injected flies at 48 h and 72 h PBM were lower (13% and 43%, respectively) than detected at the same time points when 80.5 ng of dsRNA was injected (Figure 1B). These differences could be occurring due to a still obscure feedback loop for transcription activation upon knock down, as proposed elsewhere [26]. Silencing of the PpChit1 message RNA produced a concomitant reduction in the amount of PpChit1 protein as determined by Western blots (Figure 2). Similar to the Real-Time PCR data, reduction in PpChit1 protein levels in dsPpChit1 injected flies was detected at 48 h and 72 h PBM (Figures 2A–C) when either 80.5 ng or 144 ng of dsRNA was injected. No PpChit1 expression was detected at 24 h PBM (Figure 2B). Likewise, densitometry analysis of blot developed using a chemiluminescence method displayed 95% reduction in PpChit1 protein levels at 48 h PBM when 144 ng dsPpChit1 (Figure 2C and D). Interestingly, the corresponding time point only led to 13% reduction of PpChit1 mRNA levels, as shown in Figure 1B. As injection of either 80.5 ng or 144 ng of dsRNA targeting PpChit1 transcripts are capable of significantly reducing PpChit1 expression levels in the midgut of P. papatasi (Figure 1 and 2), we assessed the effects of injecting 80.5 ng of the dsRNA on Le. major development within the injected flies. Following the injection of the PpChit1 dsRNA, flies were provided an infective blood meal, and dissected at different time points after feeding. Our results demonstrate that dsPpChit1-targeted knock-down resulted in significant reductions in parasite load within the sand fly midgut as the numbers of Le. major were reduced by 46% (or 1.85 fold) at 48 h post infection (Figure 3A) and by 63% (or 2.70 fold) at 120 h PBM post infection (Figure 3B). The injection of dsPpChit1 also affected the range of parasite loads. An analysis of the range of parasite load at 48 h and 120 h post infection points to a normal distribution of parasite numbers in the dsControl-injected flies (48 h PBM, p = 0.51, and 120 h PBM, p = 0.26, D'Agostino & Pearson omnibus normality test), whereas for dsPpChit1-injected flies this distribution was significantly affected (48 h PBM, p = 0.004, and 120 h PBM, p<0.0001, D'Agostino & Pearson omnibus normality test). Changes in P. papatasi infection levels following silencing of PpChit1 were further confirmed by comparing infection prevalence. For instance, injection of dsPpChit1 reduced the prevalence of heavy infection from 47% (dsControl-injected) to 19%, and of light infection from 19% (dsControl-injected) to 6% at 48 h post blood feeding (Figure 4A). Likewise, moderate infections levels were reduced from 57% (dsControl-injected) to 14% at 120 h post infection (Figure 4B). After a blood meal, sand flies synthesize a PM type 1 that is fully developed at approximately 36–40 h PBM [27]. In addition to compartmentalizing the blood meal and protecting the epithelia, the sand fly PM serves an additional dual role regarding Leishmania infection: as a barrier to these parasites but also as protection against proteolytic attack on transitional-stage amastigotes [8], [20], [28], [29], [30]. In order to successfully complete its cycle within the sand fly, Leishmania nectomonads must escape from endoperitrophic space, through the PM, to prevent being passed together with remnants of the digested blood meal [8]. We have previously characterized a functional, blood-induced chitinolytic system, in the midgut of P. papatasi and L. longipalpis sand flies [21], [31]. We also demonstrated that polyclonal antibodies to PpChit1 inhibit the midgut chitinolytic activity in vitro, and this effect also was shown across different sand fly species [21]. PpChit1 is presumably involved in the maturation and degradation of P. papatasi PM (as is its ortholog in L. longipalpis, LlChit1) [21], [31], and addition of allosamidin, a chitinase inhibitor to the infective blood meal of this sand fly led to entrapment of Le. major within the peritrophic space [8]. Although allosamidin may have also inhibited chitinase secreted by Leishmania, taken together, these data suggested that PpChit1 also can be involved with Leishmania escape from the endoperitrophic space. To address whether silencing of PpChit1 transcripts via RNAi-induced pathway would affect Le. major development within its natural vector, P. papatasi, we synthesized a dsRNA specifically targeting PpChit1. Injection of dsRNA targeting specific transcripts has now been widely applied in disease vectors and proven an invaluable tool for the understanding of underlying events in pathogen-vector relationships [32], [33], [34]. In sand flies, gene silencing with dsRNA was first applied to L. longipalpis cell culture [35], inducing a non-specific antiviral response. Recently, dsRNA injection of adult sand flies led to a specific reduction of Xanthine dehydrogenase expression [36], and to an effect on Le. mexicana development when a midgut trypsin produced by L. longipalpis was silenced [30]. The midgut chitinase PpChit1 is only expressed following a blood meal [21]. Thus, following injection of dsPpChit1 double-stranded RNA, sand flies were blood fed and midguts dissected at different intervals after feeding. Specific silencing of PpChit1 transcripts was detected by quantitative real-time PCR analyses (Figure 1), with concomitant knock down of PpChit1 protein levels assessed by Western blots (Figure 2). Based on the presumptive role of PpChit1 in the maturation and degradation of the PM1, we expected that silencing of this gene would lead to entrapment of Leishmania within the endoperitrophic space. Our results are consistent with this hypothesis, as Le. major load was reduced 120 h PBM in midguts of dsPpChit1 injected P. papatasi (Figures 3 and 4) suggesting that PpChit1 is indeed involved in PM1 degradation. Moreover, reduction of the Le. major load at 48 h PBM in dsChit1 compared to control-injected flies might have been caused by at least two scenarios: 1) a reduction in nutrient availability in the endoperitrophic space as the PM may be less permeable to proteolytic enzymes, or in the contrary, 2) to inability of parasites to escape leading to longer exposure to digestive enzymes inside the peritrophic space. Regardless of the mechanism, it still remains to be determined. Future studies will assess whether this is a feasible approach in preventing transmission from an infected animal to a naïve host. Moreover, the results support the targeting of PpChit1 as a mean to interfere with Leishmania development within the sand fly – a candidate transmission-blocking vaccine.

10.1371/journal.ppat.1004170

A Multifactorial Role for P. falciparum Malaria in Endemic Burkitt's Lymphoma Pathogenesis

Endemic Burkitt's lymphoma (eBL) arises from the germinal center (GC). It is a common tumor of young children in tropical Africa and its occurrence is closely linked geographically with the incidence of P. falciparum malaria. This association was noted more than 50 years ago. Since then we have learned that eBL contains the oncogenic herpes virus Epstein-Barr virus (EBV) and a defining translocation that activates the c-myc oncogene. However the link to malaria has never been explained. Here we provide evidence for a mechanism arising in the GC to explain this association. Accumulated evidence suggests that eBL arises in the GC when deregulated expression of AID (Activation-induced cytidine deaminase) causes a c-myc translocation in a cell latently infected with Epstein-Barr virus (EBV). Here we show that P. falciparum targets GC B cells via multiple pathways to increase the risk of eBL. 1. It causes deregulated expression of AID, thereby increasing the risk of a c-myc translocation. 2. It increases the number of B cells transiting the GC. 3. It dramatically increases the frequency of these cells that are infected with EBV and therefore protected from c-myc induced apoptosis. We propose that these activities combine synergistically to dramatically increase the incidence of eBL in individuals infected with malaria.

Endemic Burkitt's lymphoma (eBL) is a common tumor of young children in tropical Africa that is closely linked geographically with P. falciparum malaria. This association was noted more than 50 years ago. Since then we have learned that eBL contains the oncogenic herpes virus Epstein-Barr virus and a defining translocation that activates the c-myc oncogene. However the link to malaria has never been explained. Here we show that malaria has multiple effects that all focus on germinal center (GC) B cells that are known to be the origin of eBL. Together these effects of malaria act synergistically to dramatically increase the risk of developing eBL in individuals infected with the parasite. Clinical interventions that lessen the impact of malaria on GC B cells should dramatically decrease the incidence eBL.

Endemic Burkitt's lymphoma (eBL) is an extremely common tumor of young children in tropical Africa [1]. Genetic, phenotypic and transcriptional analysis suggests that it originates from germinal center (GC) cells [2], [3] although it actually grows in extrafollicular locations. It is defined by a well described chromosomal translocation between the c-myc oncogene and one of the immunoglobulin loci that results in constitutive activation of the oncogene leading to uncontrolled growth of the cell [4], [5], [6]. Recent studies indicate that this translocation may be mediated as a consequence of deregulated expression of the enzyme AID (Activation-induced cytidine deaminase) [7], [8], [9]. AID is highly expressed in GC B cells and is normally responsible for the processes of somatic hypermutation and class switch recombination of immunoglobulin genes as they undergo affinity maturation in the GC [10]. This restricted expression of AID further supports the notion that eBL originates in the GC. eBL is also closely associated with two infectious agents, P. falciparum malaria and Epstein-Barr virus (EBV) [1], [11], [12]. The distribution of the tumor in Africa closely matches that of hyper- and holoendemic malaria [12] while EBV was originally discovered in eBL tumor biopsies. Subsequently, we have learned a great deal about the molecular mechanism behind eBL pathogenesis and the transforming ability of EBV. EBV is a B lymphotropic herpes virus that can drive the activation and proliferation of newly infected B cells by expressing a series of latent proteins and noncoding RNAs that collectively are referred to as the growth transcription program[13], [14]. In vivo, however, EBV establishes a lifelong, quiescent, persistent infection in resting memory B cells [15], [16]. The virus makes the transition in vivo from a newly infected activated B cell blast to a resting memory B cell via passage through the GCs of the tonsillar lymphoid tissue [[17], [18]. In doing so it recapitulates the mechanism by which normal B cells become memory B cells (for a detailed description of the mechanism see [14]). Normally deregulation of c-myc expression such as is found in eBL would lead to apoptotic death of the cell; however, evidence suggests that exposure to the EBV growth program prior to entry into the GC [14], [19] and viral genes expressed in the GC [20], [21] are sufficient to convey a level of resistance to this apoptosis. Thus, the cells in the GC most likely to tolerate the c-myc translocation are the ones already latently infected with EBV. This also places EBV at the site of eBL origin, the GC. Interestingly GC cells carrying EBV express only a limited subset of the latent proteins (default transcription program) and even these become silenced as the infected cells enter the memory compartment [16], [17], [22]. Here the virus only expresses small non-coding RNAs including ∼40 miRNAs. The exception is that they also express the viral DNA tethering protein EBNA1 when the cells occasionally divide as part of normal memory B cell homeostasis [23]. Viral gene expression in eBL resembles the infected dividing memory B cells, not the GC cell: i.e. viral gene expression is limited to the viral DNA tethering protein (EBNA1) and the non-coding RNAs. This led us to propose that eBL is a tumor of a GC cell that has left the lymph node to become a resting memory B cell but is unable to do so because it continues to proliferate, driven by the deregulated c-myc oncogene. While understanding the function of c-myc and EBV in eBL has progressed, the role of P. falciparum malaria has remained poorly understood. P. falciparum malaria is immunosuppressive [24] and there is considerable evidence that this leads to much higher viral burdens of EBV [25], [26]. However it is well documented that increased EB viral loads associated with immunosuppression [27] predispose to EBV positive immunoblastic lymphoma not Burkitt's lymphoma [28]. We hypothesize that malaria plays multiple roles in eBL pathogenesis. First, we propose that malaria has the capacity to induce deregulated expression of AID thereby increasing the likelihood of the translocation event. Second, we suggest that the higher viral burdens lead to an increased frequency of newly infected B cell blasts in the tonsils. This results in more EBV infected B cells transiting the GC and consequently a higher frequency of cells in the GC able to tolerate a c-myc translocation. Taken together malaria infection would both increase the likelihood of c-myc translocations in the GC and the probability that it will occur in a cell that can tolerate it, namely an EBV infected cell. In this paper we have sought to test this hypothesis. To provide direct support for our hypothesis we have sought in vitro evidence that P. falciparum can stimulate AID expression. Tonsil B cells were incubated with malaria extract prepared by lysing red blood cells infected with P. falciparum. As controls we used the known Toll like receptor 9 (TLR9) agonist CpG and costimulation with IL-4 and CD40 ligand. Figure 1A shows a time course of AID mRNA induction with various combinations of stimulants. It is apparent that optimal induction requires a combination of T cell help (CD40 ligand and IL-4) and the TLR9 agonist CpG, with peak activation occurring after 5 days of culture. In our hands, CpG alone had minimal or no effect on AID induction although it was extremely potent in driving B cell proliferation (not shown). These results are consistent with previous reports [29]. Figure 1B shows the same experiment where extracts from lysed red blood cells infected with P. falciparum were also tested. Similarly to CpG, the parasite extract had minimal or no effect when added alone but showed strong stimulation of AID expression in combination with IL-4 and CD40 ligand. This suggests that the parasite would only stimulate AID expression when T cell help is available, i.e. in the GC. As with CpG activity peaked by day 5. At this time, the parasite extract was three times as effective as CpG in inducing AID mRNA. The parasite extract differed from CpG in that when added alone it was not able to stimulate B cell proliferation (not shown). Optimal stimulation of AID expression by CpG requires costimulation through the BCR. This is demonstrated in Figure 2, where inclusion of sIg cross-linking increased stimulation two and a half fold compared to CpG alone. Comparison to parasite extract in the same experiment demonstrated that sIg cross-linking plus CpG were about as effective as parasite extract alone and that sIg cross-linking did not significantly enhance the effect of parasite extract. Extracts from uninfected RBCs showed no activity either alone or in combination with other stimuli (Figure 2). We conclude that the parasite is able to generate a signal that is as effective as the TLR9 and BCR signals combined. To confirm that the increase in AID mRNA stimulated by the parasite was reflected in increased AID protein expression, we repeated the stimulation experiments and examined the resulting cells for AID by flow cytometry. As may be seen in Figure 3A–C, parasite extract induced comparable levels of AID protein expression to that achieved with CpG when combined with sIg cross-linking. It is noteworthy that in some cases the relative level of protein expression measured by FACS did not match the levels seen for the mRNA (see for example Unstimulated versus CD40L +IL4). This likely reflects that the cells are newly stimulated in culture and there is a lag between the synthesis of AID mRNA and the production of the protein. We conclude, therefore, that P. falciparum is a potent antigen independent stimulator of AID expression. When combined with T cell help signals it is at least as effective as the combination of CpG and BCR cross-linking. It has been reported that the P. falciparum metabolic product of hemoglobin digestion, hemozoin [30], is a ligand for TLR9, but this has only been tested with dendritic cells [31], [32]. It has not been shown for B cells. To test if hemozoin could be taken up by B cells we have incubated hemozoin and CpG with two B cell lines, BL2 (EBV negative BL line) and IM171 (spontaneous EBV positive lymphoblastoid cell line). As may be seen in Figure 4A, both were readily taken up by the B cells (Hemozoin crystals were visualized using reflection microscopy and CpG is tagged with Alexa-488). To test if hemozoin could thereby act to stimulate AID expression we have compared the stimulatory activity of hemozoin to CpG and sIg. The result is shown in Figure 4B. Hemozoin alone had no effect, however in combination with sIg cross-linking it was two and a half fold more effective than sIg alone, comparable with CpG plus sIg. No appreciable increase was observed when parasite DNA was added (3 ug/ml) either alone or together with hemozoin. This suggests that the hemozoin preparation alone was sufficient for the activity. These results demonstrate that hemozoin is one component of the P. falciparum extract that is capable of stimulating AID expression however there must be another component that is mimicked by sIg cross-linking to obtain the optimal stimulation obtained with whole parasite extracts. To test the hypothesis that malaria is associated with higher levels of AID expression in vivo, we have examined and compared purified GC cells from tonsils obtained from malaria infected and uninfected patients who were matched for age, sex and socioeconomic class. The results are summarized in Figure 5A–B. mRNA was isolated from similar numbers of GC B cells. For all samples AID and c-myc mRNA levels were normalized to β-actin and the level in the GC population is expressed relative to that in a standard naïve B cell population (calibrator) isolated from a single Boston tonsil. The level of AID mRNA in the GC cells from the control population are comparable with what has been reported previously [33]. In comparison, the levels of AID mRNA in the malaria tonsils were significantly higher, about 5 fold on average. However, although all the values for the malaria tonsil were above the range of the controls, the spread was large such that some samples had levels 8–13 fold higher than controls. The results for c-myc transcripts were less striking. Most of the malaria samples showed little or no significant difference from controls with the exception that two of the nine samples tested clearly had a markedly elevated level of c-myc mRNA. The level of c-myc mRNA did not correlate with the levels of AID mRNA. These results confirm our prediction that the levels of AID expression are higher in the tonsil GC B cells of individuals infected with malaria. To test the hypothesis that malaria is associated with higher levels of EBV infected cells in the GC, the GC cells from the same sets of tonsils described above were also analyzed for the presence of EBV. We recovered similar numbers of cells from both sets of tonsils and the fraction of B cells was also similar (Figure 6A). The fraction of GC cells in the control tonsils was ∼32% (Figure 6B), consistent with what we have seen in previous studies. However, the frequency in the malaria tonsils was unusually high ∼50%. When we analyzed the frequency of EBV infected cells in the two sets of tonsils, an even more dramatic difference was observed (Figure 6C and Table 1– note the log-scale in the Figure), with the level being ∼50 fold higher on average in the malaria tonsils compared to controls (log mean: 2,275/107 versus 51/107; median 2,580/107 versus 52/107). Taking the increased numbers of GC cells into account, this means that the malaria tonsils have approximately 70 fold more EBV infected cells in their GCs. These results confirm our prediction that the levels of EBV infected GC B cells should be higher in the tonsils of individuals infected with malaria. As may be seen in Figure 6D, although both EBV infection and AID expression were elevated in all of the malaria tonsils there was no correlation between the levels of EBV infection and the level of AID in either malaria or control samples. We conclude that individuals infected with malaria have an increased level of EBV infected cells and AID mRNA expression in their tonsil GC cells, but these levels are not correlated. We have shown above that the level of c-myc transcripts is significantly elevated in GC cells from the tonsils of a subset of individuals with malaria. However, there has been controversy as to whether c-myc is actually expressed in the GC [34], [35], [36], [37]. Therefore, to confirm that we were detecting c-myc expression in GC cells we performed several experiments as shown in Figure 7. GC cells (CD10+ tonsil B cells) were positive for c-myc expression when stained with a c-myc specific antibody and analyzed by flow cytometry (Figure 7A). The presence of c-myc protein in our GC cell preparations was confirmed by Western blot (Figure 7B), where the signal was specifically blocked by the myc specific peptide used to raise the antibody. The specificity and correct location of the protein in the nucleus was confirmed by analysis with the ImageStream (Figure 7C). We conclude, therefore, that our studies support the current opinion that c-myc is expressed in GC cells. It is now more than 50 years since the association between P. falciparum malaria and eBL was first proposed [11], [12]. Since that time, confirmation of a direct link and a mechanism to explain it has been lacking. Here we have presented studies on the effect of malaria in the context of the GC and provided evidence for a multifactorial effect of malaria that can account for the increase risk of eBL. These include the activation of AID expression, possible heightened c-myc transcription, increasing the numbers of B cells transiting the GC and increasing the fraction of these cells that are EBV infected. The common component linking these effects is the GC. The GC is the structure where immunoglobulin genes undergo somatic hypermutation and class switch recombination [38], [39], mediated by AID [10], as the cells undergo affinity maturation. It is this enzyme that is responsible for causing the c-myc translocation characteristic of eBL [7], [8], [9]. However, the GC is also the site where newly infected EBV blasts undergo the transition to become resting latently infected memory B cells [14], [16], [22]. Therefore, in the presence of malaria the number of cells able to tolerate a c-myc translocation (EBV infected) in the GC are increased and the probability of a c-myc translocation is also increased (AID activation). Thus, it is the increased probability of a fortuitous collision of AID and EBV in GC cells, both exacerbated by malaria that leads to eBL. A further observation confirming that these events occur within the GC was the finding that the P. falciparum extract alone had little or no ability to induce AID. Maximal induction required co-stimulation with CD40 ligand and IL-4. This means that the malarial parasite can only work to induce AID expression if T cell help is also provided. Since T cell help is specifically provided in the GC [40], [41], this firmly places the role of P. falciparum in the induction of AID in the GC and would seem to rule out fortuitous activation elsewhere. This would explain why eBL is uniquely a tumor of GC cells. We observed no difference in the percentage of total B cells per tonsil between the malaria-endemic and the non-malaria regions, however, GC cells are ∼2 fold higher in the malaria background. This suggests that malaria (or malaria background) does not disrupt the size of the B cell pool but increases the likelihood that more B cells, and therefore EBV-infected cells, either enter the GC or that the cells stay longer in the GC. Either way, this increases the chances that some B cells, including EBV infected cells, will develop deleterious mutations and translocations. In individuals with malaria we detected a wide range in the number of GC cells containing EBV and in the degree of AID expression. It is tempting to speculate that those with the rare fortuitous coincidence of extremely high levels for both may be the likely candidates for tumor development. We have provided compelling in vitro evidence to support the claim that P. falciparum induces AID expression, the first such evidence. Importantly, the parasite induces AID in an antigen independent, i.e. deregulated fashion. Such expression is known to be a predisposing factor for the c-myc translocation. EBV infection of B cells in vitro, as well as EBV associated proteins, have also been shown to induce AID expression [42], [43], [44]. It is conceivable therefore that EBV and malaria could even cooperate in the induction of AID. We have shown that P. falciparum is capable of eliciting AID expression at least as effectively as the combination of CpG and BCR cross-linking, suggesting that the parasite can provide both signals. We have identified hemozoin, the crystalline by product of P. falciparum digestion of hemoglobin [30] and a reported TLR9 ligand [31], [32], as one of the parasite components responsible. Consistent with its signaling through TLR9, hemozoin is only effective at inducing AID in the presence of BCR cross-linking. Thus, it is likely that there is a second, as yet unidentified, parasite derived ligand that provides the surrogate BCR signal. A likely candidate for this is the P. falciparum specific protein PfEMP1 encoded by the var gene family [45], which has been shown previously to bind to and activate B cells through the BCR [46]. This would also explain the specific link of P. falciparum with eBL since only this species of malaria expresses PfEMP1. The prediction is therefore, that a combination of hemozoin and PfEMP1 are providing the requisite TLR9 and sIg signals. Since TLR9 is a receptor for polynucleotides it was somewhat surprising that hemozoin was able to stimulate AID expression in the absence of parasite DNA. It is controversial as to whether hemozoin does [32] or does not [31] need or be associated with DNA in order to signal through TLR9. It is possible that hemozoin may be signaling via a TLR9 independent mechanism in our system since it has been reported that it can signal through other pathways including the inflammasome [47], [48]. However, there was a high death rate in our in vitro assays therefore it is possible that endogenous unmethylated DNA from the dead cells was being trafficked into the endosomal compartment by hemozoin to stimulate TLR9 and consequently lead to AID induction. Demonstration of an increased viral burden of EBV in individuals with P. falciparum malaria is not a new observation [25], [26], however in this paper we have shown a direct mechanistic consequence of this elevation that provides a risk factor for eBL namely an increased number of latently infected cells in the GC. This effect is not modest, with the mean frequency of EBV infected GC B cells from the malaria tonsils being about 50 fold more than the frequency of EBV infected GC B cells from the controls. Combined with the increase in the number of total GC cells this results on average in there being 70 times more EBV infected cells in the GCs of individuals with malaria compared to controls. We have shown previously that healthy individuals have at any time on average approximately 3 EBV infected cells per GC [49]. Our results here indicate that this would increase to around 150–200 per GC for an individual with malaria. This is a significant increase in risk since the consequence is an extremely high number of (EBV infected) cells each of which are primed to tolerate the c-myc translocation that emanates from overexpressed AID. The increased frequency of latently infected cells seen in the malaria samples (∼50 fold) is very similar to what we have reported previously for immunosuppressed patients and patients with SLE [27], [50]. Indeed, it has been reported previously that malaria is immunosuppressive for T cell responses [51], including those directed against EBV [24]. It is tempting therefore to speculate that children with malaria are immunosuppressed and that this explains the risk for eBL. However, we have pointed out previously that immunosuppression is a risk factor for post-transplant lymphoproliferative disorder (PTLD)-like disease, i.e. immunoblastic lymphoma not eBL [14], [22]. Thus, immunosuppression alone is not sufficient to explain eBL. The results presented here also provide further support for the GC model of EBV persistence. This model, which is now generally accepted, holds that EBV establishes a persistent infection by driving newly infected blasts through the GC to become latently infected resting memory B cells [14], [16], [22]. A direct prediction of this model is that a higher viral burden should produce a higher number of EBV infected cells transiting the GC and that prediction has been fulfilled in this study. Furthermore, the studies presented here provide a powerful functional confirmation of the model in that they provide an explanation for the link between malaria and eBL. Thus, the model predicts that it is the elevated rate of passage of virus infected cells through the GC, together with deregulated AID expression, that explains the origin of eBL. Our studies suggest that malaria may also induce heightened expression of c-myc in the GC, at least in some individuals. c-myc is a non-traditional transcription factor with a very complex regulation. High c-myc expression in the GC could account for the observed high rate of proliferation, as well as the high apoptotic tendency of GC B cells [34], [37]. However, the question of whether c-myc is even expressed in the GC has been controversial in the past. Martinez-Valdez et al. [37] and Cutrona et al. [34] report that c-myc is highly expressed in GC B-cells, whereas Klein et al. [36] were unable to confirm these findings. Recently, Dominguez-Sola et al. [35] have presented compelling evidence that c-myc is expressed in GC cells, specifically by B cells selected for reentry into the dark zone, a conclusion supported by our work. Furthermore, for AID to target c-myc in GC B cells, c-myc must be expressed since AID deaminates transcribed substrates and acts on selected highly transcribed genes when they are over-expressed. Thus, higher than normal levels of c-myc transcription driven by malaria could further increase the risk that AID would target the c-myc gene for a translocation event. In conclusion, we have presented the first direct evidence for a mechanism to explain the link between eBL and holoendemic malaria. If correct, these observations imply that reducing the exposure to P. falciparum malaria or the development of drugs to block the ability of malaria to induce AID should dramatically reduce the incidence of eBL in young children in tropical Africa. Specifically this should act as a spur to agencies interested in reducing the incidence of exposure to P. falciparum in young children. This study was approved by the Institutional Review Board of the Tufts Medical Center, Boston, USA and the Committee on Human Research and Ethical Publications of the School of Medical Sciences, Kwame Nkrumah University of Science and Technology (KNUST), and Komfo Anokye Teaching Hospital (KATH), Kumasi Ghana. The material used was deidentified, discarded tonsil tissue and was deemed exempt from informed consent by the IRB. The tonsil material was obtained indirectly either through the Pathology Department at Tufts Medical Center or the EENT Clinic at Komfo Anokye Teaching Hospital. The EBV-positive lymphoblastoid cell line IB4 (gift of Elliott Kieff) and Namalwa Burkitt's lymphoma cell line were used as positive controls for DNA PCR of the W-repeat region of the EBV genomes. The EBV-negative cell line CB60, a mouse T-cell hybridoma cell line (gift of Miguel Stadecker) was used as a negative control in all W-PCR experiments. The Burkitt's lymphoma cell lines Raji and Rael were used as positive controls for c-myc western blot. The EBV negative BL2 Burkitt's lymphoma cell line and SP-IM 171 spontaneous EBV lymphoblastic cell line were used in hemozoin-DNA complex internalization assays. All cell lines were cultured at 37°C with 5% CO2 in RPMI 1640 supplemented with 10% fetal bovine serum, 2 mM glutamine, 2 mM sodium pyruvate, 100 IU of penicillin-streptomycin, and 10 µg/ml ciprofloxacin hydrochloride (RPMI-complete). Palatine tonsils were obtained from patients 14 years or younger undergoing routine tonsillectomy. Twelve tonsil samples were obtained from patients at the EENT Clinic Komfo Anokye Teaching Hospital, Kumasi, Ghana. These were processed at the Kumasi Center for Collaborative Research in Tropical Medicine (KCCR), Kumasi, Ghana, stored in liquid nitrogen and shipped to Tufts University School of Medicine, Boston, USA on dry ice (under the supervision of Prof. Karen Duca, KNUST, Kumasi, Ghana). The presence of the parasite was confirmed based on detection of P. falciparum DNA and/or antigens (see below). Kumasi is an area of holoendemic malaria therefore it was not surprising/unexpected that all patient samples received tested positive for the parasite. To obtain parasite free control samples we therefore also collected twenty one tonsils from a malaria negative region (Boston MA) through the Pathology Department at Tufts Medical Center, Boston MA, USA. The identical procedure and reagents were used for harvesting tonsils in Boston and Kumasi. The malaria infected individuals who provided the tonsils are genetically distinct from the controls and could be subject to a wider range of infection and lower level of general hygiene. To minimize this possible source of variation we collected the malaria tonsils from age and sex matched donors in an area of high socioeconomic status in Kumasi where economic and medical standards were comparable to those in Boston. Tonsil tissue was cut into very small pieces in ice-cold PBSA (1x PBS +0.5% BSA) and then minced. Supernatants were pipetted through a cell strainer into 50 ml conical tubes to remove debris. Supernatants were centrifuged at 1,600 rpm at room temperature for 10 minutes and then aspirated. Pellets were re-suspended and brought to 50 ml with PBSA. About 25 ml of cells was carefully layered onto 20 mls of Ficoll-paque plus (GE Healthcare Biosciences, Philadelphia, USA) and then spun at 2,000 rpm for 30 minutes at room temperature (with no brake). Mononuclear cells were collected from the interface (buffy coat) and cell pellet discarded. The volume of mononuclear cells was adjusted to 50 ml with PBSA (and an aliquot taken for counting); cells were then washed once by spinning at 1,500 rpm for 10 minutes. After counting, cells were frozen in fetal bovine serum (FBS) (Sigma, St. Louis, USA) plus 10% DMSO (dimethyl sulfoxide) at 1×108 cells/ml. Cells were aliquoted in cryotubes and kept on ice for about 5 minutes, stored at −80°C overnight and then transferred to liquid nitrogen for long term storage. The identical procedure and reagents were used for preparing tonsil cell suspensions in Boston and Kumasi. Tonsil mononuclear cells were thawed in medium (RPMI/FBS), and spun down at 1,500 rpm for 5 minutes. B cells were either first purified using StemSep (StemCell technologies, Vancouver, Canada) according to manufacturer's instruction, or cells were re-suspended in 0.5% BSA in 1x PBS (PBSA) i.e. staining buffer. Cells were resuspended at 5×106 cells/100 µl PBSA either directly into FACs tubes or 15 ml tubes for staining. For extracellular staining the appropriate concentration of fluorochrome conjugated antibody was added to cells in appropriate tubes and incubated for 15 minutes at room temperature in the dark, after thorough mixing. Cells were washed once with PBSA, vortexed, and spun down at 1,500 rpm for 5 minutes. Finally cells were re-suspended in 300 µl PBSA and stored at 4°C until analyzed. For intracellular staining tonsil cells were pelleted, washed in Dulbecco's PBS− (without calcium or magnesium), and fixed either with 4% formaldehyde or with BD Cytofix/Cytoperm fixation and permeabilization solution (BD Biosciences, San Jose, USA) and then incubated for 20 minutes at room temperature. Cells were then washed twice either with 1x BD wash/perm buffer (BD Biosciences, San Jose, USA) or with 0.04% saponin based wash buffer and spun down at 1,500 rpm for 5 minutes. Permeabilization buffer (0.5% saponin based buffer) was used to re-suspend cells at 5×106 cells/100 µl. Two microliters (2 µl) of normal human serum (60 mg/ml, Thermo Fisher Scientific Rockford, USA) was added in order to block against non-specific antibody binding, and cells were incubated for 30 minutes at room temperature. Primary antibodies for intracellular antigens were added and incubated at room temperature for 30 minutes. Cells were then washed once and re-suspended in 0.5% saponin based buffer, 2 µl of normal human serum was again added and incubated for 30 minutes at room temperature. Fluorochrome conjugated secondary antibodies were then added at appropriate dilutions and incubated for 15 minutes in the dark. Cells were washed once with wash buffer, spun down at 1,500 rpm for 5 minutes and stored in 300 µl PBSA at 4°C until analysis. Cell sorting was performed on a MoFLo or Influx cell sorter and analysis on a FACSCalibur or Image Stream at Tufts University laser cytometry core. Sorted populations were >90% pure. A list of the antibodies and fluors used in this study is given in Table S1. The presence of the parasite was confirmed based on detection of P. falciparum DNA and/or antigens. Nested, parasite specific DNA PCR was performed for a sequence in the 2nd exon of the Chloroquine transporter (Pfcrt) gene as follows: First round (94°C, 3 min; 94°C, 30 sec; 56°C, 30 sec; 60°C, 1 min 30 cycles, 72°C, 3 min) forward primer CCGTTAATAATAAATACACGCAG, reverse CGGATGTTACAAAACTATAGTTACC (95°C, 5 min; 92°C, 30 sec; 48°C, 30 sec; 65°C, 30 sec (25 cycles); 72°C, 3 min) forward primer TGTGCTCATGTGTTTAAACTT reverse ACAAAATTGGTAACTATAGTTTTG. PCR product sizes were verified on 2% agarose gels. 1st round amplicon 527 bp, 2nd round 145 bp. P. falciparum antigens were detected employing a rapid diagnostic test (RDT) cassette (ACON Laboratories, Inc., San Diego, USA) as directed by the manufacturer. For limiting dilution analysis, GC B cell (CD19+CD10+) populations were isolated by flow cytometry and usually 10 replicates each of 2×105, 1×105, 5×104, 2.5×104, 1.25×104, 0.625×104, 0.3125×104 (higher or lower dilutions were added as needed) were placed in a 96 V bottom plate for subsequent EBV W-repeat DNA PCR. The plate was spun at 1,200 rpm for 10 minutes at 4°C and the supernatant aspirated. To each well was then added 20 µl of digestion mix (10X PCR buffer, 100 µl; Igepal (NP-40) 100 µl; Tween-20 100 µl; Proteinase K 50 µl of 20 mg/ml stock and 650 µl of water). The plate was sealed air-tight and incubated at 55°C overnight. This was followed by proteinase K deactivation at 95°C for 10 minutes. Ten microliters (10 µl) of water was then added to all the sample wells. The fraction of EBV positive wells was then assessed by W-repeat EBV DNA PCR for each replicate and the frequency of EBV infected cells in the starting population calculated using Poison statistics. The EBV-positive cell lines IB4 and Namalwa were used as positive controls and the EBV-negative cell line CB60 was used as a negative control. DNA real-time PCR specific for the W-repeat sequence of the EBV genome was performed as described [52]. For each reaction, a master mix was prepared, containing 12.5 µl IQ Supermix (Biorad cat 170-8862), 2.5 µl of 900 nM primers and 2.5 µl of 250 nM fluorogenic probe. Five microliters (5 µl) of DNA was added to 20 µl of master mix with a final reaction volume of 25 µl. (See Table S2 for primer and probe sequences). The PCR reactions were performed on a Biorad iCycler. The protocol was as follows: Step 1 (1 cycle): 3′ at 95°C; Step 2 (50 cycles): 15″ at 95°C, 1′ at 60°C. RNA was purified by TRIzol extraction (Invitrogen, Life Technologies, Grand Island, USA) and then treated with TURBO DNase (Ambion, Life Technologies, Grand Island, USA) to eliminate DNA prior to RNA amplification (where necessary). cDNA was made from the RNA using a cDNA synthesis kit (Invitrogen iScript cDNA synthesis kit). For the cDNA synthesis reaction, a master mix was prepared which included 4 µl of 5X iScript reaction mix, 1 µl of iScript reverse transcriptase, and 8 µl of nuclease-free water. Seven microliters (7 µl) of purified RNA was added to 13 µl of master mix. All reactions were performed on an Applied Biosystems PCR machine (Thermal cycler). The protocol was as follows: one cycle that included 5 minutes at 25°C, 30 minutes at 42°C, and 5 minutes at 85°C. For real time PCR a master mix was prepared, containing 12.5 µl of IQ Supermix (Bio-Rad), 2.5 µl of 900 nM primers, and 2.5 µl of 250 nM fluorogenic probe, except when Taqman pre-developed assays were being used in which case, 12.5 µl Supermix, 1.25 µl 20x primer-probe mix and 6.25 µl water. Five microliters (5 µl) of cDNA was added to 20 µl of master mix with a final reaction volume of 25 µl. All real time PCRs were performed on a Bio-Rad iCycler. The protocol was as follows: step 1, one cycle of 3 minutes at 95°C; step 2, 55 cycles of 15 seconds at 95°C and 1 minute at 60°C. All our real time PCR assays were optimized to detect down to the single cell level (See Supplemental Information for a list of primers and probes). P. falciparum parasites (3D7 line) were cultured using standard procedures as described [53]. Parasites were grown at 5% hematocrit in RPMI 1640 medium, 0.5% AlbuMAX II (Invitrogen), 0.25% sodium bicarbonate, and 0.1 mg/ml gentamicin. Cultures were incubated at 37°C in an atmosphere of 5% oxygen, 5% carbon dioxide, and 90% nitrogen. Parasite extracts were prepared by selective lysis of the host RBC membranes through the addition of saponin. Infected RBCs were suspended in 0.01% saponin in PBS and incubated at room temperature for 5 min. Host cell free parasites were pelleted by centrifugation, washed twice with PBS and stored frozen at −80 degrees C. Parasite extracts were prepared by 3 freeze-thaw cycles of frozen parasites, sonicated and stored at −80 degrees Celsius until needed. The protein concentration was determined by the Bicinchoninic assay (BCA assay – see below). Fresh tonsil mononuclear cells (MNC) were isolated as described above and re-suspended at ∼1×106 cells/ml in pre-warmed PBSA. Two microliters of 5 mM CFSE (carboxyfluorescien diacetate, succinimidyl ester) in DMSO (Invitrogen, Life Technologies, Grand Island, USA) was added per ml of cells (final concentration 10 µM CFSE). The cells were then incubated for 10 minutes at 37°C. Cells were spun down and re-suspended in 2 volumes of ice-cold culture medium and incubated on ice for 5 minutes. Cells were re-pelleted and re-suspended in 2 volumes ice cold culture medium (2 times). Washing was done one more time in 2 volumes of pre-warmed medium and cells were finally re-suspended in fresh pre-warmed cell culture medium (described above). A final concentration of 3 µM of CpG-2006 (Hycult Biotech, Plymouth Meeting, USA), 0.25 µg/ml CD40 ligand (eBioscience, San Diego, USA), 5 ng/ml IL-4 (eBioscience, San Diego, USA), 2.5 or 5 µg/ml Anti-human IgG+IgM, (Jackson ImmunoResearch, West Grove, USA) 10 µg/ml of crude parasite extract, and hemozoin (InvivoGen, San Diego, USA) 50 µg/ml were used in the tonsil mononuclear cell stimulation. Cells were harvested on days 3, 5, 7 and 10, stained with CD19 and then B cells were sorted with the MoFlo. Five microgram of human CpG-2006 DNA with a phosphorothionate (PTO) backbone bound to Alexa-488 (Integrated DNA Technologies, Coralville, USA) were mixed with 100 µg/ml of sonicated synthetic hemozoin [32]and co-incubated with rocking for 2 h followed by washing of the complex three times with PBS. Bound and unbound DNA were determined by measuring the DNA concentration in the collected supernatant, using the Nanodrop. BL2 and IM 171 cells were each plated at 1 ml (with approximately 500,000 cells per confocal plate). The CpG-Alexa-488-hemozoin complex was added to the cell lines BL2 and SP-IM 171, and incubated for 2 hours after which confocal microscopy was carried out. Confocal reflection microscopy for detection of hemozoin was combined with fluorescence microscopy to detect the Alexa-488 tagged CpG on a Leica SP2 AOBS confocal laser-scanning microscope as described in detail in [54]. Cells were spun at 1,500 rpm for 5 minutes and the supernatant was aspirated. 5×106 cells were re-suspended in 100 µl of RIPA buffer (Sigma, St. Louis, USA) with freshly added protease inhibitors (Thermo Scientific, Rochford, USA). The sample was pipetted up and down to dislodge cell clumps, and then vortexed vigorously for 15 seconds then incubated on ice for ∼7 minutes, vortexed again for 15 seconds, spun at 13,200 rpm for 16 minutes at 4°C and the supernatant transferred to a new eppendorf. Protein concentrations were determined with either the Bio-Rad protein assay reagent (Bio-Rad, Hercules, USA) or the bicinchoninic acid (BCA) protein assay reagent mix (Thermo Scientific, Rockford, USA), according to manufacturer's instruction. To each protein sample was added 12.5% beta mercaptoethanol and 1× SDS sample buffer (Boston Bioproducts, Ashland, USA). Five microliters of protein standard (Biorad, Hercules, USA) was added to designated well(s). Samples were heated at 95°C for 5 minutes and resolved on 4–20% Tris-Glycine gels (Invitrogen, Life Technologies, Grand Island, USA). Immobilon-P polyvinylidene fluoride (PVDF) transfer membrane (Millipore, Billerica, USA) was used in a semi-dry electrophoretic transfer. The membrane was blocked with either membrane blocking solution (Invitrogen, Life Technologies, Grand Island, USA) or 5% milk in 1x Tris buffered saline with Tween 20 (TBST) (Cell signaling, Danvers, USA) at room temperature for 1 hr. The PVDF membrane was then incubated with primary antibody in appropriate blocking buffer at 4°C overnight. The membrane was washed in TBST or Invitrogen wash solution three times, 10 minutes each and incubated with secondary antibody in 5% milk in TBST for 45 minutes at room temperature. The membrane was then washed with Invitrogen wash solution 4 times, 10 minutes each at room temperature and then enhanced chemiluminescence (ECL) reagent SuperSignal west femto maximum sensitivity substrate (Thermo Fisher scientific, Rockford, USA) was applied to the membrane. Hyblot CL autoradiography film (Denville scientific, Metuchen, USA) was then exposed to the membrane and developed in a Kodak X-MAT 2000 processor. Data are expressed as mean ± SD. Differences between groups were analyzed for statistical significance with Student two-tailed, unpaired t test. Significance was considered achieved when the p value was <0.05.

10.1371/journal.ppat.1003043

A Trade-off between the Fitness Cost of Functional Integrases and Long-term Stability of Integrons

Horizontal gene transfer (HGT) plays a major role in bacterial microevolution as evident from the rapid emergence and spread of antimicrobial drug resistance. Few studies have however addressed the population dynamics of newly imported genetic elements after HGT. Here, we show that newly acquired class-1 integrons from Salmonella enterica serovar Typhimurium and Acinetobacter baumannii, free of associated transposable elements, strongly reduce host fitness in Acinetobacter baylyi. Insertional inactivation of the integron intI1 restored fitness, demonstrating that the observed fitness costs were due to the presence of an active integrase. The biological cost of harboring class-1 integrons was rapidly reduced during serial transfers due to intI1 frameshift mutations leading to inactivated integrases. We use a mathematical model to explore the conditions where integrons with functional integrases are maintained and conclude that environmental fluctuations and episodic selection is necessary for the maintenance of functional integrases. Taken together, the presented data suggest a trade-off between the ability to capture gene cassettes and long-term stability of integrons and provide an explanation for the frequent observation of inactive integron-integrases in bacterial populations.

Horizontal acquisition of mobile and mobilizable genetic elements plays a major role in the development of antimicrobial drug resistance in bacteria. Despite their causal role in drug treatment failure, there is only limited understanding of how horizontal acquisitions of these elements affect bacterial fitness. A prominent group of such genetic elements are the integrons. These genetic elements harbor an integrase-gene that allows the integron to respond to environmental changes by capture and excision of gene cassettes. Here, we have experimentally determined if horizontal acquisition of an integron affect host fitness. The data demonstrate that the initial costs are substantial. However, inactivation of the integrase gene occurred rapidly by spontaneous mutation alleviating the detrimental effect of the integron on bacterial fitness. The same fitness restoring effects was also shown by targeted inactivation of the integrase gene. The inactivation results in a negative trade-off between host adaptation and loss of the ability to capture new gene cassettes. Importantly, our results explain the frequent observation of inactive integrase genes in integrons found in bacteria of different origins. Finally, we use mathematical modeling to determine the conditions necessary for maintaining functional integrases.

Horizontal gene transfer (HGT) enables bacteria to obtain alien genes and genetic elements from prokaryotic, archaeal, and eukaryotic organisms. This capacity for genetic exchange plays an important role in bacterial adaptive evolution, as exemplified by the rapid spread of antibiotic resistance determinants by HGT [1], [2]. Most often, the fitness effects of novel genes in new hosts are selectively neutral or detrimental [3], and prolonged persistence in the population requires compensatory evolution or associated linked selection [4], [5], [6], [7]. Antibiotic resistance determinants are frequently associated with mobile and mobilizable genetic elements, and they tend to reduce host fitness when newly acquired as part of mobile DNA [4], [5], [8], [9]. The magnitude of these fitness costs as well as the mode and speed of compensatory evolution are key parameters determining the frequency of resistance in bacterial populations following relaxed antibiotic selection (i.e. following interventions on drug prescription levels) [10]. From the perspective of horizontal dissemination of antibiotic resistance determinants, population dynamic studies are important to increase our insight on the evolution and reversibility of resistance [10], [11]. Several studies have described compensatory evolution and host adaptation to self-replicating plasmids [for a selection see [4], [5], [8], [9]]. However, only few studies have considered how bacteria adapt to the presence of chromosomally transferred genes and genetic elements. These studies have been limited to chromosomal allelic replacements [6], [12], transposons [13], [14] and a report on conjugative transposons [15]. Integrons are a class of genetic elements frequently involved in antimicrobial resistance dissemination where population dynamic studies are currently absent. These genetic elements have the ability to capture and excise functional gene cassettes involved in host adaptation, often including antibiotic resistance traits [16]. Typically, an integron consists of an integrase gene (intI) encoding a site-specific recombinase responsible for the recruitment and excision of gene cassettes and a promoter (PC) for the expression of captured gene cassettes. Integrases capture gene cassettes through recombination between attI (located downstream of PC) and the gene cassette-borne recombination site attC present in a circular gene cassette. Inverse correlations exist between gene-cassette promoter (PC) strength and integrase activity [17], [18] as well as expression levels [19]. Based on sequence similarity of the intI gene, five classes of “mobile integrons” have been described, for a review see [20]. Class-1 integrons are prevalent in Gram-negative clinical isolates, and harbor gene cassettes encoding resistance to the majority of clinically relevant antibiotics such as aminoglycosides, trimethoprim, and broad-spectrum β-lactams [20], [21]. Structurally, class-1 integrons are relatively diverse, but they generally consist of a 5′-conserved segment (5′-CS) including intI1, attI1, the variable regions where the gene cassettes are embedded, and a 3′-CS that includes a truncated qacE1 and sul1 [22]. Class-1 integrons are frequently linked to complete and incomplete transposons such as Tn402 [23], and Tn21-like structures [24]. Due to the often incomplete nature of the transposable elements linked to clinical class-1 integrons these structures are generally thought to be defective in terms of transposition, and for these elements to move, transposition functions need to be provided in trans. However, in clinical isolates, these integron-containing transposons are frequently located on plasmids and thus can easily spread horizontally [25], [26]. Integrons can be important factors for horizontal dissemination of novel and adaptive traits among bacteria because they facilitate “sampling” of the environmental gene-cassette-pool [27], [28]. Moreover the ability to acquire novel cassettes, or shuffle the existing ones, has shown to be increased as a response to stress [29]. Integrons with non-functional integrases are however prevalent in bacterial populations [28], [30], suggesting that the ability to acquire gene cassettes does not necessarily provide a frequent selective advantage. Thus, whereas it is clear that selection for integron-encoded traits such as antibiotic resistance determine the frequency of class-1 integrons in bacterial populations, the selection for functional integrases remains unclear. Here we show that horizontally transferred class-1 integrons from Salmonella enterica serovar Typhimurium and Acinetobacter baumannii, free of associated transposable elements, strongly reduce host fitness in Acinetobacter baylyi. We demonstrate that these fitness costs are due to an active integrase IntI1. These fitness costs were reduced during serial transfer experiments through mutational inactivation of the integrase gene, suggesting a trade-off between maintaining a functional integrase and stability of integrons in the population over time. Our results provide a rationale for why inactivated integron-integrases are frequently observed in clinical and environmental bacterial isolates. We use a mathematical model to explore the population dynamics of integrons with functional and non-functional integrases in competition with integron-free bacterial populations. We conclude that selection for pre-existing gene-cassettes acts synergistically with the ability to capture new ones [episodic selection [31]] in fluctuating environments. The model organism A. baylyi ADP1 is a close relative to the nosocomial pathogen A. baumannii and is free of integrons [32]. We constructed a set of A. baylyi strains containing cloned diverse class-1 integrons from isolates of two A. baumannii (clinical isolates) and one S. enterica serovar Typhimurium (isolated from pork). These strains allowed the investigation of the effects of newly acquired integrons on host fitness. The three integrons were inserted in an identical chromosomal locus (ACIAD3309) [32]. Mixed culture competition experiments revealed that newly acquired class-1 integrons from A. baumannii (IVS1 and IVS3) and S. enterica serovar Typhimurium (IVS2) resulted in a statistically significant reduced relative fitness (w) of 0.93 (p = 0.01**), 0.92 (p = 0.02**), and 0.89 (p<0.01**), respectively. The relative fitness of the ancestor was by default set to 1.0. The neutrality of the insertion locus (ACIAD3309) was confirmed using a pair of A. baylyi ADP1 strains that were isogenic except from the insertion of a selective/counter-selective marker pair in strain IVS4 (A. baylyi ADP1 ACIAD3309::nptII sacB) (relative fitness w = 1.01, not significantly different from 1.0 (p = 0.2)). The results are summarized in Figure 1. To verify that the relative fitness measurements were not hampered by the choice of selective antibiotic resistance markers all fitness measurements presented in this study were repeated with strain IVS4 as an integron-free competitor. The results from these experiments using sucrose (sacB) counter-selection were always congruent with kanamycin, spectinomycin, and spectinomycin/ampicillin selective platings, and the results from all parallel competition experiments were pooled before statistical analyzes. The class-1 integrons inserted into A. baylyi ADP1 differed in their gene cassette promoter sequences, located in the intI1 open reading frame. Sequence alignments of the three intI1 sequences inserted into A. baylyi revealed that the integrons with the highest (IVS2), and lowest (IVS1) fitness costs contained cassette promoters identical to the recently described weak (PCW) and strong (PCS) promoters, respectively [17]. The difference in relative fitness between strains IVS1 and IVS2 was statistically significant in independent sample t-tests (p = 0.03*), suggesting a correlation between integrase activity and the fitness cost of harboring an integron. The integron with the intermediate fitness cost (strain IVS3, w = 0.92±0.04) contained a hybrid PC promoter. In three integron-containing A. baylyi strains (IVS1, IVS2, and IVS3), the intI1 integrase genes were inactivated by insertions of either cat (strain IVS1 intI1::cat) or nptII sacB cassettes (strains IVS2 intI1::nptII sacB and IVS3 intI1::nptII sacB). These intI1 knockout mutants displayed no significant reduction in relative fitness in mixed competition experiments with the ancestral A. baylyi ADP1 (Figure 1). To test the hypothesis that strains with inactivated integrases increased fitness when compared to their functional counterparts, independent sample t-tests were performed. For all pairs, the intI1 inactivation restored fitness completely: IVS1 vs. IVS1 intI1::cat (p = 0.015**), IVS2 vs. IVS2 intI1::nptII sacB (p<0,001**), and IVS3 vs. IVS3 intI1::nptII sacB (p = 0,003**). These data further demonstrate that the initial fitness cost of integron-carriage was due to the presence of an active integrase. Expression of the integrase genes in the chromosomal insertion locus was verified by reverse transcription PCRs (RT-PCR) in IVS1, IVS2, and IVS3. No transcripts were detected in strains IVS1 intI1::cat, IVS2 intI1::nptII sacB, and IVS3 intI1::nptII sacB (Figure S1). A total of 20 A. baylyi IVS1 cultures were subjected to daily 1∶100 dilutions in fresh LB medium. During the serial transfer experiments the evolving populations were screened for colonies of increased size on LB agar plates, a method regularly used to identify fitness compensated mutants [33], [34]. Twice a week agar plates were visually inspected and the first colony of increased size appeared after 30 days in one of the populations. This colony was isolated and frozen down for further analyses. At day 42 we isolated two additional colonies from different populations. These isolates were analyzed and they all contained mutations in the intI1. Complete integrons from these three evolved A. baylyi IVS1 genetic backgrounds were transferred back into the ancestral A. baylyi ADP1 strain, yielding strains IVS1EV-1, IVS1EV-2, and IVS1EV-3. To test the hypothesis that the evolved integrons increased fitness, they were competed against the ancestral ADP1. Mixed culture competition experiments revealed that fitness was completely restored in these strains (Figure 1). Independent sample t-tests further verified that the relative fitness of the each evolved integron was significantly different from its intI1-functional ancestor IVS1, (p = 0,001**, for all three comparisons). Subsequent characterizations of these three transformants by DNA sequencing revealed frameshift mutations close to the start codon of the intI1 open reading frame rendering the integrase non-functional (Figure S2). RT-PCR of evolved strain IVS1EV-1 yielded no transcript (Figure S1). We hypothesized that functional integrases are maintained by episodic selection provided by fluctuating environments [31]. To test this hypothesis in silico we parameterized a mathematical model with our own experimental data, and relevant parameters from the literature. Parameters related to resource utilization (e and km) were calibrated to yield population sizes close to what we observed in the laboratory. The MIC values were based on our own experiments, parameters on growth characteristics were derived from our own study (fitness values) combined with values from the literature. For a complete list of parameters used in these serial transfer simulations, see Table 1. Fig. 2A shows the predicted population dynamics of strains harboring a newly acquired integron with a functional integrase with one (I1 – blue line) and two (I2 – black line) gene cassettes, the integron free susceptible wild type (P – green line), and two fitness ameliorated integrase- mutants (M1 - light blue and M2 - grey). The predicted in silico population dynamics, before “shift” in Figure 2A, mirrors our experimental data form the serial transfer cultures where integrase specific fitness compensating mutants were isolated after 30 and 42 days of serial transfers. These single mutants were selected on antibiotic free agar plates with approximately 100 colonies, suggesting an approximate frequency of 1/100. Fluctuating environments are simulated by a probability of encountering antibiotic A for a period of 40 transfers, and then antibiotic B for the remaining time period, both at a 10% probability per transfer. The results shown in Fig. 2A are the median values for 100 simulations. Our simulations show that functional integrases are descending when only one antibiotic is present. However, the switch to a second antibiotic B allows the pre-existing two-gene cassette integron (I2) to rapidly ascend to high frequency. During this ascent I1 and M1 are driven extinct. Without further environmental change, the mutated integrase M2 outcompetes its less fit counterpart I2. As shown in Fig. 2B, persistence of integrons with functional integrases strongly depends on when the switch to antibiotic B occurs. To assess the robustness of the model predictions scenario in Fig. 2A we explored different parameter ranges for the gene cassette acquisition rate (λ), mutation rate for inactivated integrase (π), and the mutation rate for restoration of functional integrase (θ). We performed 500 additional simulations where π and θ were varied over 10 values each, and λ over 5 levels (ranges provided in Table 1). As illustrated in Figure S3 the model predictions were robust for a wide range of these parameter combinations. Further, we explored the extreme values of the 95% CI of the relative fitness parameter V as experimentally determined (w = 0.91 and 0.95) alone and in combinations with different parameter values. These values and the mean fitness value (w = 0.93) for VI were tested when π, θ, and λ varied over a small range (±.2.5%) to assess changes in model predictions. Qualitatively all additional simulations (n = 581) were consistent with the scenario presented in Figure 2A providing generality to the model predictions (data not shown). We show for the first time that newly acquired integrons can substantially reduce relative fitness of its new bacterial host. Following the insertion in a selectively neutral chromosomal locus, the three class-1 integrons from isolates of A. baumannii and S. enterica serovar Typhimurium reduced fitness in the A. baylyi recipient by 7–11%. For comparison, these fitness costs are in the range of mutations conferring antimicrobial resistance through modifications of housekeeping genes such as par/gyr mutations (fluoroquinolone resistance) in Streptococcus pneumoniae [35], and some rpoB mutations in E. coli [36]. Direct insertional inactivation of the three intI1 alleles completely mitigated the initial fitness reductions, clearly suggesting that the fitness costs observed were due to the presence of a functional integrase gene (intI1). Non-functional integrase genes due to frameshift- and nonsense-mutations are frequently encountered in surveys [28], [30], [37]. We asked whether functional intI1 genes would be inactivated during experimental evolution. After 30–42 days of daily serial transfers we observed colonies of increased size on agar plates, representing putative fitness compensated mutants. Integrons from evolved isolates were subsequently introduced into the ancestral genetic A. baylyi ADP1 background, and in these strains they no longer reduced fitness of the host bacterium (Figure 1). Sequence analyses of the three intI1 genes revealed the presence of frameshift mutations in the first quarter of intI1 resulting in premature stop codons, rendering these integrases inactive (Figure S2). The emergence of non-functional intI1 genes during experimental evolution with mutational inactivation patterns identical to those reported from bacterial isolates of environmental and clinical origins [28], [30], [37] strongly suggests that integrase pseudogenes may ascend to high frequencies in bacterial populations by natural selection. It was recently demonstrated that intI expression is under the control of the SOS response through the presence of LexA binding sites in the integrase promoters (including class-1 intI1) [29], [37]. These authors proposed that LexA repression reduce the potential detrimental effects of intI expression, and that SOS induction allows expression of the integrase gene when new gene cassettes could provide a response to stressful and potentially lethal environmental conditions [29], [37]. It was also suggested that integrase inactivation is correlated with absence of LexA regulation [37], and that this is a key factor explaining the high proportion of pseudo-intI-genes found in integron-containing bacteria [28], [30], [37]. The experimental data reported here are the first to support both these hypotheses. The majority of Acinetobacter species, including our model organism A. baylyi and the clinically relevant A. baumannii all lack lexA homologues [38], [39]. Thus, intI1 is most likely not under LexA repression in our model system, and the newly acquired integrons reduced fitness in A. baylyi, despite the presence of native LexA binding sites in two out of three integrons. The mutational inactivation of intI1 completely mitigated the fitness costs of integron carriage, and in the absence of repression the inactivation could very well mimic tight repression of integrase expression. The serial transfer experiments were performed in nutrient-rich LB medium, as opposed to minimal medium for the competition experiments. The emergence of fitness compensated A. baylyi with non-functional integrases during experimental evolution strongly suggests that the fitness costs of integron carriage are not limited to specific growth conditions. Consequently, the fitness restoration due to intI1 inactivation leads to stabilization of the cassette arrays in the bacterial population, and integron-borne antibiotic resistance determinants will not be reduced following relaxed selective antibiotic pressures. Previous reports indicate an inverse correlation between gene-cassette promoter (PC) strength and integrase activity [17], [18] as well as expression levels [19]. From the results presented in these reports it could be hypothesized that a strong gene-cassette promoter would decrease the overall activity of the integrase gene, and that the cost of integron carriage should be reduced. Our results favor this hypothesis. However, the results should be interpreted with some caution since we achieved significance at the alpha level, but not when Bonferroni correction was applied. The newly acquired integron from a clinical A. baumannii strain (∼7% fitness cost) contained a cassette promoter sequence identical to the “strong promoter” (PCS) whereas the cassette promoter of the integron from the S. enterica serovar Typhimurium strain (∼11% fitness cost) displayed a “weak promoter” (PCW), as reported by Jove et al. [17]. Moreover, the integrase sequence from S. enterica serovar Typhimurium revealed amino acids in positions 32 (R), and 39 (H) consistent with the highest recombination activity demonstrated in [17]. Jove and co-workers suggested that increased expression of gene cassettes, leading to higher levels of resistance, would be selected for in environments with strong antibiotic selective pressures. Our results add complexity to that hypothesis insofar that the increased expression of gene-cassettes also could lead to reduced integrase activity, and thus stabilize functional integrons in non-selective environments. Two lines of evidence support that the mechanistic basis for the observed fitness effects of functional integrases is reduced genomic stability. First, IntI1 can catalyze recombination events between attI/attC sites and frequently encountered non-canonical sites in the genome, as demonstrated by Recchia and co-workers [40]. Secondly, purified IntI1 enzyme possesses all functions necessary for target recognition and recombination, as shown in in vitro strand transfer assays [41], [42]. Consequently, when newly acquired and in the absence of tight regulation, expressed integrase would be able to form recombination junctions between the integron and sequence-regions elsewhere in the genome. Resolution of such single strand crossovers ultimately leads to potentially lethal deletions of the genomic region between the recombination sites either following replication or IntI1 activity, as demonstrated in co-integrate resolution experiments [40]. We hypothesized that environmental fluctuations and episodic selection [31] are key to the maintenance of functional integrases, and explored this in computer simulations. According to our hypothesis selection for pre-existing gene cassettes in integrons (type-1 episodes) acts synergistically with the ability to capture new cassettes that provide bacteria with a selective benefit in changing environments (type-2 episodes). Type-1 episodes favoring pre-existing gene cassettes allow integrons to reach high frequencies in the population but during these conditions, due to the fitness cost of the active integrases, non-functional integrases rapidly ascend in the population. Type-2 episodes select for new gene cassettes acquired by the active integrase. Our simulations show that maintenance of functional integrases depends on the time between the different episodes (i.e. the frequency of environmental change), as well as the continuous availability of new and adequate gene cassettes. Of course the selective episodes could be other favorable traits encoded by gene-cassettes, and are not limited to antibiotic resistance determinants. In conclusion, the presented data suggest that in the absence of intI1 repression, a fitness trade-off exists for the maintenance of integrons with functional integrases. The initial high fitness cost of the integrase can only be outweighed by selection for gene cassette dynamics. The bacterial strains, plasmids, and primers used in this study are listed in Tables 2, S1 in Text S1, and S2 in Text S1. Strains were grown in S2-minimal medium, amended with 2% lactate [43], or Luria Bertani (LB) agar or broth at 30°C or 37°C under aeration. Plasmid pTM4 is derived from the pGT41 [44] and was used for in vitro insertion of integrons into a chromosomal locus. pTM4 contains segments identical to upstream and downstream segments of the 5′-region of the chromosomal A. baylyi ACIAD3309 open reading frame for homologous recombination, interrupted by a SacI/Ecl136II restriction site, and was constructed as follows: The downstream segment (707 bp) was PCR-amplified with primers ACIAD3309-down-F (including a 5′-heterologous tail containing an Ecl136II/SacI site) and ACIAD3309-down-R (Table S2, in Text S1) with Phusion DNA polymerase (Finnzymes, Espoo, Finland) according to the manufacturer's instructions but with 10% DMSO added, and inserted into the KspAI site of pGT41, giving pTM1. The upstream segment (785 bp) was amplified with primers ACIAD3309-up-f and ACIAD3309-up-r (with 5′-Ecl136II/SacI tail) and inserted into the OliI site of pTM1, giving pTM2. From pTM2, two unwanted segments were removed as follows: A 2.7 kbp insert containing an nptII (kanamycin resistance) gene was excised by cleavage with SacI (has 2 sites in pTM2) and re-circularization of the large fragment, resulting in pTM3 which has the two segments for homologous recombination ligated immediately upstream and downstream of an Ecl136II/SacI restriction site. From pTM3, the bla (ampicillin resistance) gene was truncated and rendered non-functional by cleavage with XmnI (contains 2 sites in pTM3) and re-circularization of the large fragment, giving pTM4. A. baylyi IVS1 was constructed as follows: The integron of A. baumannii Ab064 (Table 2) including the 5′- and 3′-CS was PCR-amplified with Phusion polymerase using 5′-phosphorylated primers IntF2 and OrfRev3 (Table S2, in Text S1) and ligated to Ecl136II-cleaved (blunt-ended linear) pTM4, respectively. The ligation assay was used as donor DNA to naturally transform (see below) A. baylyi ADP1. Transformants were selected on medium containing kanamycin (25 µg/ml). One transformant was generated from a PCR product covalently joined to a vector molecule at both ends and that substituted the 5′-end of ACIAD3309 with the integron from A. baumannii Ab064, and termed IVS1. Co-integrates were excluded by screening for chloramphenicol sensibility, and the desired insertion was verified by PCR. The strains IVS2 and IVS3 were constructed as described for IVS1 with integrons of S. enterica serovar Typhimurium 490 and A. baumannii 47-42 (Table 2), respectively, using primers IntF2/OrfRev2 and employing corresponding selection and PCR controls. The three class-1 integrons differed in the variable regions (Table 2) as well as in the integrase sequences (different gene cassette promoters and SNPs). The integrase accession numbers are JX041889 (A. baumannii Ab064), AM991977 (S. enterica serovar Typhimurium 490), and JX259274 (A. baumannii 47-42). Strain IVS4 (locus neutrality control) was obtained by transformation of A. baylyi ADP1 by pTM2 (kanamycin-resistant, sucrose-sensitive, verified by PCR). The intI1 gene of IVS1 was disrupted by natural transformation with HincII-linearized pACYC177-int-cat as substrate for natural transformation (Table S1, in Text S1). This plasmid contains an internal fragment of the intI1 gene of A. baumannii AB064 with a cat (chloramphenicol resistance) gene inserted. The resulting strain was PCR-verified and termed IVS1 intI1::cat. The intI1 genes of IVS2 and IVS3 were insertion-inactivated in a corresponding manner by pACYC177-int-nptII-sacB, which contains the nptII sacB marker pair (kanamycin resistance/sucrose susceptibility) from pTM2 (cloned as Ecl136II fragment) instead of cat [45], [46]. The resulting strains were verified phenotypically, and by PCR and termed IVS2 intI1::nptII sacB and IVS3 intI1::nptII sacB, respectively. Strain IVS1 with a class-1 integron from A. baumannii Ab064 was subjected to daily one hundred-fold dilutions in 10 ml LB broth in 20 independent parallels for 30–42 days. Aliquots were plated every third day on LB agar plates to screen for fitness-compensated mutants by increased colony size. Evolved integrons were transferred back into the ancestral A. baylyi ADP1 background by PCR-amplification including surrounding regions of homology using homologous transformation (yielding strains IVS1EV-1, IVS1EV-2, and IVS1EV-3) (Table 2). Integron-containing and -free A. baylyi ADP-1, otherwise isogenic, were subjected to mixed competition experiments as previously described [12], [47] with the following modifications: Competing strains were pre-grown in S2 minimal media for 24 hours before diluting 1∶10 in NaCl (0,9%), and 150 µl of each competitor was transferred and mixed into 2.7 ml S2 medium supplied with 0.1% DNase (to exclude natural transformation in the assays). Initial (N0) and final densities (N24) of competing strains were measured before the onset of competitions and after 24 hours by selective and non-selective plating. Selective traits exploited were antibiotic resistance markers or a counter-selective marker (nptII or aadB, kanamycin resistance; aadA, spectinomycin resistance; blaOXA-30 ampicillin resistance; sacB, sucrose susceptibility). From these densities, the Malthusian parameter (m) of each competitor was determined using the equation m = ln (N24/N0). Relative fitness (w) was estimated as the ratios of each competitor's Malthusian parameter (m1/m2) [47]. To avoid potential marker-bias m1 and m2 were estimated by selective plating on antibiotics (kanamycin/spectinomycin/ampicillin) in one genetic background followed by sucrose selection in the other. Results were always congruent for the antibiotics and concentrations chosen, and data from both selective regimes were pooled. Estimates of w were based on 12–24 parallel experiments for each competition experiment. Preparation of competent cells and transformation assays were performed as described previously [12], [48] with some modifications. Briefly, competent cells were prepared by diluting an overnight culture of A. baylyi 1∶100 in fresh LB. The culture was incubated at 30°C with vigorous shaking until the cell titer reached 1×109 ml−1. The cells were chilled on ice, pelleted by centrifugation at 5000×g and 4°C for 15 min, and re-suspended in LB supplemented with 20% glycerol. Aliquots were stored at −80°C until use. For transformation, competent cells were thawed on ice and diluted 1∶40 in LB medium containing the donor DNA. The assays were aerated for 90 min at 30°C and plated on selective media plates in appropriate dilutions. The plates were incubated at 30°C until visible colonies had formed (16–40 hours). The minimal inhibitory concentrations (MICs) of the donor, recipient and transformant strains were determined for sulfamethoxazole, kanamycin, streptomycin, spectinomycin, gentamicin, and ampicillin, by E-test according to the instructions of the manufacturer (BioMeriux, France). Nucleic acids were isolated with QIAGEN Genomic/Plasmid DNA kits (QIAGEN, Germany), according to the manufacturer's instructions. The transformation assay using A. baumannii 064, S. enterica serovar Typhimurium 490 and A. baumannii 47-42 as donors, resulted in a number of transformants that were analyzed phenotypically (MIC values, Table S3, in Text S1) and genotypically. Primers IntF2/OrfRev3 and IntF2/OrfRev2 were used to amplify the entire integron region in both transformants and donor strains, giving approximate sizes of 4 kb, 5 kb, and 6 kb for A. baumannii 064, S. thyphimurium 490 and A. baumannii 47-42 transformant strains, respectively. Primers 5CS′/3CS′ were used to verify the size of the variable regions in both donor and test strains; primers UpF/DownR as well as IntF2/3CS′ and 5CS′/OrfRev2/OrfRev3 were used to confirm the correct position of the aquired integrons in the ADP1 genome. Primers IntF2/OXA303R and IntF2/aacC1-OrfP-R were used to verify the position of the gene cassettes within an integron in the strains IVS2 and IVS3, respectively. Primers aadBF/aadBR, OXA305F/OXA303R and aacC1-F2/aacC1-orfP-R (Table S2, in Text S1) were used for gene cassettes identification within the integrons. The unknown regions surrounding the integron in the donor were sequence determined by direct genomic DNA sequencing (primer walking) as described previously [12] with the following modifications: 20 µl sequencing reactions consisted of 4 µl BigDye v3.1 sequencing mix (Applied Biosystems), 4 µl of the primer at a concentration 10 mM, 4 µl of a sequencing buffer, and ∼4 µg of the purified chromosomal DNA. The sequences of the integrons in the donor strain and transformants were determined by sequencing (BigDye Chemistry) of the PCR products obtained from the primers IntF2/OrfRev3, or IntF2/OrfRev2 (Table S2, in Text S1). The sequence of the integrase gene was determined by sequencing of the PCR products amplified with the primers IntF2/aadBR, IntF2/GCS1RevComp, and INCINTF/IntI1F. PCR products were purified by adding a mix of exonuclease 1 (0.2 U/µl PCR product) (New England Biolabs) and shrimp alkaline phosphatase (0.01 U/µl PCR product) (Roche) followed by 30 minutes incubation at 37°C and 5 minutes at 95°C in a PCR machine. The obtained sequences were analysed by the Sequencher v.4.2.2 programme (GeneCodes, USA) and compared to previously published sequences (GenBank). RNA was isolated using the Total RNA Isolation KIT (Macherey-Nagel, Germany), and cDNA was synthesized using MonsterScript 1st-strand cDNA synthesis Kit (Epicentre Biotechnologies, USA), both according to the manufacturer's instructions. The generated cDNA was amplified using primers INCINTF/IntI1F (Table S2, in Text S1). To investigate the conditions that favor maintenance of integrons in bacterial populations, we used a mathematical model and numerical solutions, based on [31] [49]. This serial passage model included five populations. Populations I1 and I2, contain functional integrases where I1 has captured a single cassette encoding resistance to antibiotic A, I2 has captured two gene cassettes and is resistant to both antibiotics A, and B. Populations I1 and I2 can acquire frameshift mutations in intI1 and turn into populations M1 and M2 with non-functional integrases, respectively. Population P is the antibiotic susceptible, integron-free wild type. The growth rates of I, M, and P populations are determined by the pharmaco-dynamic function developed by Regoes and co-workers [49], where a Hill-function determines the growth rate or death rate (negative growth rate) of the populations in the presence of antibiotics [50], [51]. Briefly, the growth rate depends on the concentration of resource (R), antibiotics (A and B), and antibiotic susceptibility (MIC). In this model each population have two different growth rates; and . In the simulations is chosen if antibiotic A is present, and when B is present, such that . Thus, the model does not simulate events where both antibiotics are present. With these definitions the changes in the population densities during one serial transfer event of I, M, and P populations are given by the following equations:where e µg/ml is the conversion efficiency (the resource concentration necessary to produce one new cell) [52], da and db are the decay rates of the antibiotics, π is the mutation rate for generating defective integrases, and θ is the mutation rate for restoring functionality of defective integrases. λ is the rate at which populations with functional integrases acquire gene cassettes. An illustration of the model with respect to π, θ, λ is given in Figure S4. A list of parameter values is given in Table 2. Following each simulated dilution (1∶100) 50 µg/ml of the resource was added, and the introductions of antibiotics were stochastic events. Each transfer was assigned a random value (range 0 to 1) from a uniform distribution of numbers. When this value was above a defined probability of 10%, antibiotics were added at 2× (antibiotic A) and 10× (antibiotic B) the MIC concentration of the susceptible populations in order to ensure proper selective effects of the added antibiotics. To investigate the temporal effect of fluctuating environments on the population dynamics of integron containing populations the temporal switch from antibiotic A to B was set at days 20, 40, 50, 60, 70, 75, 80, 100. A total of 100 simulations were performed at each frequency. To qualitatively test the robustness of the model predictions 500 additional simulations were run for different combinations of θ (10), π (10), and λ (5) within the ranges provided in Table 2. We also tested the model behavior where the parameters θ, π and λ were combined with a small variation around the original selected model parameter (±2.5%) for three levels of the relative fitness of integron carriage parameter (Vx). These levels of Vx included the extreme values from the 95% confidence intervals provided in the experimental measurements. For a numerical solution of the differential equations and to simulate the experimental conditions, the open source computer program R version 2.14.1 was used [53]. Dilutions as well as introduction of resource and antibiotics were determined by the events argument in the lsoda function from the deSolve package version 1.10-3 [54]. We assume that gene cassettes are available for the populations with functional integrase. Further, the resistance genes are assumed to be selectively neutral, as supported by the experiments conducted in this study. We model the use of two antibiotics to show the principle of a heterogeneous environment and the antibiotics are assumed to have no interactions. In these simulations a cut off was set at 1 CFU per ml where all growth and interactions were stopped. All populations were diluted 1∶100 every 24 hours. For simplicity, gene cassette reshuffling (the order of resistance genes) or loss of single gene cassettes was not considered. For each set of environmental variables the median population densities from 100 simulations were calculated for each time point and the logarithm of the densities plotted at 24-hour intervals until I2 population reaches 1 CFU/ml. Parameter estimation and statistical tests were performed in SPSS vs. 17. In addition to significance at the alpha level (0.05*), multiple testing issues were addressed by Bonferroni corrections of significance levels (indicated as ** throughout the text). IntI1 from A. baumannii Ab064: JX041889. IntI1 from S. enterica serovar Typhimurium 490: AM991977. IntI1 (partial) A. baumannii 47-42: JX259274.

10.1371/journal.pcbi.1004669

Brain Connectivity Dissociates Responsiveness from Drug Exposure during Propofol-Induced Transitions of Consciousness

Accurately measuring the neural correlates of consciousness is a grand challenge for neuroscience. Despite theoretical advances, developing reliable brain measures to track the loss of reportable consciousness during sedation is hampered by significant individual variability in susceptibility to anaesthetics. We addressed this challenge using high-density electroencephalography to characterise changes in brain networks during propofol sedation. Assessments of spectral connectivity networks before, during and after sedation were combined with measurements of behavioural responsiveness and drug concentrations in blood. Strikingly, we found that participants who had weaker alpha band networks at baseline were more likely to become unresponsive during sedation, despite registering similar levels of drug in blood. In contrast, phase-amplitude coupling between slow and alpha oscillations correlated with drug concentrations in blood. Our findings highlight novel markers that prognosticate individual differences in susceptibility to propofol and track drug exposure. These advances could inform accurate drug titration and brain state monitoring during anaesthesia.

Though scientific understanding of how brain networks generate consciousness has seen rapid advances in recent years, application of this knowledge to accurately track transitions to unconsciousness during general anaesthesia has proven difficult due to considerable variability in this gradual process across individuals. Using high-density electroencephalography, we studied changes in these networks as healthy adults were sedated using propofol. By measuring their behavioural responsiveness and amount of sedative in their blood, we found a striking pattern: the strength of their brain networks before sedation predicted why some participants lost consciousness while others did not, despite registering similar blood levels of drug. By uncovering underlying signatures of this variability, our findings could enable accurate brain monitoring during anaesthesia and minimise intra-operative awareness.

Understanding how the human brain reversibly generates and loses consciousness, through complex interactions of neural activity at multiple spatial and temporal scales, is a grand challenge for modern neuroscience. Recent theoretical advances have argued that consciousness changes when the balance between integrated and differentiated neural activity is affected [1–4]. However, accurately tracking these changes in brain dynamics remains a key research challenge with potentially wide-ranging applications, and is complicated by the significant individual variability in the trajectory along which consciousness is lost and regained. The process of reversibly inducing unconsciousness using anaesthetic drugs like propofol is commonplace in clinical medicine [5]. However, tracking brain activity to accurately assess the depth of anaesthesia in an individual is currently not a universal component of clinical practice. Indeed, surface electroencephalography (EEG) is relatively easy to measure from the scalp and has long been known to index changes in brain dynamics induced by anaesthetic action [6], but it is still not universally used in the clinical setting. This is despite the fact that intraoperative awareness during surgery continues to result in pain and distress [7], highlighting the need for reliable depth of anaesthesia monitoring in the operating room. The absence of ubiquitous brain monitoring during general anaesthesia is, in part, due to the lack of robust EEG markers derived from current advances in neuroscience [8–12], which can accurately track the loss and reestablishment of reportable consciousness. Monitoring of brain states is currently limited to proprietary systems with mixed results [13–15]. Crucially, one reason for this is the considerable individual variability in susceptibility to anaesthetic dosage [16], which adversely affects the accuracy of these systems [17]. To better understand the factors underlying this variability, we combined the measurement of high-density resting state EEG from healthy volunteers sedated with propofol with measurement of drug concentrations in blood, in addition to objective assessment of behavioural responsiveness. With this aim in mind, we administered propofol at dosages expressly aimed at engendering varying degrees of mild to moderate sedation across our participant group, rather than complete unconsciousness in all of them. Employing modern functional EEG tools to assess spectral power and connectivity, we identified key changes in brain networks using graph-theoretic tools, and linked these changes to individual variability in drug concentrations and loss of behavioural acuity during sedation. Drawing upon previous research [18–21], we hypothesised characteristic impairments in the strength and topography of EEG power and connectivity, especially manifesting in the slow and alpha frequency bands, alongside administration of propofol. In addition to confirming these hypotheses, our findings highlight valuable EEG-derived signatures that can not only track the actual amount of propofol in blood, but also predict loss of responsiveness even before any drug is administered. These findings contribute to the current interest in identifying consistent markers of the loss and recovery of consciousness during propofol sedation. In the clinical context, these findings could lead to more accurate drug titration and brain state monitoring during anaesthesia. The behavioural changes accompanying the administration of progressively increasing amounts of propofol (Fig 1A) are shown in Fig 1B, which plots the hit rate of participants as a function of the level of sedation. Based on binomial modelling of their hit rates (see Materials and Methods), we identified a subgroup of 7 participants who became behaviourally impaired at this simple task during moderate sedation; 13 others remained responsive throughout, though their reaction times were impaired during sedation (Fig 1C). We designate these two groups as drowsy (green triangles) and responsive (blue triangles) in the following descriptions. As expected, we found a highly significant interaction between group and sedation level in hit rates (Fig 1B; F(3) = 38.4, p = 9e-09). Further, in the responsive group, there was a significant effect of sedation on reaction times (Fig 1C; F(2) = 14.6, p = 0.0002). In comparison to the relative distinction between the two groups in their hit rates, there was considerably more overlap in drug concentrations measured in blood plasma (Fig 1D). We found a relatively weaker interaction between group and level of sedation in drug concentrations: F(2) = 4.7, p = 0.0242, and the difference between drug concentrations in the two groups reached significance only during moderate sedation (p = 0.0181). This finding points to the well-studied inter-individual variability in pharmacodynamic impact of propofol [16, 17], and motivates the development of more accurate signatures of responsiveness that can be measured passively and non-invasively during propofol sedation. Connectivity between EEG channels was assessed to directly investigate the impact of propofol on the structure of brain networks of oscillatory neural interactions, using the debiased weighted Phase Lag Index (dwPLI, see Fig 2 and [22]). Here, we define brain networks as the characteristic patterns of scalp-level connectivity observable in human EEG at different frequencies, generated by underlying cortical networks [23] with firing rates oscillating at their natural frequencies [24]. We employed the dwPLI connectivity matrices in each band to construct such EEG-derived brain networks, and used graph-theoretic algorithms to quantitatively compare their topological properties. By representing the EEG channels as nodes of a network and the strength of dwPLI between them as weighted, undirected links between them, we calculated four measures that captured micro-scale (clustering coefficient), meso-scale (modularity and participation coefficient) and macro-scale properties (characteristic path length) of each participant’s network at each level of sedation (see bottom right panel of Fig 2 for a visual description of these properties). Importantly, these metrics were chosen a priori to summarise key network properties that we expected to be modulated during propofol sedation. In the alpha band, median dwPLI across all channel pairs was significantly more reduced in the drowsy group during mild (p = 0.003) and moderate sedation (p = 0.01). Further, the clustering coefficient [25, 26], which measures local efficiency, was significantly lower (Fig 3A) in the frontal alpha networks of the drowsy group during mild (p = 0.007) and moderate sedation (p = 0.04). Furthermore, within the responsive group, clustering during moderate sedation tended to decrease linearly alongside increasing reaction times (Fig 3B), though this effect only approached significance. Conversely, characteristic path length (Fig 3C), the inverse of global efficiency, was significantly higher during mild (p = 0.0004) and moderate sedation (p = 0.0035), and tended to increase with slower reaction times among responsive participants (Fig 3D). Taken together, small-worldness, a combined measure of a network’s local and global efficiency (calculated as the ratio of clustering to path length [26, 27]), was significantly reduced in the drowsy group during mild (p = 0.005) and moderate sedation (p = 0.03). At the meso-scale, these drowsy alpha networks were also more modular at moderate sedation (Fig 3E, p = 0.02), and hence more separable into relatively disconnected topological modules [28]. Crucially, these modules lacked hub nodes that connected them into an integrated network, as evidenced by statistically lower standard deviation (p = 0.002) of participation coefficients [29] in the drowsy group (Fig 3F). Together, these network differences demonstrated that the frontal alpha connectivity in the drowsy group did not have the network capacity of the occipital alpha network commonly observed in human resting EEG during wakefulness. These changes in alpha networks can be understood more visually with Fig 4A. At baseline, both groups had prominent frontocentral and occipital modules of strong connectivity. While these modules persisted through moderate sedation in the responsive group, the structure of connectivity networks in the drowsy group shifted to qualitatively distinct state comprising of coherent, frontally centered oscillations that manifested as a frontal module (Fig 4B), before reverting back to the typical pattern of baseline connectivity during recovery. On the whole, this shift in alpha connectivity mirrors the frontal shift in alpha power (Fig 5) commonly observed during propofol sedation [18, 19, 30–32]. In contrast to these changes in alpha networks, no differences were observed between delta networks in the two groups (see S1 Fig). Spectral connectivity in the alpha band identified a prospectively valuable determinant of the variability in susceptibility to propofol seen in the behavioural data. During the baseline period before sedation, though there were no differences in the topography or relative strength of alpha power between the responsive and drowsy groups (Fig 5A and 5B), there were significant differences in median dwPLI (p = 0.0085) and key network properties that captured the topological structure of connectivity in the alpha band. Specifically, alpha networks in the drowsy group were already less clustered (Fig 3A; p = 0.04) and less small-worldy (p = 0.0187) at baseline. They were also more modular (Fig 3E; p = 0.04), and had fewer hubs (Fig 3F; p = 0.0018). Remarkably, these baseline alpha network differences were evident when the two groups of participants were indistinguishable, both in terms of behavioural hit rates (Fig 1B) and occipital alpha power (Fig 5B). Furthermore, this predictive value of brain connectivity was unique and specific to the alpha band, and not evident in other frequency bands (see S1 Fig). In line with previous findings [31, 33, 34], sedation selectively increased beta/gamma power and connectivity among responsive participants, but baseline power or connectivity in these bands was not significantly different between the two groups. To explicate this result further, Fig 6A depicts a scatter plot of alpha network small-worldness in each participant measured during pre-drug baseline, against their consequent behavioural hit rates and drug concentrations measured during moderate sedation. Though there was considerable variability in small-worldness across the responsive group at baseline, the drowsy group already had relatively lower small-worldness in comparison. To directly test whether participants who already had less robust brain networks at baseline later became drowsy or unresponsive during moderate sedation, Fig 6B plots the individual hit rate trajectories of the participants separated based on whether their baseline small-worldness was above or below the median. Those in the group with high baseline small-worldness remained responsive, and had significantly higher hit rates during moderate sedation (Fig 6B, inset; p = 0.0093). This predictive role of alpha brain networks in characterising individual variability in susceptibility to propofol is exemplified in Fig 6C, which depicts their evolution in two ‘drug concentration-matched’ participants. Despite registering relatively similar drug concentrations at moderate sedation, one of them remained responsive while the other became completely unresponsive. As is evident, the latter participant already had a comparatively less robust alpha network already at baseline, which then evolved into a frontally alpha module at moderate sedation. In comparison, the responsive participant had a relatively more small-worldy, less modular network at baseline, which was sustained during moderate sedation. These differences potentially explain why the drowsy group, whose alpha networks were already compromised to some degree, became behaviourally impaired while the responsive group did not, despite both groups registering overlapping levels of propofol as measured in their blood at moderate sedation. It is important to note that these differences observed in the baseline alpha networks were abolished at recovery (see Fig 3A and 3C). This suggested that these differences between the two groups were essentially dependent on the latent alpha network state of the participants at the beginning of the data collection rather than any individual trait, and were ‘reset’ after the washout of the drug. We found that, at baseline, participants in both responsive and drowsy groups had similar temporal coupling between the phase of slow oscillations and alpha power, with negative values of phase-amplitude coupling (PAC; Fig 7A) over occipital channels (delineated in Fig 5A, top left). This pattern persisted during mild sedation and only changed during moderate sedation within the drowsy group, in whom it shifted toward positive PAC values, before reverting back to negative PAC at recovery. There was a significant interaction in occipital PAC between level of sedation and group (F(3) = 3.8, p = 0.021). Fig 7C provides more detail on this, using angular histograms of alpha power distributed over slow phase, for a pair of representative participants, one in each of the two groups, responsive and drowsy. At baseline, occipital alpha power was either evenly spread over slow phase, or was greater near the trough of the slow oscillation, resulting in a trough-max distribution and negative PAC. During moderate sedation, only the drowsy participant’s distribution shifted towards peak-max positive PAC with greater alpha power near slow oscillation peaks. At recovery, this distribution reverted back to a trough-max pattern with negative PAC. Further, we also found a highly significant positive correlation between PAC and drug concentrations in blood during moderate sedation (Fig 7B). This correlation did not manifest during mild sedation or recovery, when drug concentrations were relatively low. Importantly, there was no significant correlation between PAC and reaction times. This was in contrast to the correlations between alpha power/connectivity and reaction times (Figs 3B and 3D and 5C), and highlights a novel dissociation between phase-phase and phase-amplitude coupling: while the former correlated with responsiveness as measured by hit rates and reaction times, the latter correlated drug concentrations in blood. Juxtaposed with previous research, our findings are convergent with existing evidence for characteristic changes in PAC alongside propofol induction. Trough-max slow-alpha PAC has been shown to accompany transitions to unconsciousness in frontal EEG channels, which then switches to a peak-max pattern in the same channels following loss of consciousness during deep sedation [18, 35]. While we have highlighted complementary changes in occipital channels, we also replicated these previous findings. In frontal channels, slow-alpha PAC values were close to zero at baseline, and progressed to a trough-max pattern during moderate sedation (see S2 Fig). This resulted in a significant interaction between level of sedation and group in frontal PAC values (F(3) = 4.1, p = 0.0136), with the drowsy group showing a significantly stronger trough-max pattern than the responsive group during moderate sedation (p = 0.011). Further, as with occipital PAC, there was a significant correlation between frontal PAC and drug concentrations in blood during moderate sedation (S2 Fig). Our experimental design used propofol sedation to engender transitional states of responsiveness that varied across participants. The levels of drug administered produced a variable pattern that spread the participant group along a spectrum of varying behavioural impairment, rather than resulting in complete unconsciousness in all of them. Using EEG to track brain activity and measuring actual levels of drug in blood alongside this spectrum of impairment has enabled us to identify neural markers that dissociate conscious report from drug exposure [2], and makes the results presented here distinctive in their contribution to advancing understanding of the neural markers of loss of consciousness due to propofol. We have built upon previous research that has shown that while occipital alpha power progressively drops as participants become behaviourally compromised as measured by reaction times, the qualitatively dissimilar onset of frontal alpha power is a characteristic marker of the loss of consciousness [18, 19, 30, 32, 36]. Confirming our hypotheses, while this frontal alpha generates meso-synchronous modules, brain network connectivity as a whole is nevertheless impaired. Graph-theoretic measures quantify this loss of the capacity of individual brain networks in the alpha band, linking them to concomitant variability in behavioural impairment across participants. Small-worldness is commonly seen as a measure of the cost-versus-efficiency optimality of a network configuration, and our findings converge with previous evidence [37] highlighting the reduction in the efficiency of cortical networks during loss of consciousness during propofol sedation, potentially due to dysfunctional modulations in thalamocortical connectivity [8, 38, 39]. It is worth noting that a similar breakdown in the capacity of alpha networks has also been reported with other anaesthetic agents like sevoflurane and ketamine [40–42]. This is despite the fact that these distinct anaesthetic agents had varying effects on EEG oscillations and, unlike propofol, did not always produce increases in frontal alpha. Hence the observed changes in alpha networks due to sedation cannot be explained as a shift of alpha power and connectivity from posterior to anterior areas. Rather, our results, along with these previous findings, point toward a broader understanding of characteristic signatures of connectivity in alpha networks as potentially reliable correlates of reportable consciousness [43]. Measurement of drug concentrations at each level of sedation dissociated a principal clinical pharmacodynamics target per se (sedation and consequent behavioural unresponsiveness) from incidental pharmacodynamic consequences of drug exposure during propofol sedation. The considerable individual variability in the susceptibility to anaesthesia has been documented [16], and is evident in the large overlap between blood levels of drug in our responsive and drowsy groups. While our measurement of modulations in phase-phase coupling in delta and alpha bands during sedation showed clear correlations with behavioural impairment, we have also demonstrated a latent relationship between slow-alpha phase coupling and individual variation in drug concentrations. It is important to distinguish these dynamic slow oscillations from stable slow cortical potentials observed during propofol anaesthesia [12], and from delta oscillations during sleep [44]. This link between PAC and individual levels of drug in blood was not observed in the delta or alpha bands separately, in either power or connectivity. Analytical approaches used for estimating Bispectral Index (BIS, see [45] that do not take phase information into account are unlikely to detect this key marker of individual drug concentration [18]. Hence our findings are relevant to the challenge of engendering an appropriate level of unconsciousness by accurately tailoring drug concentrations to individuals, a key consideration with significant implications for clinical anaesthesia. Finally, by tracking individual brain networks across levels of sedation, we have shown that the quantifiable robustness of alpha connectivity networks in the awake state before sedation predicts susceptibility to propofol. Specifically, given two behaviourally indistinguishable individuals undergoing administration of sedative, the one with the more robust, small-worldy alpha network with well-connected hubs is likely to require a greater amount of drug to render them unresponsive to the same degree. It is important to note that this latent variability in the state of alpha connectivity at baseline could be detected despite the lack of any significant differences in behavioural performance or alpha power at that time. Orthogonally, slow-alpha PAC complements this predictive capability by tracking the concentration of propofol in blood plasma. This set of results, if replicated and verified in the clinical context, could contribute to reliable applications of brain monitoring for tracking and accurately modulating consciousness with anaesthetics during routine surgery. All healthy controls gave written informed consent. Ethical approval for testing healthy controls was provided by the Cambridgeshire 2 Regional Ethics Committee. All clinical investigations were conducted in accordance with the Declaration of Helsinki. A convenience sample of 22 neurologically healthy adults participated in the study. Data from two participants could not be used due to technical issues, leaving 20 participants (9 male; 11 female) (mean age = 30.85; SD = 10.98) whose data were analysed. Each experimental run began with an awake baseline period lasting 25–30 minutes (Fig 1A) following which a target-controlled infusion of propofol [46] was commenced via a computerized syringe driver (Alaris Asena PK, Carefusion, Berkshire, UK). With such a system the anesthesiologist inputs the desired (“target”) plasma concentration, and the system then determines the required infusion rates to achieve and maintain the target concentration (using the patient characteristics which are covariates of the pharmacokinetic model). The Marsh model is routinely used in clinical practice to control propofol infusions for general anesthesia and for sedation. Three blood plasma levels were targeted– 0.6μg/ml (mild sedation), 1.2μg/ml (moderate sedation), and recovery from sedation. The state of mild sedation was aimed to engender a relaxed but still responsive behavioural state. At each target level, a period of 10 minutes was allowed for equilibration of plasma propofol concentrations to attain a steady state, following which behavioural tests and EEG measurements were commenced. After cessation of infusion, plasma propofol concentration exponentially declined toward zero. Computer simulations with the TIVATrainer pharmacokinetic simulation software revealed that plasma concentration of propofol would approach zero in 15 minutes leading to behavioural recovery; hence behavioural assessment was recommenced 20 minutes after cessation of sedation. Blood samples of 1cc each were taken at the beginning and end of the mild and moderate sedation states, and once at recovery, as indicated in Fig 1A. In total, 5 blood samples were taken during the study. These samples were analysed offline for characterising the significant inter-individual variability in actual propofol levels in blood plasma. We confirmed that the samples taken at the beginning and end of mild and moderate sedation had similar values of propofol concentration. The average of the two values, along with the value at recovery, were used as distinct covariates for EEG data analysis. At each of the 4 steady-state levels above, participants were requested to perform a simple behavioural task involving a fast discrimination between two possible auditory stimuli (Fig 1A). Specifically they were asked to respond with a button press to indicate whether a binaurally presented stimulus was a buzz or a noise. These stimuli constituted either broadband noise or a harmonic complex with a 150Hz fundamental frequency (buzz). Forty such stimuli, twenty of each kind, were presented in random order over two blocks, with a mean inter-stimulus interval of 3 seconds. We calculated a participant’s cognitive processing of these stimuli at each sedation level based on their hit rates, i.e., percentage of correct responses. In addition, we measured reaction times based on the delay between auditory tone onset and correct button press. We employed binomial modelling to distinguish participants who became behaviourally impaired during moderate sedation, from those who remained responsive, albeit with slower reaction times. Specifically, we fitted a binomial distribution to each participant’s hit rates at baseline and during moderate sedation. With each fitted model, the distribution parameter p, the probability of a correct response, and its 95% confidence intervals were estimated. For a given participant, if the confidence interval at moderate sedation was lower than and non-overlapping with that at baseline, they were considered to have become significantly impaired, and we designated them as drowsy. If the confidence intervals overlapped, we designated them as responsive. From each participant, approximately 7 minutes of 128-channel high-density EEG data were collected at each level of sedation. EEG was measured in microvolts (uV), sampled at 250Hz and referenced to the vertex, using the Net Amps 300 amplifier (Electrical Geodesics Inc., Eugene, Oregon, USA). Participants had their eyes closed in a resting state during data collection. Data from 91 channels over the scalp surface (Fig 2) were retained for further analysis. Channels on the neck, cheeks and forehead, which tended to contribute most of the movement-related noise, were excluded. Retained channels were filtered between 0.5–45Hz, and segmented into 10-second long epochs. Each epoch thus generated was baseline-corrected relative to the mean voltage over the entire epoch. Data containing excessive eye movement or muscular artefact were rejected by a quasi-automated procedure: abnormally noisy channels and epochs were identified by calculating their normalised variance and then manually rejected or retained by visual inspection. After pre-processing, a mean (SD) of 38 (5), 39 (4), 38 (4) and 40 (2) epochs were retained for further analysis in the baseline, mild sedation, moderate sedation and recovery conditions, respectively. An ANOVA revealed no statistically significant difference between the numbers of epochs retained. Finally, previously rejected channels were interpolated using spherical spline interpolation, and data were re-referenced to the average of all channels. These processing steps were implemented using custom MATLAB scripts based on EEGLAB [47]. Fig 2 depicts the data processing pipeline employed to calculate spectral power and connectivity measures from the clean EEG datasets. Spectral power values within bins of 0.25Hz were calculated using Fourier decomposition of data epochs using the pwelch method. At each channel, power values within canonical frequency bands, namely delta (0–4Hz), theta (4–8Hz), alpha (8–15Hz), beta (12-25Hz) and gamma (25–40Hz), were converted to relative percentage contributions to the total power over all five bands. Alongside, cross-spectrum between the time-frequency decompositions (at frequency bins of 0.49Hz and time bins of 0.04s) of every pair of channels was used to calculate debiased weighted Phase Lag Index (dwPLI, see [22]). For a particular channel pair and frequency band, mean dwPLI across all time at the peak frequency within each band was recorded as the ambient amount of connectivity between those channels. dwPLI is a sensitive measure of connectivity between cortical regions that has been shown to be robust against the influence of volume conduction, uncorrelated noise, and inter-subject variations in sample size [22], and has previously be used to characterise connectivity in pathological [48] and pharmacological [49] alterations in consciousness. However, as pointed out by Vinck, Oostenveld [22], dwPLI is relatively insensitive to true connectivity at phase differences close to 0 or 180 degrees. Further, the actual locations of brain sources producing dwPLI connectivity between a pair of sensors might not necessarily be spatially proximal to those sensors. Nevertheless, for the purposes of this study, it provides a robust measure for estimating how this indirect connectivity is affected by propofol sedation. Phase-amplitude coupling (PAC), also referred to as cross-frequency coupling [50], was used to measure the propofol-induced changes in the relationship between the phase of ongoing oscillations in the slow (0.5–1.5Hz) and alpha (8–15Hz) bands at each channel. Calculation of PAC was based on the Direct PAC estimator formally defined by Ozkurt and Schnitzler [51] and implemented in the Brainstorm 3.2 toolbox [52]. Purdon, Pierce [18] and Mukamel, Pirondini [35] previously identified changes from trough-max to peak-max PAC during propofol sedation, as determined by whether the slow oscillation is at its trough (at a phase angle of pi) or its peak (phase angle of 0) when alpha power is maximal, respectively. Such variations were measured by assigning a negative or positive sign to the amplitude of the complex-valued Direct PAC estimator depending on whether its phase angle was closer to pi or 0 radians, to indicate trough-max and peak-max coupling respectively. The 91x91 subject-wise, band-wise dwPLI connectivity matrices were thresholded to retain between 50–10% of the largest dwPLI values. They were then represented as graphs with the channels as nodes and non-zero values as links between nodes. The lowest threshold of 10% ensured that the average degree was not smaller than 2 * log(N), where N is the number of nodes in the network (i.e., N = 91). This lower boundary guaranteed that the resulting networks could be estimated [26]. Similar ranges of graph connection densities have been shown to be the most sensitive to the estimation of ‘true’ topological structure therein [53, 54]: higher levels of connection density result in increasingly random graphs, while lower levels result in increasingly fragmented graphs. At each step of the connection density between 50% and 10% in steps of 2.5%, the thresholded graphs were submitted to graph-theoretical algorithms implemented in the Brain Connectivity Toolbox [55]. These algorithms were employed to calculate metrics that captured key topological characteristics of the graphs at multiple scales, and avoided the multiple comparisons problem entailed by comparing large numbers of network connections. These included the micro-scale clustering coefficient and macro-scale characteristic path length [26], alongside meso-scale measures like modularity and community structure [56], and participation coefficient [29]. Here, this functional notion of modularity measures the extent to which the nodes of a graph can be parcellated into topologically distinct modules with more intra-modular links than inter-modular links [28]. Modularity as calculated by the heuristic Louvain algorithm, and all measures derived therefrom, were averaged over 50 repetitions. Next, each graph metric thus derived was normalised by the average of 50 null versions of the metric similarly derived, but after repeatedly phase-randomising the original cross-spectra and recalculating dwPLI for each channel pair. Finally, the small-worldness index of a graph was calculated as the ratio of normalised clustering coefficient to characteristic path length [57]. Metrics were compared using two-way ANOVAs with one non-repeated (group) measure and one repeated (level of sedation) measure. The obtained p-values were corrected for violations of sphericity using a Greenhouse-Geisser correction. Pairwise tests between groups were corrected for multiple comparisons using Tukey’s HSD test. The ability of graph metrics of individual participants to predict their behaviour was tested using robust linear regression [58] to calculate R2 and p-values.

10.1371/journal.ppat.1005333

Intravital Imaging of Vascular Transmigration by the Lyme Spirochete: Requirement for the Integrin Binding Residues of the B. burgdorferi P66 Protein

Vascular extravasation, a key step in systemic infection by hematogenous microbial pathogens, is poorly understood, but has been postulated to encompass features similar to vascular transmigration by leukocytes. The Lyme disease spirochete can cause a variety of clinical manifestations, including arthritis, upon hematogenous dissemination. This pathogen encodes numerous surface adhesive proteins (adhesins) that may promote extravasation, but none have yet been implicated in this process. In this work we report the novel use of intravital microscopy of the peripheral knee vasculature to study transmigration of the Lyme spirochete in living Cd1d-/-mice. In the absence of iNKT cells, major immune modulators in the mouse joint, spirochetes that have extravasated into joint-proximal tissue remain in the local milieu and can be enumerated accurately. We show that BBK32, a fibronectin and glycosaminoglycan adhesin of B. burgdorferi involved in early steps of endothelial adhesion, is not required for extravasation from the peripheral knee vasculature. In contrast, almost no transmigration occurs in the absence of P66, an outer membrane protein that has porin and integrin adhesin functions. Importantly, P66 mutants specifically defective in integrin binding were incapable of promoting extravasation. P66 itself does not promote detectable microvascular interactions, suggesting that vascular adhesion of B. burgdorferi mediated by other adhesins, sets the stage for P66-integrin interactions leading to transmigration. Although integrin-binding proteins with diverse functions are encoded by a variety of bacterial pathogens, P66 is the first to have a documented and direct role in vascular transmigration. The emerging picture of vascular escape by the Lyme spirochete shows similarities, but distinct differences from leukocyte transmigration.

Lyme disease is the most common vector-transmitted infection in North America and Europe. Diverse clinical manifestations of Lyme disease result from the dissemination of the spirochetes causing the disease into a variety of tissue sites. Dissemination results from invasion of the vasculature by the bacteria, followed by exit into virtually all tissue types. The mechanism of vascular transmigration by Lyme disease spirochetes remains uncharacterized. Here we describe a novel approach to study transmigration of Lyme disease spirochetes using intravital microscopy of the peripheral knee vasculature in living mice. Our studies have identified an adhesin, P66, and its integrin-binding function as playing important roles in Lyme spirochete transmigration and dissemination.

Lyme disease is a spirochetal illness caused by various members of the genus Borrelia, and the most prevalent vector-borne illness in North America and Europe [1–5]. The disease is transmitted to humans during the feeding of infected hard-shelled ticks that have acquired the spirochetes during an earlier blood meal on infected reservoir animals. Once inoculation of the skin has occurred, the highly motile Borrelia species multiply and migrate in the skin, often resulting in an erythema migrans lesion or “bulls-eye rash”. As the disease progresses (see Fig 1), spirochetes invade the vasculature, which provides a mechanism for dissemination throughout the body, followed by extravasation into a variety of tissue types. Dissemination of spirochetes can result in non-specific illness, arthralgia, carditis and neuroborreliosis. Persistent untreated infection can result in acrodermatitis and a variety of neurological problems, as well as Lyme arthritis, a common feature of the disease in North America that results from the inflammatory response to spirochete invasion into the joints [6]. Although hematogenous dissemination is important for disease development, little is known regarding the mechanisms involved in this process. Hematogenous dissemination is a multi-step process likely involving multiple Borrelia proteins as well as host proteins and macromolecular cell receptors in the host vasculature [7, 8]. Direct analysis and quantification of the transmigration process has been complicated by a variety of factors. Generalized dissemination assays by culture of spirochetes from mouse tissues or by direct observation do not provide quantitative information [9]. More recently visualization of dissemination by whole body bioluminescent imaging [10] provides quantitative information. The above methods provide useful information, but do not reveal the causes for a lack of dissemination or the step along the transmigration pathway where defects occur. A failure to disseminate may be related to a lack of spirochete survival due to metabolic defects [11], clearance of spirochetes by the host immune system [12, 13], or a defect in one of the spirochetal components directly required for early or later steps along the transmigration pathway. We have previously used intravital microscopy to study early events in dissemination [14–16]. The high dose needle inoculation and short-term observation period precludes the use of our intravital system for investigation of many aspects of the natural infectious cycle of B. burgdorferi in mice. However, it does provide an exceedingly powerful approach for investigation of mechanistic aspects of spirochete-host interactions in their natural setting. In particular, interactions of B. burgdorferi with the host microvasculature can be imaged in real time at high resolution under shear force in a living animal. This has allowed us to effectively image B. burgdorferi interactions with post-capillary venules [14–16]. Spirochetes in the vasculature have been observed to participate in initial tethering interactions, followed by longer-lived dragging interactions and finally stationary adherence to the endothelium. These interactions can be quantitatively analyzed and mechanistically dissected. However, invasion of surrounding tissue is a rare event compared with vascular adhesion and the number of transmigrating spirochetes observed by intravital microscopy was exceedingly low during the course of one-hour experiments. Moreover, transmigrated spirochetes were highly motile in flank skin, where the experiments were performed, and rapidly escaped the site of transmigration. Therefore, enumeration of transmigration required direct observation of rare escaping spirochetes, making the use of intravital microscopy ineffective for studying this process. In the current work we report the development of a high resolution intravital imaging transmigration assay that allows enumeration of transmigrated spirochetes in the peripheral knee vasculature of living mice. In knee joint-proximal tissue in the mouse, iNKT cells surround the outside of blood vessels and block extravasation; they also eliminate those spirochetes that do successfully transmigrate into the joint-proximal tissue [17]. The use of Cd1d-/- mice, which lack iNKT cells, results in a 25-fold increase in spirochetes disseminating into the mouse joint [18], providing numbers of spirochetes amenable for direct visualization. Moreover, fluorescent transmigrated B. burgdorferi persist in the joint-proximal tissue and can be enumerated at 24 hours post intravenous (iv) inoculation. Using this new approach, coupled with an analysis of vascular clearance, we show that B. burgdorferi lacking BBK32, an important fibronectin and glycosaminoglycan adhesin [7, 19–21] display wild-type levels of transmigration. In contrast, spirochetes lacking the integrin adhesin and porin P66 [22–24] or those specifically defective in integrin binding display almost no vascular transmigration into the mouse knee joint-proximal tissue, indicating a direct role for P66-integrin interactions in the transmigration process. In our previous work we noted a 25-fold increase in spirochete burden in the knee joints of Cd1d-/- mice versus isogenic wild-type mice. More recently, our use of intravital microscopy of the peripheral knee vasculature of infected Cd1d-/- mice revealed readily observable spirochetes in the joint-proximal tissue [17], prompting us to further develop and exploit this system to study the transmigration process in the absence of spirochete killing by iNKT cells. The area exposed for intravital imaging in this study was over the patellar ligament and to the medial side of the right hind limb. We examined post-capillary venules that drain blood from the anterior tibialis muscle and join with the anterior tibial vein that drains into the great saphenous vein. The location of the venules ranges from 0 to a maximal depth of approximately 200 μm. These venules are located near the surface, rather than in the interior of the joint and transmigrating spirochetes are found predominantly in the anterior tibialis muscle. The venules display properties that are unique compared with venules in other tissue. The iNKT cells appear to be the first line of defense against B. burgdorferi in only the peripheral joint tissue; they have an extravascular location and are mobile compared to in the liver where they are intravascular and stationary [17, 18]. Mice were inoculated intravenously with 4 x 108 spirochetes through the tail vein, and at 24 hours post-inoculation intravital microscopy was performed (see Materials and Methods) on the right knee of each infected mouse. As shown in Fig 2, intravital microscopy of GFP-expressing B. burgdorferi by spinning disk laser confocal microscopy revealed spirochetes that had escaped from the peripheral joint vasculature, which was labeled with a fluorescent anti-PECAM-1 antibody. These spirochetes displayed a to-and-fro translational motility (S1 and S2 Videos) and remained within the knee joint-proximal tissue, possibly as the result of a chemotactic response [25] to localized chemical stimuli. A kinetic analysis of spirochete transmigration (Fig 3) revealed that although low levels of transmigrated spirochetes were visible at a variety of time points, the highest observed levels were found at 24 hours post-inoculation, the necessary endpoint of these experiments due to constraints imposed on our animal protocol. This time point was used for subsequent analysis of spirochete transmigration. BBK32 is a B. burgdorferi adhesion that binds to fibronectin (Fn) and glycosaminoglycans (GAGs) [7, 19–21]. BBK32 promotes vascular tethering and dragging interactions in the vasculature of living mice [16] through endothelial interactions mediated by its Fn and GAG binding domains, respectively [15]. To determine if BBK32 plays an essential role in B. burgdorferi transmigration into joint-proximal tissue, we infected Cd1d-/- mice with wild-type and Δbbk32 spirochetes expressing GFP (see S1 Table for strain information). Intravital microscopy was performed at 24 hours post-inoculation. The host vasculature was labeled with fluorescent antibody to PECAM-1 and video footage was collected on at least five areas in each knee during the one hour intravital experiment. Spirochetes that had transmigrated into the joint-proximal tissue were counted. As shown in Fig 4A, no difference was observed in the transmigration of bbk32 versus wild-type spirochetes. In contrast, no detectable transmigration was observed using a fluorescent, non-infectious high-passage strain. In our previous intravital studies on vascular adhesion, we noted that different B. burgdorferi strains and mutants sometimes displayed widely variable clearance rates from the vasculature [15]. Spirochete concentrations in the blood are certainly expected to influence early vascular interaction levels. However, their effect upon transmigration is currently unknown because transmigration is believed to be preceded by early stationary adhesion to the endothelium [14]. At 24 hours post-inoculation the number of circulating spirochetes has decreased by three orders of magnitude and is difficult to accurately determine. Moreover, with the low numbers of spirochetes in circulation at 24 hours, it is not possible to quantify the level of stationary adhesions. Nonetheless, large changes in spirochete clearance in mutant strains can be accurately assessed at one hour by direct microscopic counting. This was done by comparing blood samples obtained at five and 60 minutes post-inoculation. Sixty minutes was chosen as a general time point as the number of spirochetes was usually high enough for accurate counting and it allowed discrimination between very rapidly cleared and slowly cleared strains. Direct counting of spirochetes in plasma (see Materials and Methods) was performed by dark-field microscopy. Spirochete counts for each individual mouse at five minutes post-inoculation were defined as 100% and the percent spirochetes remaining for each mouse at 60 minutes was calculated. Fig 4B shows that the disruption of bbk32 did not result in a change in the clearance rate of the mutant versus the wild-type, with approximately 80% of the spirochetes remaining in the circulation at one hour. In contrast, the high-passage non-infectious strain GCB706 was rapidly cleared. It is currently unknown if an increased clearance rate would affect transmigration levels, however, the absence of a change in the clearance rate of the bbk32 mutant versus wild-type spirochetes eliminates the potential influence of circulating spirochete numbers from consideration as a factor in the influence of BBK32 on transmigration. P66 is a B. burgdorferi integrin adhesin [26–28] and porin [29–32] that is required for mouse infectivity [24]. The underlying reason for the lack of infectivity displayed by B. burgdorferi carrying a p66 gene disruption is unknown. To determine if P66 plays an essential role in B. burgdorferi transmigration into joint-proximal tissue in the knee, we inoculated Cd1d-/- mice with wild-type infectious and Δp66 spirochetes expressing GFP. Intravital microscopy was performed at 24 hours post-inoculation. As shown in Fig 5A, a dramatic and highly significant decrease in transmigration into the knee joint-proximal tissue was observed. This effect was specific to the p66 mutation as demonstrated by restoration of transmigration by replacement of the disrupted p66 gene with the wild-type allele. Analysis of the rate of clearance of the p66 deletion strain revealed a clearance rate about twice as fast as the wild-type strain, with restoration of the wild-type clearance rate in the complemented strain. The mechanism of the more rapid clearance of the Δp66 mutant is unknown and the effect of a clearance rate doubling on the formation of stationary adhesions leading to transmigration is unknown, but not expected to be responsible for the dramatic decrease in transmigration observed. However, it does raise a possible concern for the strength of the conclusion that can be drawn from this experiment. To further investigate the mechanism underlying the decreased transmigration of the p66 mutant, we assessed the transmigration of two site-directed mutants, p66D205A,D207A and p66Δ202–208 [33]. The proteins encoded by these genes are either full length or have a seven amino acid deletion, respectively, in contrast to the Δp66 strain, which has no P66 on the cell surface. Both mutant proteins are specifically defective in integrin-binding activity, but are localized on the surface of B. burgdorferi and retain channel-forming activity [33]. As shown in Fig 6, both P66 integrin binding mutants displayed a dramatic and highly significant reduction in transmigration into joint-proximal tissue by the intravital microscopic assay 24 hours after inoculation of Cd1d-/- mice. Moreover, in contrast to the P66 full deletion mutant, both site-directed mutants displayed clearance rates from the vasculature that did not differ significantly from that of wild-type B. burgdorferi. The two site-directed P66 mutants, therefore, provide more convincing data for an essential role for P66, and in particular the integrin-binding residues, in the transmigration process. The corollary of the conclusions from Fig 6 is that integrins bound by P66 are expressed on the surface of the endothelium in post-capillary venules. To explore this, we visualized β3 integrins on the surface of endothelial cells by multi-channel spinning disk intravital microscopy. Fig 7 shows micrographs with spirochetes and PECAM-1 (green and red, respectively) visualized in the top panels. The bottom panels show the same micrographs with spirochetes and β3 integrins (green and blue, respectively) visualized. In general, a uniform pattern of staining with the anti-integrin antibody is seen, along with areas of more pronounced staining, similar to that observed for PECAM-1. However, specific areas of integrin concentration differ from those observed with PECAM-1. Spirochete binding did not occur in areas of integrin concentration, but was found in areas with uniform integrin expression. From the data in Figs 5, 6 and 7 we conclude that P66 is required for efficient transmigration into the joint-proximal tissue of the knee and that the integrin-binding activity of the protein plays an integral role in this process. Because of the requirement for P66 and its integrin binding sequences for transmigration in peripheral knee vasculature, it was of interest to determine whether P66 can promote vascular adhesion, an early stage of the transmigration process that can be monitored by intravital microscopy in a living mouse [14–16]. A complicating factor in designing such an experiment was that early B. burgdorferi-microvasculature interactions are mediated predominantly by BBK32 and one or more unknown fibronectin-independent adhesins [14, 15]. In the presence of these adhesins, the effect of other adhesins with less dramatic endothelial-binding properties is obscured. We, therefore, introduced the Δp66 mutation into the non-infectious strain GCB706, which is a fluorescent version of the high passage strain B31-A (see S1 Table) that lacks expression of bbk32 [15, 16]. This wild-type non-infectious strain background, therefore, displays only low levels of spirochete-endothelial interactions in post-capillary venules. For this experiment, the spirochetes were grown without blood supplementation to avoid induction of other possible adhesins. As shown in Fig 8, intravital microscopy was used to monitor tethering plus dragging (Fig 8A) and stationary attachment (Fig 8B) of the wild-type parent, the Δp66 mutant and the complemented mutant in the peripheral knee vasculature. No difference in microvascular interactions (tethering and dragging interactions combined) was observed between wild type and the Δp66 mutant or the Δp66 mutant and the Δp66/complemented strain (Fig 8A). Similarly, there was no difference in stationary adhesions between the wild-type and the mutant strain or the mutant strain and the complemented mutant (Fig 8B). Finally, the vascular clearance rates of all three strains were the same (Fig 8C). These results indicate that P66 does not promote the early transient or stationary interactions that withstand the shear force of vascular flow. Instead, B. burgdorferi relies upon other adhesins to promote initial microvascular-endothelial interactions, which facilitates subsequent P66-integrin contacts leading to transmigration. In this work we report a new approach to study vascular transmigration. This approach utilizes assessment of transmigration in the mouse peripheral knee vasculature by intravital microscopy in Cd1d-/- mice. The intravital methodology allows direct visualization and enumeration of transmigrated spirochetes in living mice. The use of Cd1d-/- mice makes this assay possible by eliminating iNKT cells, a component of the innate immune response that disrupts dissemination of B. burgdorferi into mouse joints through a granzyme dependent pathway [17, 18]. The peripheral knee vasculature also provides a desirable target for observation since transmigrating spirochetes remain in the local environment, in contrast to flank skin where they rapidly leave the transmigration site [14]. The general applicability to other tissues of conclusions drawn herein from intravital data obtained from the peripheral knee vasculature is currently unknown as this work is the first reported study of spirochete transmigration using intravital microscopy and the peripheral knee vasculature currently the only location where quantitative imaging of this process is currently viable. Using intravital microscopy of Cd1d-/- mice inoculated with GFP-expressing B. burgdorferi, we have monitored the effect of two important spirochete adhesins on transmigration into the joint-proximal tissue of the mouse knee. The Fn and GAG binding protein, BBK32, is believed to be involved in early steps of the dissemination process. It is sufficient to restore microvascular interactions to a non-adherent, non-infectious, high-passage B. burgdorferi strain [16]. BBK32 mediates tethering and dragging interactions with the host microvasculature through its Fn and GAG binding domains, respectively [15]. These interactions are believed to precede transmigration. Interestingly, disruption of the gene for BBK32 did not result in a decrease in transmigration into the mouse knee joint-proximal tissue and the mutant spirochetes were not cleared significantly faster than the wild-type parent. The absence of a transmigration phenotype for the bbk32 mutant is likely the result of duplicity of function; in the absence of BBK32, an unknown Fn-independent adhesin (or adhesins) maintain microvascular interactions in the mouse peripheral knee vasculature at 50% of the wild-type level [15]. We predict that the combined disruption of both BBK32 and the unknown Fn-independent adhesin(s) would dramatically impact transmigration; however, such an experiment is precluded until the identity of the unknown Fn-independent adhesin(s) is established. It is also noteworthy that in contrast to a lack of an effect of a bbk32 mutation on transmigration by our intravital assay, it has been previously reported that a bbk32 mutation results in a decreased ability to invade joints as determined by bioluminescent imaging, qPCR and spirochete recovery by culture [10]. An explanation for the discrepancy with our results is that the previous study utilized intradermal needle inoculation of mice with 103−105 spirochetes and followed outcomes up to 14 days post-infection. Although not natural tick inoculation, this mouse model has been used for decades to study Borrelia infections. In contrast, to obtain mechanistic data on the transmigration process, we have used the reductionist approach of intravenous inoculation of 4x108 spirochetes in Cd1d-/- mice for intravital imaging of transmigration at 24 hours. Our approach seeks to identify molecules directly involved in the transmigration process and seeks to eliminate other possible roles of BBK32 in the early infection process. The variance between our results and those previously published [10] suggests that BBK32 may have other functions in the infection process in wild-type mice that our assay has sidestepped, but which are noticeable in the traditional animal model. Alternatively, the variance may result from differences in gene expression associated with host adaptation of the spirochetes within the timeframes of the two experiment [34–38]. In contrast to the BBK32 result, disruption of the gene encoding the integrin adhesin and porin, P66, resulted in a dramatic reduction in transmigration. Moreover, mutants with either alanine substitutions for the aspartic acids at position 205 and 207 or deletion of residues 202–208, also resulted in a dramatic loss of transmigration ability. The noted mutations specifically target the integrin-binding residues of P66 without affecting its porin activity [33]. Our data therefore suggest that integrin binding is an essential feature of B. burgdorferi transmigration. Analysis of the effect of p66 disruption on vascular adhesion in a low-adherence high passage strain (lacking BBK32) revealed that P66 does not promote early transient interactions or stationary adhesions. Therefore, the transmigration process appears to be temporally choreographed with adhesins such as BBK32 and the unknown fibronectin-independent adhesion(s) playing an initial role of establishing B. burgdorferi-microvascular interactions that subsequently facilitate P66-integrin communication leading to transmigration through an as yet undefined process. In agreement with the findings presented here are the recent results that the p66Δ202–208 mutant displays undetectable bacterial burdens at two weeks after iv inoculation and reduced migration across microvascular endothelial monolayers in vitro [33]. A variety of bacterial pathogens encode integrin-binding proteins (see [39, 40]). Integrin binding plays diverse roles for bacterial pathogens, including as an adherence target on host cells and extracellular matrix (ECM). Integrin binding also functions to promote pathogen-host signaling pathways that may lead to cellular invasion [39, 40]. However, a function for an integrin-binding protein in bacterial vascular transmigration has not been previously reported. The slowly emerging picture of B. burgdorferi transmigration is that it is a multi-step process involving several B. burgdorferi proteins and host molecules [14–16] and has general similarities to leukocyte transmigration [41–43], although very significant deviations exist. In both cases the process is initiated by tethering interactions that withstand the shear stress of vascular flow. In B. burgdorferi this occurs between the Fn-binding domain of BBK32 and endothelial surface localized Fn. Subsequently, rolling (leukocytes) or dragging (B. burgdorferi) occurs. In Lyme disease spirochetes the dragging stage involves additional interactions between the GAG-binding domain of BBK32 and glycosylated endothelial receptors. Thereafter, stationary adhesion occurs, which involves at least one additional spirochete factor [14–16]. Our current work suggests that P66 is involved upon the establishment of a stationary interaction, where P66-integrin interactions can occur within an established B. burgdorferi-endothelial complex. Tethering, dragging and stationary B. burgdorferi interactions can be observed within minutes of iv spirochete injection. However, transmigration occurs only at background levels until about 24 hours after iv injection. The velocity of B. burgdorferi in liquid or 2–3% gelatin is about 3.5–5 μm per second [44, 45]. The rate of forward motion of spirochetes escaping postcapillary venules is much slower, about 3.4 μm per minute. Even at this slower velocity the average time span of spirochetal escape is about 11 minutes [14]. This raises the question of why only background levels of transmigration occur until 24 hours post-inoculation when both adhesion and physical escape can occur within a much shorter time-frame. A possible explanation for this that there may be a requirement for stationary B. burgdorferi adhesions to promote localized endothelial activation resulting in an increased permeability of the endothelium that facilitates spirochete transmigration. Conversely, stable adherence of B. burgdorferi to the endothelium may promote activation of the spirochete by inducing the production of factors that are necessary for transmigration. In the case of leukocytes, activation of integrins precedes an increase in ligand affinity, resulting in stationary adhesion and migratory arrest on the endothelial surface, although integrins also play other roles in leukocyte transmigration [41–43]. In contrast, B. burgdorferi does not contain integrins, but encodes P66, an integrin adhesin and porin. We show here that the integrin-binding region of P66 is required for transmigration into the mouse knee joint-proximal tissue. Whether the porin activity is also required for transmigration remains unknown at present since P66 mutants affecting the porin but not the integrin-binding activity do not currently exist. The mechanism by which P66 is involved in the transmigration process awaits further characterization. Finally, it is important to note that the intravital imaging studies reported here provide a powerful tool for studying the role of P66 in vascular interactions and transmigration, however, they do not shed light on other roles of P66 in the natural infectious process as recently reported [33]. The work reported here describes a new approach to study vascular transmigration of the Lyme disease spirochete, which has resulted in the identification of the first B. burgdorferi adhesin believed to be directly involved in vascular transmigration. All animal experimentation was carried out in accordance with the principles outlined in the most recent policies and Guide to the Care and Use of Experimental Animals by the Canadian Council on Animal Care. The animal protocol (AC12-0218) was approved by The Animal Care Committee of the University of Calgary. Wild-type BALB/c mice were purchased from Charles River (Wilmington, MA) and CD1d-/- mice in a BALB/c background (Jax #2962) were bred in-house at the Clara Christie Centre for Mouse Genomics at the University of Calgary. Mice of both genders between 5–16 weeks of age were used. The GFP-expressing B. burgdorferi strains used in this study are described in S1 Table. pTM61-strep (SmR/SpR) was constructed by digesting pTM61 [14] with AvrII and MluI (New England Biolabs, Beverly, MA), then dephosphorylated using alkaline phosphatase. The spectinomycin/streptomycin resistance cassette [46] was amplified from pTAspc (provided by Dr. D. Scott Samuels) using the primers flgBpAvrII (CCTAGGTAATACCCGAGCTTCAAGGAAG) and aadAMluI (ACGCGTGACGTCATTATTTGCCGACTACC). The PCR product was cloned directly in pGEM T Easy (Promega, Madison, WI), then excised with AvrII and MluI and ligated into the pTM61 prepared as described above. The ligation mixes were used to transform E. coli strain Top10 (Life Technologies, Grand Island, NY) by electroporation [47]. Selection of the desired recombinant plasmid was performed by plating on Luria broth agar supplemented with spectinomycin at 80 μg/ml. For further experiments in B. burgdorferi, the plasmid was selected for using 80 μg/ml of streptomycin in Barbour-Stoenner-Kelly II (BSK-II) medium (with 6% rabbit serum, Cedarlane Laboratories ltd., Burlington, ON). Plasmid content was analyzed for all constructed strains by multiplex PCR [48]. Spirochete cultures in BSK-II medium were inoculated from frozen glycerol stocks. Levels of antibiotics used are indicated in S1 Table. The spirochete cultures were grown at 35°C to a concentration between 1–5 x 107/ ml. Spirochetes were then diluted to 1 x 105/ml in 100 ml BSK-II medium and grown for 24 hours, followed by 48 hours growth with 1% BALB/c mouse blood (vol/vol) and 1x Borrelia antibiotic mixture (final concentrations of 20 μg/ml phosphomycin, 50 μg/ml rifampicin, and 2.5 μg/ml amphotericin B). The bacterial densities were not allowed to exceed 5 x 107/ml. Spirochetes were harvested by centrifugation at 6000xg for 15 minutes at 4°C and washed twice with 100 ml of cold phosphate-buffered saline (PBS). BALB/c or CD1d-/- mice were anesthetized by intraperitoneal injection of 200 mg/kg ketamine hydrochloride (Bimeda-MTC, Animal Health Inc., Cambridge, ON) and 10 mg/kg of xylazine hydrochloride (Bayer Inc., Toronto, ON). For vascular transmigration assays, anesthetized CD1d-/- mice were secured in a mouse restrainer and inoculated by tail vein injection with 4 x 108 spirochetes in 150 μl of phosphate buffered saline (PBS). For vascular adhesion assays, anesthetized mice were intravenously injected with 4 x 108 spirochetes via a jugular vein catheter. Mice were anesthetized as described above. Additional intraperitoneal injection of anesthetic was done as necessary to maintain anesthesia for the total duration of the procedure. For direct spirochete counts in the vascular transmigration assays, blood samples (35 μl, collected in an equal volume of 20mM sodium citrate) were taken from the saphenous or tail vein at five and 60 minutes post-inoculation. For counts in vascular adhesion assays, samples were collected at three and 18 minutes post-inoculation. Samples were drawn into glass capillary tubes, the bottom of the capillary tubes were blocked with vacuum grease and incubated vertically at 4°C overnight to allow separation of blood cells by gravity. The next day, the glass capillaries were cut with a diamond pencil near the white blood cell-plasma interphase and 25 μl of plasma was collected and centrifuged at 6000xg for 15 minutes at 4°C to pellet the spirochetes. The supernatant fluid was carefully decanted and discarded. The pellet was resuspended in 6 μl of PBS, diluted as necessary, and loaded on a Petroff-Hausser counting chamber and counted. Spirochete numbers for each mouse at a later time point (18 or 60 minutes post-inoculation) were expressed as a percentage of the spirochete numbers observed at an initial time point (3 or 5 minutes post-inoculation) for each individual mouse. Deep surgical anesthesia in BALB/c or CD1d-/- mice was induced as described above and maintained by regular iv injections of anesthetic in sterile normal saline. A home-made catheter made of PE10 polyethylene tubing (I.D. 0.011" O.D. 0.024", Intramedic, Becton Dickinson and Company, Sparks, MD) was used as a jugular vein catheter. The catheter was connected to a 1ml syringe filled with 100 I.U. / ml of heparin in physiological saline solution. After confirmation of induction of surgical anesthesia, the knee was prepared for intravital microscopy as previously described [49]. The hair over the jugular vein area was clipped and the mouse was positioned in dorsal recumbency, and secured over a regulated heating pad. The skin was incised at the ventral cervical region right to the midline at the level of the clavicle bone. Under a dissection microscope, the right jugular vein was separated carefully from adjacent fat and connective tissues by blunt dissection and the anterior aspect of the jugular vein was tied with 4.0 surgical sutures and secured tightly to the imaging platform with surgical tape. A slightly bent 30½ G needle was inserted parallel in the jugular vein and the vein was slightly lifted up gently with help of the inserted needle. A PE10 polyethylene tube was introduced into the vein with simultaneous removal of the needle from the vein. The jugular vein patency was confirmed by withdrawal of blood in the tubing and the catheter was secured at cranial and distal locations with a double surgical knot around the vein. The jugular vein was used for maintaining the anesthesia level, injecting fluorescent antibodies, and injecting spirochetes. For knee preparation, the right hind limb was slightly flexed and secured to a home-made plastic knee holder using a surgical tie just above the metatarsal joint. The medial aspect of the knee joint was slightly oriented upwards, placed, and immobilized in the plastic knee holder. The hair was clipped, the area was swabbed with mineral oil, and the skin over the medial aspect of the knee joint was carefully removed by incision. The connective tissues over the exposed area was carefully removed without injuring blood vessels. The exposed area was always immersed in isotonic saline to keep it moist. A glass cover slip was positioned over the exposed joint area and secured to the knee-holder using vacuum grease. Spinning disk confocal microscopy was performed using an Olympus BX51W1 base (Olympus, Center Valley, PA) fitted with 10x/0.30 UPlanFLN air and 20x/0.95 XLUMPlanFI water immersion objectives. The microscope was equipped with a confocal light path (WaveFx, Quorum, Guelph, ON, Canada) based on a modified Yokogawa CSU-10 head (Yokogawa Electric Corporation, Tokyo, Japan). The mice were injected with antibodies immediately before i.v. inoculation with B. burgdorferi, which corresponded to roughly 10–20 minutes before imaging, depending upon the experiment. The endothelial cells were stained with phycoerythrin (PE)-conjugated anti-PECAM-1 (5 μg per mouse; clone MEC 13.1; BD Biosciences, Mississauga, ON) or PECAM-1 antibody (50 μg per mouse; clone MEC 13.3; BD Biosciences, Mississauga, ON) conjugated in-house to Alexa Fluor 555 (Invitrogen, Burlington, ON) or Alexa Fluor 594 anti-mouse CD31 (5 μg per mouse; clone MEC 13.3; BioLegend Inc, San Diego, CA). β3 integrin staining was achieved using Alexa Fluor 647 anti-mouse/rat CD61 antibody (7.5 μg per mouse; clone 2C9.G2; BioLegend Inc, San Diego, CA). The spirochetes were engineered for GFP expression as described earlier. Vascular transmigration was evaluated by exciting the fluorophores with two laser wavelengths (488 nm and 561 nm; Cobolt, Stockholm, Sweden) and visualizing with proper band-pass filters (Semrock, Rochester, NY). A back-thinned electron-multiplying charged-coupled device camera (512 x 512 pixels; C9100-13, Hamamatsu, Bridgewater, NJ, USA) was utilized for detecting emitted fluorescence. Volocity software (version 6.0.1, Improvision, Lexington, MA) was used to control the microscope, image acquisition, and analysis. The sensitivity settings for the red and green laser used was 255 with the autocontrast setting turned on. All laser power settings were set to the highest, i.e. 1.84 mv. For video acquisition, the shutter setting was selected as ‘do not manage shutters’ and binning settings were set to 2x for GFP and RFP channels, which allowed data acquisition at a rate of between 14–20 frames per second. The vascular transmigration assay for B. burgdorferi was done in deeply anesthetized CD1d-/- mice with the medial aspect of the knee surgically prepared for imaging as described above. For transmigration assays, the CD1d-/- mice were intravenously inoculated with spirochetes 1–24 hours before imaging as described above. Images were acquired using a 10x objective lens and spirochetes were counted manually in the field of view (FOV). The data were pooled from a total of 6–16 mice per group from two independent experiments. Five fields of view or more were counted for each mouse. Because of the variability of intravital results that can be observed in different experiments, all B. burgdorferi strains being compared were analyzed in the same intravital session (eg, in Fig 5 one mouse was analyzed for each B. burgdorferi strain in each intravital session). The vascular adhesion assay was done in deeply anesthetized CD1d-/- mice with the medial aspect of the knee surgically prepared for imaging as described above. For vascular adhesion assays, BALB/c mice were intravenously inoculated via the jugular vein catheter with spirochetes 10 minutes before imaging. Vascular interactions were recorded 18–45 minutes post-inoculation from straight, unbranched post-capillary venules of 15–25 μm diameter in the knee. Videos were recorded through 20x/0.95 XLUMPlanFI water immersion objectives. Spirochete-vascular interactions were counted in a length of 100 μm along the blood vessel manually as described before [14–16]. Because of the variability of intravital results that can be observed in different experiments, all B. burgdorferi strains being compared were analyzed in the same intravital session (eg., in Fig 5 one mouse was analyzed for each B. burgdorferi strain in each intravital session). GraphPad Prism 5.0 (GraphPad Software Inc., San Diego, CA) was used for statistical data analysis. Error bars in all graphs denote the standard deviation. Statistical significance was computed for experiments as noted in each figure legend.

10.1371/journal.ppat.1003942

‘Death and Axes’: Unexpected Ca2+ Entry Phenologs Predict New Anti-schistosomal Agents

Schistosomiasis is a parasitic flatworm disease that infects 200 million people worldwide. The drug praziquantel (PZQ) is the mainstay therapy but the target of this drug remains ambiguous. While PZQ paralyses and kills parasitic schistosomes, in free-living planarians PZQ caused an unusual axis duplication during regeneration to yield two-headed animals. Here, we show that PZQ activation of a neuronal Ca2+ channel modulates opposing dopaminergic and serotonergic pathways to regulate ‘head’ structure formation. Surprisingly, compounds with efficacy for either bioaminergic network in planarians also displayed antischistosomal activity, and reciprocally, agents first identified as antischistocidal compounds caused bipolar regeneration in the planarian bioassay. These divergent outcomes (death versus axis duplication) result from the same Ca2+ entry mechanism, and comprise unexpected Ca2+ phenologs with meaningful predictive value. Surprisingly, basic research into axis patterning mechanisms provides an unexpected route for discovering novel antischistosomal agents.

Schistosomiasis (Bilharzia) is one of the most burdensome parasitic worm infections, encumbering third world economies with an annual loss of several million disability-adjusted life years. The key treatment for schistosome infections is the drug praziquantel but the mechanism of action of this drug remains controversial hampering targeted development of next generation antischistosomal agents. Here we provide fresh insight into the signaling pathways engaged by PZQ, by resolving commonalities in the action of PZQ with the process of regenerative signaling in free-living planarian flatworms. A similar calcium-dependent network is engaged in both model systems, but with divergent phenotypic outcomes. This relationship provides predictive insight such that basic research on signaling pathways involved in tissue regeneration reveals novel drug leads for schistosomiasis, and reciprocally schistosomal drug screens reveal targets involved in regenerative signaling. We believe this phenology will be helpful for uncovering new antischistosomal drug targets by exploiting broader vulnerabilities within the PZQ interactome.

Over a third of the world's population is estimated to be infected with parasitic worms. One of the most burdensome infections underpins the neglected tropical disease schistosomiasis (Bilharzia), caused by parasitic flatworms of the genus Schistosoma. The debilitating impact of schistosomiasis results from the host's immune response to schistosome eggs, which are deposited in prolific numbers in the liver, intestine and/or bladder where they elicit granuloma formation and fibrosis [1]. Clinical outcomes span gastrointestinal and liver pathologies, anaemia, undernutrition, growth retardation, genitourinary disease and a heightened risk for co-morbidities. This burden encumbers third world economies with an annual loss of several million disability-adjusted life years [2]–[4]. The key treatment for schistosome infections is the drug praziquantel (PZQ). PZQ is a synthetic tetracyclic tetrahydroisoquinoline derivative discovered over 30 years ago to confer anthelminthic activity [5]–[7] by evoking a spastic paralysis of the adult worms [8]. The low cost (∼$0.07/tablet) yet high cure rate associated with PZQ underpins current strategies for increasing PZQ distribution to reduce the burden of schistosomiasis [9], but obviously continued efficacy of PZQ is critical for the success of these initiatives. From a drug development perspective, it remains problematic that despite three decades of clinical use, the target of PZQ remains ambiguous and synthesized structural derivatives prove consistently less efficacious [5]–[7], [10], [11]. Resolution of the target and effector mechanisms of PZQ would be massively helpful for identifying new drug targets that exploit vulnerabilities within the broader PZQ interactome. Recently, we have attempted to bring fresh insight into the mechanism of action of PZQ by studying an unusual impact of this drug on regeneration of a free living planarian flatworm (Dugesia japonica), a representative of a model system widely utilized by basic scientists as a model for regenerative biology [12], [13]. This line of investigation grew from the serendipitous finding that PZQ exposure invariably caused regeneration of worms with two heads (‘bipolar’), rather than worms with normal anterior-posterior (‘AP’, head to tail) polarity [14]. The capacity of PZQ to evoke this complete AP axis duplication was phenocopied by several Ca2+ signaling modulators, a relationship underpinned by the demonstration of PZQ-evoked Ca2+ uptake in native planarian tissue [14], [15]. The tractability of planarians to in vivo RNAi methods allowed mechanistic interrogation of various Ca2+ entry pathways, and this approach revealed the bipolarizing efficacy of PZQ depended on the expression of neuronal voltage-operated Ca2+ channel (Cav1) isoforms [14], [15]. These observations were intriguing in the context of schistosome biology, as PZQ is well documented to cause Ca2+ entry in schistosomes [8], [16], [17] and PZQ has been shown to activate Ca2+ entry via modulation of a heterologously expressed schistosome Cav accessory subunit [18], [19]. But how Ca2+ entry engages acute and chronic [20]–[22] downstream signaling pathways in either planarians or schistosomes is less clear, with resolution of this broader PZQ interactome key for identifying new druggable targets and vulnerabilities for chemotherapeutic exploitation [17]. Here, we evidence a Ca2+-dependent phenology of PZQ action between these two quite different models. We propose the same Ca2+ entry and downstream pathways are engaged by PZQ in planarians and schistosomes, and the mechanistic interrelationship underpinning these different outcomes (death in schistosomes, axis duplication in planarians) augers predictive value for discovery of new anti-schistosomal agents. For example, in planarians, we demonstrate the planarian AP axis duplication phenotype results from coupling of Cav1A activity to bioaminergic signaling. Modulators of regenerative polarity which impact dopaminergic and serotonergic pathways in planarians are effective against schistosomes, and reciprocally recently discovered drug leads active against schistosomes (for example, PKC and GSK3 modulators) regulate AP specification in planarians. As unexpected phenologs [23], this discovery underscores the utility of basic research on axis patterning mechanisms in the tractable planarian system for the discovery of novel antischistosomal drug leads, and more broadly mechanistic insight into the signaling pathways engaged by PZQ, a key human therapeutic. Exposure of excised trunk fragments to PZQ caused regeneration of viable, two-headed flatworms (Figure 1A), an effect previously shown to relate to modulation of neuronal voltage-operated calcium (Cav) channels [14], [15]). Given the role of Ca2+ entry in synaptic and dendritic exocytosis [24], [25], we hypothesized that PZQ-evoked Ca2+ entry impacted neurotransmission and thereby stem cell behavior, consistent with a ‘neurohumoral’ model for regulation of planarian stem cell proliferation proposed two decades ago [26]. To test this idea, we used loss-of-function (in vivo RNAi) and pharmacological methods to interrogate whether different planarian neurotransmitters mimicked the PZQ-evoked bipolarity effect. Figure 1B schematically summarizes the major neurotransmitter classes in flatworms [27]–[29], of which neuropeptides predominate by number. A recent characterization of planarian bioactive peptides revealed >50 prohormone genes, the vast majority being neuronally expressed with over 250 discrete peptides generated from these precursors [30]. Further, bioinformatic prediction supports at least 130 planarian neuropeptide targeted G protein coupled receptors [31]. This expansive neuropeptidergic arsenal co-exists with several ‘classic’ neurotransmitter families more familiar to mammalian neurophysiologists. The largest group of these transmitters are the biogenic amines, a group of protonated amines including serotonin, histamine, catecholamines (notably dopamine) as well as tyramine and octopamine, two phenolamines widely used as invertebrate neurotransmitters [27], [32]. Roles for acetylcholine (ACh) and amino acids (glutamate, GABA) are also evidenced [27], [32]. To test the involvement of these different neurotransmitter families as PZQ effectors, we used in vivo RNAi to knockdown key enzymes involved in their synthesis. Knockdown of prohormone convertase 2 (PC2, [33]), an enzyme required for motility [34] and neuropeptide processing [30], failed to impact the penetrance of PZQ-evoked bipolarity (Figure 1C). Similarly, knockdown of glutamate decarboxylase (GDC, to decrease planarian GABA levels [35]), and choline acetyltransferase (CAT, to deplete ACh [36]), failed to modulate the penetrance of PZQ (Figure 1C). Negative results were also obtained following RNAi of tyramine-β-hydroxylase (TBH) and tyrosine/histidine decarboxylase (T/HDC). These data were also consistent with the outcomes of pharmacological experiments where application of the phenolamines tyramine and octopamine failed to perturb AP polarity (Table 1). In contrast, results with other biogenic amines were more intriguing – knockdown of tyrosine hydroxylase (TH) attenuated the ability of PZQ to evoke two-headed worms, whereas knockdown of tryptophan hydroxylase (TPH) increased PZQ-evoked bipolarity (Figure 1C). TH is the rate-limiting enzyme of catecholamine synthesis, catalyzing the conversion of tyrosine to L-dihydroxyphenylalanine (L-DOPA), whereas TPH converts tryptophan to 5-hydroxytryptophan, the first step in 5-HT synthesis. Knockdown of TH in D. japonica decreases dopamine without impacting 5-HT production [37], while knockdown of TPH decreases 5-HT but not dopamine [38]. These RNAi results suggest that PZQ activity is mimicked by dopaminergic activity (TH RNAi) to promote head regeneration, and this action is opposed by serotonergic signaling (TPH RNAi). On the basis of this hypothesis, we proceed to screen modulators of dopamine and 5-HT receptors: dopaminergic stimuli should phenocopy the bipolarizing activity of PZQ, while PZQ action should be opposed by serotonergic agonists. While this is a reasonable approach, care must be taken in assuming the specificity of agents established in mammal models transfers to flatworm systems. Flatworms may express more bioaminergic receptors than humans [31], and the few flatworms receptors that have been successfully expressed and pharmacologically profiled [39] underscore the risks of assuming similar drug activities to those assigned in mammals. Keeping this caveat in mind, we nevertheless used a pharmacological approach but accrued evidence with multiple ligands and used secondary validation assays to best mitigate this problem. Below, we first describe results of drug assays assuming specificities based upon mammal data, and then we return to the issue of validating ligand specificity against particular neurotransmitter pathways. A range of compounds were screened for effects on AP polarity (Table 1), and these investigations yielded the following observations. First, the exclusion of individual neurotransmitter families on the basis of RNAi results (Figure 1C) received further support from pharmacological screening, as most modulators of adrenergic, GABAergic, glutaminergic, histaminergic and cholinergic pathways failed to impact regenerative polarity (Table 1). Second, bromocriptine, a potent D2 agonist in mammalian systems, produced two-headed regenerants at high penetrance (maximal effect ∼85±5% bipolar, Figure 1D & 1E), with an EC50 of 220 nM compared with an EC50 of ∼40 µM for PZQ (Figure 1F). Other dopaminergic modulators yielded a low, but robust, proportion of two headed worms including apomorphine (a non-selective dopaminergic agonist in mammals) and dopamine itself (Figure 1D & 1E). Third, haloperidol, a traditional antipsychotic and known inhibitor of dopaminergic signaling in planaria [40], blocked the bipolarizing activity of both bromocriptine and PZQ (Figure 1F, inset). Fourth, 5-HT blocked head regeneration, an effect observed with 5-HT, the synthetic ligand 8-OH DPAT (a mammalian 5-HT1A agonist) and a serotonin-specific reuptake inhibitor (SSRI, fluoxetine, Figure 1D & 1E), all of which blocked the bipolarizing effect of PZQ (IC50 ∼147 µM, 111 nM and 230 nM, Figure 1G). In contrast, mianserin (a 5-HT antagonist in flatworm [41]–[43] and mammalian systems) yielded a small proportion of two-headed worms (Figure 1D&E). Given the effects of bromocriptine, we further investigated the characteristics of bromocriptine efficacy in planaria. First, bromocriptine exhibited a similar kinetic action to that observed with PZQ (Figure 2A), suggesting a similar action early in regeneration. Second, while knockdown of Cav1A attenuated PZQ-evoked bipolarity, bromocriptine-evoked bipolarity persisted in Cav1A RNAi worms (Figure 2B). This surprising result is consistent with the idea that bromocriptine activation of head signaling pathways occurs downstream of Cav1A function. For example, if PZQ-evoked Ca2+ entry [15] activates neurotransmitter release, then the bipolarizing efficacy of bromocriptine should persist at downstream receptors even if Ca2+ entry is impaired. Third, given concerns about presumptions of similar pharmacological effects between mammalian and flatworm systems, we investigated whether bromocriptine exhibited affinity for dopaminergic systems in planaria by performing 3H-dopamine displacement assays. Specific 3H-dopamine binding, defined by complete displacement with cold dopamine (IC50 = 1.5±0.5 µM), was inhibited by bromocriptine and other head-promoting agents (haloperidol and apomorphine, Figure 2C). The extent of 3H-dopamine displacement by maximally effective concentrations of haloperidol and apomorphine was greater (>80% of specific binding) than observed with bromocriptine (∼40% of specific binding at 10 µM). This indicated bromocriptine may exhibit selectivity for only a subset of dopaminergic targets compared to the broader and more complete binding inhibition observed with the other agents. Finally, we investigated the impact of agents presumed to impact neurotransmitter levels (reserpine, fluoxetine) via HPLC. Figure 2D shows that fluoxetine (a 5-HT reuptake inhibitor on the basis of mammalian and schistosome literature [44], [45]) increased 5-HT levels in regenerating planarian trunk fragments, consistent with the inhibitory effects of 5-HT (and fluoxetine) on head regeneration (Figure 1G). In contrast, reserpine exposure depleted 5-HT in regenerating fragments (Figure 2C), an opposing outcome consistent with the differential polarity effects of these drugs on head regeneration (reserpine vs fluoxetine, Figure 1D&E). Collectively, these pharmacological data support the model derived from RNAi data (Figure 1C) where dopaminergic signaling mimics and serotonergic activity opposes PZQ action. The distinct phenotypic outcomes of dopaminergic and serotonergic modulation are also consistent with observations that these neurotransmitter networks in planarians are morphologically discrete [28]. These discoveries piqued our interest since dopaminergic and serotonergic ligands have recently emerged as hits in drug screens against various schistosome life cycle stages [46], [47]. Figure 3A collates examples of recent drug screening data to show how efficacious drug hits are distributed relative to the functional representation of drugs screened [46], [47]. The top three functional categories represent dopaminergic and serotonergic ligands followed by regulators of ion channel activity, notably Cav channel modulators. This triumvirate parallels the PZQ-engaged components in planarians in this study (bioaminergics, Figure 1) and previously (Ca2+ channels, [14], [15]). As such, we propose the distinct phenotypes - PZQ-evoked bipolarity in planarians and PZQ-evoked toxicity against schistosomes - represent unexpected yet orthologous phenotypes (‘phenologs’, [23]) resulting from engagement of the same fundamental Ca2+-triggered interactome in each system. Although PZQ-evoked Ca2+ entry is evoked via similar mechanisms (Cav1A) it is harnessed in the two organisms to yield differential outcomes (‘death’ versus ‘axes’). The utility of this phenology is its predictive value. As both outcomes derive from the same effector network, basic research on axis patterning in planarians may harbor potential for discovering new agents effective as antischistosomals. This assertion can be tested by asking whether other antischistosomals cause planarian bipolarity, and reciprocally, whether bipolarizing agents in planarians are active against schistosomes. Do other antischistosomal compounds cause planarian bipolarity? To test this, we identified the next most prevalent category from the schistosome drug screening datasets, which was the ‘phosphorylation’ category (Figure 3A). The predominant group of compounds within this category were several drugs that target protein kinase C (PKC), and a couple of singleton kinase inhibitors, including one targeting glycogen synthase kinase-3 (GSK3). We investigated the role of both kinases to resolve any impact on planarian regenerative polarity (Figure 3A). First, the PKC activators phorbol-12-myristate-13-acetate (PMA), phorbol-12,13-dibutyrate (PDB) and oleoyl-acetyl-glycerol (OAG) all produced bipolar worms (penetrance ∼5–55% respectively, Figure 3B), while the PKC inhibitor calphostin C [48] inhibited PZQ-evoked bipolarity (Figure 3C). To complement the pharmacological data with molecular insight, we cloned several planarian PKC isoforms and diacylglycerol kinase (DAGK) and investigated their roles in PZQ-evoked bipolarity by RNAi. Knockdown of DAGK, which opposes PKC activity via the degradation of DAG, potentiated the penetrance of sub-maximal doses of PZQ; while RNAi of a conventional PKC isoform, but not a novel and atypical PKC, attenuated PZQ evoked bipolarity (Figure 3C). The involvement of a Ca2+-regulated PKC was also consistent with the observation that the penetrance of PMA in yielding bipolar regenerants was Ca2+ dependent (Figure 3D). Similarly, alsterpaullone (ALP), a GSK-3 inhibitor also phenocopied PZQ in regenerative assays, producing a low frequency of two headed worms and synergistically potentiating sub-maximal doses of PZQ (Figure 3E). The small molecule GSK3 agonist DIF-3 [49] displayed the opposing action, inhibiting PZQ-evoked bipolarity (Figure 3E). Therefore, both these targets in the ‘phosphorylation’ category prioritized from the schistosomal screening literature (Figure 3A) were resolved to miscue planarian AP polarity during regeneration. Are drugs that miscue planarian regeneration deleterious to schistosomes? To investigate this issue, schistosomules (juvenile parasites) were exposed to compounds first identified in planarian regenerative assays (Figure 4A). Schistosomules normally exhibit a basal level of spontaneous contractile activity (Figure 4B), which provides a simple phenotype for assaying drug action and paralysis, an outcome integral to the elimination of schistosome infections [46]. Bromocriptine caused a rapid paralysis of schistosomules, an effect that phenocopied the action of PZQ (Figure 4B). This effect was dose-dependent (Figure 4B&C). Other compounds that yielded planarian bipolarity were also found to impair schistosomule contractility, including apomorphine, mianserin and reserpine (Figure 4B). In contrast, application of exogenous serotonin and other ligands that inhibited planarian head regeneration (e.g. 8-OH DPAT and fluoxetine) resulted in hyperactivity (Figure 4C). Quantification of the action of these agents which inhibited and stimulated schistosomule activity is collated in Figures 4D&E respectively. Therefore, not only were both classes of bioaminergic compounds efficacious against schistosomules, but the dopaminergic and serotonergic ligands evoked divergent phenotypes in each model: paralysis versus hyperactivity (schistosomules), compared with ‘two-headed’ versus ‘no-head’ regenerants (planaria). Beyond the conservation of single genes as nodes in a signaling pathway, broader network architectures are conserved between diverse organisms. While the phenotypic outputs of these networks are diverse, their common architecture provides the mechanistic basis for predictive phenology [23]. We suggest these divergent PZQ-evoked outcomes (death versus axes) represent unexpected Ca2+-dependent phenologs initiated by small molecule activation of a signaling node (Cav1A) within a shared bioaminergic interactome (Figure 5A). This conservation infers reciprocal predictive value for both discovery of new antischistosomal compounds, and reciprocally new signalling pathways impacting anterior-posterior signaling in planarians. We illustrate this principle here by highlighting de novo new compounds effective against schistosomules (bromocriptine) and new druggable targets (bioaminergic signaling) as the downstream PZQ-evoked interactome is revealed in the more tractable planarian model. PZQ engages similar pathways in these different platyhelminths such that chemical/functional genetic approaches in planarians can assist in discovering next generation antischistosomals and resolving their molecular action. This line of reasoning is analogous to a longer history of studies exploiting C. elegans for comparative insight into new drugs targeting parasitic nematodes, and this experience underscores both the utility of this approach but also the frustration in harvesting viable clinical leads from a large number of efficacious compounds in both nematode models [50], [51]. Reciprocally, this unexpected phenology can reveal new modulators of AP patterning from the schistosome screening literature (e.g. PKC, GSK3). Such insight from schistosome life cycle drug screens will be of utility for understanding the process of in vivo stem cell differentiation and CNS regeneration in response to injury that are inherent to the remarkable regenerative prowess of planarians. Indeed, resolution of the coupling of specific neuronal Cav channels to defined neurotransmitters integrates our studies of PZQ-evoked Cav activity [14], [15] with an older literature supporting a role for bioamines in planarian regeneration [52]. But how is small molecule activation of Cav1A in one organism deleterious, but the same Ca2+ influx process harnessed physiologically in another to regulate polarity during regeneration? We speculate the same PZQ-evoked interactome differentially couples to these outcomes because of the different ionotropic channel portfolio supporting cellular excitability in the two organisms. Planarians express a surprisingly broad array of voltage-gated entry channels - five unique Cav channels in addition to Nav channels (Figure 5B). This broad channel repertoire likely permits subfunctionalization of Cav1A activity within a broad organismal complement of voltage-gated channels in planarians to yield a physiological exploitable Cav1A dependent Ca2+ influx. In contrast, schistosomes express a more limited portfolio of voltage-sensitive channels, lacking both Nav and LVA Cav channels (Figure 5B). The more limited gene repertoire of these parasites imparts a dependency and thereby vulnerability to Cav1A activity within their smaller ionotropic channel portfolio. In this context, it is intriguing that both muscle contraction and tegumental damage are Ca2+ triggered phenomena in adult schistosomes (reviewed in [17]), such that Ca2+ dysregulation may serve as a common nexus predictive of in vivo antihelmintic activity. Further insight into this problem will be provided by understanding how acute Ca2+-dependent effects evoked by PZQ in different schistosome tissues regulate both acute downstream targets (bioaminergic receptors and their second messenger coupling) and the relevance of more chronic Ca2+ dependent transcriptional effects [20], [22], e.g. CamKII [21], that have emerged from recent mRNA profiling analyses. In conclusion, exploitation of this Ca2+ dependent phenology should rekindle interest in drugs such as bromocriptine, and the druggability of their cognate bioaminergic receptors, as an avenue for resolving novel antischistosomals and modulating in vivo stem cell behavior during regeneration. A clonal line of Dugesia japonica (GI strain) was maintained at room temperature and fed strained chicken liver puree once a week [53]. Regenerative assays were performed using 5 day-starved worms in pH-buffered Montjuïch salts (1.6 mM NaCl, 1.0 mM CaCl2, 1.0 mM MgSO4, 0.1 mM MgCl2, 0.1 mM KCl,1.2 mM NaHCO3, pH 7.4 buffered with 1.5 mM HEPES) as described previously [53]. Drugs were sourced as indicated in Table 1, and used either at the highest concentrations which did not impact worm viability, or at lower concentrations if such treatments elicited an effect of maximal penetrance. Planarian regenerative phenotypes were archived using a Zeiss Discovery v20 stereomicroscope and a QiCAM 12-bit cooled color CCD camera. Total RNA was isolated from 50 intact planarians using TRIzol® and poly-A purified using a NucleoTrap mRNA mini kit. cDNA was synthesized using the SuperScript™ III First-Strand Synthesis System (Invitrogen). Gene products were amplified by PCR (LA Taq™ polymerase), ligated into the pGEM®-T Easy vector (Promega) for sequencing, and subcloned into the IPTG-inducible pDONRdT7 RNAi vector transfected into RNase III deficient HT115 E. coli. In vivo RNAi was performed by feeding [53], and a Schmidtea mediterranea six-1 (Smed-six-1) construct, which did not yield a phenotype in D. japonica, was used as a negative control. RNAi efficiencies varied between different genes, but mRNA knockdown typically ranged anywhere between 20–80%. Targeted sequences: tyrosine hydroxylase (NCBI accession numbers AB266095.1, 136–1657 bp), tryptophan hydroxylase (AB288367.1, 4–1623 bp), tyramine beta-hydroxylase (671–1629 bp), tyrosine/histidine decarboxylase (FY934632.1, 26–685 bp), glutamate decarboxylase (AB332029.1, 154–1937 bp), choline acetyltransferase (AB536929.1, 74–1175 bp), prohormone convertase 2 (PC2 (1–2285 bp), Cav1A (HQ724315.1, 2229–4133 bp), Smed-six-1 (AJ557022.1, 1–506 bp). Protein kinase C (PKC) sequences and DAGK were cloned from planarian ESTs displaying homology to Schistosoma mansoni PKC isoforms - cPKC (FY950278.1, FY947802.1, FY970060.1), aPKC (FY933556.1, FY941429.1), nPKC (FY934640.1) and DAGK (FY953983, FY959647.1, and BP187372.1). Biomphalaria glabrata snails exposed to mirarcidia (NMRI Puerto Rican strain of Schistosoma mansoni) were obtained from the Biomedical Research Institute (Rockville, MD) and maintained at 26°C for 4 to 6 weeks. Matured cercaria were shed into aged tap water (40 ml) by exposure to light (1.5 hrs) and subsequently transformed into schistosomules [54]. Briefly, cercaria were separated from debris by filtration (47 µm) and then captured onto a 25 µm filter prior to resuspension in aged tap water with an equal volume of DMEM. Cercaria tails were sheared by three rounds of vortexing (45 sec), each followed by incubation on ice (3 min) prior to tail removal by Percoll column centrifugation (24 ml Percoll, 4 ml 10× Eagle's minimum essential medium, 1.5 ml penicillin-streptomycin, ml of 1M HEPES in 0.85% NaCl, 9.5 ml distilled water) at 500 g (15 mins, 4°C). The tail-containing supernatant was discarded and the pellet-containing bodies were washed three times in DMEM (400 g, 10 mins), resuspended in modified Batch's media [55] and transformed into schistosomules (incubation at 37°C/5% CO2). For contractility assays, drugs were solubilized in DMSO and diluted in pre-warmed modified Batch's media. While detailed protocols for quantifying aspects of worm dynamics in adult worms [21], or higher throughput screening of schistosomules [56] have been developed, the effects on schistosomule activity were simply quantified here using a custom written plugin (wrMTrck) in ImageJ to using resolve schistosomule body length (major axis of an ellipse) over time following drug exposure (30 min), just as in [39].Videos were captured using a Nikon Coolpix 5700 camera affixed to a Nikon Eclipse TS100 microscope. Typically, for a single video ∼7–10 schistosomula were measured within the field of view (10× microscope objective) over a 2 minute recording period. Data represent means for analysis of results from three independent treatments. Planarian membrane fractions were prepared by homogenizing worms on ice (∼1000 worms/prep) in HEPES (20 mM) supplemented with cOmplete™ protease inhibitor cocktail (Roche). Cellular debris was pelleted by centrifugation (8000 g for 5 mins) and the resulting supernatant was centrifuged (56,000 g for 45 mins) to yield a pelleted membrane fraction. This material was resuspended (20 mM HEPES, with protease inhibitors) to a final protein concentration of ∼5 µg/µl and stored at −80°C. Binding assays were performed on planarian membrane protein (50 µg) with 26 nM 3H-dopamine (specific activity 21.2 Ci/mmol, Perkin Elmer). Indirect binding assays were performed with various ligands (bromocriptine, 10 µM; haloperidol, 100 µM; apomorphine, 10 µM) in TE buffer (1 mM EDTA, 50 mM Tris-HCl, pH 8.3; final volume of 500 µl). Samples were incubated on ice for 15 minutes, after which time 500 µl PEG (30%) and 20 µL IgG (25 mg/ml) were added and samples centrifuged (20,000 g for 5 mins). The resulting pellet was washed (PEG, 15%), centrifuged (20,000 g for 5 mins) and solubilized in TE (200 µL, containing 2% Triton X-100). Displacement was measured by liquid scintillation counting and nonspecific binding assessed by subtraction of values in samples incubated with cold dopamine (1 mM). All centrifugation steps were performed at 4°C. Thirty planarian trunk fragments were amputated and incubated with or without specific drugs for 24 hrs, after which time media was removed and replaced with ascorbic acid (300 µl, 1% m/v). Samples were then lysed by three successive freeze-thaw cycles and cellular debris pelleted by centrifugation (10,000 g for 5 mins). The resulting supernatant was then filtered (0.45 µm filter plate, Millipore) by centrifugation (3,000 g for 10 mins) and the filtrate (180 µL) supplemented with 0.5M HClO4 (20 µl, 500 mM final concentration). The samples were mixed and injected by an autosampler into an Agilent 1200 HPLC apparatus, with a 5 µm, 4.6×150 mm Eclipse XDB C18 column attached to a Waters 2465 electrochemical detector with a glassy carbon-based electrode. The current range was set at 50 nA with a working potential of 0.7 V versus an in situ Ag/AgCl reference electrode. The mobile phase mixture (13 mg/L of the surfactant sodium octylsulfate, 170 µL/L dibutylamine, 55.8 mg/L Na2EDTA, 10% methanol, 203 mg/L sodium acetate anhydrous, 0.1M citric acid, and 120 mg/L sodium chloride) was ran at a flow rate of 2 ml/min. The area underneath the peaks was analyzed for total amount of serotonin and dopamine. Results were normalized to sample protein concentration determined by Bradford assay (Thermo Scientific). Data were analyzed using two-tailed, unpaired t-tests, and presented as mean ± standard error of the mean from at least three independent assays, except where indicated. Differences were considered significant at p<0.05 (*), p<0.01 (**).

10.1371/journal.pcbi.1006969

A global map of the protein shape universe

Proteins are involved in almost all functions in a living cell, and functions of proteins are realized by their tertiary structures. Obtaining a global perspective of the variety and distribution of protein structures lays a foundation for our understanding of the building principle of protein structures. In light of the rapid accumulation of low-resolution structure data from electron tomography and cryo-electron microscopy, here we map and classify three-dimensional (3D) surface shapes of proteins into a similarity space. Surface shapes of proteins were represented with 3D Zernike descriptors, mathematical moment-based invariants, which have previously been demonstrated effective for biomolecular structure similarity search. In addition to single chains of proteins, we have also analyzed the shape space occupied by protein complexes. From the mapping, we have obtained various new insights into the relationship between shapes, main-chain folds, and complex formation. The unique view obtained from shape mapping opens up new ways to understand design principles, functions, and evolution of proteins.

Proteins are the major molecules involved in almost all cellular processes. In this work, we present a novel mapping of protein shapes that represents the variety and the similarities of 3D shapes of proteins and their assemblies. This mapping provides various novel insights into protein shapes including determinant factors of protein 3D shapes, which enhance our understanding of the design principles of protein shapes. The mapping will also be a valuable resource for artificial protein design as well as references for classifying medium- to low-resolution protein structure images of determined by cryo-electron microscopy and tomography.

Proteins are the primary workers in a living cell, involved in transportation, catalysis, signaling, energy production, and many other processes. Classification of protein structures provides fundamental information for our understanding of the principles that govern and determine protein structures, which is one of the essential goals of structural biology and protein bioinformatics. Understanding the repertoire of protein structures is also of practical importance for artificial protein design, which has broad applications in therapeutics such as designing inhibitors [1] and small peptide drugs [2], as well as the development of biomaterials [3]. Conventionally, protein structures have been classified based on their main-chain conformations and evolutionary history [4–6]. Such classifications led to several important observations including the number of different protein folds in nature [7–9], distributions of folds in genomes [10,11], and the relationship between sequence and structure conservations [12]. The discovery of the limited number of folds yielded stimulating discussions on the mechanism behind it [13,14]. Furthermore, such studies contributed to the birth of a very successful paradigm of threading [15] and more recent fragment-based approaches [16] in protein structure prediction. Some recent studies mapped protein structures into a low-dimensional space to reveal high-level organization of the variety of protein structures. Kim and his colleagues computed structural similarity with DALI, a residue-contact map-based structure comparison method [17], and mapped representative proteins into a 3D space using multidimensional scaling [18,19]. Osadchy and Kolodny represented protein structure domains as a vector indicating the occurrence of fragments in the structure [20]. In both works, the maps exhibited a trend where structures formed clusters according to their fold classes, α, β, α/β, and α+β, and others, which is reasonable but expected. Here, we present a global mapping of 3D surface shapes of single proteins and complexes. In contrast to the previous works [18–20] that considered main-chain conformation to define the structural similarity, the use of surface shape representation led to findings of previously undescribed relationships between protein shape, fold class, and assemblies. We perform a thorough analysis of surface shapes in consideration of the rise of medium- to low-resolution structures determined by electron tomography [21] and cryo-electron microscopy (cryo-EM) [22]. Classifying protein structures by shape would be more relevant to functional classes of proteins than using conventional main-chain conformations since protein functions such as binding and catalysis occur at the surfaces of proteins. As shown in our previous study [23], functionally related proteins often share similar global surface but with low sequence and backbone conformation similarity. An illustrative example is DNA topoisomerase I from human and E. coli. Despite their low sequence identity and structure similarity, both of them share a characteristic pore to encircle DNA double strand. This function similarity can be easily captured by shape descriptors, but not captured by conventional main-chain conformation approach. Protein surface shapes were represented with 3D Zernike Descriptors (3DZD), mathematical moment-based invariants of 3D functions [23]. 3DZD has been demonstrated efficient for various biomolecular structure comparisons [24], including comparisons of EM maps [25]. Another critical difference between the current study and the previous works is that we analyzed protein complexes in comparison with single proteins. The shape mapping of single-chain and complex protein structures with 3DZD yielded a unique landscape of protein structure space that was not explored before. Dominant features that characterize protein shape are the eccentricity, which is the degree of elongation of shapes, and the number of domains. Symmetry groups are another feature that affects the shape in the case of protein complexes. A detailed analysis of the principal axis corresponding to the elongation of protein shape has suggested that proteins are required to form multimers if their shape is elongated over a certain degree. Overlapping the shape space occupied by single proteins and complexes identified shapes that are only possible in complexes. The unique view obtained from the current shape mapping leads to a more comprehensive understanding of building mechanisms, evolution, and design principles of proteins. We first discuss the protein surface shape space for single chain proteins, followed by the analysis of shapes of protein complexes. Fig 1 overviews the 3D space mapping of 6,841 representative single-chain protein shapes. The surface shape of each protein was represented with the 3DZD, a rotation-invariant mathematical descriptor of 3D protein surface shape, and mapped to a 3D space using principal component analysis (PCA). 3DZD is based on a series expansion using 3D basis functions, which represents the target 3D shape by a weighted combination of the basis functions. The rotation-invariance is achieved by computing a norm of the coefficient values that are assigned to the basis functions (see Methods). PCA locates similar protein shapes close to each other in the space. The color of points indicates the eccentricity of the shapes, which quantifies how much a shape deviates from a sphere, with a higher value (red) assigned for more elongated structures (the maximum value is 1) and 0 for a perfect sphere (blue). A video clip of the 3D distribution (Appendix, S1 Movie) is also provided, along with an interactive PyMOL file (Appendix, S1 Pymol File) to help readers better understand and further investigate the 3D shape distribution. Many entries of the Protein Data Bank (PDB) [26] contain only a fraction of the whole structure; thus, we thought it may be possible that the distribution we see in Fig 1A is biased toward surface shapes of structure fragments. For comparison, we also show in the inset figure of Fig 1A the distribution of 2,366 almost complete protein structures, which have at least 95% structure coverage of the whole proteins. The projection was made with PCA independently for this high-coverage dataset. As shown, the distribution of the high-coverage protein dataset is very similar, indicating that partial structures do not bias the distribution of the single-chain dataset. Next, we discuss the shape space of protein complexes (Fig 6, S2 Movie, S4 Pymol File). The dataset of protein complexes contains 5,326 non-redundant structures. We obtained the biological units of complexes from PISA. As in Fig 1, the color indicates the eccentricity of shapes. The complex shape space is overall very similar to the single-chain shape space (Fig 1), with the majority of structures located around the globular region near the origin of the axes and a tail region dominated by elongated shapes (the region with many red points). On the other hand, some differences were observed between the complex and single-chain distributions. The protein complexes have more spherical shapes than the single-chain distribution (data points in dark blue in the mapping) (Fig 6A and 6B). The eccentricity histograms for the single-chain and complex datasets (S2 Fig) verify this observation, which shows that the complex dataset contains highly spherical shapes with a low eccentricity. While there are no single-chain proteins with an eccentricity below 0.2, the complex dataset includes 72 such cases. The differences between the shape spaces of the single-chain proteins and complexes become apparent when they are superimposed (Fig 7, S5 Pymol File). To compare the size of the spaces occupied by the two datasets, the space was segmented into cubes of 1 axis unit edge length, and cubes were counted if they were occupied by the proteins in the datasets. Among all the cubes (3,895 cubes) that were occupied by at least one protein, 24.5% were filled by both single-chain and complex structures while 26.1% and 49.4% were occupied by only single-chain proteins and complex structures, respectively. Thus, the complex structure dataset occupies a larger space than the single-chain protein dataset. Fig 7C shows two example structures each from single-chain specific and complex-specific areas in the shape space. In the single-chain dataset, structures with a flexible tail (e.g. 3gzrA) were observed. Another example shown is 3e7kA, a narrow, elongated shape with a single helix, which is obviously very unique in single-chain proteins. On the other hand, highly spherical or symmetrical shapes are unique in protein complexes. 1yzv shown in Fig 7C has a spherical shape with the octahedral symmetry and 4ldm has a two-layer tube-like structure. The wide spread of complex structures suggests that assembling subunits into complexes can increase the range of attainable structures. Fig 6C and 6D annotate representative structures in the complex shape space. The outskirts of the distribution in the first quadrant (i.e. top right) in Fig 6C includes shapes of the almost perfect sphere (e.g. 1yzv, 2y3q), two layers of circular ring-like arrangements (e.g. 1lnx), and cube-like shapes (e.g. 3hsh). In the second quadrant (top left) several symmetrical “spiky” shapes with multiple protrusions are observed (e.g. 4fdw, 3r88). Close to the origin (0, 0, 0), dimeric complexes (e.g. 1hzt, 2zum) are observed. Fig 6D views the complex shape mapping from a different direction, showing the tail region occupied by structures with elongated shapes. They include protein structures of different fold classes, e.g. long α helices (e.g. 4cqi, 3okq), β structures (e.g. 3aqj), mixtures of them (e.g. 1rfx), and tube-like shapes (e.g. 2wie, 2zbt). In Fig 10, we examined how the eccentricity, the size of pockets, and the Vp/Vc ratio distribute relative to the number of amino acids for protein structures in the single-chain and the complex structure datasets. The first panel (Fig 10A) shows that very low eccentricity, i.e. highly spherical shapes, are achieved only by complex structures, which confirms the observation in earlier sections. Complex structures tend to have larger pockets as shown in Fig 10B. Naturally, larger protein complexes are capable of having larger pockets. Furthermore, a closer look at the plot around the protein length of up to 1,000 residues indicates that complex structures tend to have larger pockets than single-chains even when proteins of the same size are compared. Fig 10C examines the Vp/Vc ratio, the ratio of the protein volume relative to the convex hull of the protein. Overall, single-chain proteins and complex structures show similar distributions, but there are more complex structures observed in the lower end of the Vp/Vc ratio. Panels D, E, F illustrate the difference of shapes with a small Vp/Vc ratio between the two datasets. In the case of single-chains, a small Vp/Vc ratio occurs for flexible proteins such as 3ag3I (Fig 10D) while for complexes typical such shapes are symmetrical ones with protrusions (Fig 10E) and shapes with a large hollow inside as shown in Fig 10F. In this study, we have constructed a mapping of the protein structure space for the first time by considering the overall surface shape of both single-chain and complex proteins. The shape space visualized in this work would give an impression that the protein shape space is continuous, but this is not specific to the protein surface shape representation. Indeed, earlier works that mapped protein structures considering main-chain conformations also show continuous structure distributions [17–20]; and moreover, there exists active discussion on the continuity [29] or the many-to-many similarity relationship [30] of the protein structure space. Analogous to well-established protein main-chain structure classifications, such as SCOP [5] and CATH [4], this work will lead to a new classification for protein shapes at a medium to low resolution, which are being accumulated at an increasing pace by cryo-electron tomography and cryo-EM. By establishing the classification from the distribution of the protein shapes, for example, we will be able to take a census of protein shapes, that is, to count the number of specific protein shapes in organisms and compare across different organisms [31]. The observed variety of protein shapes in this work will also be useful for designing protein representations used in a cell-scale physical simulation of biomolecules [32]. Rather than using an overly simplified molecular representation, as is usual for such a simulation, one could diversify protein shapes in the simulation box by sampling structures from different locations in the shape space (Fig 1 and Fig 6). Last but not least, this work has strong implications for protein design. Our study indicates that a protein shape can be realized with utterly different backbone conformations that even belong to different fold classes as shown in Table 1 and S1 Fig. Also, the shape mappings of single chains and complexes revealed regions in the shape space that are not occupied by either of them, or are occupied only by complex shapes (Fig 7). Shapes that correspond to the former may be difficult to construct with proteins, and other materials such as DNAs or polysaccharides may be required, while those in the latter region may be better designed using complexes rather than a single-chain protein. In the coming age of medium- to low-resolution biomolecular structures, protein design needs a novel way of viewing biomolecular shapes. We expect that this work makes a unique and significant contribution by providing a foundation of understanding the protein shape universe. The representative set of single-chain protein structures was selected from a PISCES culled list with a resolution cutoff of 2.2 Å, an R factor cutoff of 0.2, and a pairwise sequence identity cutoff of 25% [33]. From 7,260 chains in the list, we removed short chains with less than 40 amino acids. We have also removed proteins that have a large spatial gap, i.e. structures having more than one cluster when Cα atoms were clustered with a 9 Å cutoff. We further removed 82 chains were further removed from the list because their sequences had more than 25% sequence identity to other chains. This process yielded a dataset of 6,841 non-redundant protein structures. From this dataset, we prepared another dataset by pruning structures that include less than 95% of residues relative to the whole chain length. The protein lengths were obtained from UniProt [34]. There are 2,366 chains in this high-coverage single chain dataset. For each chain, fold class was assigned following CATH. Also, by referring to PISA [27], we assigned biological unit information. This pruned dataset was shown in inset of Fig 1A and 1B. From PDB, we identified structures that exist as a complex as defined in PISA and downloaded the first biological unit (BU). The same resolution, R factor, and length cutoffs as in the single chain dataset were applied. A complex is considered as redundant if there is another complex with the same number of chains and corresponding chains between them have over 25% sequence identity. Among redundant complex entries, we chose the one with the highest resolution and the lowest R factor. This procedure yielded 5,326 complexes. Symmetry information for complexes was obtained from PDB if the BU of the complex considered has the same composition as in PDB. Out of the 5,326 complexes, 2,876 of them acquired symmetry information. We used 3DZD, mathematical rotation-invariant moment-based descriptors, to represent the surface shape of single-chain proteins and complexes. For a protein structure, a surface was constructed using the MSMS program [35] and then mapped to a 3D cubic grid of the size of N3 (N was set to 200). Protein size is not explicitly considered in 3DZD calculation. But in our previous study [23], we have shown that it is rare for proteins with very different sizes to share similar global surface. Moreover, in Fig 4C and 4D, we have also analyzed the chain length distribution in the single-chain shape space. MSMS failed to generate surface for two cases each in the single-chain dataset and the complex structure dataset, for which we used the MSROLL program [36] instead. Each voxel (a cube defined by the grid) is assigned either 1 or 0; 1 for a surface voxel that locates closer than 1.7 grid interval to any triangle defining the protein surface, and 0 otherwise. This 3D grid with 1s and 0s was considered as a 3D function f(x), for which a series is computed in terms of the Zernike-Canterakis basis [37] that is defined by the collection of functions Znlm(r,ϑ,φ)=Rnl(r)Ylm(ϑ,φ) (1) with −l<m<l, 0≤l≤n, and (n−l) even. Ylm(ϑ,φ) are spherical harmonics. Rnl(r) are radial functions defined by Canterakis, constructed so that Znlm(r,ϑ,φ) are homogeneous polynomials when written in terms of Cartesian coordinates. 3D Zernike moments of f(x) are defined as the coefficients of the expansion in this orthonormal basis, i.e. by the formula Ωnlm=34π∫|x|≤1f(x)Z¯nlm(x)dx. (2) To achieve rotation invariance, the moments are collected into (2l+1)-dimensional vectors Ωnl=(Ωnll,Ωnll−1,Ωnll−2,Ωnll−3,…,Ωnl−l), and the rotationally invariant 3D Zernike descriptors Fnl are defined as norms of the vectors Ωnl. Thus Fnl=∑m=−lm=l(Ωnlm)2 (3) Index n is called the order of the descriptor. The rotational invariance of 3D Zernike descriptors means e.g. that calculating Fnl for a protein and its rotated version would yield the same result. We used 20 as the order because it gave reasonable results in our previous works on protein 3D shape comparison [23,38–40]. A 3DZD with an order n of 20 represents a 3D structure as a vector of 121 invariants [23]. The similarity between two proteins X and Y was measured by the Euclidean distance dE between their 3DZDs, dE=∑i=1121(Xi−Yi)2, where Xi and Yi represent the ith invariant for protein X and Y, respectively. To illustrate the characteristics of 3DZDs, we compare it against two other structure similarity measures, the Procrustes distance [41] and TM-Score [42]. The Procrustes distance is a root-mean square deviation (RMSD) between corresponding points in two objects after an appropriate optimization of translation, rotation, and scaling. The smaller the Procrustes distance, the more similar the shape are. On the other hand, TM-Score is one of the common measures of the similarity of the main-chain conformations of proteins. TM-Score ranges from 0 to 1, with 1 for identical protein structures. Proteins within the same fold usually have a score above 0.5. The Euclidean distance of 3DZD is usually below 10 for proteins of the same shape [23,39]. In S3 Fig, the Euclidian distance of 3DZD and the Procrustes distance were compared in two datasets. Panel A compares pairs of 20 ellipsoids with increasing eccentricities, while panel B shows results on 1,278 single-chain protein pairs that have the same number of vertices in the surface representation. The two measures correlated well with a correlation coefficient of 0.9784 for the ellipsoid dataset (S3A Fig), because surface points were systematically distributed in the same fashion for all the ellipsoids and thus corresponding points are easily matched for aligning two ellipsoids. On the other hand, the two measures often have very different distances in protein shape cases (S3B Fig), which typically happened when point correspondences do not even allow appropriate scaling of the two structures. In S3B Fig, there are many protein pairs that have different surface shapes with a 3DZD Euclidean distance of over 10 but with a small Procrustes distance of around 0.2. S3C and S3D Fig show such protein pairs. As shown, proteins in these pairs have very different shapes, which indicates that 3DZD performs more reasonably for comparing protein shapes. Indeed, for protein shape comparison, The Procrustes distance has difficulty because corresponding surface points in two proteins need to be determined prior to the distance computation, which are not available in general for protein surface comparison. This is more difficult when two proteins have a different number of surface points to be compared. Apparently, 3DZD does not have such a problem because it does not align points to points. S4A and S4B Fig show the comparison between 3DZD and TM-Score. As shown, these two measures have virtually no correlation. The correlation coefficient was -0.1735 for these two measures. Panel B shows the density of the two measures. The highest density (yellow) was observed at around 3DZD distance of 5 to 10 and TM-score of 0.3, which is the score range for proteins with similar surface shape but with different main-chain fold. As also shown in Table 1, there are cases that proteins of the different fold class have a small 3DZD Euclidian distance. S4C and S4D Fig shows two such examples, where two structures have a similar surface shape to each other according to 3DZD but have a very large difference in their main-chain conformations. These results are consistent with our earlier work where we extensively compared 3DZD with conventional protein structure comparison methods [23]. The 3DZD files of the single-chain and the complex datasets are made available at S1 Data. 3DZD can be also computed for PDB files at the benchmark page of 3D-SURFER (http://kiharalab.org/3d-surfer/batch.php) [25,38]. We used principal component analysis (PCA) to project 3DZDs of 121 value vectors of protein structures into 3D. Three eigenvectors were chosen for the mapping because the scree plots (S5 Fig) showed that adding more eigenvalues does not contribute much to explaining data variance, and also to be consistent with the previous related works [18–20]. The three eigenvalues explained 52.64% and 47.76% of the total variation in the single-chain and the complex structure datasets, respectively. In order to quantify how elongated a structure is, we have defined the term eccentricity, which is calculated from the minimum volume enclosing ellipsoid (MVEE) of a structure. Given all atoms in a structure, protein MVEE is the ellipsoid with minimum volume that encloses all atoms. From MVEE, the eccentricity is defined as (2−b2/a2−c2/a2)/2, where a, b, and c are the length of longest, the second longest, and the third longest semi-principal axes of the ellipsoid, respectively. Elongated structures have an eccentricity close to 1, while spherical structures have an eccentricity close to 0. The volume of proteins was computed using MSROLL with a probe radius set to 0. For 42 cases in the single-chain dataset and 82 cases in the complex dataset where the MSROLL failed, we used the ProteinVolume program [43] instead. The volume values computed by these two programs were very consistent; the difference of volume values for ten randomly selected protein structures was on average 1.04%. The convex hull of a protein structure and its volume was computed using the ConvexHull function in the scipy.spatial package [44]. A pocket on a protein surface was identified and its volume was computed with VisGrid [45]. The average size of the pocket volume in the single-chain proteins was 6,302.9 Å3. We analyzed the location of proteins with a large pocket whose size is within the top 10% (12,219 Å3 or larger) in the single-chain protein surface space (Fig 1D). Donut-shaped structures were identified by first screening structures with genus > 0 and then with the conditions of 0.9≤b/a≤1.0 and 0≤(c2/a2+c2/b2)/2≤0.6, where a, b, and c are the parameters of MVEE of the structures. Then, structures that passed the criteria were visually examined. The genus number was computed with the Euler-Poincaré Formula, which states the following relationship between the number of vertices (V), edges (E), faces (F), loops (L), shells (S), and genus (g) of a manifold: V + F–E–(L–F) = 2 (S–g). To obtain these values of a protein surface, we used triangular meshes computed by EDTSurf [46]. L is equal to F for triangle meshes since triangular faces have exactly 1 loop. S was computed as the number of disconnected groups of faces.

10.1371/journal.ppat.1002730

Evolution of an Eurasian Avian-like Influenza Virus in Naïve and Vaccinated Pigs

Influenza viruses are characterized by an ability to cross species boundaries and evade host immunity, sometimes with devastating consequences. The 2009 pandemic of H1N1 influenza A virus highlights the importance of pigs in influenza emergence, particularly as intermediate hosts by which avian viruses adapt to mammals before emerging in humans. Although segment reassortment has commonly been associated with influenza emergence, an expanded host-range is also likely to be associated with the accumulation of specific beneficial point mutations. To better understand the mechanisms that shape the genetic diversity of avian-like viruses in pigs, we studied the evolutionary dynamics of an Eurasian Avian-like swine influenza virus (EA-SIV) in naïve and vaccinated pigs linked by natural transmission. We analyzed multiple clones of the hemagglutinin 1 (HA1) gene derived from consecutive daily viral populations. Strikingly, we observed both transient and fixed changes in the consensus sequence along the transmission chain. Hence, the mutational spectrum of intra-host EA-SIV populations is highly dynamic and allele fixation can occur with extreme rapidity. In addition, mutations that could potentially alter host-range and antigenicity were transmitted between animals and mixed infections were commonplace, even in vaccinated pigs. Finally, we repeatedly detected distinct stop codons in virus samples from co-housed pigs, suggesting that they persisted within hosts and were transmitted among them. This implies that mutations that reduce viral fitness in one host, but which could lead to fitness benefits in a novel host, can circulate at low frequencies.

The latest human influenza pandemic highlights the ability of influenza viruses to jump species barriers and emerge in new hosts, as well as the role of pigs in generating viruses with pandemic potential. The mutational power of influenza virus, caused by intrinsically error-prone viral polymerases, has been directly linked to viral emergence, as adaptive mutations present in the reservoir host are likely to be key to the evolution of sustained transmission in new hosts. Hence, studying how mutations are generated, maintained and transmitted in and among pigs is critical to understanding how novel viruses could emerge. Here we characterized the evolution and mutational spectra of influenza virus populations within naïve and vaccinated pigs linked by natural transmission, by analyzing multiple viral sequences obtained at different times post-infection. We show that the genetic make-up of influenza viruses in pigs is highly dynamic: the frequency of particular mutations, including those that could potentially alter host specificity or result in vaccine escape, fluctuated markedly, including one rapid fixation event. We also show that co-infections are common and multiple viruses – even defective ones – were transmitted between pigs despite being vaccinated. Our results provide empirical evidence of the complex dynamics of influenza viral populations in pigs and provide insight on the evolutionary basis of RNA viral emergence.

Influenza viruses are archetypical emerging viruses, as illustrated by the four human pandemics that have taken place since 1918. Although the natural reservoir of influenza viruses is wild waterfowl, the establishment of human lineages derived directly from birds is rare. The pig is therefore thought to play an important role in the adaptation of avian viruses to humans [1]. Despite the ongoing debate over whether the 1918 pandemic virus was transferred into humans directly from birds or if the pig was an intermediate host [2], the ecological importance of the latter in the generation of pandemic viruses is underscored by the latest H1N1 human pandemic [3], [4]. Although the 1957 and 1968 pandemics provide compelling evidence for the importance of segment reassortment in influenza emergence [5], [6], this process is not always a necessary requirement for the establishment of a novel lineage in a new host population. In particular, the emergence of Eurasian avian-like swine influenza virus (EA-SIV) in the late 1970's and the recent emergence of canine influenza virus (CIV) constitute examples of direct (i.e. without reassortment) host transfers from birds and horses into pigs and dogs, respectively [5], [7]. Clearly, during those host-switching events that do not involve reassortment, the rate at which adaptive mutations appear within individual animals is of critical importance. Most of our knowledge on influenza virus evolution and emergence is based on the analysis of either partial or complete consensus sequence of genomes derived from samples collected in surveillance studies. Although fundamentally important, this only constitutes a partial picture of the processes that drive their epidemiology and evolution. Studies focusing on the drivers of viral diversity at other scales are therefore required to provide an integrated picture of influenza phylodynamics [8]. Recent studies have focused on the viral genetic diversity present within infected individuals using a variety of influenza viruses in diverse hosts [9]–[11]; this, in turn, provides an empirical framework for the quantitative analysis of host-pathogen interactions. Such studies are key to understanding how virus and host-associated traits influence the generation of viral genetic and phenotypic diversity and their impact on biological properties such as host range, antigenicity, antiviral resistance and virulence. Further, by studying intra-host viral diversity in the context of transmission experiments it is possible to infer the epidemiological consequences of within-host evolution by examining how transmission bottlenecks mediate the structure and extent of viral genetic diversity in the recipient host. Studies that explore the within-host evolutionary dynamics of swine influenza viruses in pigs are lacking. EA-SIVs were first detected in 1979 [5], although it has been estimated that they may have originated as early as 1963 [12]. As noted above, this lineage is thought to have originated from a direct host-switch transfer from avian influenza A viruses. After its first isolation in 1979, EA-SIV became enzootic among pig populations in Western Europe and Asia and replaced the classical swine lineage that had been circulating for decades [13]. Of note, the neuraminidase (NA) and matrix (M) gene segments of the recently emerged human H1N1/2009 were derived from the EA-SIV lineage [3], [4]. The influenza hemagglutinin (HA) is a major surface glycoprotein that binds to host-cell receptors, and is also the main target for neutralizing antibodies [14]. As influenza infection results in partial cross-protection against novel variants, the HA is subject to strong immune selection. While mutations at antigenic sites can result in antigenic drift, amino acid changes at the receptor-binding domain (RBD) can result in expanded host-range [15], [16]. To determine the evolutionary dynamics of an Eurasian avian-like swine influenza virus in its natural host and how prior immunity impacts the mutational spectra of viral populations, we examined the intra- and inter-host genetic variation of the hemagglutinin 1 (HA1) gene of A/swine/England/453/2006 (H1N1) from a recent transmission study that included naïve and vaccinated pigs [17]. We examined the intra-host genetic variation of influenza virus present in daily nasal swabs obtained from two previously published transmission studies [17]. The “naïve” study consisted of a transmission chain among pairs of naïve pigs, while the “vaccinated” study involved the use of both naïve and vaccinated pigs (the latter immunized with an heterologous commercial bivalent vaccine containing A/New Jersey/8/76 [H1N1] and A/Port Chalmers/1/73 [H3N2]). This commercial vaccine is most broadly used in Europe. It was chosen to recreate the immune status of vaccinated pigs in the field as this would be the immune pressure that circulating viruses could face in nature. An outline of the experimental design is illustrated in Figure 1A. We generated 50 individual data sets, each derived from daily nasal swabs, and containing from 6 to 81 sequences of the first 939 nucleotides of the hemagglutinin 1 (HA1) gene. All antigenic sites and the receptor-binding domain were present in the HA1 region under study. Within these data we observed 3129 mutations out of a total of 2,402,901 nucleotides sequenced, of which 684 were unique (i.e. occurred once in the whole data set). The estimated mutation frequency ranged from 2.8×10−4 (when only unique nucleotide changes were counted and assuming that repeated mutations resulted from viral replication) to 1.3×10−3 mutations per nucleotide site (when all mutations were considered independent events). The analysis of intra-host variation for the naïve and the vaccinated study is summarized in Tables 1 and 2, respectively. Synonymous (syn) and nonsynonymous (nonsyn) mutations were distributed throughout the HA1 segment without a clear regional clustering (Figure 1B). The overall frequency and distribution of mutations suggested that most of the observed nucleotide changes were due to random viral polymerase errors during virus replication. Consistent with this, the dN/dS for the data set as a whole was 0.77 (95% CI = [0.72,0.84]), although we observed statistically significant evidence of positive selection (i.e. dN>dS) at codon positions 207 and 254. Interestingly, the former position lies within putative antigenic site Ca1 [18], [19]. Daily viral populations were composed of a mixture of genomes closely related to a predominant or consensus sequence. However, the consensus was dynamic due to marked changes in the relative frequency of mutations within viral populations. For example, pig 109 in the naïve chain exhibited two different consensus populations on the two days that it was sampled (Figure 2a): on day 7 the consensus exhibited a nonsynonymous mutation at position 553 (Asn168Asp in the mature HA1), whilst on day 8 the majority of the sequences displayed two mutations: A696G (syn) and G914A (nonsyn, Ser288Asn), which in turn were different from the predominant sequence in all the previous animals. Interestingly, Asn168Asp is located at antigenic site Ca1, likely altering the overall antigenicity of that viral population on that day. This mutation is transmitted from pig 109 to pig 105 (Figure 3). Notably, however, these mutated consensus sequences were not fixed down the transmission chain. A median joining network for the naïve chain as a whole is shown in Figure S4. Most striking of all, in the vaccinated chain synonymous mutation A696G (which was present at low frequency in three out of the four seeder pigs) became dominant in pig 417 and was then fixed along the transmission chain (Figure 3 and Figure S5). This is the only fixation event in our study. The reason why A696G displayed a transient high frequency in the naïve chains but was fixed only in the “vaccinated” transmission chain is unclear, but its appearance in both strongly suggests that this synonymous mutation has a marked impact on viral fitness. As it has been hypothesized that pigs act as “mixing vessels” in which avian viruses adapt to infect humans, we searched for mutations that could affect host-range. Accordingly, we detected two independent nonsynonymous mutations at amino acid position 133 within the receptor-binding domain (RBD). While Thr133Ile was observed in two consecutive days in naïve pig 104, Thr133Ala was observed in two pigs (410 and 401) in the vaccinated chain. Of note, when we examined this site at the epidemiological scale, the only residues observed were Thr and Ser, suggesting that Ile133 and Ala133 likely have a major impact on fitness, including altering host range. We also detected another amino acid change in the RBD: His180Arg was present in multiple animals along the transmission chain that included vaccinated pigs. An analysis of 2091 publicly available HA1 sequences from swine H1 viruses (Dataset S1) reveals that His180 is strictly conserved among them, and in 3671 HA sequences from a diversity of species only one isolate from mallards exhibited Asp180 and two isolates from humans displayed Pro180 (Dataset S2). Similarly, we detected nonsynonymous changes at antigenic sites in both data sets (Tables 1 and 2). Interestingly, some of these were observed along multiple days and/or in different animals suggesting that they have been transmitted among them (Tables S1 and S2). For example, mutation A758G (Gln236Arg, Ca1 site) was present in pigs 401, 405, 412, 415 and 416 in the vaccinated study (Table S4), whereas mutation T623C (Leu191Pro, Sb site) was present in pigs 104, 108, 112 and 115 of the naïve study (Table S3). Moreover, we detected up to three mutations at antigenic sites in individual sequences. For example, a single sequence derived from naïve pig 111 displayed three mutations at antigenic sites (G270A, A607C, and A619G), two of which resulted in amino acid changes at antigenic sites Sb and Cb. Similarly, a sequence derived from vaccinated pig 417 exhibited two nonsynonymous mutations (A272G and G634A) located at antigenic sites Cb and Sb, respectively. These results therefore suggest that significant antigenic variation can be generated along the course of infection even in the presence of pre-existing immunity. Glycosylation of the HA can impact both antigenicity and receptor binding [20]–[22]. There are five predicted N-glycosylation sites (N-X-S/T) in the HA1 segment under study, and we detected 40 mutations that disrupted glycosylation motifs, of which only two (Asn11Asp, Asn23Asp, Table S4) were possibly transmitted since they were present in animals linked by direct contact (pigs 400–414 and 412–413, respectively). We also detected six mutations that created glycosylation sites but they all constituted singletons. We inferred maximum likelihood (ML) and median joining (MJ) trees from sequences derived from individual animals. Both the ML and MJ trees from the animals that were naturally infected down the transmission chain displayed a characteristic star-like structure. In contrast, more complex trees with multiple branching events were inferred for the viruses sampled from the inoculated animals that seeded the transmission chains, and which are indicative of more complex intra-host evolution (Figure 2 and Figure S2). Hence, it is possible that the type of the infection (natural vs. experimental) could have an impact on the nature of intra-host viral evolution. In this particular case such differences could be due to the large inoculation dose used to ensure infection and/or to the appearance of egg-adaptive mutations in the inocula (as the virus was egg-grown). A previous study suggested that transmission bottlenecks for equine influenza virus (EIV) in co-housed horses are not particularly tight, based on the observation of shared mutations among different horses in a transmission chain [9]. An analysis of the current pig data resulted in a similar observation; in particular, we observed multiple clones sharing identical mutations among different pigs (Tables S3 and S4). Indeed, we observed distinct mutations in multiple links of the transmission chain, including sequences sharing two mutations in different pigs, mostly in the transmission chain that included vaccinated pigs (Table 3). For example, mutation T867C was detected in vaccinated pigs 401, 415 and 403 (all linked by direct contact) and was linked to mutation A696G in pigs 415 and 403. We also observed sequences sharing three mutations between pigs: C447T, A824G and G844A were all present in the same clone, derived from pigs 410 and 412 in the vaccinated chain. Although it is possible that some of these mutations arose de novo in different pigs, this is not likely for sequences sharing multiple mutations. By including pairs of pigs in each link of the chain we were also able to test if mixed infections were common; if this was the case, recipient pigs would harbor mutations present in both donor pigs. Notably, in both transmission studies we observed that mutations were likely to have been transmitted from both donors to single recipients, even in vaccinated pigs, thereby supporting the hypothesis of loose bottlenecks despite the presence of prior immunity (Figure 3). A complete description of the transmitted mutations (i.e. mutation type, motif, detection in multiple links in the transmission chain) is provided in Figure S3. One of the most striking observations of our study was that the detection of 12 and 11 sequences carrying stop codons in the naïve and vaccinated transmission chains, respectively (Tables 1 and 2). In addition, three stop codon mutations – C361T, C487T and G420A – were at a frequency >1, while C361T and G420A were present in successive days of pigs 111 and 412, respectively (Table S5). Furthermore, C361T, C487T, and G420A were also present in multiple animals (some of them linked by direct transmission, Figure S3, and Table S5). For example, we observed mutation C361T (Gln104Stop) on day 5 of naïve pig 104, again on days 7 and 8 of naïve pig 111 and then again on day 15 of naïve pig 106. The fact that pigs 104 and 111 were co-housed and that this mutation was detected on two consecutive days in the latter, strongly suggests it was maintained throughout the course of infection and further transmitted (although the possibility of appearing de novo cannot be excluded entirely). Similarly, mutation C487T (Arg146Stop) was present on days 3 and 6 of co-housed pigs 115 and 116, respectively. Hence, these findings suggest that influenza viruses carrying low-fitness or deleterious mutations can persist and even be transmitted between pigs. During the transmission experiments, vaccinated pigs shed less virus than naïve ones [17]. To assess the impact of vaccination on intra-host viral populations we compared the sequences derived from only naturally infected naive and vaccinated pigs. The mean pairwise distance (MPD) was greater in the naïve group (MPDnaives: 0.0018, MPDvaccinees: 0.0013) and this difference was statistically significant (t = 74.3461, p<2.2e-16) although small in absolute terms. Likewise, the proportion of singletons was higher in naïve pigs than in vaccinated pigs (0.30 and 0.17, respectively, W = 159, p-value = 0.004). Notably, with the exception of mutation A696G, we detected persistent mutations - i.e. those present in multiple days of a single animal - only in naïve pigs across both studies (Tables S1 and S2). However, we did not find significant differences in selective pressures between the two groups as the dN/dS values for naive and vaccinated pigs were 0.76 and 0.70, respectively. Finally, we hypothesized that viral populations from vaccinated pigs would exhibit a greater proportion of mutations at antigenic sites – a function of immune selection – and that transmission bottlenecks in this group of pigs would be tighter. Unexpectedly, we did not detect differences between the two groups. Similarly, we did not detect significant differences in the proportion of transmitted nonsynonymous mutations for each transmission chain. However, more mutations were transmitted between vaccinated (n = 64) than in naïve co-housed pigs (n = 52) even though the number of possible transmissions was lower in the former group (12 vs 16 respectively, see Figure 3). Understanding the biological mechanisms that shape viral genetic diversity is essential to unravel fundamental aspects of influenza evolution, such as the generation of antigenic variation and the successful adaptation to new host species. Within a phylodynamic framework, experimental studies on intra- and inter-host influenza evolution are critical to link the dynamic processes that shape viral phylogenies from individual hosts to epidemiological-scale meta-populations. Here, we determined the genetic variation of an Eurasian avian-like influenza virus along two transmission chains, one that included only naïve pigs and another that included both naïve and vaccinated pigs. The choice of the virus was based on the fact that this lineage established in the pig population following a complete genome interspecies transfer from birds [5], such that mutation accumulation rather than reassortment is likely to be central to host adaptation. We detected changes in the frequency of variants during the course of infection, revealing a complex pattern of within-host evolution. Our observation of transient changes in the consensus sequence of daily viral populations is of particular importance because it highlights the time-frame in which genetic (and potentially antigenic) novelty can be generated. This result is also consistent with that observed for canine influenza virus in dogs [10]. The estimated mutation frequency was approximately one order of magnitude higher than that previously reported for equine and avian influenza viruses in their natural hosts [9], [11]. This is noteworthy given the fact that in those studies the methodology was similar to that applied here. It is theoretically possible that the genetic structure of avian-like viral populations would change when infecting mammals, possibly as a result of virus adaptation to the host. Parallel studies comparing the evolutionary dynamics of long host-adapted swine viruses in pigs, human viruses in humans, avian viruses in pigs and in birds (i.e. ducks or chickens) could be used to address this hypothesis. It is also clear that some of the observed mutations could result from RT-PCR errors. Although the total proportion of artifact mutations cannot be estimated due to the impossibility of directly determining the number of mutations introduced during reverse transcription, we had previously estimated that the PCR enzymes used here under our laboratory conditions could introduce up to one mutation every 25,600 nucleotides [9]. If this was the case, less than 3% of the observed mutations would have been the result of PCR errors. The possibility that such mutations will alter the distributions of the sequences examined over time and/or along the transmission chains is highly unlikely, as we have previously shown using a Bayesian statistical framework to analyze within-host viral populations [23]. In contrast with our previous study on the intra- and inter-host evolutionary dynamics of EIV [9], a change in the consensus sequence (A696G) became fixed in the transmission chain that included vaccinated pigs. Fixation on such a rapid scale can only realistically be explained by positive selection of the A696G mutation, or as a result of a hitchhiking effect (i.e. A696G is linked to a beneficial mutation somewhere else in the genome). The fact that G696 became fixed along the vaccinated but not the naïve chain despite being detected is suggestive of selection due to differences in immune pressure, even though it is a synonymous change. Indeed, at the epidemiological scale, G696 is present in 2.7% of the H1 sequences, confirming that viruses carrying this mutation have circulated. More generally, observing the process of allele fixation in vivo within days is of fundamental importance because it shows how rapidly natural selection can act on influenza viruses. Although the precise function of the A696G mutation is unknown, it is reasonable to think that a similarly rapid evolutionary process could apply to cases of vaccine escape, enhanced virulence or expanded host-range. The estimated dN/dS for the intra-host data set was higher to that calculated for swine influenza H1N1 at the epidemiological scale (dN/dS: 0.34, 95% CI = [0.345,0.346]), in agreement with other studies on within-host viral evolution [9], [10] and reflecting the fact that intra-host genetic diversity frequently contains transient deleterious mutations that have yet to be removed by purifying selection. Despite this, we also observed that trees inferred from sequences derived from inoculated animals exhibited more complex topologies from those inferred from animals that were infected through natural transmission. This was likely due to the large inoculum doses used to ensure infection, and hence the transmission of more lineages, and the growth of virus in eggs. The latter process is likely to select for adaptive mutations, which is probable as the inoculum used was an avian-like swine virus. As such, our results suggest that caution should be taken when studying intra-host evolution in experimentally infected animals since the combination of large inocula and adaptive mutations generated during virus growth (either in cell culture or in eggs) could lead to artificially altered mutational spectra. For example, recent work has shown that the number of transmitted variants is correlated with the inoculum dose in a rhesus macaque infection model for HIV [24]. The observation of multiple genetic variants transmitted between pigs is consistent with what we observed in the case of EIV in horses [9]. However, the observation of loose transmission bottlenecks among vaccinated pigs is particularly striking as we had anticipated that the immune status of the host would impact on the size of the transmission bottleneck. Although uncertain, this could be a function of the phylogenetic (and likely antigenic) distance between the strains included in the vaccine and the challenge virus (Figure S1). Overall, our findings suggest that vaccination does not have a major effect on the genetic structure of intra-host viral populations through immune selection, nor on the size of transmission bottlenecks, at least for the combination of challenge virus and vaccine used in this study. Further experiments using homologous challenge and vaccine virus and different contact methods are required to address this point. Moreover, that we repeatedly detected the same mutations in recipient pigs as in both donor pigs indicates that mixed infections are common. The significance of this finding is enormous if we consider the structure of pig populations, where large numbers of piglets are often housed in warehouse-like buildings for several weeks until they reach their target weight at approximately 22–30 weeks. Although the all-in-all-out swine production system minimizes the transmission between groups of pigs, our results show that very high levels of genetic variation could be generated during growing/finishing stages even in vaccinated herds. Indeed, vaccination could result in undetected virus circulation as vaccinated pigs are likely to show very mild clinical signs of disease. Also of note was the observation that mutations detected on multiple days were present in naïve but not in vaccinated pigs, consistent with similar studies we have performed in vaccinated horses (Murcia and others, unpublished). This could be due to a more efficient viral clearance in vaccinated pigs. Alternatively, since viral shedding was lower in this group [17], it is possible that persistent mutations were present but undetected. This would suggest that minor subpopulations could persist and be transmitted even in the presence of prior immunity, thus allowing natural selection to act more rapidly. Future work using ultra-deep sequencing technologies will address this issue. One of our most notable observations of this study was that stop codon mutations, which are presumably defective, were both present within individual pigs and also transmitted among them. To the best of our knowledge, this is the first observation of the transmission of defective influenza viruses in vivo, although it has been reported in other RNA viruses [25]. In theory this could be achieved by trans-complementation, a mechanism that has been described during influenza replication in vitro [26], and implies that co-infection of single cells is a common feature during infection in vivo. This observation may have important implications for viral emergence, since it clearly means that low fitness mutations can be maintained within a host population, flattening the fitness valleys that separate donor and recipient hosts [27], such that mutations that are deleterious in the donor host may be advantageous in the recipient host. In sum, the combination of loose bottlenecks, mixed infections, rapid allele fixation, common cellular co-infections and trans-complementation observed in this experimental study not only reveals the complex mechanisms at work during influenza evolution, but also provides a mechanistic framework to better understand the evolutionary basis of viral emergence. The transmission studies from which the samples were obtained have been published previously [17]. All animal work was done under GB Home Office license following full ethical approval. Naïve transmission study: five- to six-week-old piglets seronegative to influenza viruses of the H1N1, H1N2 and H3N2 subtypes were used. Two “seeders” were inoculated intranasally with 106.8 EID50 of A/swine/England/453/2006. Upon confirmation of virus excretion using the Directigen test, two other piglets (N1) were introduced into the same pen. Upon detection of virus excretion in N1 pigs, the seeders were removed and two further piglets were introduced into the pen. This procedure was repeated in order to establish a transmission chain (Figure 1A). Nasal swabs were collected for up to four days after infection or contact, immersed in viral transport medium (VTM, PBS supplemented with 2% tryptose phosphate buffer broth, 2% penicillin/streptomycin and 2% amphotericin B), vortexed, aliquoted and stored at −80°C. Transmission in vaccinated pigs: piglets were vaccinated with two doses of Gripovac, the first one at the age of four-to-five weeks and the second dose four weeks later. The viral strains contained in the vaccine (A/New Jersey/8/76 [H1N1] and A/Port Chalmers/1/73 [H3N2]) are different lineages from the challenge virus. Vaccinated pigs were tested by hemagglutination inhibition (HI) at regular intervals until the antibody titers reached a target value of ≤40 HIUs in order to allow natural infection. Vaccinated pigs reached the target antibody levels when they were approximately five months old. Two unvaccinated pigs were inoculated with 106.1 EID50 of A/swine/England/453/2006 as described above. Nasal swabs were collected on a daily basis and an aliquot was RNA-extracted and subject to qPCR for virus detection on the same day. Upon detection of virus shedding the transmission chain was established as described above. Vaccinated pigs entered the transmission chain in order of antibody titer (lower first). Two pairs of naïve pigs were added to the end of the chain when virus shedding could no longer be detected in vaccinated pigs (Figure 1A). Details on the kinetics of viral shedding, clinical signs and gross lesions can be found in [17]. RNA was extracted from 280 µl-aliquots of nasal swabs using the QIAamp viral RNA mini kit (Qiagen). A two-step RT-PCR was performed to amplify a segment of HA1 starting from the 5′ terminal region to nucleotide position 1115. Reverse transcription was performed in 20 ul-reactions and PCR products used as a template 5 ul of cDNA. For most samples, cDNA from a single RT-PCR reaction was sufficient to generate enough clones for sequencing. cDNAs of the viral genomic HA gene was generated using Superscript III reverse transcriptase (Invitrogen) and primer Bm-HA1 [28]. Reverse transcription was performed at 55°C for 90 min, followed by incubation at 70°C for 10 min. PCR amplification was performed using Platinum Taq High Fidelity (Invitrogen) using primers Bm-HA1 [28] and HA1115 (5′RCTGTCCATCCCCCCTCAATYAANCCYGCAAT 3′). PCR amplification was performed for 35 cycles (94°C for 15 sec, 51°C for 30 sec and 68°C for 2 min) after 2 min of initial denaturation at 94°C and followed by a final extension at 68°C for 10 min. Samples with low copy numbers were amplified using a hemi-nested PCR. The first PCR was performed using universal primers for the amplification of the full length HA gene [28], followed by a second reaction as described above. PCR products were gel-purified using the QIAquick Gel Extraction Kit (Qiagen) and further cloned using the Zero-Blunt TOPO PCR Cloning kit for sequencing (Invitrogen) following the manufacturer's instructions. Clones were sequenced at the Wellcome Trust Sanger Institute using fluorescent sequencing chemistry and ABI 3730xl capillary sequencers. Forward and reverse sequencing reads from each clone were trimmed of vector sequence and poor quality regions. Reads with an average Phred quality score below 20 were rejected. We merged the forward and reverse reads into a single contig using the consensus HA1 sequence of stock A/swine/England/453/2006 (kindly provided by Ian Brown) as a framework reference. The alignment, assembly and analysis were performed using a bioinformatics tool specifically developed for mutation detection in viral sequences (Ramirez-Gonzalez, Hughes and Caccamo, in preparation). This software aligns the forward and reverse reads to the provided reference using a Smith-Waterman approach followed by a base-by-base inspection of the alignment. For every mismatch, it calls a mutation in the sample if the quality Phred score in the sequencing read is above 25 (this threshold is a parameter in the analysis). If the score is below the threshold it is assumed that the mismatch was induced by a sequencing error and the base in the provided reference is selected for the consensus. If the sequencing reads overlap, the base with the best quality is selected for the comparison with the reference. This analysis also computes the mutation type (i.e. whether they induce synonymous, nonsynonymous or missense alleles), which is reported by the tool together with the assembled contig. For the analysis reported here we only considered contigs longer than 935 nt and sequences with good quality insertions or deletions (introducing frameshifts) were also discarded. A total of 2559 intra-host EA-SIV sequences isolated from experimentally infected animals (GenBank accession numbers JQ520376–JQ522934), and 2091 epidemiological-scale publicly available H1 HA sequences from swine (Dataset S1, obtained from the Influenza Virus Resource [http://www.ncbi.nlm.nih.gov/genomes/FLU/FLU.html]) were collated. All sequences ran from the start codon of the HA open reading frame to nucleotide position 939. Sequence alignments were generated from the assembler output for the intra-host EA-SIV sequences, and using MAFFT [29] for the SIV epidemiological-scale sequences. Because of the very small genetic distances involved, the mean pairwise genetic diversity within each sample was estimated from the uncorrected pairwise distance matrix (p-distance) between taxa (available from the authors upon request). Maximum likelihood (ML) trees were estimated using PhyML [30] under the best-fit model of nucleotide substitution determined using MrAIC (http://www.abc.se/~nylander/mraic/mraic.html). For the very large data sets combining all sequences with the epidemiological-scale data, the phylogeny was estimated using RAxML [31] and the GTRGAMMA substitution model with 500 bootstrap replicates. Mean numbers of nonsynonymous (dN) and synonymous (dS) substitutions per site (ratio dN/dS) and the estimated 95% confidence intervals were estimated using the Single Likelihood Ancestor Counting (SLAC) algorithm available in the HyPhy software package [32]. Gaps were removed from the alignment prior to calculation of dN/dS. Finally, median joining networks were calculated from the sequence data using the median joining algorithm available in NETWORK 4.6.00 (http://www.fluxus-engineering.com/sharenet.htm).

10.1371/journal.ppat.1002567

Comparative Genomics of the Apicomplexan Parasites Toxoplasma gondii and Neospora caninum: Coccidia Differing in Host Range and Transmission Strategy

Toxoplasma gondii is a zoonotic protozoan parasite which infects nearly one third of the human population and is found in an extraordinary range of vertebrate hosts. Its epidemiology depends heavily on horizontal transmission, especially between rodents and its definitive host, the cat. Neospora caninum is a recently discovered close relative of Toxoplasma, whose definitive host is the dog. Both species are tissue-dwelling Coccidia and members of the phylum Apicomplexa; they share many common features, but Neospora neither infects humans nor shares the same wide host range as Toxoplasma, rather it shows a striking preference for highly efficient vertical transmission in cattle. These species therefore provide a remarkable opportunity to investigate mechanisms of host restriction, transmission strategies, virulence and zoonotic potential. We sequenced the genome of N. caninum and transcriptomes of the invasive stage of both species, undertaking an extensive comparative genomics and transcriptomics analysis. We estimate that these organisms diverged from their common ancestor around 28 million years ago and find that both genomes and gene expression are remarkably conserved. However, in N. caninum we identified an unexpected expansion of surface antigen gene families and the divergence of secreted virulence factors, including rhoptry kinases. Specifically we show that the rhoptry kinase ROP18 is pseudogenised in N. caninum and that, as a possible consequence, Neospora is unable to phosphorylate host immunity-related GTPases, as Toxoplasma does. This defense strategy is thought to be key to virulence in Toxoplasma. We conclude that the ecological niches occupied by these species are influenced by a relatively small number of gene products which operate at the host-parasite interface and that the dominance of vertical transmission in N. caninum may be associated with the evolution of reduced virulence in this species.

Coccidian parasites have a major impact on human and animal health world-wide and are among the most successful and widespread parasitic protozoa. They include Neospora caninum which is a leading cause of abortion in cattle and one of its nearest relatives, Toxoplasma gondii. Despite its close phylogenetic relationship to Toxoplasma, Neospora has a far more restricted host range, does not infect humans and its epidemiology depends predominantly on efficient vertical transmission. The divergent biology of these two closely related species provides a unique opportunity to study the mechanisms of host specificity, pathogenesis and zoonotic potential not only in these, but other Coccidia. We have sequenced the genome of Neospora and the transcriptomes of both species to show that despite diverging some 28 million years ago, both genome and gene expression remain remarkably conserved. Evolution has focused almost exclusively on molecules which control the interaction of the parasite with the host cell. We show that some secreted invasion-related proteins and surface genes which are known to control virulence and host cell interactions in Toxoplasma are dramatically altered in their expression and functionality in Neospora and propose that evolution of these genes may underpin the ecological niches inhabited by coccidian parasites.

Toxoplasma gondii and Neospora caninum are closely related tissue-dwelling Coccidia – intracellular protozoan parasites of the phylum Apicomplexa. T. gondii can infect essentially any warm-blooded vertebrate and is found in nearly one third of humans, arguably being the world's most successful zoonotic parasite [1]; it causes neonatal mortality, spontaneous abortion and blindness [2]. T. gondii is most often transmitted horizontally following ingestion of environmentally resistant oocysts excreted by its definitive host (cats), or via ingestion of persistent asexual stages (bradyzoites) residing in the tissues of intermediate hosts. The biology of T. gondii has been intensively studied, but despite advances in understanding host cell invasion, the role of secreted kinases in parasite virulence [3], [4] and its population and evolutionary biology [5], [6], the molecular mechanisms responsible for its highly promiscuous nature remain unknown. Neospora caninum is a close relative of T. gondii [7]. They are both tissue-dwelling Coccidia and share many common morphological and biological features [8]. Each is able to develop in intermediate hosts, reproducing asexually, or to move between intermediate and definitive hosts, reproducing sexually. Neospora was initially misidentified as Toxoplasma, but was subsequently differentiated based on host preferences, etiology, morphological and genetic differences [8]. Despite these similarities the two species differ in their definitive host: while Toxoplasma completes its sexual cycle in felids, Neospora does so exclusively in canids [9]. Unlike Toxoplasma, Neospora appears not to be zoonotic, having a more restricted host range [10], [11] in which it occupies a unique ecological niche showing a striking capacity for highly efficient vertical transmission in bovines [12]. N. caninum is one of the leading causes of infectious bovine abortion, resulting in significant economic losses to the dairy and beef industries [13]. The molecular determinants of host specificity and in particular zoonotic capability in the Apicomplexa are not known. It is possible that a large part is played by the host cell invasion machinery common to all Apicomplexa which involves surface antigens and specialized apical secretory organelles named rhoptries, micronemes and dense granules [14], but this is yet to be substantiated by experimental evidence. The process of host cell invasion has been well studied in Toxoplasma and components of the invasion machinery are also involved in host cell modification and interaction with the host immune system [3], [4]. Attachment to host cells is mediated by a family of highly abundant surface antigens [15], after which the micronemes release adhesins which engage an actin-myosin motor to provide the driving force for host cell invasion [16], [17], [18]. Rhoptry neck proteins are then released to form a tight region of contact with the host cell, known as the moving junction, which acts as a scaffold for the parasite to enter the cell and form the parasitophorous vacuole (PV) in which it resides [19]. The rhoptries also release a range of proteins that modulate host cell function [20], [21], [22], in particular, virulence-related rhoptry kinases interact with host defenses; for example, ROP18 inactivates host immunity-related GTPases (IRGs) that would otherwise rupture the PV membrane and kill the parasite [23], [24]. Whilst it is anticipated that the overall process of host-parasite interaction in Neospora is likely to be similar, we hypothesize that the molecular characteristics of this interface are likely to be the key determinant in the unique biological features of the two parasites. In fact, small but defining differences in the biology of these two closely related organisms provide a unique opportunity to identify the mechanisms which underlie the basis of host specificity, pathogenesis and zoonotic potential not only in these important parasites, but also in the wider members of the phylum. This includes several groups of organisms of key interest to human and animal welfare (e.g. Plasmodium, Cryptosporidium and Eimeria). In order to exploit this opportunity we have sequenced the genome of N. caninum and the transciptomes of both N. caninum and T. gondii, undertaking the first comparative transcriptome analysis of any apicomplexans at single base-pair resolution. We show that Neospora caninum and Toxoplasma gondii have very similar genomes with largely conserved gene content and synteny. As predicted, differences are most common amongst groups of genes which interact with the host. We find that surface antigen gene families are expanded in N. caninum suggesting that larger repertoires of such genes may be important in becoming a more host-restricted coccidian parasite, although data from a more extensive range of related parasites would be required to test this hypothesis. We also find that some rhoptry genes are highly variant between species and demonstrate that the pseudogenisation of ROP18 in N. caninum leads to a functional change in the interaction of the parasite with host immune mechanisms. We propose that such mutations may be associated with changes in transmission strategy. In addition to these biological insights, our data provides a vital community resource for comparative genomics in this important phylum of medical and veterinary parasites. We sequenced the genome of N. caninum Liverpool strain using Sanger sequencing to ∼8-fold depth. It was assembled into 585 supercontigs with an N50 of 359 kb totaling 61 Mb (Table 1). We constructed a set of N. caninum pseudochromosomes by aligning 242 supercontigs to the fourteen publicly available T. gondii Me49 chromosomes [25] based on predicted protein sequence similarity (Figure 1A). It has been shown previously using our partially assembled sequencing data that N. caninum and T. gondii genomes are largely syntenic [26]. Here we show that for almost all regions where conservation of gene order (synteny) is interrupted, corresponding orthologous regions are found elsewhere in the genome. This suggests that while there may have been a small number of chromosomal rearrangements, there has been very little net gain or loss of genetic content (Figure 1B). The N. caninum Liverpool genome sequence has been added to the European Nucleotide Archive as project CADU00000000. To determine gene expression differences between species and to improve genome annotation we sequenced the transcriptome of the invasive stage (tachyzoite) of N. caninum Liverpool and T. gondii VEG using mRNA sequencing (mRNAseq) on an Illumina GAIIx machine (Tables S5 & S6). The parasites were grown asynchronously for a period of six days in cell culture. We took samples of RNA at days three, four and six. We found that days three and four showed fairly similar expression profiles within species and so we have pooled this data for most analyses. We found however that day six N. caninum parasites were showing expression of bradyzoite (quiescent stage) marker genes (Text S1). These parasites had not fully converted into bradyzoites, but may be preparing to do so. We did not observe expression of these markers at day six in T. gondii, so we did not seek to compare transcriptomes of the two species at this timepoint. Transcriptome sequencing data has been submitted to ArrayExpress with accessions E-MTAB-549 for N. caninum sequences and E-MTAB-550 for T. gondii sequences. Combining de novo gene predictors and mRNAseq evidence we identified 7121 protein-coding genes in N. caninum and produced a revised T. gondii ME49 gene count of 7286, a reduction of 9% from previous predictions (Table 1). This was achieved predominantly by merging adjacent genes based on mRNAseq evidence. In N. caninum we detected the expression of 74% of genes during the tachyzoite stage. In T. gondii 80% were expressed, significantly more than the 49% recently reported using microarrays suggesting greatly improved sensitivity [27]. Using a combination of automated orthologue identification and manual curation we identified a small number of unique (i.e. organism-specific) genes in both genomes that might underlie their phenotypic differences (Figure 2A). Excluding surface antigen families (discussed later), we found 231 genes unique to T. gondii and 113 to N. caninum, i.e. with no orthologue or paralogue based on our orthologue analysis. Of these, 72 from T. gondii and 43 from N. caninum had Pfam domains or proteomics-based evidence (Table S1). These genes represent good candidates for understanding organism-specific differences and are enriched for those involved in host-parasite interactions. The remainder had no detectable homologues or proteomics-based evidence, although most had good transcriptome evidence. Only one organism-specific multigene family, with no homologues in the other species was identified: a family that we have named Toxoplasma-specific family (TSF; Figure S1). This family is located largely in chromosomal regions with no similarity to N. caninum (regions 3, 5 and 17 in Figure 1B) and varies in size between T. gondii strains. We found that all ten members of TSF from T. gondii Me49 were expressed during the tachyzoite stage. No significant domains, motifs or signal peptides were identified; however a putative transmembrane helix was predicted between amino acids 195 and 217 on TgTSF1. Another previously unidentified family was present in N. caninum, but was expanded in T. gondii (Figure S2). This family comprises three genes from N. caninum and seven from T. gondii Me49. Sequences were scanned using InterProScan [28] but no significant domains or motifs were identified. Lysine-arginine rich motifs are however present in the sequences suggesting possible nuclear localization signals. We have therefore named this family Lysine-Arginine rich Unidentified Function (KRUF). KRUF genes appear to be highly expanded in the GT1 strain of T. gondii, with up to twenty members [25]. Two of the three N. caninum members are expressed in the tachyzoite and early bradyzoite stages (NCLIV_002020 and NCLIV_002030). Most of the T. gondii members are expressed, some at very high levels (esp. TGME49_051170). While 32% of the genes shared by T. gondii and N. caninum have orthologues in a range of eukaryotes, we found that ∼39% of the shared genes do not have orthologues in other apicomplexans sequenced to date (Figure 2B). Furthermore, while ∼29% of the shared genes not found outside apicomplexans have orthologues in at least one apicomplexan, only 0.3% are shared between all apicomplexans with completed genome sequences. Due to the assumptions behind this analysis we have likely underestimated the similarity between Apicomplexa and more detailed manual analysis will no doubt reveal more divergent orthologues. However our results suggest that the genome content of apicomplexans is very diverse and that many novel and divergent genes are found within the Coccidia. To determine whether their divergent lifestyles are associated with differences in metabolism we compared the predicted repertoires of metabolic enzymes and pathways of N. caninum to those of T. gondii [25]. The pathways identified in N. caninum appeared identical to those in T. gondii and we found no single metabolic gene specific to either species suggesting that changes in metabolism do not play a large role in host restriction and zoonotic compatibility in these species. Although a small number of metabolic genes were differentially expressed between species, we found little evidence that these were clustered in any particular pathway, although there is some evidence that nitrogen metabolism may be upregulated in N. caninum and porphyrin metabolism may be upregulated in T. gondii (Figure 2C). Previous estimates from rRNA analysis have suggested that N. caninum and T. gondii diverged between 12 and 80 million years ago (mya) [6], [29]. To gain a more accurate estimate we examined a large number of orthologue alignments, determining synonymous substitution rates between N. caninum and T. gondii and between malarial parasites of human and non-human homonidae: Plasmodium falciparum and P. reichenowi, respectively. We assumed constant evolutionary rates between the Plasmodium spp. and Coccidia, excluding genes which were found to have evolved in a non-clock-like manner. We used a previously determined estimate of 2.49 mya for the split between P. falciparum and P. reichenowi [30]. This allowed us to date the speciation of N. caninum and T. gondii to 28.0 mya or between 21.7 and 42.7 mya using the confidence intervals of the P. falciparum and P. reichenowi divergence time. This suggests that speciation of N. caninum and T. gondii occurred after the speciation of their definitive hosts (estimated at 54–67 mya) [31]. The ability to reject non-clock-like genes is dependent on gene length and so we also calculated the divergence time using only the longest 25% of the orthologous groups. This led to a divergence time of 26.9 mya, very close to that calculated using all groups, suggesting that a tendency to exclude longer genes using the clock test has not biased our results. Examination of gene gain and loss and differential expression implicated two host-interaction gene families: SAG1-Related Sequence (SRS) and ROPK, as among the most divergent features of the two species (Figure 2D, Figure 3). SAG1 was the first SRS protein identified and is the major surface antigen of Toxoplasma. SRS proteins localize to the cell surface of both T. gondii and N. caninum. They are thought to play a key role in attachment to host cells, modulation of host immunity and regulation of parasite virulence [32]. Wasmuth et al. (submitted) found 109 functional genes and 35 pseudogenes in T. gondii Me49 with similar numbers across several different strains. They are present sometimes in single copies, often in tandem arrays. They are dispersed across all chromosomes rather than showing a preference for subtelomeric regions as is found for some large gene families in Plasmodium, Babesia and Theileria (Figure 1A). It has been suggested that the large number of SRS genes is present in T. gondii to accommodate the wide spectrum of potential host-cell molecular interactions presented by its exceptionally large host range [33], [34]. However, our data refute this; we found the SRS gene family to be substantially expanded in N. caninum compared to T. gondii with a total of 227 N. caninum SRS genes (NcSRSs) and 52 NcSRS pseudogenes (Figure 1A). Expression data suggest however that T. gondii expresses a greater number of its SRS repertoire (55 vs. 25 in N. caninum) during the tachyzoite stage (Figure 1A). In N. caninum we found in most cases that only a single SRS gene was expressed at a multigene locus, whereas in T. gondii we often found several. Extending our gene expression studies beyond the rapidly growing and invasive tachyzoite stage, we noticed that N. caninum cultures maintained until day six showed expression of known bradyzoite-specific genes (e.g. BAG1, SRS13, SAG4), suggesting they were beginning to convert into the slow-growing stage (Text S1). We observed a greater number of NcSRS genes (36 vs. 25) expressed at day six than at earlier points in the culture. Despite this it remains unclear whether most members of this expanded family in N. caninum are expressed and further expression data are required from all life-stages before the role of these genes can be better understood. SRS genes consist of one or more copies of the SAG domain family, which has been classified into eight subfamilies (Figure S3; Wasmuth et al., submitted). The doubling of SRS gene numbers in N. caninum compared to T. gondii is largely accounted for by expansion of a particular subfamily with a 7–8 domain architecture. No novel SAG domain subfamily has evolved in either lineage, however several domain combinations are found in low copy numbers in only one or the other species (Table S2). Since a particular SRS locus tends to contain genes with the same domain architecture in both species, expansion has likely occurred by tandem duplication. We found evidence that gene conversion may have occurred at, at least, one locus (SRS19; Figure S4A), whereas one of the most highly expressed loci in both organisms (SRS29, containing the SAG1 gene) showed no evidence of gene conversion (Figure S4B), perhaps due to functional constraints. SUSA genes (SAG-Unrelated Surface Antigen genes) are a superfamily of surface antigens unrelated to SRSs but which are also postulated to interact with the host immune system [35]. In common with the SRS superfamily we found that N. caninum had an expanded number of SUSA genes but that a greater number were expressed in T. gondii (Figure 1A). In fact none of the NcSUSA genes were expressed in the tachyzoite stage. Two NcSUSA genes (NCLIV_067570 and NCLIV_067920) were however expressed at day six of culture. The apical complex is the defining characteristic of the Apicomplexa. It includes the rhoptry, microneme and dense granule secretory organelles, which are essential for cell invasion. Figure 3 shows how the repertoires and expression of gene products known or predicted to be localized to these organelles differs between T. gondii and N. caninum. Several T. gondii rhoptry genes (ROP18, ROP16 and ROP5) have been implicated in virulence based on a genetic cross between the type II and III [3] and type I and III [4] lineages of T. gondii. N. caninum differs from T. gondii at each of these loci, but shares some similarities with low virulence strains. Expression of TgROP18 is associated with virulence in mice [4] and in some hosts high ROP18 expression may reduce parasite fitness by causing rapid host death [36]. It is involved in preventing the host interferon-gamma (IFN-γ) response, during which the host loads immunity-related GTPases (IRGs) onto the parasitophorous vacuole (PV) leading to its disruption and parasite cell death in avirulent strains [37]. Virulent T. gondii strains express high levels of ROP18, which phosphorylates and inactivates IRGs to safeguard the PV [24], [37]. We found that in Neospora Nc-Liv ROP18 is a pseudogene due to several interrupting stop codons in the sequence syntenic with the Toxoplasma gene. We confirmed the presence of these stop codons in a further four strains of the parasite isolated from different geographic locations and hosts (Table S3). To determine whether N. caninum is able to phosphorylate IRGs without a functional copy of ROP18 we examined the loading of Irga6 (a member of the host IRG GTPase family) onto the PV by immunofluorescence studies. We observed that, in both N. caninum and T. gondii infections, host cells responded by loading Irga6 onto the PV but only T. gondii was able to phosphorylate Irga6 and thereby presumably inactivate the IRG protein (Figure 4A). This suggests that N. caninum is unable to prevent its host from using IRGs to attack the PV. In T. gondii ROP16 directly interferes with host signaling pathways (e.g. Stat3, Stat6) to modulate the proinflammatory host cytokine IL-12 [21], [38]. A single polymorphic residue on TgROP16 determines the strain-specific activation and phosphorylation of Stat3 [39]. We found that ROP16 was highly expressed in T. gondii VEG (type III) tachyzoites and it has been shown elsewhere to be highly expressed in all T. gondii strain types [27]. While its orthologue in N. caninum, NcROP16, possesses the key active-site leucine residue for Stat3 activation, the gene was not expressed in our experiments. Although it is possible that NcROP16 could be expressed in other cell types, our experiments predict that N. caninum infection does not activate Stat3 due to its lack of expression. Several additional T. gondii rhoptry genes are missing from N. caninum (Figure 3), most notably the entirety of the locus which encodes ROP2A, ROP2B and ROP8. The TgROP5 multigene locus accounts for 50% of inherited variation in Toxoplasma virulence [40]. The relationship between ROP5 genotype and virulence in T. gondii is however not clear. The most virulent, type I T. gondii strain (e.g. RH) has six copies, while the less virulent type II T. gondii strain (e.g. Me49) has around ten copies and the least virulent type III strain (e.g. VEG) has four. N. caninum Liverpool encodes only two copies of the ROP5 gene both of which are highly expressed in the tachyzoite stage (Figure 4B). The secreted proteins of the microneme organelles play a crucial role in host cell attachment and invasion by mediating gliding motility [16]. We identified thirteen previously undescribed genes putatively encoding micronemal proteins by virtue of conserved domain architectures. Of these newly identified genes, MIC26 (a MIC2 paralogue) and MIC19 (a PAN domain-containing gene) are unique to N. caninum. Some differences also exist between the species in dense granule genes which are involved in the modification and function of the parasitophorous vacuole (PV) [41]. Dense granule genes GRA11 and GRA12 were absent from the N. caninum genome sequence. Serine proteases are important to the maturation of both rhoptry and microneme proteins and their inhibition blocks parasite replication and rhoptry formation [42]. TgSUB2, a subtilisin-like serine protease has been identified as a likely processor of several rhoptry proteins [43] and whilst T. gondii is vulnerable to a variety of protease inhibitors, including serine protease inhibitors, N. caninum invasion is inhibited only by aspartyl protease inhibitors [44]. We found that while all 12 identifiable T. gondii subtilases had orthologues in N. caninum, there was a significant decrease in expression of these proteases (hypergeometric test; p = 0.003) compared with T. gondii. This suggests that subtilisin-like serine protease activity may not be used to the same extent in N. caninum as in T. gondii and may explain why N. caninum is less susceptible to its inhibition. The ApiAP2 family represents the major group of apicomplexan transcription factors. They have been implicated, for example, in control of the intraerythrocytic development cycle (IDC) [45] and sporozoite development [46] of malaria parasites. Twenty-nine such genes have been identified in Plasmodium and 68 in Toxoplasma [47]. We found N. caninum orthologues for all 68 TgAP2 genes but detected significant differences in the expression of eleven of them (Figure 3), which in turn may be responsible for expression differences we have observed in other genes. It has been suggested for instance that rhoptry genes are regulated by AP2 transcription factors in Plasmodium [48]. We found that 54 of 68 NcAP2s and 61 of 68 TgAP2s were expressed during the tachyzoite stage, more than a previous study [49]. This is surprising considering that one would expect the principal family of transcription factors in organisms with a complex life cycle to be highly specific to different life stages. As expected, the repertoire of ncRNA genes of known function (e.g. t-RNAs, snoRNAs, snRNAs etc.) is almost identical between Toxoplasma and Neospora. However, we were able to identify an expansion of a previously unidentified candidate structured non-coding RNA family in N. caninum. This suggests that ncRNA repertoires are divergent in these species, although the functions of these RNAs remain to be identified (Text S2, Figures S5 & S6). We have used genome and transcriptome sequencing to probe the apicomplexan parasites Toxoplasma gondii and Neospora caninum for differences which might underlie their divergent host ranges, transmission strategies and zoonotic potential. We have demonstrated that the two genomes show a high degree of synteny, with a one-to-one correspondence between most protein-coding genes. We calculated that speciation occurred around 28 mya, after the divergence of their respective definitive hosts, the cat and dog. This is consistent with two possibilities: 1) one or both parasite species may have switched to a new definitive host since their divergence, 2) a common ancestor used both cats and dogs as definitive hosts but during divergence N. caninum and T. gondii eventually became restricted to their present day definitive hosts. Our data clearly show that genes interacting most closely with the host have diverged to the greatest extent and we have therefore been able to narrow investigations to a relatively small number of candidate gene families and individual genes. Although many genes of unknown function remain to be characterized in these organisms, the majority of these are conserved. We have identified two novel protein-coding gene families (TSF and KRUF) and a putative ncRNA family which differs between species and that warrant further experimental characterization. The principle surface antigen gene family, the SRSs, was the most divergent family. This result was expected because in all Apicomplexa examined so far, including several malaria parasite genomes, the surface antigens are the largest, most rapidly evolving of all gene families [50]. However, the observation that N. caninum has more than twice as many SRS genes as T. gondii is striking and rather unexpected. It had been assumed that T. gondii requires a large number of these genes to accommodate its extraordinarily large host range and cover all potential host-cell molecular interactions with corresponding parasite proteins [33], [34]. The much smaller host range of N. caninum would suggest that this hypothesis is not supported and that perhaps conversely, a larger number of SRS genes might be advantageous in evolving a narrower host range. Transcriptome evidence however suggests that N. caninum uses fewer SRSs than T. gondii during the tachyzoite stage, suggesting that they may be of more importance in other parts of the life cycle. In fact, in N. caninum there is rarely more than one SRS gene expressed at each locus, while in T. gondii there are frequently multiple genes expressed. This implies there have been significant changes in regulation of these host-interacting genes between species, although the mechanisms of regulation of these genes remain unknown. Interestingly it is a small number of subfamilies which have been expanded in N. caninum, in particular the fam7-8 architecture, the most common in both species. It may be that the more limited host range of N. caninum is related to specialization of this subset of the SRS genes. In common with SRS genes, important species-specific differences were identified in rhoptry organelle genes where the divergence of key genes of known function may help to explain phenotypic differences between Toxoplasma and Neospora. In particular ROP18 is a key virulence determinant in T. gondii which protects the parasitiphorous vacuole from attack by the mouse immune system [23]. We showed that this gene is pseudogenised in N. caninum and that N. caninum is unable to perform ROP18-mediated inactivation of immunity-related GTPases (IRGs) in murine cells. Our data suggest a reduced role for T. gondii virulence factor orthologues in N. caninum, for example, in relation to the virulence-associated rhoptry proteins ROP18, ROP16 and ROP5. The loss of ROP18 function in N. caninum might be adaptive, preventing killing of its host and promoting parasite survival in the species to which it is restricted. Intriguingly, it has been proposed and shown experimentally in viruses that reduced virulence is associated with the evolution of vertical transmission [51], [52], one of the most striking characteristics of N. caninum transmission in cattle. N. caninum may have increased successful vertical transmission from cow to calf by reducing virulence mechanisms, thus reducing the likelihood of host mortality. Alternatively, if ROP18 is only relevant to a subset of T. gondii intermediate hosts, its loss in N. caninum may reflect the fact that these intermediate host species are less important to N. caninum. Indeed, the importance of the cat-mouse cycle in the epidemiology of T. gondii may explain the evolution of ROP18-mediated inactivation specifically of murine IRGs, a mechanism which is certainly less relevant to N. caninum in which canids rather than felids are the definitive host. In fact IRG homologues are known to be present in the bovine genome [53] and novel N. caninum rhoptry genes could mediate IRG-defense in these hosts. Both of these scenarios suppose that N. caninum has become more specialized in its host range, suggesting that the common ancestor of N. caninum and T. gondii had a wide host range. In order to test this hypothesis it will be necessary to examine the genomes of coccidian outgroups such as Sarcocystis and Eimeria and to better characterize the function of those rhoptry proteins specific to N. caninum. The genomic resource we present will be useful in generating further understanding of apicomplexan genome evolution in general and coccidian biology in particular. Furthermore our description of these parasites will help to kick-start large-scale population-based studies to understand how genetic variation affects their biology. Neospora caninum Liverpool strain was originally isolated from the cerebrum of a congenitally infected dog [54]. N. caninum Liverpool and Toxoplasma gondii VEG tachyzoites were maintained as described previously [55]. Paired end reads of N. caninum DNA were generated from random subclone libraries and additional reads were directed to close gaps and improve the data coverage of low quality regions. All sequencing was performed using BigDye terminator chemistry and used AB 3730xl analyzers (Life Technologies). In total, 920k reads were obtained, quality-clipped and screened for contamination. 92% of reads were used in the final assembly. Based on an estimated genome size of 62 Mb for N. caninum the sequencing coverage was ∼8x. Sequence reads were assembled using PHRAP (P. Green, unpublished) into 960 supercontigs with an N50 of 354 kb. To reanalyse the ROP5 locus these reads were reassembled using Arachne [56]. N. caninum pseudochromosomes were generated by aligning supercontigs to T. gondii ME49 chromosomes using PROmer [57]. 242 contigs (90.4% of the sequence) aligned successfully to the 14 chromosomes of T. gondii. Of the remaining 718 contigs, 375 were removed due to contamination, poor quality or if they were <1 kb in length. The remaining 343 contigs were grouped as UnAssigned Contigs (UACs) and used in further analysis alongside the pseudochromosomes. Telomeres were identified by examining chromosome ends for the typical TTTAGGG septameric repeat. A genome resequencing library for N. caninum were prepared as in [58]. Sequencing was performed on an Illumina GAIIX as for transcriptome libraries. Illumina paired-end reads were mapped using SSAHA2 [59]. The ROP18 region of five N. caninum isolates (Table S3) was amplified in two overlapping sections using the following primer pairs: F1+R3 and F3+R1 (exp. product 1268 bp and 890 bp respectively) (supplied by Eurofins). ROP18_F1 – GAGTGCCACGGTCCTCTAAG, ROP18_R3 – ATTTGTCCGACGCAAAATTC, ROP18_F3 – GGCTTCTGCTCCAGTATTCG, ROP18_R1 – GCCTTATAAACCACCCGTCA. PCR Reagents were supplied by Qiagen. Toxoplasma genome sequences and gene models were downloaded from ToxoDb v5.2 (http://www.toxodb.org); they were generated at the J. Craig Venter Institute and have been kindly provided by the Toxoplasma research community. N. caninum gene models were created using several algorithms [60]–[66] trained on T. gondii Me49 (ToxoDB v4.2) and using ESTs from N. caninum Liverpool and NC-1 strains collected by the Gene Index Project [67]. The models were examined using the Artemis Comparison Tool [68] and where possible corrected based on evidence from synteny and sequence conservation with T. gondii and transcriptome sequencing evidence. We found a large number of erroneously unfused gene models. Using our T. gondii transcriptome data a total of 354 pairs of adjacent, same-strand genes were linked by reliably mapped bridging read-pairs. A further 449 genes in T. gondii were found to be parts of an adjacent gene but did not have spanning read pairs, usually being likely UTR segments. This resulted in a large drop in the predicted T. gondii gene count and we incorporated these corrections into our subsequent analysis. We used orthoMCL [69] to identify an preliminary set of orthologous groups between T. gondii and N. caninum. These results was modified using 679 manually identified orthologue pairs. We identified 6348 one-to-one orthologous gene pairs, which we then used to determine whether genes in these organisms tend to be shared with other apicomplexan species. We performed an orthoMCL with representative genes for the one-to-one (core) T. gondii/N. caninum set as well as predicted protein sequences for the following species: the plant Arabidopsis thaliana, the piroplasmic apicomplexan Babesia bovis, the apicomplexan Cryptosporidium parvum, the slime-mould Dictyostelium discoidium, human, the haemosporian apicomplexan Plasmodium falciparum, the yeast Saccharomyces cerevisiae, the piroplasmic apicomplexan Theileria annulata, the kinetoplastid Trypanosoma brucei and the diatom Thalassiosira pseudonana. Where a core gene was found to have an orthologue in three or more non-apicomplexan eukaryotes, we defined it as eukaryotic. Where it was not eukaryotic, but conserved amongst all apicomplexa, we defined it as conserved apicomplexan, If a gene was not conserved apicomplexan, but found in one or more apicomplexan species, other than T. gondii and N. caninum it was defined as apicomplexan. Remaining genes were considered specific to T. gondii/N. caninum. SRS genes and SAG domains were identified as in Wasmuth et al. (submitted). We identified pseudogenes as clusters of significant BLAST hits which did not overlap valid gene models. Putative pseudogenes were manually checked to determine whether rational gene models could be made and whether Ilumina resequencing data supported any stop codons. N. caninum SAG domains were clustered as in Wasmuth et al., to identify any novel domain subfamilies. N. caninum ROP, MIC, GRA and AP2 genes were initially determined by manually identifying orthologues of known T. gondii genes with reference to various studies [70]–[77]. Where homologous families of proteins fell into these groups, e.g. ROPK family for ROP, TRAP and MAR for MIC, novel members were sought using BLAST and HMMer. Poly A+ mRNA was purified from total RNA using oligo-dT dyna bead selection followed by metal ion hydrolysis fragmentation with the Ambion RNA fragmentation kit. 1st strand cDNA was synthesized using randomly primed oligos followed by 2nd strand synthesis to produce dscDNA. Fragments were selected for 200–250 bp inserts amplified by PCR to enrich for properly ligated template strands. Libraries were sequenced using the Illumina Genome Analyzer IIX in paired end mode for 2×76 cycles using proprietary reagents according to the manufacturer's recommended protocol. RNA-seq reads were aligned against the reference genomes using SSAHA2 [59]. Reads were included only where one end of the pair aligned uniquely to the genome and the distance between the pairs was within the expected insert size range, plus the expected intron length (80–4000 bp). We used Reads Per Kilobase of exon model per Million mapped reads (RPKM) normalised by the unique length of the gene as a measurement of expression level. We excluded positions which were non-unique from the length calculation using a kmer window of 75 bp, 37 bp either side of that position. Non-uniquely mapped reads were excluded by removing reads with a score <10. In order to determine whether or not a gene was expressed we calculated an RPKM threshold (Figure S7). We used DESeq to determine differentially expressed genes [78]. In each pairwise comparison of two conditions A and B (e.g. N. caninum day 4 tachyzoites with T. gondii day 4 tachyzoites) biological replicates were used for both A and B to gain more accurate estimates of experimental variation. Genes with an adjusted p-value of <1e-5 were considered differentially expressed. When considering differential expression between species rather than between different time points in the same species we considered only genes identified as pairwise orthologues. Orthologues of N. caninum and T. gondii are often different lengths and therefore we normalised the read counts for T. gondii genes based on the gene length of the N. caninum orthologue. We determined orthologous relationships between N. caninum, T. gondii, P. falciparum and P. reichenowi using orthoMCL [69]. Orthologous groups containing a single gene from each species were aligned using muscle [79] and those with less than 50% conserved positions across all four species (including gaps) were excluded, leaving 184 alignments. We further excluded those orthologous groups which we determined have not evolved in a clock-like manner. To do this we used a likelihood ratio test for a constant rate of evolution [80]. Likelihood computations on a fixed species tree under a model where branch lengths are free to vary, and under a model in which branch lengths were constrained to be clock-like were performed in PAUP 4b10 [81], under a general time-reversible model [82] incorporating both a proportion of invariant sites and a gamma distribution of rates across sites [83], [84]. The GTR+I+G model of evolution was applied for each locus independently. We performed this test for single-copy orthologous sequences, determined as above, from T. gondii, N. caninum, P. falciparum, P. reichenowi and P. berghei. Twelve of the 184 alignments were excluded because they were determined to have evolved in a nonclock-like manner. As clock-like evolution is the null hypothesis in this test, failure to reject a molecular clock can be either due to the true process of evolution being clock-like, or very close to clock-like for a particular locus, or because of a lack of power in the data to reject the clock. To test that this was not introducing a significant bias in our results, we looked at the estimate of divergence times from two different sets of loci: all loci that fail to reject the clock model, and only those genes in the top quartile of alignment length (which should have the most statistical power to reject the clock). Codeml [85] was used to calculate the maximum likelihood value of dS in pairwise runmode with the JTT model allowing 2 or more dN/dS ratios for branches. Using all 172 alignments the median 4-fold coding site synonymous substitution rate (dS) across pairs of N. caninum/T. gondii orthologues was 0.856 substitutions per site. Between P. falciparum and P. reichenowi this was 0.076, similar to that calculated by Neafsey et al. [86] (0.068; 95% CI [0.060–0.077]). We assume that these Plasmodia diverged 2.49 mya (95% CI [1.93–3.79]) [30]. We thus dated the split for N. caninum and T. gondii to 28 mya, or taking into account the confidence intervals for the Plasmodium divergence estimate, between 21.7 and 42.7 mya, after the divergence of the definitive hosts around 52.9 mya [87]. If we calculate the median dS values using only those longest 25% of the 172 orthologous groups, we get a dS of 1.230 for N. caninum/T. gondii and 0.114 for P. falciparum/P. reichenowi. This translates to a divergence time of 26.9 mya, This value is very close to that calculated using all 172 alignments. This suggests that a any tendency to exclude longer genes using the clock test has not biased our results. Enzyme Commission (EC) number mappings were extracted from the KEGG database [88] from 23 different species covering prokaryotes, archaea and eukaryotes and were mapped on to the corresponding genes in OrthoMCL database [69]. All N. caninum proteins that shared orthology with these enzymes were transitively assigned one or more EC number. KEGG pathway mapping/coloring tools were used to map EC numbers to pathways. The final set of N. caninum metabolic pathways was compared to that of T. gondii (EC numbers assigned and used in similar fashion to Neospora). Pathways containing significantly high numbers of gene expression differences were determined as discussed in Statistical analysis. Cell culture was performed as described in [37]. The following immunoreagents were used (dilutions in parentheses). From J.C. Howard (University of Cologne): mouse anti-Irga6 monoclonal antibody (mAb) 10E7 (1∶500) [89], anti Irga6 phosphopeptide Ab T102-555 (1∶5000) [23] Alexa 350/488/546/555/647-labelled donkey anti-mouse, rabbit and goat sera (Molecular Probes), donkey anti-rabbit- (GE Healthcare), donkey antigoat- (Santa Cruz Biotechnology) and goat anti-mouse- HRP (horseradish peroxidase) (Pierce) antisera (all 1∶1000). From P. Bradley (UCLA): mouse anti-N. caninum ROP2 family member monoclonal antibody 20B5D5 (1∶2000). 4′,6-Diamidine-2′-phenylindole dihydrochloride (DAPI, Invitrogen) was used for nuclear counterstaining at a final concentration of 0.5 mg ml−1. Saponin permeabilization and immunostaining were performed as described in [90], [91], except for slides stained with T102-555 which were permeablized in ice cold methanol for 20 min and blocked in 1% BSA in PBS for 30 min. N. caninum proteins were purified from a tachyzoite pellet and resolved into 127 contiguous bands using acrylamide gel electrophoresis. Bands were excised and digested with trypsin. LC MS/MS was carried out using an LTQ ion trap mass spectrometer (Thermo Fisher Scientific Inc, Waltham, MA, USA) with an electrospray ionization source, Tryptic peptides were eluted using a linear gradient of 0–50% (v/v) acetonitrile/0.1% (v/v) formic acid over 140 minutes followed by 100% (v/v) ACN/0.1% formic acid for 20 minutes and a further 20 minutes of 0% (v/v) acetonitrile/0.1% (v/v) formic acid. Protein identifications were made as in [92], those above 1% false discovery rate were discarded. 1053 proteins were found to have at least one significantly matching peptide. To determine whether certain gene functions were overrepresented in differentially expressed genes we assigned GO terms using InterPro2GO [93]. The hypergeometric test was used to determine overrepresented GO terms in pooled day three and four expression data with a DESeq q-value cutoff of 1e−5. The Benjamini-Hochberg method was used to correct for multiple hypothesis testing. Values of P<0.05 were considered significant. The hypergeometric test was also used in the same way to determine whether KEGG metabolic pathways were enriched in differentially expressed genes.

10.1371/journal.pcbi.1000226

Neutrality and Robustness in Evo-Devo: Emergence of Lateral Inhibition

Embryonic development is defined by the hierarchical dynamical process that translates genetic information (genotype) into a spatial gene expression pattern (phenotype) providing the positional information for the correct unfolding of the organism. The nature and evolutionary implications of genotype–phenotype mapping still remain key topics in evolutionary developmental biology (evo-devo). We have explored here issues of neutrality, robustness, and diversity in evo-devo by means of a simple model of gene regulatory networks. The small size of the system allowed an exhaustive analysis of the entire fitness landscape and the extent of its neutrality. This analysis shows that evolution leads to a class of robust genetic networks with an expression pattern characteristic of lateral inhibition. This class is a repertoire of distinct implementations of this key developmental process, the diversity of which provides valuable clues about its underlying causal principles.

The diversity of life is a consequence of changes in the genotype (genes and their interdependence), but it is upon the observable organism's morphology (phenotype) that natural selection acts. Thus, the study of genotype–phenotype mapping can reveal key mechanisms driving life's capacity of continuous evolution and resilience in diverse environments. In this context, it has been observed that small numbers of genes form robust functional developmental modules, hierarchically reused throughout development. Here we analyze the evolution of small genetic modules toward higher diversity and robustness. Given the small size of the gene network, we can afford to analyze all possible topologies and thus the entire fitness landscape. This exhaustive study as well as simulations of evolutionary processes uncover a set of genetic interactions producing robust and diverse phenotypes. We single out the distinctive features of these networks responsible for their stability against environmental and structural perturbations. More precisely, all these robust genotypes can be related to the key mechanism of lateral inhibition for which a cell of a given type inhibits its neighbors to keep them from adopting the same type. Their distinctive features can thus shed light on the underlying mechanisms leading to pattern formation through lateral inhibition.

The evolution of life forms on our planet has led to the generation of an enormous variety of living structures. How such patterns of organization emerge [1]–[3], how contingency [4] and constraints [5],[6] shape them and how they acquire robustness [7] are unanswered questions that have been at the forefront of biology for more than a century and are still open key questions. The research field encompassing these fundamental issues is referred to as evolutionary developmental biology or in short evo-devo [1]. With the increasing capacity of mathematical modeling to provide fresh insight into the biological processes [8], computer simulations and experimental approaches in this field have recently reached common ground (see [9],[10] as recent reviews). A major conceptual problem for the modeling approach to evo-devo is the mapping between genotype (hereditary genetic information) and phenotype (the physical characteristics of the resultant organism). It is the phenotype that determines the organism's chances of survival (fitness), as it is on it that natural selection acts. The set of all genotypes, their resultant phenotypes and associated fitness is called fitness landscape. Since Wright's pioneering idea in the early 30's [11] that the hill-climbing process of population's adaptive evolution intimately depends on how smooth or rugged the fitness landscape is, numerous theoretical works have been contributing to what now can be considered as the theory of fitness landscapes [5], [12]–[14]. Moreover, empirical studies of fitness landscapes can nowadays be performed in the laboratory [15]–[17], revealing the real evolutionary paths undertaken by the organisms, and thus opening a previously-unavailable window on the actual evolution process. The extensively studied theoretical case that has become the classic example of evolution in a fitness landscape is provided by RNA folding [18]–[20]. Here the genotype is defined by the nucleotide sequence, whereas the phenotype consists of the secondary structure formed by the (planar) pattern of the base pairs. Within the RNA context, the existence of iso-phenotypic genotypes (or neutrality) has significant implications in evolution, in general [21]–[24] and evo-devo, in particular [25]. More precisely, neutrality is hypothesized to allow a more exhaustive search in the genotype space and consequently, better accessibility to diverse and potentially fitter phenotypes [13],[26]. The neutrality feature has been encountered and studied in other works of similar nature to the RNA's, such as in the origin and complexification of the protein universe [27], or in tunable-neutrality models of abstract molecular species [28], but also in other fields of very different nature. An example is provided by a model of feed-forward signaling networks [29]. Here, a minimal Boolean network receives a set of input signals, and computes the output. The genotype is defined by the wiring diagram (the network topology plus the weight of each interaction), whereas the phenotype is specified by the Boolean computation being performed. An example closer to the current study is a Boolean model of genetic networks [30], a study that inquires on the requirement of “genetic flexibility” or more precisely, of phenotype continuity in evolution, and the subsequent constraints it may pose to species evolution in a changing environment. In a more recent work, the same group developed an evolutionary model of network evo-devo [31] that adds to the same approach as the current study, with the two works providing complementary clues on the evolution of minimal developmental modules. Again under the Boolean approach, Andreas Wagner's studies ranging from the “epigenetic stability” of developmental pathways [32] to bridging robustness and evolvability by means of neutrality features in models of gene networks [33] complete the framework in which the present work is formulated. Moreover, the present formulation constitutes a continuation of the model introduced in [34], as well as a Hawk's eye view of an isolated genetic sub-system. Its exhaustive study allows uncovering of features that are generally not accessible from statistical large-scale studies of similar nature. As far as we know, no parallel exhaustive analogies of Boolean approaches have been applied within the context of spatially-explicit evo-devo. We have addressed here the role of neutrality and robustness in the evolution of minimal developmental modules. It is now apparent that the genetic networks responsible for major events in the development of organisms present significant robustness to a wide range of perturbations [35]. Moreover, experimental works reveal that certain genes and their interactions are recurrently encountered in very diverse organisms (e.g. Homeobox genes [1],[36]), suggesting that minimal genetic modules may underlie fundamental developmental pathways. The current work is inspired by the pioneering theoretical and empirical analysis of developmental genetic regulatory networks in long-germ-band insects (Drosophila melanogaster) ([37]–[39] and references therein) and plant (Arabidopsis thaliana) development [40]. As anticipated by [34], Drosophila is a suitable model organism to inquire on small gene modules that control specific parts of the development process. The goal of the current work is not a precise explanation of a specific genetic module, but a description of possible underlying principles of network assemblage and evolution. In this context, our guiding questions are: what classes of spatial expression patterns can possibly emerge from signals mediated by juxtacrine (intra or inter-cellular) interactions in a minimal genetic network? Are there intrinsically robust modules and what are their defining characteristics? Our approach addressing these questions is organized as follows. We introduce the model of gene interactions whose dynamics provides the gene expression pattern. We present the minimal set of genes producing a specific, biologically-relevant expression pattern, and the exhaustive analysis of all possible gene interactions and their associated expression patterns. Among all these topologies, we identify those providing a robust expression pattern, being thus the candidates for the developmental modules discussed above. Ultimately, an evolutionary study of populations of such networks conditioned on diversity is presented, revealing rapid evolution towards robust stripe-like expression patterns. We show that the structure of the encountered minimal robust networks relates to the phenomenon of lateral inhibition, a widespread mechanism of biological pattern formation, emphasizing thus the importance of these minimal development-driving modules. As mentioned in the introduction, the present work has as biological reference existent information on the logic of early development in two model systems: in long-germ-band insects and plants. For insects, during the syncytium phase, a series of chemical stripes forms, which are actually alternating evenly-spaced bands of transcription factors encoded by the pair-rule genes. These different cell states, defined also by the subsequent expression of the segment polarity genes, will determine the future body segments. The mechanisms responsible for the expression stripes have been the object of numerous studies, initiatives that have emphasized the necessity to uncover the gene circuitry or gene network topology [41]. Even though the importance of temporal and spatial expression of genes in development [42] has been addressed and demonstrated prior to the introduction of the gene circuit method [41],[43], only in the last decade has become apparent (also experimentally- and computationally-feasible) that the unification of topological, positional and dynamical information of gene expression is compulsory [44],[45]. In parallel with this unifying view on the mechanisms of stripe formation, the search for the underlying developmental bricks, the key driving interactions responsible for the robustness and accessibility of the segment polarity developmental pattern, has been the object of several studies [46],[47]. The same approach was successfully applied to uncover the structural robustness of the neurogenic gene network also in Drosophila embryo development [48]. In our modeling, we have approached the minimal module issue from a different perspective: on a more basic level of pattern formation mechanisms, on one hand, and on a more general level than the particularities of the segment polarity or neurogenic gene network, on the other hand. In so doing, we have searched for an organizing module of robust pattern formation within the features inspired from the observed modules in developmental biology. In the calculation of the expression patterns of the genetic networks, we have continued the Boolean approach of [34], and we have inspired also from more recent studies and extensions of the reaction-diffusion (continuous) connectionist model 49–52. By the existence of these two approaches, continuous and discrete (Boolean), or analog and digital, respectively, the resultant conclusions can pinpoint gradient-specific and topological mechanisms responsible for specific processes. In this sense, both approaches are needed and thus necessary for a complete understanding. The Boolean modeling approach has been widely employed in modeling the logic of genome architecture, of which development is a constitutive part [31],[34]. These models have been shown to successfully recover the same expression patterns as those resultant from continuous models [37],[53]. Even though we emphasize here the literature on Boolean modeling in evo-devo, the continuous approach of reaction-diffusion models constitutes the standard tool for evo-devo. Since the revolutionizing work of Alan Turing on pattern-formation and morphogenesis [54], there have been rapid and continuous advances in our understanding of what are now called Turing patterns (see [55] for a recent review). As in the case of Boolean modeling, this approach too is constantly employed for addressing new questions in this field. Until recently there has been a significant emphasis on the analyses providing answers to how gene networks work, an answer being mechanisms such as Turing bifurcations. With the advances in computational methods, the issue of increasing interest is why the gene networks have the topology observed, an issue that needs to be addressed in the light of evolution. Again, it is a problem whose resolution is facilitated by applying both approaches, continuous ([56],[57] just to mention a few) and discrete [31],[52]. In the current model, the network is composed of N genes whose state can be active (state = 1), or inactive (state = 0). Among these genes, a number G are local genes that code for intra-cellular molecules, and the rest H, are hormones [50] that code for short range, diffusible paracrine molecules (see Figure 1). More precisely, the first group of genes interact intra-cellularly with all the genes, while the short-range signaling proteins coded by hormones interact only inter-cellularly with the local genes, affecting thus their expression in neighboring cells. In the previous formulation of [34], the two types of interactions, local and non-local, are referred to as the internal and external gene network, respectively. In this context, a standard term in evo-devo for “hormone” is morphogen [58],[59], whose gradient concentration determines the fates of surrounding cells. Intimately related to the already mentioned concept of positional-information, the diffusion-controlled concentration and residence-time of a morphogen are interpreted by cells as committing signal for a certain state. We have chosen to employ here the term morphogen instead of hormone [50], even though our Boolean approach does not distinguish gradients of concentration. We consider one-dimensional organisms composed of a collection of C cells. In our case, C = 8, with larger values having no substantial influence on the results presented here. The equations determining the time evolution of the pattern are(1)(2)where denotes cell index, , local gene's index, , morphogen index, and(3)(4)Here the G×N matrix A (internal network; see also Figure 2) includes the intra-cellular interactions (continuous arrows in Figure 1) and the H×G matrix B (external network), the inter-cellular interactions (dashed arrows in Figure 1). The interactions consist of either activation or inhibition, with the values of the matrix being +1 or −1, respectively. The function ∨ is the “OR” function (the result is 1 if either of the short-range signals from neighboring cells is active, and 0, otherwise). For the two extremes of the organism (the anterior and posterior poles), the cells have a single neighbor. The function Θ is the threshold function yielding 1 if the argument is positive, and 0, otherwise. As initial condition, a maternal signal is considered at the anterior pole (leftmost cell), with only the first gene being active, i.e. ; , where δij = 1 if i = j and zero, otherwise. For this initial condition and the chosen interaction matrices, we determined the steady states. More precisely, we only consider the one-state attractors (fixed point attractor), discarding thus the unstable and the oscillatory cases (see Methods). Using the previous definitions, we can define the mapping between wiring and pattern (Figure 2) as:(5)implying that for each genotype (genes' wiring) Wa = (Aij, Bkl)∈W, we have a phenotype (expression pattern) . As we shall see, this system shares with other genotype-phenotype mappings a set of interesting features. On one hand, one-point mutants of a given genotype can generate very diverse phenotypes, and on the other, multiple genotypes can generate the same phenotype (Figure 2). One can also see that the Boolean approach allows a direct relationship between genotype and phenotype, a discretization that would have been hampered in a continuous modeling. As a first approach, we have studied the diversity of expression patterns with the aim of characterizing this genotype-phenotype mapping for a specific case of (G,H). Additionally, by introducing a fitness function, we have studied how adaptation proceeds through the nature of the mapping Ω. In order to select a model for study, we have sought the existence of a specific expression-pattern feature that appears in all developmental modules studied so far. It consists in a stripe-like pattern of a one-cell-wide alternating active-inactive values. For the rules defined above, we found that the minimal number of genes capable of producing such an expression pattern is composed of 2 local genes and 2 morphogenes. Four-element networks have already been shown, through slightly different model assumptions, to be the minimal nets able to generate all possible types of Boolean spatial arrangements [52]. One can exhaustively study all the possible interaction networks of (N,H) = (4,2) as it is a tractable number: . For larger networks, the number of configurations becomes intractable for an exhaustive study, but we shall address the statistical study of larger networks as a continuation of the present work. Among the configurations for (N,H) = (4,2) and through the approach presented in Methods, there are 405 908 genetic networks that reach point attractors, giving rise to 457 different organism (or tissue) patterns produced by 43 distinct gene patterns. Some patterns are very common, as they can be produced by many distinct networks, while other patterns result from very specific topologies. Ordering or ranking by decreasing frequency associates thus a rank to the patterns, resulting into the distribution N1(r) of a rank r. It has been reported to follow Zipf's law, N1(r)∝a(b+r)−γ, for both RNA folding [19] and feed-forward signaling nets [52]. In the present case, the observed distribution follows a power law N1(r)∝r−γ, with γt = 2.3 for tissue patterns frequency (Figure 3A) and γg = 3.8 for gene patterns frequency (Figure 3B). The system of local-genes and morphogenes as described above presents symmetry with respect to the latter (Figure 3C). The symmetry in the local genes is broken by the initial condition–first gene is active. Therefore, a significant majority of the interaction matrices have a symmetric pair that is equivalent in terms of the interactions and thus resultant expression pattern (457 tissue patterns reduce to 263 unique or non-degenerate patterns). However, in the present study we have addressed also the issue of evolution, for which the totality of possible networks has to be employed in order to allow for different evolutionary paths. Thus, we chose to maintain this degeneracy. Robustness and evolution have been shown to be closely linked, even though there is no consensus on this correlation being entirely positive or rather a positive-negative trade-off [62],[63]. The developmental scheme has to be robust enough to guarantee a reliable organism but not too robust to impede evolutionary changes and thus improved adaptive solutions. In this direction, theoretical studies of gene networks can shed light on the mechanisms responsible for this trade-off. Such a task is difficult to assign to experimental approach but perfectly assignable to theoretical modeling, even though the inspiration and final results relate to the fossil record [64] and experimental work [65]. In this context and for the evolutionary part of our study, we have associated a fitness function weighting pattern complexity. The fitness function associated to a given phenotype is inspired in previous works on the RNA folding landscape [66] and it is:(6)with β = 0.01 and , where the parameters H and A are the entropy and activity measure, respectively. Networks giving rise to unstable expression patterns are attributed a minimum fitness value, F = 0.01. The measure of activity, A, is defined as the fraction of the genes active in at least one cell. As we study the case (N,H) = (4,2), the activity A takes the values 0.0,0.25,0.75,1. The activity A is introduced in order to guarantee that all genes are used at least once through development. The entropy of the resultant gene expression is a measure of the heterogeneity of the pattern and is defined in terms of(7)where H(gi) is the (spatial) entropy of the i-th gene. Since only ON-OFF states are allowed, it reduces to(8)being p1 the probability that gi takes the ON state, i.e. . As defined, H(gi) = 0 for a fully homogeneous pattern and H(gi) = log 2 for a pattern with equal number of ON and OFF states. Having defined the fitness function, the entire fitness landscape for the case-study of N = 4 can be calculated. A glance at the fitness landscape shows that only through one-link mutations, a given expression pattern and/or fitness value can be maintained in long neutral paths. For illustrative purposes, we arbitrarily chose an example of such neutrality in diversity in Figure 8 by a path of one-link mutations maintaining the expression pattern (fitness) and robustness. In our evolutionary study, we have used a constant population model (N = 500 networks) of non-overlapping generations, with the individual networks replicating according to their fitness. By simulating the temporal evolution of this population initiated by identical networks of only one link, we have witnessed the increase in the average fitness of the population as more diverse patterns appear. A couple of examples of such evolutionary paths is shown in Figure 9. We display both the time evolution of the mean fitness () and robustness (), and the corresponding path in the (L+,L−) space. As a general trend in our evolutionary experiments, we have noticed that the population rapidly becomes dominated by stripe networks, constituting a stable almost-unitary fraction of the total population. It is interesting to remark that, even though the mean robustness varies, it fluctuates around a high value. This behavior is expected, as one can infer from the robustness distribution of the stripe networks (Figure 5B). Even so, it remains an important result, as it intrinsically relates high robustness with stripe patterns. As a general characteristics for the evolutionary paths, we have noticed that all networks increasingly acquire positive interactions (Figure 9A and 9C) which provide an increase in diversity, and implicitly in the entropic measure H. The last steps prior to reaching the maximum fitness are characterized by the acquisition of negative regulatory interactions, stabilizing and diversifying the expression pattern. We wondered about the particularities of the networks of maximum fitness together with maximum robustness. First of all, there exist several such networks characterized by a proper balance between activating and inhibiting interactions (Figure 10). In average, this proper balance results to be L+/L−≈1, and ensures their robustness and the maximum diversity of expression pattern. Interestingly enough, all these networks present the same expression pattern, a stripe pattern in all genes. We expected that maximum fitness networks could be of non-stripe pattern, as maximum diversity can be obtained through other patterns too (e.g. an all-active half plus an all-inactive half; similar to the example of gene pattern in Figure 3B). Unexpectedly, no other pattern of maximum diversity other than the all-stripes one exists among the stable patterns. This points to the fact that, in such a minimal pattern-formation module, there is a tight inter-dependence between the stripe-like pattern and high robustness values. This is supported by two issues. The first indication is related to the robustness distribution of the stripe networks (Figure 5B) where it can be seen that they are biased towards high robustness values, with more than 60% of them having maximum robustness. The second argument, as mentioned above, relates to the fact that non-stripe networks of maximum entropy do not exist among stable networks. Even though individual genes may be present in an organism in the form of all-active half and an all-inactive second half, these individual gene patterns do not combine into an H = 1 stable organism. In fact, we notice that such a gene pattern exists in stable organisms only in combinations with null gene pattern. Finally, in support of the tight relationship between robustness and stripe networks, the neurogenic network in Drosophila embryo has been shown to present such inter-dependence [48],[67], and we shall come back to this issue shortly. Moreover, we remarked that all these robust stripe networks form a connected meta-graph or a neutral meta-graph, where connections imply one-link mutation. A relevant conclusion from this observation relates to the stability of the expression pattern against changes in the interaction rules. The robustness to the interaction rules relates to genetic robustness, in which gene knock-outs are contemplated. Such robustness has been observed for the developmental module that underlies the ABC model of floral organ specification in A. thaliana [53], consistent with an overall floral plan widely conserved among flowering plants. Similarly, structural alterations (gene knock-outs) of the neurogenic gene regulatory network in Drosophila appear to be well tolerated by the system from the point of view of the resultant gene expression [48]. In the general context of genome architecture, there is undeniable evidence of redundancy (or multiple backup circuits) [68] as a key element, though not unique, responsible for this structural robustness property. This type of robustness manifests itself by the resilience of circuit designs to the removal or loss of a given unit. In the relationship between robustness and modularity too, it is interesting to mention the distinction between redundancy and degeneracy [69], where degeneracy refers to different units performing a given function, while the redundancy relates to the presence of multiple identical copies of a unit. Several models have explored the role of evolution in driving the formation of backup circuits [70],[71], emphasizing the gene duplication processes as the primary dynamical building-block of innovation [72]. It is worth noticing that there exist several minimal networks at the root of all these particular best networks. By minimal we refer to the minimum number of genetic interactions leading to this robust fittest phenotype. For visualizing the relationship between them, we have represented all the fittest networks in the form of an inclusion directed meta-graph in Figure 11. Nodes represent networks, and we considered that network A is connected to network B if one link has been added to the network A to produce network B. As a detail, the size of the node is an indication of the number of constitutive interactions of the associated network. All these networks have in common the same gene expression pattern, a pattern characterized by stripe-like expression for all the genes (Figure 11). In addition to these symmetry considerations, we also noticed that pairs of these minimal networks (brackets in the upper part of Figure 11) share common construction of the stable expression pattern from the initial condition. For illustrative purpose, the steps necessary to reach the stable patterns have been drawn in Figure 12 for two minimal networks (networks indicated by an asterisk in Figure 11). One can identify a connection motif as the key element responsible for the robustness and diversity, a motif emphasized in Figure 12. By isolating this interactions in the colored boxes we emphasize also the fact that the inhibitory interaction can be provided either by a morphogen or a local gene (see the pairs under brackets in Figure 11). The resultant robust configurations and the interaction motif recall a key process in pattern formation, especially in developing tissues: lateral inhibition with feedback [55]. Lateral inhibition refers to a type of cell-cell interaction in which a cell that adopts a particular fate inhibits its immediate neighbors from doing likewise. The modeling of the neurogenic genes Notch and Delta, and their associated trans-membrane proteins sheds light on the mechanism of amplification of differences between adjacent cells [73]. Moreover, it has been shown that for the neurogenic network in Drosophila embryo, the lateral inhibition buffers the expression pattern against perturbations (knock-outs) [48], resulting in a tight correlation between robustness and stripe-like pattern mentioned above. We have considered in the present study the one-dimensional organisms, as this approach provides a clarifying perspective on the basics of pattern formation in such minimal networks, and thus a faster identification of the underlying key features for robustness and diversity. Preliminary results on the 2D (N,H) = (4,2) case yield interesting comparisons with the 1D case. Among these, slightly more than 10% of the most-robust fittest 1D networks constitute the set of most-robust fittest networks (according to eq. 6) in the 2D case. In this context, although most of our qualitative trends are also observed, the number of non-null stable patterns is slightly reduced (189 658 in 1D compared to 165 856 in 2D). This decrease is consistent with the higher degrees of freedom allowed by the dimensional increase. It also opens the possibility of increased instability, and thus less robustness. In future works, we shall inquire on the necessary features of the interaction network leading to the maintenance of robustness and diversity independently of the spatial framework. Embryonic development is a particular field in biology characterized by a constant feedback between theoretical analysis and experimental work. Even though experimentalists still remain cautious on the predictive power of the former, there have been important advances in clarifying the organizational principles of embryonic pattern formation [1],[3],[55],[58]. Restricting ourselves to studies on Drosophila development (even though the conclusions seem universal), extensive simulations have shown that topology constrains the possible behavior of a regulatory network [74]. Similar studies on plant development also support this conclusion [40]. Moreover, in the context of development and not only, a crucial relationship has been proved to exist between topology and robustness [39],[74]. It is thus apparent that under the requirements of a given phenotype, selection will ensure that increasingly stable networks of interactions evolve towards it. In this direction, developmental modules appear to play the major organizing role. These kernels of the entire developmental genetic network perform distinct regulatory functions and constitute information-processing units in the correct and precise unfolding process of development [75]–[77]. Thus, two of the central key topics of developmental biology are the evolution and robustness of patterning mechanisms, and the still unsettled relationship between them. In this context, we have studied small epigenetic networks that could behave evolutionarily as minimal modules capable of producing a stripe expression pattern similar to those common in early embryonic development. In the present approach the minimal number of genes capable of producing such an expression pattern is N = 4, number that allows an exhaustive analysis of the genotypic space. Considering both topological and robustness issues, we have determined the space of expression patterns produced by such module using a dynamical modeling inspired from previous related studies of Boolean and continuous models [50],[52],[78],[79]. Among all possible expression patterns, we have identified those presenting enhanced reliability in maintaining their expression through perturbations. From performing evolutionary experiments (Figure 9), we can conclude that the paths towards the most robust and diverse expression pattern are short. In other words, the optimal modules are rapidly encountered in the landscape. We find necessary a comparison between the above-mentioned continuous models and the currently employed discrete approach. The former works are related to a different class of assumptions, both on the dynamical side (namely, Michaelis-Menten kinetic description of gene-gene interactions) and in the type of questions being considered (namely, a statistical study of the parameter space and network structure). In these works, search algorithms explored extended regions of the parameter space and, once a pattern-forming network was found, a network reduction process was applied in order to find minimal modules. The leading mechanisms pervading the formation of stripes cannot be directly compared with our study (where the equivalent nonlinearities would be of higher order, Hill-like class). Moreover, we have concentrated here on a well-defined, small-sized network such that the calculation of the entire space of possibilities could be feasible. Exploring the landscape structure in such a systematic way would be much more difficult (if possible at all) under the continuous approximation, and thus our conclusions need to be restricted to the discrete level. Nevertheless, we consider that a direct comparison of results between continuous and discrete models requires a detailed dedicated study. At least in the segment polarity network in Drosophila, there is general agreement between continuous and discrete models. That is, comparison has been conducted between approaches associated to a given system and thus characterized by similar assumptions. A general comparison of capabilities and limitations of discrete versus continuous models has not been addressed, as far as we know, and it is thus an important open question. Here the analysis of the most robust modules uncovered a set of networks, all forming a meta-graph where links are one-point mutations between networks. The existence of this meta-graph is an indication of structural robustness of such networks, as many mutations can be neutral. Also associated to this set, there exist certain minimal networks responsible for robustness and diversity, and many additional interactions provide a back-up mechanism or alternative pathways. The generic properties of the optimal modules indicate thus that lateral inhibition is likely to be a generic form of creating ON-OFF spatial patterns, although the exact structure of the generating module might differ, given the observed neutrality. Future work will explore how these modules might emerge and evolve within larger gene regulatory webs, the underlying phylogenetic patterns as well as the impact of network topology on evolvability and developmental plasticity. The equations determining the evolution of genes' state in time are:where ∨ is the “OR” function. Similarly, genes coding for short-range signaling molecules receive inputs only from the first set,with specific equations at the boundaries reading: The function Φ(x) is a threshold function, i.e. Φ(x) = 1 if x>0 and zero (inactive) otherwise. Given the initial condition and after transient time T ( = N*C, with C the number of cells) time steps, we check on the stability of the resultant pattern, considering only the fixed-point attractors and not the oscillatory ones.We consider such a relatively short transient time as relevant to the evolutionary studies that we shall introduce in the following section. As defined, the phenotype in our model is given by the steady state defined by the N×C matrix P* given by:where and indicate the stationary values of each regulatory element after the transient. With our previous definitions, we can properly define the mappingwhere for each genotype Wa = (Aij, Bkl)∈W, we have a phenotype . The distance between two genotypes, Wa and Wb is defined by(9)where δa,b(Aij) = 0 if , and it is 1 otherwise. If d(Wa,Wb) = 1, the networks Wa and Wb are connected in a meta-graph (see Figure 11). The Python code developed for the calculation of the fitness landscape and for the evolution experiments is available as Protocol S1. The dataset corresponding to the landscape of the study case (N,H)  =  (4, 2) is also hosted online as Dataset S1.

10.1371/journal.pmed.1002471

Cell salvage and donor blood transfusion during cesarean section: A pragmatic, multicentre randomised controlled trial (SALVO)

Excessive haemorrhage at cesarean section requires donor (allogeneic) blood transfusion. Cell salvage may reduce this requirement. We conducted a pragmatic randomised controlled trial (at 26 obstetric units; participants recruited from 4 June 2013 to 17 April 2016) of routine cell salvage use (intervention) versus current standard of care without routine salvage use (control) in cesarean section among women at risk of haemorrhage. Randomisation was stratified, using random permuted blocks of variable sizes. In an intention-to-treat analysis, we used multivariable models, adjusting for stratification variables and prognostic factors identified a priori, to compare rates of donor blood transfusion (primary outcome) and fetomaternal haemorrhage ≥2 ml in RhD-negative women with RhD-positive babies (a secondary outcome) between groups. Among 3,028 women randomised (2,990 analysed), 95.6% of 1,498 assigned to intervention had cell salvage deployed (50.8% had salvaged blood returned; mean 259.9 ml) versus 3.9% of 1,492 assigned to control. Donor blood transfusion rate was 3.5% in the control group versus 2.5% in the intervention group (adjusted odds ratio [OR] 0.65, 95% confidence interval [CI] 0.42 to 1.01, p = 0.056; adjusted risk difference −1.03, 95% CI −2.13 to 0.06). In a planned subgroup analysis, the transfusion rate was 4.6% in women assigned to control versus 3.0% in the intervention group among emergency cesareans (adjusted OR 0.58, 95% CI 0.34 to 0.99), whereas it was 2.2% versus 1.8% among elective cesareans (adjusted OR 0.83, 95% CI 0.38 to 1.83) (interaction p = 0.46). No case of amniotic fluid embolism was observed. The rate of fetomaternal haemorrhage was higher with the intervention (10.5% in the control group versus 25.6% in the intervention group, adjusted OR 5.63, 95% CI 1.43 to 22.14, p = 0.013). We are unable to comment on long-term antibody sensitisation effects. The overall reduction observed in donor blood transfusion associated with the routine use of cell salvage during cesarean section was not statistically significant. This trial was prospectively registered on ISRCTN as trial number 66118656 and can be viewed on http://www.isrctn.com/ISRCTN66118656.

Given that cesarean section rates are rising worldwide, the need for promoting alternatives to blood transfusion in cesareans, such as harnessing the patient’s own reserves where feasible, is well recognised. While cell salvage in operations outside of obstetrics has been shown to reduce the need for donor blood transfusion, its effectiveness in cesarean section is unclear and is based on only two small controlled trials of cell salvage in cesarean section, with imprecise and inconclusive findings. Current obstetric guidelines on cell salvage have focused attention on the lack of high-quality research, recognising that although there is potentially a place for cell salvage in emergency blood loss during cesarean section, its use remains controversial. We conducted a multicentre randomised controlled trial, including 3,028 women at risk of haemorrhage during cesarean section, to test the effect of the routine use of cell salvage compared to the current standard of care on the need for donor blood transfusion. While donor blood transfusion rates were lower in the cell salvage group than in the control group (2.5% versus 3.5%, meaning that, on average, 1 in every 100 mothers given cell salvage avoided a donor blood transfusion), the difference between the groups was not statistically significant. Additionally, the study showed that in women with RhD-negative blood type who gave birth to RhD-positive babies, cell salvage was associated with increased maternal exposure to fetal blood. These findings indicate that routine cell salvage does not lead to a statistically significant reduction in donor blood transfusion rates in all women at risk of haemorrhage during cesarean section.

Childbirth by cesarean section is on the rise worldwide [1]. Excessive blood loss (haemorrhage) is an important cause of maternal death [2], emergency hysterectomy [3], and maternal critical care admission [4] among women undergoing a cesarean birth [5]. The treatment of major haemorrhage, in addition to optimising red cell mass and managing anaemia, includes strategies to minimise blood loss. Donor (allogeneic) blood transfusion is employed when the operative loss is life-threatening or when the mother has severe anaemia following arrest of haemorrhage. Red cell concentrates used in donor transfusion are a finite, nationally pooled resource in demand simultaneously by many clinical services [6]. Such transfusions also carry risks for recipients [7]. To promote alternatives to donor transfusion, harnessing the patient’s own reserves where feasible, is a recognised need [8,9]. Along with surgical expedience and medical therapy (including tranexamic acid [10,11]), the use of intraoperative cell salvage may reduce the pressure on transfusion services. Cell salvage, which collects, processes, and returns the woman’s own blood lost during surgery, is increasingly being deployed during cesareans. In theory, it reduces the infectious and allergenic risks associated with donor blood transfusion. It has also been shown to reduce the need for such transfusions in a wide spectrum of surgical disciplines [12,13]. However, obstetric practitioners remain concerned about the risk of amniotic fluid embolism and red cell isoimmunisation with the use of cell salvage [14,15]. Evidence for its effective, safe use in obstetrics is limited [16–18], and our systematic review [16] identified only 2 small randomised controlled trials with inconclusive findings [19,20]; thus, opinion about its value is not yet solidified [18]. We conducted a large, pragmatic, multicentre randomised trial to determine whether the routine use of cell salvage during cesarean section in women at risk of haemorrhage could safely reduce the need for donor blood transfusion in comparison to the current standard of care, where salvage is not routinely used. The SALVO study was designed as a pragmatic, multicentre individually randomised controlled trial with cost-effectiveness analysis. The study protocol was approved by the UK National Research Ethics Committee (North West–Haydock, approval number 12/NW/0513), and local permission was obtained in all participating obstetric units. The study protocol is available as S1 Text and can also be accessed at https://njl-admin.nihr.ac.uk/document/download/2007068. The trial was conducted in 26 UK obstetric units. No changes to the protocol design, statistical parameters, outcomes, eligibility criteria, or intervention were introduced during the study. Three substantial amendments to the protocol concerned changes to recruitment materials and strategies as well as clarifications. The findings are reported as per CONSORT guidelines (S2 Text). Our sample consisted of women who were admitted to the labour ward for delivery by emergency or elective cesarean section, with an identifiable increased risk of haemorrhage, who were at least 16 years of age, and able to understand written and spoken English for informed consent. We defined increased risk of haemorrhage as any emergency cesarean or as an elective cesarean for any reason other than maternal preference or known breech presentation, i.e., we excluded women undergoing an elective first cesarean due to either maternal preference or known breech presentation. We also excluded women with contraindications to either cell salvage or donor blood transfusion, such as active malignancy; sickle cell disease or trait; cultural, religious, or social beliefs against donor blood transfusion; or rare antibodies restricting the use of cross-matched donor blood. All study participants were provided with antenatal information about the study, and gave informed consent before being enrolled. All participants admitted for elective cesarean section gave written informed consent before enrolment. In the case of participants undergoing emergency cesarean sections, either written consent was obtained before enrolment, or, if this was not possible due to the urgency of the operation, verbal consent was obtained before enrolment, and written informed consent was then sought after delivery. Participating women were randomised by entry into a bespoke online system, using random permuted blocks of variable sizes to maintain allocation concealment, to either intervention or control, at a ratio of 1:1. Randomisation was stratified by treatment centre, indication for cesarean (emergency versus elective), placentation (abnormal versus normal), and multiple birth (twins or more versus singleton). Classification of indication for cesarean was based on urgency of delivery [21,22] as follows: emergency cesareans had varying levels of urgency based on the threat or potential threat to the life of the woman or fetus, whereas elective cesareans had no maternal or fetal compromise. Abnormal placentation was defined as a pathologically low-lying placenta (placenta praevia) or abnormally invasive placenta (placenta accreta, increta, or percreta) [23]. Allocation concealment with third-party randomisation helped minimise selection bias. However, given the nature of the intervention, it was not possible to blind local treatment staff to the allocation post-randomisation, but in general the staff caring postpartum were different to those involved in intraoperative care. Performance bias as a result of knowing the participant’s allocation could lead transfusion rates to vary. Pragmatically, the need for donor blood transfusion postpartum was determined according to the policies of each participating hospital, and donor blood transfusion rates and transfusion thresholds were monitored for compliance with these. Participants were allocated either to cesarean section with routine use of cell salvage (intervention group), i.e., salvage equipment set up at the outset of cesarean to collect, process, and return blood lost at surgery after delivery of baby, or to cesarean section with the usual standard of care (control group), i.e., without routine use of cell salvage. In life-threatening acute haemorrhage, women were managed in line with the standard of care for such an emergency [2,23], which potentially included the use of cell salvage in the control group. The intervention was delivered by staff (anaesthetists, operating department practitioners, midwives, or nurses as per local policy) who had been formally trained in the use of the cell salvage equipment, in accordance with local procedures and requirements for competence. In line with the pragmatic nature of the trial, no specific cell saver model was prescribed, and both standard and continuous transfusion models were in use. For patients randomised to the intervention, full cell saver set-up for both collection and processing was mandated as part of the study protocol, as was the return of any volume of processed blood. Other process factors, such as swab washing [24], leukocyte depletion filter use, or number of suckers used, were left to local policy—although swab washing was encouraged as it was expected to increase the volume of blood available for processing and thus for re-transfusion. The above process factors and adherence or non-adherence to the allocated intervention were captured on case report forms. For participants allocated to the intervention, we documented whether non-adherence was due to technical failure of the equipment or whether cell salvage was not set up, in violation of the protocol. For participants allocated to the control group, we documented whether non-adherence was due to acute emergency blood loss or whether cell salvage was set up from the beginning of the procedure, in violation of the protocol. As part of the continuous central trial oversight, sites with high rates of deviation were contacted and encouraged to review their procedures and equipoise. Participants were followed up until hospital discharge. Postnatal investigations captured the outcomes listed below. RhD-negative women with RhD-positive babies were assessed for anti-D dose given after delivery and exposure to fetal blood by a screening acid elution test (Kleihauer) to determine if additional anti-D was needed. Confirmatory flow cytometry tests were documented for Kleihauer tests indicating a fetomaternal haemorrhage of >2 ml. If additional anti-D was indicated or where fetomaternal haemorrhage was >4 ml, the results of repeat testing undertaken after 72 hours were documented to establish clearance of fetal cells from the maternal circulation [25]. Adverse events were monitored, investigated, classified (serious or not; related or not), and reported to capture data on the safety of cell salvage. The primary outcome was the rate of women receiving donor blood transfusion to manage haemorrhage and its consequences, either during cesarean section or between surgery and hospital discharge. The primary outcome was assessed at sites from medical records and subsequently verified by cross-checking transfusion laboratory records. Secondary outcomes included units of blood transfused, time to first mobilisation, length of hospital stay, pre- and postoperative serum haemoglobin, fetomaternal haemorrhage measured by Kleihauer acid elution test, maternal fatigue captured using the Multidimensional Fatigue Inventory (MFI) [26], safety outcomes (including transfusion reactions), costs of resources and service provision, and process outcomes (including volume of salvaged blood returned and technical failure of cell salvage). A sample size of 3,050 women (1,525 per group) was planned to detect an absolute difference in transfusion rate of 2% (5% in the standard care group, 3% in the cell salvage group, relative risk 0.6) with a power of 80% for a 2-sided test, and a type I error rate of 5% (for rate assumptions see the SALVO protocol). All analyses were performed using Stata version 12 and on an intention-to-treat basis. For each primary and secondary outcome, we analysed all participants with non-missing data for that outcome. This approach is valid if data are missing at random (MAR) [27]. Our analysis plan specified that if more than 5% of primary outcome data were missing, we would conduct sensitivity analyses to investigate the impact of departures from the MAR assumption on our conclusions. Numbers of participants with missing outcome data are recorded in the results. Univariate and multivariable regression were used to estimate crude and adjusted odds ratios (ORs) for binary outcomes and mean differences for continuous outcomes, along with 95% confidence intervals (CIs). Adjusted risk differences for the primary outcome were calculated from multivariable logistic regression results using the ‘nlcom’ procedure in Stata. Number needed to treat (NNT) was calculated as 100 divided by the risk difference in percent. ‘Time to event’ variables were analysed using Cox proportional hazard regression to estimate the hazard ratio (HR). Multivariable models adjusted for stratification factors (with treatment centre as a random effect) and factors identified a priori to be prognostic for the primary outcome. The adjusted analysis was pre-specified as primary: such adjustment typically achieves substantial improvements in power, even when covariates are balanced [28]. We performed 2 pre-specified subgroup analyses: analyses of treatment effect by indication for cesarean section (elective versus emergency) and by treatment centre. The first of these was analysed by statistically testing for an interaction between indication for cesarean section and treatment. The second was analysed by testing for a random regression coefficient for the effect of treatment at different centres, in addition to a random intercept. Post hoc we conducted an analysis of treatment effect by normal versus abnormal placentation, also by testing for an interaction term. We conducted 2 pre-specified sensitivity analyses: first, the primary analysis was redone excluding cases of placental abruption; second, we analysed the primary outcome where return of cell salvaged blood in the control group was reclassified as receiving a donor blood transfusion. Post hoc we also restricted the second sensitivity analysis so that only participants who received cell salvaged blood in the control group in an emergency setting were reclassified as having received donor blood. A cost-effectiveness analysis was carried out from the perspective of the healthcare provider (UK National Health Service) [29] based on the principal clinical outcome of the trial, with the results expressed as cost per unit of donor blood transfusion avoided. A decision tree model was used that collated all the relevant resource use, cost, and outcome data collected prospectively during the trial to compare the overall cost-effectiveness of cell salvage with standard care. The resource use for both groups of the trial was estimated by prospectively evaluating the individual components of cell salvage and standard care (bottom-up costing). Unit cost data were then attached to the resource use. A probabilistic sensitivity analysis was carried out to explore the effects of the inherent uncertainty in parameter estimates on model results [30]. A trial steering committee and an independent data monitoring committee provided oversight to the study. The UK National Childbirth Trust collaborated in the project by providing patient and public input through involvement in trial design and protocol development. Prior to this trial, a survey was conducted among women who received cell salvage, showing that they perceived the intervention as reassuring, safe, and preferable to donor blood transfusion (our primary outcome). A patient representative was a member of the trial steering committee to provide oversight and advice regarding recruitment, dissemination, and general trial management. We are planning to disseminate findings to participants in the form of a newsletter following primary publication of these results. Between 4 June 2013 and 17 April 2016, 3,054 participants were recruited. The trial ended after inclusion and treatment of the originally planned sample of participants, with the discharge of the last patient on 21 April 2016. After exclusions for eligibility and consent issues, 3,028 participants were randomly allocated to either control or intervention. Of these, 1,672 were scheduled for emergency and 1,356 for elective cesarean section. After excluding further participants due to vaginal delivery or transfer to another hospital, 1,492 participants remained in the control group and 1,498 in the intervention group for analysis (Fig 1). Baseline characteristics of participants were similar in the 2 groups (Table 1; additional characteristics are available in Table A in S1 Appendix). Adherence to the assigned intervention was 96.1% (1,434 participants) in the control group and 95.6% (1,432 participants) in the cell salvage group. In the cell salvage group, 50.8% had salvaged blood returned, averaging 259.9 ml (Table 2); there were 24 cases (1.6%) where the salvage machine was unavailable or out of order and 42 cases (2.8%) where the machine was simply not set up, in deviation from the protocol. In the control group, 15 participants (1.0%) had cell salvage used in an emergency and 43 participants (2.9%) had it set up from the start of the procedure, in deviation from the protocol. All participants had complete data on the primary outcome and on those characteristics specified as covariates in adjusted analyses. Overall, the transfusion rate was 3.5% in the control group versus 2.5% in the intervention group (adjusted OR 0.65, 95% CI 0.42 to 1.01, p = 0.056; adjusted risk difference −1.03, 95% CI −2.13 to 0.06; NNT 97, at the lower limit of 95% confidence NNT was 47 and at the upper limit the number needed to harm was 1,667) (Table 3). In the planned subgroup analysis, the transfusion rate was 4.6% in women assigned to control versus 3.0% in those assigned to cell salvage among emergency cesarean sections (adjusted OR 0.58, 95% CI 0.34 to 0.99), whereas it was 2.2% in women assigned to control versus 1.8% in women assigned to intervention among elective cesarean sections (adjusted OR 0.83, 95% CI 0.38 to 1.83) (interaction p = 0.46). The test for heterogeneity of treatment effect across treatment centres (random regression coefficient for centre) was non-significant (p = 0.09). In the exploratory subgroup analysis, the transfusion rate was 2.9% in women assigned to control versus 1.8% in those assigned to cell salvage among cesarean sections with normal placentation (adjusted OR 0.56, 95% CI 0.34 to 0.94), whereas it was 8.9% in women assigned to control versus 9.6% in women assigned to intervention among cesarean sections with abnormal placentation (adjusted OR 0.83, 95% CI 0.38 to 1.83) (interaction p = 0.28). The planned sensitivity analysis assuming that any return of cell salvaged blood in the control group was in place of a donor blood transfusion showed a reduction in the rate of participants requiring donor blood transfusion from 5.6% to 2.5% (adjusted OR 0.39, 95% CI 0.26 to 0.59, p < 0.001). A reduction was also observed when the sensitivity analysis was restricted to reclassifying only those who received salvaged blood in the control group for acute emergency blood loss (4.0% versus 2.5%, adjusted OR 0.56, 95% CI 0.36 to 0.86, p = 0.008). All secondary outcomes had less than 5% missing data, except for fetomaternal haemorrhage (Table 4). There were small differences between groups for time to mobilisation (median 0.74 versus 0.72 days for the control and intervention group, respectively, adjusted HR 1.11, 95% CI 1.03 to 1.19, p = 0.006) and length of hospital stay (2.131 versus 2.126 days, adjusted HR 1.08, 95% CI 1.00 to 1.16, p = 0.050). For the subgroup of RhD-negative mothers with RhD-positive babies, women assigned to the intervention group had a greater rate of fetomaternal haemorrhage ≥2 ml than women assigned to the control group (10.5% [n = 9] versus 25.6% [n = 21], adjusted OR 5.63, 95% CI 1.43 to 22.14, p = 0.013). When blood was returned in this subgroup during cell salvage, 48% of participants (n = 15) experienced exposure to fetal blood, compared to 13% (n = 6) when blood was not returned (see Table D in S1 Appendix). There were no differences between groups in other secondary outcomes, including adverse events. Of 18 events related to cell salvage, 16 were associated with leukocyte depletion filter use. Two serious adverse reactions were reported: one patient experienced tachycardia and difficulty breathing following re-transfusion of cell salvaged blood, and another patient experienced sudden hypotension after transfusion of 600 ml of cell salvaged blood. Both events were classed by the local investigator as life-threatening and potentially related to the use of cell salvage, in particular to the use of a leukocyte depletion filter (which was not mandated by the study protocol). In both instances, patients recovered fully after cell salvage was discontinued. There was not a single case of amniotic fluid embolism in any instance of cell salvage use, with or without leukocyte depletion filters (details of adverse events are available in Tables E–G in S1 Appendix). The result of the cost-effectiveness analysis, based on the intention-to-treat analysis, showed an incremental cost-effectiveness ratio (ICER) of £8,110 (US$10,303; €9,711) per transfusion avoided for cell salvage compared to standard care. The probabilistic sensitivity analysis shows that although cell salvage was more effective than standard care for avoiding donor blood transfusion, it is uncertain whether it was less or more costly than standard care. Overall, if a decision-maker was willing to pay £50,000 (US$63,520; €59,869) to avoid a donor blood transfusion, the probability of cell salvage being cost-effective was 62% (see S1 Appendix for more detailed data). This large, pragmatic, multicentre randomised trial showed that the routine use of cell salvage during cesarean section did not lead to a statistically significant reduction in the rate of donor blood transfusion in all women at risk of haemorrhage during cesarean section. Cell salvage was associated with increased maternal exposure to fetal blood among RhD-negative mothers. No other clinically relevant differences were observed in secondary outcomes. No cases of amniotic fluid embolism were observed, with or without leukocyte depletion filters. The cost-effectiveness of cell salvage is uncertain. To our knowledge, our study is the largest randomised controlled trial in the area of cell salvage in obstetrics, and the only large-scale exploration of the clinical effectiveness and cost-effectiveness of cell salvage in cesarean section. It was prospectively registered, robustly conducted, independently monitored, rigorously analysed, and transparently reported. We recruited to target with independent data monitoring, had minimal patient or data loss, and achieved comparability at baseline. Compliance with assignment was generally excellent, but the deployment of cell salvage in the control group was a weakness as it could have potentially averted the use of donor transfusion, reducing the control event rate. We could not ethically prevent such action in emergencies, but we performed a sensitivity analysis reclassifying such cases as having experienced the primary outcome (donor blood transfusion), which showed an effect consistent in direction with the main result. Our audit to evaluate the risk of performance bias did not show imbalance in compliance with local transfusion policies. Our primary analysis followed the a priori statistical analysis plan, written before unblinding the randomised allocation, as agreed upon with our trial steering and data monitoring committees. It adjusted for the variables pre-specified. The usual rule of thumb for sample size in multivariable logistic regression of 10 cases per variable [31] was met in the adjusted analysis model. These methodological features should provide confidence in the validity and reliability of the findings. The diversity of our sample, in terms of cesarean indication, age, ethnicity, and geographic spread across many treatment centres, adds to generalisability. A p-value that is in the region of 0.05, regardless of the side of the significance threshold on which it lies, deserves careful consideration. It would be incorrect to conclude that the addition of further data would push the p-value below the threshold [32]. We believe our observations can justifiably be classed as modest [32], but not certain, evidence that can be useful in decision-making. Our finding concerning the safety of cell salvage in cesarean sections shows that concerns about the risk of amniotic fluid embolism should not be a barrier to its deployment. The 2 serious adverse reactions observed are in keeping with known effects of leukocyte depletion filters [33]. If cell salvage is to be used, avoidance of these filters should be considered in order to reduce the risk of adverse reactions. Our finding concerning fetomaternal haemorrhage should be interpreted with caution. There were fewer than 10 events per variable in the model. This was in part because of large rates of missing data. Sensitivity analyses that assume worst- or best-case scenarios would inevitably give divergent results in this situation. It is debatable whether one could rely on accurately imputing missing outcomes from the data that were available. Despite these limitations, the risk of maternal exposure to fetal blood is a key issue for policies concerning obstetric use of cell salvage. There is a need to put mechanisms in place for maximising adherence to anti-D prophylaxis guidelines for the prevention of RhD red cell isoimmunisation. UK guidelines recommend a dose of 1,500 IU of anti-D following birth of an RhD-positive baby to an RhD-negative mother after cell salvage, with tests for fetomaternal haemorrhage to check if additional doses are needed [15]. The findings around the secondary outcome of fetomaternal haemorrhage highlight a need not only for long-term vigilance but also for research to determine the efficacy of anti-D prophylaxis, given that our study does not provide long-term follow-up data on RhD-negative mothers. The UK Serious Hazards of Transfusion haemovigilance scheme has flagged up the risk of sensitisation in women who do appear to have received appropriate prophylaxis [34]. Investigation is needed to determine if greater amounts of routine anti-D administration are required where cell salvage has been used in RhD-negative mothers. Additionally, the rate and severity of red cell isoimmunisation to rarer, non-RhD antibodies following cell salvage is unknown [16] and merits further study. Concerning policy-making for deployment of cell salvage, its cost-effectiveness is going to be an issue for funders of services. Even if routine use of cell salvage was shown to be clinically effective, it is currently unlikely to be considered cost-effective for routine use in all indications for cesarean sections. Emergency cesarean sections have higher blood loss, and in these, cell salvage is not currently in routine use in practice. The potential for benefit in this group merits confirmation through additional research. The future benefit will depend on the extent to which cell salvage represents good value for money when changes occur in the rate of cesarean section, the rate of donor blood transfusion, the quality of the supply chain of donor blood for transfusion, and the contingency to address shocks on the supply of donor blood. Further delineation of cost-effectiveness in high-risk subgroups, particularly in settings with a limited supply of blood for transfusion, will be helpful in guiding decision-making.

10.1371/journal.pntd.0005187

Engineered Aedes aegypti JAK/STAT Pathway-Mediated Immunity to Dengue Virus

We have developed genetically modified Ae. aegypti mosquitoes that activate the conserved antiviral JAK/STAT pathway in the fat body tissue, by overexpressing either the receptor Dome or the Janus kinase Hop by the blood feeding-induced vitellogenin (Vg) promoter. Transgene expression inhibits infection with several dengue virus (DENV) serotypes in the midgut as well as systemically and in the salivary glands. The impact of the transgenes Dome and Hop on mosquito longevity was minimal, but it resulted in a compromised fecundity when compared to wild-type mosquitoes. Overexpression of Dome and Hop resulted in profound transcriptome regulation in the fat body tissue as well as the midgut tissue, pinpointing several expression signatures that reflect mechanisms of DENV restriction. Our transcriptome studies and reverse genetic analyses suggested that enrichment of DENV restriction factor and depletion of DENV host factor transcripts likely accounts for the DENV inhibition, and they allowed us to identify novel factors that modulate infection. Interestingly, the fat body-specific activation of the JAK/STAT pathway did not result in any enhanced resistance to Zika virus (ZIKV) or chikungunya virus (CHIKV) infection, thereby indicating a possible specialization of the pathway’s antiviral role.

Dengue has represented a significant public health burden for a number of decades, and given the lack of dengue-specific drugs and limited availability of licensed vaccine, new methods for prevention and control are urgently needed. Here, we investigated whether genetic manipulation of the mosquitoes’ native JAK/STAT pathway-mediated anti-DENV defense system could be used to render mosquitoes more resistant to infection. We generated Ae. aegypti mosquitoes overexpressing the JAK/STAT pathway components Dome and Hop under the control of a bloodmeal-inducible, fat body-specific vitellogenin (Vg) promoter. These genetically modified mosquitoes showed an increased resistance to DENV infection, likely because of higher expression of DENV restriction factors and lower expression of DENV host factors, as indicated by transcriptome analyses. Expression of the transgenes had a minimal impact on mosquito longevity; however, it significantly impaired the mosquitoes’ fecundity. Interestingly, bloodmeal-inducible fat body-specific overexpression of either Hop or Dome did not affect mosquito permissiveness to either ZIKV or CHIKV infection, suggesting a possible specialization of JAK/STAT pathway antiviral defenses. Thus, our study is the first to provide a proof-of-concept that genetic engineering of the mosquitoes’ JAK/STAT immune pathway can be used to render this host more resistant to DENV infection.

Despite decades of attempts at disease control, dengue remains a major mosquito-borne arboviral disease, causing an estimated 390 million infections annually [1]. Without drugs and with only limited availability of a licensed vaccine, vector control has remained the most important approach to reduce disease transmission. Dengue virus (DENV: Flavivirus) is maintained in a population through a horizontal transmission cycle between Aedes mosquitoes and humans. After mosquitoes acquire an infectious bloodmeal, the virus needs to complete its infection cycle and end up in the mosquito’s salivary glands for transmission to occur. Three major DENV infection barriers have been described in various refractory Aedes aegypti strains. The midgut infection barrier does not allow the virus to establish infection after ingestion of an infectious bloodmeal, and the disseminated infection barrier does not allow the virus to escape from the midgut tissue and disseminate to other parts of the insect; salivary gland infection and escape barriers have also been described. If the virus can overcome these impediments, it can then be injected into a human host in the mosquito’s saliva, thus transmitting the disease [2]. The replication cycle of DENV from midgut to salivary glands in Aedes mosquitoes takes 7–14 days, but this time interval can vary depending on the mosquito and virus strains as well as the temperature [3–7]. The Janus kinase/signal transducer and activator of transcription (JAK/STAT) pathway is a conserved immune signaling pathway that regulates developmental processes and antiviral immunity in both mammals and insects. We have previously shown that the JAK/STAT pathway controls DENV infection in Ae. aegypti [8]. Transient activation of the JAK/STAT pathway through RNAi-mediated gene silencing of the protein inhibitor of activated STAT (PIAS) renders mosquitoes more resistant to DENV infection of the midgut, whereas silencing of the receptor Dome or the Janus kinase Hop renders the mosquitoes more susceptible to DENV infection [8]. The JAK/STAT pathway controls DENV infection as early as 3 days post-infectious bloodmeal (dpibm), suggesting that genetic engineering of the pathway for earlier activation after a bloodmeal might result in a DENV resistance phenotype, and therefore offers a likely strategy to reduce dengue transmission. Activation of the JAK/STAT pathway is triggered by cytokine binding to the extracellular domain of the receptor, Dome. This binding changes the conformation of Dome, resulting in a dimerization of the receptor and self-phosphorylation of the Janus kinase Hop. Activated Hop then phosphorylates the cytoplasmic tail of Dome to generate a docking site for the transcription factor STAT. Once STAT is recruited to the receptor, it is phosphorylated, leading to dimerization. Dimerized STAT is then translocated to nucleus to activate the transcription of JAK/STAT pathway-regulated genes [9]. The JAK/STAT pathway is also negatively regulated at various steps by the suppressor of cytokine signaling (SOCS) and PIAS proteins [10]. We hypothesized that activation of the JAK/STAT pathway prior to, or immediately upon DENV ingestion could significantly restrict virus infection, perhaps to a degree that would adversely affect DENV transmission. To activate the JAK/STAT pathway, we generated genetically modified Ae. aegypti that expressed Dome or Hop under the control of the bloodmeal-inducible, fat body-specific vitellogenin (Vg) promoter. These transgenic Ae. aegypti showed greater resistance to DENV infection than did wild-type (WT) mosquitoes, and they have enabled further characterization of the molecular interactions between DENV and the mosquito vector. Interestingly, while the JAK/STAT pathway-hyperactive mosquitoes showed increased resistance to two DENV serotypes (DENV2 and DENV4), transgenic pathway activation did not confer resistance to two other important arboviral pathogens, Zika virus (ZIKV: Flavivirus) and chikungunya virus (CHIKV: Alphavirus), suggesting that the mosquito’s innate immune system and the JAK/STAT pathway deal differently with different viruses. This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Mice were used only for mosquito rearing as a blood source, according to the approved protocol. The protocol was approved by the Animal Care and Use Committee of the Johns Hopkins University (Permit Number: M006H300). Commercially obtained anonymous human blood was used for DENV, CHIKV, and ZIKV infection assays in mosquitoes, and informed consent was therefore not required. A schematic of the gene constructs used to generate the VgDome and VgHop transgenic Ae. aegypti lines is shown in Fig 1A. The Ae. aegypti Dome and Hop genes were PCR-amplified from Ae. aegypti cDNA using the primers listed in S6 Table and cloned downstream of the vitellogenin promoter [11]. Ae. aegypti Dome (AAEL012471) was PCR-amplified from cDNA in two segments: bp 1–1531 and bp 1532–3432; full-length Dome was then obtained through PCR using the Dome1F_PstI and Dome2R_PstI primers, with equal proportions of each segment as template. Dome was cloned into the pBluescript II KS vector (Stratagene) at the EcoRV site. A 392-bp sequence from the putative terminator region of Anopheles gambiae trypsin was PCR-amplified from the vector pENTR-carboxypeptidase P-antryp1T [12,13] and cloned into pBluescript downstream of Dome at the XhoI/Klenow-filled site. A 2085-bp fragment from the promoter region of Ae. aegypti vitellogenin [11] was PCR-amplified from genomic DNA and cloned into pBluescript at the SmaI site upstream of Dome. The terminator sequence from the An. gambiae trypsin gene was cloned downstream of Dome. The AeVg-Dome-TrypT cassette was excised from pBluescript with FseI and cloned into the FseI site of the pBac[3xP3-EGFPafm] vector [14]. The resulting vector was used for embryo microinjections to generate the VgDome line. Ae. aegypti Hop (AAEL012533) was PCR-amplified from cDNA in two segments: bp 1–1516 and bp 1517–3408. Each segment was separately cloned into pBluescript at the EcoRV site. The 5’ and 3’ segments were digested out with EcoRI/SacI and SacI/SalI, respectively, and re-ligated into pBluescript at the EcoRI/SalI sites to obtain full-length Hop. The trypsin terminator sequence was cloned at the XhoI/Klenow-filled site downstream of Hop, and the Vg promoter sequence was cloned at the XbaI/Klenow-filled site upstream of Hop. The AeVg-Hop-TrypT cassette was excised from pBluescript with FseI and cloned into the FseI site of the pBac[3xP3-DsRedafm] vector. The resulting vector was used for embryo microinjections to generate the VgHop line. Embryo microinjections and initial screening for transformants were performed by the Insect Transformation Facility at the University of Maryland Biotechnology Institute using Orlando (Orl) strain Ae. aegypti. To generate the VgDome transgenic line, 565 embryos were injected with the transformation vector and the phsp-pBac helper plasmid. Of these, 279 survived to adults and were backcrossed to WT Orl adults in 19 pools. G1 larvae were screened for EGFP eye fluorescence (S1 Fig), and one pool was found to contain positives. Similarly, to generate the VgHop transgenic line, 613 embryos were injected with the transformation vector and the phsp-pBac helper plasmid. Of these, 132 survived to adulthood and were backcrossed to WT Orl adults in 10 pools. G1 larvae were screened for DsRed eye fluorescence (S1 Fig), and one pool was found to contain positives. Positive larvae were reared to adults and intercrossed to G5 to ensure homozygosity of the transgene. PCR confirmation of each line was performed with the VgPro R and ITRR2’ primers for the VgDome line and the AeVgPro R and DsRed S primers for the VgHop line (S2 Fig). When characterizing the transgenic lines generated in the Orl background, we discovered that the Orl strain was in fact highly refractory to DENV infection [5]. Since the low levels of dengue infection in this strain would make it difficult to observe the impact of transgenic JAK/STAT activation on the virus, we undertook the additional step of backcrossing the transgene into the DENV-susceptible Rockefeller/UGAL (Rock) strain Ae. aegypti for five generations. After outcrossing with the Rock strain, both the VgDome and VgHop transgenic lines were bred within the same strain for another five generations to ensure homozygosity. The WT Orl strain was mated with the WT Rock strain in parallel to serve as a control. These Rock-introgressed VgDome, VgHop, and WT (OrlxRock) lines were used for subsequent gene expression analyses and infection experiments. In an attempt to increase the induction of the JAK/STAT pathway, we crossed homozygous transgenic VgDome male mosquitoes with homozygous transgenic VgHop female mosquitoes in a ratio of 1:5 to generate a heterozygous hybrid VgDomexVgHop line overexpressing both Dome and Hop after blood feeding. All adult mosquitoes were maintained on 10% sucrose solution in a controlled environment at 27°C and 80% humidity with a 12 h light/dark cycle. Ae. aegypti mosquito lines was reared using standard rearing procedures and mouse blood (BALB/c 028, Charles River Laboratories) was used for blood feeding. The Ae. albopictus C6/36 cells (ATCC CRL-1660) were maintained in MEM medium (Gibco, USA) supplemented with 10% heat-inactivated FBS, 1% L-glutamine, 1% penicillin-streptomycin, and 1% MEM non-essential amino acids at 32°C and 5% CO2. Baby hamster kidney cells (BHK-21, ATCC CCL-10) were maintained on DMEM medium supplemented with 10% FBS, 1% penicillin-streptomycin, and 5 μg/ml Plasmocin at 37°C and 5% CO2. Green monkey kidney (Vero) (Sigma-Aldrich) cells were maintained in DMEM with 5% FBS and 1% penicillin-streptomycin at 37°C and 5% CO2. DENV serotype 2 New Guinea C strain (DENV2), DENV serotype 4 strain Dominica/814669 (DENV4), ZIKV strain FSS 13025, and CHIKV strain 99659 were used as indicated in experiments. Mosquitoes were orally infected with DENV2 or DENV4 via artificial membrane feeding, as previously described [8,15]. In brief, C6/36 cells grown to 80% confluence were infected with DENV2, DENV4, or ZIKV at a multiplicity of infection (MOI) of 3.5 and incubated at 32°C and 5% CO2 for 6 days. The infected cells were then harvested and lysed through 3 cycles of freezing and thawing (between dry ice and a 37°C water bath). CHIKV was amplified on Vero cells at an MOI of 0.01 and harvested approximately 36 h later. The propagation yielded virus titers of 106 to 107 PFU/ml. The viruses were then mixed 1:1 v/v with commercial human blood and supplemented with 10% human serum and 1 mM ATP. The bloodmeal was offered to mosquitoes via an artificial membrane feeding system. Each experiment was performed in at least two to three biological replicates, as indicated. Plaque assays for DENV2 were performed in the BHK cell line, while CHIKV and ZIKV were titrated on Vero cell monolayers, and plaques were visualized by staining with 1% crystal violet. TCID50 assays for DENV4 were performed in C6/36 cells and visualized using peroxidase immunostaining, with monoclonal antibody 4G2 (a flavivirus group-specific monoclonal antibody) [16] as the primary antibody and a goat anti-mouse horseradish peroxidase (HRP) conjugate as the secondary antibody. All procedures involving DENV and ZIKV infections were performed in a BSL2 environment, and procedures involving CHIKV infections were performed in a BSL3 environment. To determine transcriptomic changes in the fat body and midgut after blood feeding in the transgenic VgDome and VgHop mosquito lines, RNA samples of the transgenic lines were compared to WT mosquitoes at 24 h post-naïve bloodmeal (hpbm) using Agilent-based oligonucleotide microarrays, as previously described [5]. Each transcriptomic comparison was performed in 3 or 4 biological replicates. In brief, pools of abdominal fat body or midgut tissue from 10–15 WT or transgenic mosquitoes were collected at 24 h after a naïve bloodmeal. We used 200 ng of total RNA from each pool to generate cy3- and cy5-labeled dCTP probes. Hybridizations were performed according to the manufacturer’s instructions, and the arrays were scanned with an Agilent SureScan microarray scanner; spot intensity was extracted using Agilent Feature extraction software. The expression data were processed and analyzed as described previously [5]. Self-self hybridizations have been used to determine a cutoff value on these microarrays for the biological significance of fold changes in transcript abundance: 0.75 on a log2 scale, which corresponds to a 1.68-fold regulation [17]. Numeric microarray gene expression data are presented in S1 Table (fat body transcriptomes) and S3 Table (midgut transcriptomes). Gene Expression Omnibus (GEO) accession number for raw microarray data is GSE90515. Gene Ontology enrichment analysis was performed with the GOstats package in R (http://www.bioconductor.org/packages/release/bioc/html/GOstats.html) [18]. Over-representation of gene functional category based on previous classification [15] was performed using the hypergeometric test with the phyper package in R [19]. The results from the hypergeometric test are presented in S2 Table. We used RNAi-mediated gene silencing to study the function of candidate host and restriction factors in WT mosquitoes as previously described [15]. Primers used to generate the dsRNAs are listed in S6 Table. EGFP dsRNA was used as a negative control for all experiments, and gene silencing efficiency was determined 3 days after dsRNA injection by using real-time PCR with gene-specific primers (S6 Table). Mosquito longevity and fecundity assays were performed in three biological replicates as previously described [12]. Longevity assays with mosquitoes maintained on sucrose solution were performed with 3- to 4-day-old adult male or female mosquitoes. For the longevity assays involving JAK/STAT pathway activation, female mosquitoes were provided a single naïve human bloodmeal, followed by maintenance on a 10% sucrose solution. The number of dead mosquitoes was then monitored daily. For the fecundity assays, 3- to 4-day-old adult female mosquitoes were fed on human blood via an artificial membrane. Fed mosquitoes were individually transferred to oviposition tubes, and the number of eggs laid was monitored until 5 days post-bloodmeal. Pantoea spp. and Bacillus cereus isolated from a field site in Zambia [20] were used to represent Gram-negative and Gram-positive bacteria, respectively. Bacteria were cultured in Luria-Bertani (LB) medium at 30°C at 250 rpm for 12–14 h. Overnight cultures were washed twice with 1X PBS buffer, then resuspended in 1xPBS buffer to OD600 = 0.01. For bacterial challenge, we blood-fed mosquitoes with a naïve bloodmeal to activate the JAK/STAT pathway, then injected 69 nl of resuspended bacteria (approximately 400 bacteria per injection) into the thorax of each cold-anesthetized mosquito at 24 hpbm. Mosquitoes injected with 1X PBS were used as a negative control. To activate the JAK/STAT pathway in female mosquitoes upon bloodmeal acquisition, we generated the homozygous transgenic lines VgDome and VgHop, which over-express the pathway receptor Dome or the Janus kinase Hop under the control of the bloodmeal-inducible, fat body-specific Vg promoter (Fig 1A, S1 Fig). The Vg promotor has been shown to be activated after a bloodmeal and to reach its peak level of activity 24–48 h after blood ingestion [21]. Aedes mosquitoes usually acquire multiple bloodmeals during their gonadotropic cycle, especially when blood feeding is interrupted by a physical response from the host or probing in a non-optimal skin area [22–24], and we therefore hypothesized that transgene-mediated activation of the immune pathway by the selected promoter would likely prime the mosquito’s JAK/STAT-mediated anti-DENV defense for the next potentially infectious bloodmeal. To generate a hybrid transgenic line over-expressing both Dome and Hop simultaneously, male homozygous VgDome and female homozygous VgHop were mated in a ratio of 1:5. The offspring were then screened for the expression of both EGFP and DsRed (S1 Fig), checked for transgenic Dome and Hop expression in the fat body, and used for subsequent experiments to test their susceptibility to DENV. In the VgDome line, fat body expression of Dome was rapidly induced relative to WT mosquitoes, peaking as early as 6 h post-bloodmeal (hpbm), and again at 48 hpbm (Fig 1B). Dome induction in the hybrid line followed a similar pattern, albeit with an approximately 2-fold higher peak at 6 and 24 hpbm (Fig 1B). In the VgHop line, Hop expression was induced more gradually, peaking at 24 hpbm. Hop induction in the hybrid line followed a similar pattern, but with an earlier peak at 12 hpbm (Fig 1B). It is of course also possible that differences in transgene expression patterns could be due to position effects arising from transgene integration in different genomic locations. Expression of dengue virus restriction factor 1 (DVRF1; AAEL008492), a putative anti-DENV effector molecule known to be transcriptionally regulated by the JAK/STAT pathway [8], peaked at 24 hpbm in all three lines (Fig 1B), indicating pathway activation. Interestingly, DVRF1 expression was not increased in the hybrid line (Fig 1B), suggesting that the pathway may have become maximally activated by each of the two transgenes, and that there might be limiting factors downstream of Dome and Hop. We first investigated the effect of the transgene-mediated activation of the JAK/STAT pathway on DENV infection. Mosquitoes were first fed a naïve bloodmeal to activate the JAK/STAT pathway; 2 days later, they were orally infected with DENV2 via a second (infectious) bloodmeal. We determined midgut infection at 7 dpibm (Fig 2A), disseminated infection at 14 dpibm (Fig 2B), and salivary gland infection at 21 dpibm (Fig 2C). VgDome and VgHop mosquitoes showed significantly lower midgut DENV2 titers than did the WT mosquitoes (78.18% and 83.63% reduction in median titers for VgDome and VgHop, respectively). Both lines displayed an 87.37% (VgDome) and 94.21% (VgHop) reduction in median disseminated DENV2 titers; more importantly, the transgenic mosquitoes also displayed significantly lower virus titers in the salivary glands. Interestingly, the VgDome mosquitoes showed a significant reduction in DENV2 infection prevalence (percentage of mosquitoes with a detectable infection), but only at the stage of disseminated infection (13.74% reduction in prevalence), and not in the midgut or salivary glands (Fig 2G). VgHop mosquitoes had significantly lower levels of DENV2 prevalence at the disseminated infection stage (27.46% reduction) and in the salivary glands (39.14% reduction) but not in the midgut (Fig 2G). Next, to determine if pathway activation at the point of DENV2 infection was sufficient for mediating systemic resistance, we omitted the initial naïve bloodmeal, and offered VgDome and VgHop mosquitoes a single DENV2-infected bloodmeal. Only the VgHop line showed significantly lower midgut DENV2 titers compared to WT (42.86% reduction in median titers); midgut titers in the VgDome line were not significantly different from WT (Fig 2D). DENV2 midgut prevalences in both lines without initial bloodmeal were comparable to WT (Fig 2G). The hybrid line, which over-expresses both Dome and Hop, also displayed significantly reduced DENV2 titers and prevalence in the disseminated infection compared to WT when they were given a naïve bloodmeal before te DENV2 infection. However, this reduction was not significantly different from the homozygous VgDome and VgHop lines (Fig 2E and 2G). This result points to the possible existence of a limiting factor downstream of Dome and Hop, and is consistent with the induction patterns of DVRF1 transcripts described above. Since we observed no difference in DENV2 susceptibility between the hybrid and the VgDome and VgHop transgenic lines, we chose to use only the two homozygous lines for subsequent experiments. Previous studies on the role of the JAK/STAT pathway during DENV infection in Ae. aegypti [8,25] have only been performed with DENV2. To determine if the inhibitory activity of the JAK/STAT pathway on DENV infection is conserved for different DENV serotypes, we challenged the VgDome and VgHop lines with DENV4 and assessed disseminated infection (Fig 2F and 2G). Both lines showed a significantly lower DENV4 titers and prevalence compared to WT mosquitoes. Immune system activation and transgenic over-expression of certain immune-related genes have been associated with fitness trade-offs [26,27]; transgenic JAK/STAT activation may be particularly prone to this because the pathway also functions in insect development and other processes [28–30]. For this reason, we examined the impact of transgene expression on certain fitness parameters in our transgenic lines. We first examined the impact of Dome and Hop transgenesis on the longevity of male and female mosquitoes maintained on 10% sucrose solution (i.e. without a bloodmeal that would induce transgene expression). In male mosquitoes, the longevity of the VgDome line was comparable to WT, while the longevity of the male VgHop line was greater (by 4 days) than WT (Fig 3A). The longevities of the female VgDome and VgHop lines were comparable to WT, suggesting a minimal impact of the transgenes on the mosquitoes’ life span in the absence of a bloodmeal. We next examined the effect of bloodmeal-inducible transgene expression on female Ae. aegypti longevity. The longevity of the female VgDome and VgHop lines after blood feeding was comparable to WT, suggesting minimal fitness trade-offs in terms of mosquito life span when the JAK/STAT pathway is transiently activated. The VgDome and VgHop lines both produced significantly fewer eggs than did WT mosquitoes (Fig 3B), suggesting that transgene expression compromises fecundity. The lower egg production may in part be due to competition between the endogenous and transgenic Vg promoters for transcriptional machinery such as transcription factors and RNA polymerase. This is supported by our observation of reduced Vg gene expression in transgenic mosquitoes compared to WT after blood feeding (Fig 3C). The JAK/STAT pathway-regulated antiviral effectors responsible for suppressing DENV infection are largely unknown, except for two genes, DVRF1 and DVRF2, that encode putative secreted and membrane-bound proteins, respectively, of unknown function [8]. In an effort to comprehensively characterize the impact of JAK/STAT activation, we used whole-genome oligonucleotide microarrays to compare fat body transcriptomes of the transgenic and WT lines, at 24 hpbm. DVRF1 expression peaks at this time point, suggesting peak JAK/STAT pathway activity. As expected, DVRF1 transcripts were enriched in both transgenic lines relative to WT (S1 Table), an indication of pathway activation. The fat body transcriptomic analysis identified hundreds of JAK/STAT pathway-regulated transcripts belonging to various functional groups. Genes with diverse (DIV) and unknown (UKN) functions were particularly prominent (Fig 4A), which was not unexpected since the JAK/STAT pathway regulates a variety of biological processes. In VgDome mosquitoes, 130 transcripts (0.75% of the whole transcriptome) were enriched, and 71 (0.47%) were depleted compared to WT mosquitoes (Fig 4A and 4B). In VgHop, 254 transcripts (1.46%) were enriched compared to WT, and 204 (1.18%) were depleted (Fig 4A and 4B). Only 50 transcripts were commonly enriched and 18 commonly depleted in the fat bodies of both the VgDome and VgHop transgenic lines compared to WT mosquitoes (Fig 4A and 4C). Functional pathway analysis using the GOstats package in R [18,31] revealed that gene ontology (GO) terms related to cell cycle regulation were over-represented among these 68 commonly regulated transcripts (Table 1). This small overlap suggests the presence of complex regulatory mechanisms, such that over-expression of different pathway components activates different subsets of genes. Because of the incomplete gene ontology assignment of Ae. aegypti transcripts in Vectorbase (i.e., several immune-related genes have not been assigned immunity-related ontologies), we also performed an over-representation analysis based on a previously annotated gene functional category list [15] using the phyper package in R [19,31]. In both transgenic lines, transcripts with immune-related (IMM) functions made up the largest specific class of regulated transcripts (excluding those with diverse [DIV] or unknown [UKN] function), and were significantly over-represented (Fig 4B and S2 Table). Of the 659 Ae. aegypti immune-related genes (IMM), 2.58% were induced and 1.52% were repressed in the VgDome line compared to WT; in the VgHop line, 4.40% of all IMM genes were enriched, and 2.28% were repressed. The percentage of regulated genes in the IMM category was 2 to 3-fold higher than the average percentage across all functional categories (S2 Table). These results emphasize the importance of the JAK/STAT pathway in mosquito immune regulation, and are consistent with our observations that the transgenic lines control DENV infection better than the WT. Several JAK/STAT pathway-induced IMM transcripts may encode potential DENV restriction factors (RFs), i.e. proteins that inhibit DENV replication in the mosquito. Among the 50 genes commonly enriched in VgDome and VgHop, the IMM category represented the largest class (9 genes, 1.37% of the total IMM) (Fig 4A, S1 Table). These were: three C-type lectins (CTLs; AAEL005482, AAEL011610, and AAEL014390), three fibrinogen and fibronectin-related proteins (FBNs; AAEL006704, AAEL011400, and AAEL013417), two transferrins (TFs; AAEL015458, and AAEL015639), and a cathepsin b (CatB; AAEL015312). Super oxide dismutase (AAEL006271) was the only IMM gene among 18 genes commonly depleted in both lines. Over-expression of Dome and Hop also regulated specific subsets of IMM transcripts (S1 Table). Eight IMM genes were enriched in VgDome but not in VgHop mosquitoes, including three serine proteases (AAEL003279, AAEL000030, and AAEL006434), two Niemann-Pick Type C2 molecules (AAEL012064, and AAEL004120), a cathepsin b (AAEL007599), and a lysozyme C (AAEL017132). Twenty IMM transcripts were enriched in VgHop but not in VgDome mosquitoes. These included four cathepsin b genes (AAEL009637, AAEL009642, AAEL007585, and AAEL012216); four serine proteases (AAEL007969, AAEL007006, AAEL015430, and AAEL003625); a thioester-containing protein (TEP22; AAEL000087); and several anti-microbial peptides (AMPs) such as cecropins (AAEL000621, AAEL000625), defensins (AAEL003832, AAEL003841), a gambicin (AAEL004522); and a lysozyme P (AAEL003723). Upregulated IMM transcripts could potentially encode as-yet uncharacterized DENV restriction factors. FBNs, for example, are thought to play pattern recognition roles in Drosophila and in Anopheles mosquitoes [32–35], but their function in Ae. aegypti has yet to be elucidated. TEP22, which encodes a complement factor-like protein, was previously reported to be involved in the mosquito's anti-fungal response [9,32,36,37] but remains unstudied in the context of DENV infection. Finally, while the antiviral activities of AMPs from the defensin and cecropin families have previously been reported [38,39], the role of gambicin in anti-DENV defense remains unknown. Hundreds of transcripts belonging to functional classes other than IMM were differentially expressed in the fat body of VgDome and VgHop lines compared to WT (Fig 4A), reflecting the JAK/STAT pathway’s important roles in other biological processes such as cell development and homeostasis, as well as lipid metabolism [9,36,37]. Transcript abundances of several previously reported putative DENV host factors (HFs; genes that facilitate virus replication in the host) were significantly depleted in the transgenic lines compared to WT mosquitoes (S1 Table); these included vacuolar ATP synthase subunit ac39 (vATPase-ac39; AAEL0011025), sterol carrier protein 2 (SCP2; AAEL012697), and DEAD-box ATP-dependent RNA helicase (DDX; AAEL004978). This suggests possible mechanisms for the increased resistance to DENV observed in these lines. Down-regulation of virus HFs may act in parallel with the induction of virus RFs to limit virus infection. vATPase-ac39 transcripts were depleted 2.707 log2-fold in VgHop compared to WT mosquitoes. Knockdown of vATPase-ac39 and several other vATPase subunits, as well as chemical inhibition of vATPase activity with bafilomycin, have been shown to inhibit DENV replication in Ae. aegypti [40]. Sterol carrier protein 2 (SCP2; AAEL012697) transcripts were depleted 3.624 log2-fold in the VgHop line compared to WT; proteins of this family are thought to mediate intracellular trafficking of cholesterol and other lipids. Lipid metabolism is known to be influenced by the JAK/STAT pathway [36], and also plays important roles in the replication of DENV, an enveloped virus. DENV is thought to facilitate its own replication by altering the expression of lipid-binding proteins and enzymes involved in lipid biosynthesis, such as fatty acid synthases and Niemann-Pick type C protein family members [41,42]. Transcripts of the DDX gene were depleted 0.79 log2-fold in VgHop compared to WT. DDX gene family members are required for the replication of hepatitis C virus (HCV) [43,44], retroviruses [45,46], and Japanese encephalitis virus (JEV) [47]. DDX proteins are used by these viruses to regulate the translational machinery and for viral RNA transport to favor virus replication. However, the role of this gene family in the context of DENV infection in Ae. aegypti remains unstudied. Based on their expression patterns and previous reports of their function, we selected five candidate DENV restriction factors for further functional characterization through RNAi-mediated gene knockdowns: the immune-related genes FBN, TEP22, and gambicin, and two genes of unknown function, Ukn7703 and Ukn566 (Table 2). Ukn7703 (AAEL007703) encodes a putative secreted protein with a C-terminal beta-propeller domain distantly related to WD-40 repeats, which are involved in protein-protein interactions in several biological processes, including signal transduction [48]. Ukn566 (AAEL000566) is predicted to be a transmembrane protein with conserved cysteine positions; cysteine repeats have previously been reported to be important for the three-dimensional structure and function of receptor proteins such as LDL [49] and scavenger receptors [50]. We further selected two candidate host factors—DDX and SCP2—from among the depleted transcripts for further characterization (Table 2). The potential modes of action of these genes have been elaborated on in the previous section. Across the candidate genes, silencing efficiencies varied from 22% to 85% (S4 Fig). Our screen confirmed Ukn7703 as a putative DENV restriction factor (31.82% increase in median DENV titers when compared to the GFP dsRNA-injected group), and SCP2 as a putative host factor (85.71% decrease in DENV titers when compared to the GFP dsRNA-injected group) (Fig 5). Silencing of DDX also reduced DENV2 titers in the carcass by 61.43%, although this result was not statistically significant by a small margin (p = 0.0555) (Fig 5). This may be a result of the lower silencing efficiency (22%) achieved for this gene. An interesting feature of the VgDome and VgHop transgenic lines was that activation of the JAK/STAT pathway in the fat body also restricts DENV2 infection in the midgut. This is unlikely to be due to leaky activation of the Vg promoter in the midgut, since both transgenes were induced to much higher levels in the fat body compared to the midgut (31-fold for Dome; 7-fold for Hop) (Fig 6A). Further, DVRF1 transcripts were not bloodmeal-induced in VgDome midguts, and induced only two-fold in VgHop midguts (Fig 6B). To further investigate this, we again used whole-genome oligonucleotide microarrays to compare the midgut transcriptomes of transgenic and WT mosquitoes, at 24 hpbm. Intriguingly, Vg promoter-driven JAK/STAT activation regulated the expression of a larger number of transcripts in the midgut than in the fat body (Fig 6C, S3 Table): 415 transcripts (2.39% of the transcriptome) were enriched and 299 (1.72% of the transcriptome) depleted in VgDome midguts compared to WT; 365 transcripts (2.1% of the transcriptome) were enriched and 299 (1.72% of the transcriptome) depleted in VgHop midguts compared to WT (Fig 6C and 6D). Among these, 92 were commonly enriched and 57 commonly depleted in the midguts of both transgenic lines (Fig 6C). GO representation analysis of commonly depleted transcripts indicated over-representation of genes involved in proteolysis, protein metabolic processes, and lipid localization and transport, while no gene ontology was significantly over-represented in the enriched genes (Table 3). While IMM transcripts were over-represented in the fat body transcriptome, transcripts belonging to the digestion (DIG) functional category were over-represented in the midgut (Fig 6D). Certain putative host factors identified in the fat body transcriptome analysis were also depleted in the VgHop midgut; these included SCP2, and vATPase-ac39. VgDome midguts displayed a higher transcript abundance of Unk7703, a novel putative DENV restriction factor that was also induced in the fat body (S4 Table). These data suggest that JAK/STAT activation in the fat body can have a profound impact on distal organs, possibly through uncharacterized signaling mechanisms. Since JAK/STAT pathway activation resulted in the upregulation of numerous immune-related transcripts, we investigated whether transgenic mosquitoes also showed altered resistance to systemic bacterial infection. Independent transgenic mosquito cohorts were injected with either the Gram-negative bacterium Pantoea spp., the Gram-positive bacterium Bacillus cereus, or sterile PBS, after blood feeding. We found no resulting differences in mortality between the VgDome or VgHop lines and WT mosquitoes (S3 Fig). This is consistent with data from our previous study, in which transient silencing of PIAS, a negative regulator of the JAK/STAT pathway, had no effect on mosquito mortality upon bacterial infection [8]. It is possible that the regulated AMPs may have more specialized anti-DENV function or may not have anti-microbial activity against these particular bacteria. Similarly, a previous study of defensins from humans has also suggested that the anti-bacterial activity of certain AMPs is highly specific [51]. Further, since the JAK/STAT pathway is active against different DENV serotypes [52,53], we also investigated its role in mosquito immunity against CHIKV and ZIKV. Only the VgHop line was used in these experiments since it showed a more profound DENV resistance phenotype than the VgDome line. VgHop mosquitoes were provided with a naïve bloodmeal to activate the pathway, then orally infected with CHIKV or ZIKV via a second bloodmeal, as done for DENV. We measured both midgut and disseminated infection at both 7 and 14 dpibm (Fig 7; descriptive statistics are presented in S5 Table). At 7 dpibm, CHIKV titers were significantly higher in VgHop midguts compared to WT, but no differences in disseminated infection levels were observed (Fig 7A and 7B). At 14dpibm, infection levels did not differ between VgHop and WT in either tissue (Fig 7C and 7D). CHIKV infection prevalence did not differ between WT and VgHop cohorts at either time point or tissue (Fig 7E). ZIKV infection intensity did not differ significantly between the VgHop and WT lines at either time point or tissue (Fig 7F–7I). While disseminated ZIKV prevalence was significantly reduced at 7 dpibm in VgHop mosquitoes as compared to WT, this difference was absent by 14 dpibm (Fig 7J). Interestingly, in both the transgenic and WT lines, ZIKV disseminated much less efficiently from the midgut as compared to DENV and CHIKV (Fig 7J). Taken together, these data suggest that the antiviral activity of transgenic JAK/STAT pathway activation is restricted to DENV2 and DENV4. While previous studies have linked the JAK/STAT pathway with DENV restriction in Ae. aegypti, the biology and translational potential of this relationship remains poorly understood. To examine the impact of JAK/STAT pathway activation on mosquito biology and identify possible genes and mechanisms mediating the inhibition of DENV infection, we generated transgenic Ae. aegypti that activated the JAK/STAT pathway in the fat body after bloodmeal. We reasoned that this spatially and temporally controlled JAK/STAT pathway activation, as opposed to the more random RNAi-mediated activation, would enable a much more detailed analysis. Over-expression of Dome and Hop in the fat body prior to viral challenge controls DENV infection both in the midgut and systemically. Activation of the JAK/STAT pathway via a naïve bloodmeal prior to DENV exposure is required to maximize systemic resistance to the virus. In the absence of this immune pre-activation, only the VgHop line showed an increased resistance to midgut infection, and this effect was not as profound as when a naïve bloodmeal was provided. Pre-activation may boost the insect’s anti-viral defense, perhaps by priming uninfected cells and/or by maximizing the concentration of anti-viral effectors. Besides its impact on DENV, transgenic JAK/STAT activation also profoundly affected the mosquito’s transcriptome, resulting in differential expression of hundreds of transcripts in the fat body and midgut. This reflects the pathway’s known involvement in a variety of biological processes, including development, cell proliferation, lipid homeostasis, and immunity. In the fat body, the site of Dome and Hop over-expression, genes implicated in cell cycle regulation and kinase activity were over-represented. Most differentially regulated transcripts were specific to either Dome or Hop over-expression; these line-specific gene subsets suggest further complexities in JAK/STAT pathway regulation, such as novel branches, fine-tuning mechanisms, and multiple roles for the two transgenes. That profound transcriptomic changes, along with increased resistance to DENV, were also observed in the midgut despite transgene over-expression in the fat body suggests possible JAK/STAT pathway-mediated inter-tissue signaling and immune priming. The Drosophila JAK/STAT pathway exerts a similar systemic effect [5,54], and the mammalian JAK/STAT pathway also plays critical roles in systemic type I interferon-mediated antiviral responses [51,55]. Chakrabarti et al. recently demonstrated JAK/STAT pathway-mediated inter-tissue signaling in Drosophila [56], where septic injury triggers hemocytes to secrete the cytokine Upd3, which then activates the JAK/STAT pathway in the fat body and gut, resulting in gut stem cell proliferation and an antimicrobial response [56]. Our transcriptomic analyses provide insights into how JAK/STAT signaling may control DENV infection in the mosquito. JAK/STAT activation resulted in the broad induction of numerous transcripts encoding immune recognition and effector molecules, including several that have previously been shown to restrict DENV. In addition, functional pathway analyses revealed that digestion- and lipid transport-related transcripts were differentially regulated in transgenic mosquitoes; these processes have previously been shown to impact DENV replication [41,42,57]. Finally, JAK/STAT activation also down-regulated transcript abundances of several genes that have known or potential roles as DENV host factors. These include vATPase (required for viral genome entry into host cells), DDX (translation of viral proteins), and SCP2 (lipid trafficking and homeostasis). That the pathway appears to impact DENV through diverse mechanisms bodes well for its use in transmission control, since this reduces the likelihood of the virus populations evolving resistance. Through RNAi-mediated gene silencing assays, we provided further evidence that UNK7703 and SCP2 function as novel putative DENV restriction and host factors, respectively. UNK7703 is also induced in Wolbachia-infected Ae. aegypti [39,58], and is conserved among Aedes, Culex, and Anopheles mosquitoes (S4 Fig). It encodes a putative secreted protein with a C-terminal beta-propeller domain that is distantly related to WD-40 repeats; we speculate that it may be involved in cell signaling. SCP2 encodes an intracellular sterol carrier protein that facilitates cholesterol uptake in Ae. aegypti cells [59]. Knockdown or chemical inhibition of SCP2 was recently shown to inhibit DENV replication in Ae. aegypti Aag2 cells [60]. Here we confirm, for the first time, a role for SCP2 as a DENV host factor in vivo. Taken together with our transcriptomic data, which revealed regulation of lipid transport processes in both transgenic lines, and with previous studies [41,42], these results emphasize important roles for lipid homeostasis during DENV infection in Ae. aegypti. While our transcriptomic analyses have yielded interesting insights into JAK/STAT pathway biology, we recognize that immune signaling in a WT genetic context may differ from our transgenic setup. In this study, we have made an effort to further characterize several candidate genes (identified through analysis of our transcriptomic data) through gene silencing assays in WT mosquitoes; more studies of this nature are needed to better understand the role of the JAK/STAT pathway in natural settings. Although transgenic JAK/STAT pathway activation in the fat body effectively controlled both DENV2 and DENV4, it had no effect against two other important arboviral human pathogens, the alphavirus CHIKV and flavivirus ZIKV. While more detailed studies are required, this suggests a more complex nature for Ae. aegypti defenses against different arboviruses, and cautions against generalizing certain pathways as pan-antiviral. In Drosophila, hop function is required but not sufficient for the activation of Drosophila C virus (DCV)-induced immune genes [53]; similarly, it is possible that the JAK/STAT pathway is involved in the mosquito response against CHIKV and ZIKV, but that additional signals or regulators are also required to efficiently limit infection. CHIKV, in particular, does not appear to activate, and may even suppress, classical insect immune pathways in mosquito cells [8,15,61], suggesting that the mosquito mounts a very different response to this virus compared to DENV. Finally, since CHIKV and ZIKV reached higher infection levels than DENV in our WT mosquitoes, the replication rate of each virus could also have affected the pathway’s efficiency at controlling infection. It is important to evaluate the fitness impact of any potential transgenic strategy. While the transgenic and WT lines did not differ in longevity, it should be noted that the insects were maintained under laboratory conditions, with an abundant food supply and minimal environmental stress; further experiments will be necessary to fully evaluate the effect of transient JAK/STAT pathway activation on longevity in natural settings. Both transgenic lines showed impaired fecundity compared to WT mosquitoes. Reduced egg production has also been observed in transgenic An. gambiae lines in which the Vg promoter is used to drive gene expression [12]; the use of alternative fat body-specific promoters may help minimize this fitness disadvantage. In sum, our transgenic mosquito lines have provided valuable insights into the biology of the JAK/STAT pathway and its anti-DENV action, and allowed the identification of novel putative host and restriction factors. Further, this study serves as a proof-of-concept that genetic engineering of the Ae. aegypti JAK/STAT pathway has potential to increase resistance to DENV and further development and optimization, prior to extensive field-testing, could contribute towards the development of novel dengue control strategies. For example, it may be possible to achieve improved or total resistance by expressing additional transgenes that block the virus through different mechanisms, and/or by using more effective promoters. Recently developed powerful mosquito gene-drive systems [62,63], used circumspectly, are likely to make it possible to spread pathogen resistance genes in mosquito populations in a self-propagating fashion, even at a certain fitness cost.

10.1371/journal.ppat.1002442

Nef Decreases HIV-1 Sensitivity to Neutralizing Antibodies that Target the Membrane-proximal External Region of TMgp41

Primate lentivirus nef is required for sustained virus replication in vivo and accelerated progression to AIDS. While exploring the mechanism by which Nef increases the infectivity of cell-free virions, we investigated a functional link between Nef and Env. Since we failed to detect an effect of Nef on the quantity of virion-associated Env, we searched for qualitative changes by examining whether Nef alters HIV-1 sensitivity to agents that target distinct features of Env. Nef conferred as much as 50-fold resistance to 2F5 and 4E10, two potent neutralizing monoclonal antibodies (nAbs) that target the membrane proximal external region (MPER) of TMgp41. In contrast, Nef had no effect on HIV-1 neutralization by MPER-specific nAb Z13e1, by the peptide inhibitor T20, nor by a panel of nAbs and other reagents targeting gp120. Resistance to neutralization by 2F5 and 4E10 was observed with Nef from a diverse range of HIV-1 and SIV isolates, as well as with HIV-1 virions bearing Env from CCR5- and CXCR4-tropic viruses, clade B and C viruses, or primary isolates. Functional analysis of a panel of Nef mutants revealed that this activity requires Nef myristoylation but that it is genetically separable from other Nef functions such as the ability to enhance virus infectivity and to downregulate CD4. Glycosylated-Gag from MoMLV substituted for Nef in conferring resistance to 2F5 and 4E10, indicating that this activity is conserved in a retrovirus that does not encode Nef. Given the reported membrane-dependence of MPER-recognition by 2F5 and 4E10, in contrast to the membrane-independence of Z13e1, the data here is consistent with a model in which Nef alters MPER recognition in the context of the virion membrane. Indeed, Nef and Glycosylated-Gag decreased the efficiency of virion capture by 2F5 and 4E10, but not by other nAbs. These studies demonstrate that Nef protects lentiviruses from one of the most broadly-acting classes of neutralizing antibodies. This newly discovered activity for Nef has important implications for anti-HIV-1 immunity and AIDS pathogenesis.

Nef is a pathogenic factor expressed by primate lentiviruses. HIV-1 virions produced by cells that express Nef acquire unknown modifications that allow them to infect new target cells with higher efficiency. We hypothesized that Nef might alter the structure or function of the HIV-1 Env glycoproteins. In this study we tested whether Nef alters the sensitivity of HIV-1 to several agents that inhibit HIV-1 by binding to different parts of Env. We found that Nef confers 10 to 50-fold resistance to neutralization by two antibodies (2F5 and 4E10) that belong to one of the most powerful classes of neutralizing agents, which are active against a wide range of HIV-1 isolates. We established that Nef decreases the recognition of the virus particles by these antibodies, which bind to a domain of the Env adjacent to the retroviral membrane (MPER). Env from diverse HIV-1 isolates are equally sensitive to this activity, and Nef proteins derived from both HIV-1 and SIV retain the activity. By protecting lentiviruses from one of the most broadly-acting classes of neutralizing antibodies, this new activity of Nef might make a significant contribution to AIDS pathogenesis.

Nef is a multifunctional pathogenicity factor expressed by primate lentiviruses. Disruption of nef is associated with defective virus replication in vivo and delayed pathology [1]–[3]. At the cellular level, Nef has well-documented activities that include the ability to downregulate cell-surface molecules CD4 [4]–[6] and MHC-I [7], [8], and to modulate the threshold activation state of T-cells and macrophages [9]–[12]. Nef alleles derived from most SIVs also down-regulate the TCR/CD3 complex [13]–[15]. In addition, SIV Nef was recently found to counteract the restriction factor BST-2 [16], [17]. Perhaps the least understood of the many Nef functions is its requirement for the production of virion particles with maximal infectivity [18], [19]. The magnitude of this activity is greatest when particles are generated from lymphoid cells [20], though it is not a consequence of CD4 downregulation by Nef during virion production [18], [21]–[26]. Nef can be found in virions, but there is no evidence that Nef encapsidation is required to promote HIV-1 virion infectivity [27], [28]. Other virion modifications, then, must account for the higher infectivity of virions produced in the presence of Nef. Additionally, clues about Nef function might be gleaned from future comparative studies with glycosylated-Gag from gammaretroviruses; despite the absence of sequence homology with Nef, this protein substitutes fully for Nef in promoting virion infectivity [20]. Nef has a well documented ability to interact with adaptor protein complexes and to alter vesicular transport and the selection of vesicle cargo [29]. In addition, we have found that Nef interacts with the cellular GTPase dynamin 2 and requires intracellular vesicle formation which depend on both dynamin 2 and clathrin to increase viral infectivity. Incidentally, the cytoplasmic tail of Env from HIV and other retroviruses contains sorting motifs that interact with components of the intracellular vesicle transport system [30]–[32], so it is reasonable to suppose that Nef might influence the trafficking and incorporation of Env, as has been reported [33]. Nonetheless, previous studies have failed to detect an effect of Nef on the quantity of HIV-1 Env incorporated into virions [20], [34]. Therefore, in the present study we considered the possibility that Nef confers a qualitative, rather than a quantitative effect, on Env encapsidation. To probe for these putative modifications to virion-associated Env, we took advantage of neutralizing antibodies (nAbs) and other reagents that target distinct features of the Env glycoprotein, postulating that Nef-induced alterations would influence their binding to Env and therefore their neutralizing potency. To explore potential links between HIV-1 Nef and Env, we posited that, if Nef modifies the retroviral glycoprotein, this putative modification might alter susceptibility to Env-specific neutralizing agents. For this purpose, a panel of HIV-1 entry inhibitors was collected, each of which probes different features of Env. Reagents included dextran sulphate, which targets the V3 loop of SUgp120 [35], soluble CD4 and the monoclonal antibody b12 [36], which interact with the CD4 binding site on gp120, monoclonal antibodies 17B [37] and E51 [38], which recognize CD4-induced epitopes on gp120, monoclonal antibody 2G12 [39], which recognizes a carbohydrate-dependent antigen on gp120, a goat polyclonal antiserum raised against the entire gp120 protein, monoclonal antibodies 2F5, 4E10 and Z13e1 [40]–[42], each of which target residues within the membrane proximal extracellular region (MPER) of TMgp41, and the peptide T20 [43], which inhibits fusion via association with the 6-helix bundle. The sensitivity of wild-type HIV-1NL4-3 to each of these reagents was compared to that of Nef-defective HIV-1NL4-3. Pairs of these otherwise isogenic viruses were collected from the supernatant of acutely infected Jurkat T cells, normalized by exogenous reverse transcriptase activity, and inoculated onto TZM-bl reporter cells in the presence of the specified inhibitors. These cells bear a Tat-responsive, β-galactosidase reporter, and infectious events were enumerated by in situ X-gal staining of the target cell monolayer. Under these conditions, Nef increased the infectivity of lab strain HIV-1NL4-3 by up to 30-fold [20]. The Nef-positive virus inoculum would therefore have to contain 30-fold less virus particles than the Nef-defective counterpart in order to produce the same amount of infected cells. This would result in a 30-fold increase of nAb/antigen ratio for the wild-type virus compared to the Nef-defective virus. To normalize the virus infectious titres, while maintaining a similar antibody/antigen ratio between samples, viruses were first equalized based on their RT-activity, which estimates the physical amount of virions. The infectious titers were then normalized by adding 15 µM of the reverse transcriptase inhibitor AZT to the wild-type virus, a concentration that reduced virion infectivity 30-fold (Figure S1A). Nef had no detectable effect on the sensitivity of HIV-1NL4-3 to any of the reagents targeting gp120 (Figure 1A). In contrast, wild-type HIV-1NL4-3 was significantly less sensitive than its Nef-defective counterpart to neutralization by 2F5 and 4E10, two of the three monoclonal antibodies which target the MPER of TMgp41 (Figure 1B and Figure S1B). IC50 values derived from the fitted sigmoidal curves (Table S1) revealed that Nef increased the concentrations of 2F5 and 4E10 required to neutralize HIV-1NL4-3 by 5 to 10 fold (Figure 1C). However, Nef did not affect HIV-1NL4-3 sensitivity to Z13e1, a monoclonal antibody that also targets the MPER, nor did it alter sensitivity to the fusion inhibitor T20. The reproducibility of these results was tested under a variety of experimental conditions. The specific effect of Nef on susceptibility to neutralization by 2F5 was equally apparent by measuring Tat transactivation of a luciferase reporter, rather than β-galactosidase, in TZM-bl transduced cells (Figure S1C). Differential sensitivity of the two viruses to 2F5 was observed when wild-type and nef-defective HIV-1NL4-3 stocks were generated by transfection of proviral DNA, rather than by infection, of Jurkat T cells (Figure S1D). The effect was also reproduced with a different reporter cell line (Ghost-X4-R5) in which infection activates a GFP reporter (Figure S1E). Though the data shown with HIV-1NL4-3 in Figure 1 was obtained with infectivity normalized by AZT, the relative effect of Nef on neutralization sensitivity was evident whether or not infectivity was normalized with AZT (Figure S1F) and remained significant when inocula were normalized according to infectious titre, disregarding RT activity (Figure S1G). In the above experiments, virus stocks were produced in CD4+ Jurkat T cells. Since Nef downregulates CD4 from the cell surface, Env incorporation into virions in the absence of Nef might be perturbed by interference from CD4 [44], [45]: the requirement for a higher concentration of antibody to neutralize wild-type virus could result from greater quantity of Env incorporation into the wild-type than into the Nef-negative virus. This explanation seems unlikely since Nef caused decreased sensitivity to only two of the eleven neutralizing agents that target Env tested here. Nonetheless, to investigate this possibility directly, the effect of Nef on the efficiency of Env incorporation into virions was analyzed. Virions produced in Jurkat T cells by wild-type and Nef-negative HIV-1 were pelleted through a sucrose cushion and subjected to immunoblotting with a polyclonal antibody against SUgp120. A control sample expressing env in the absence of gag excluded that the Env signal in the virion pellet was due to contamination from free Env protein. Comparable amounts of Env (both gp120 and gp41) were associated with wild-type and Nef-defective virions, as visualized by Western blot (Figure 2A). For a more precise quantification of the Env incorporated into particles, the same virus samples were tested by an ELISA assay to quantify virion-associated gp120 (Figure 2B), which revealed that a similar amount of Env was present in virus samples, irrespective of nef. The effect of Nef on neutralization, therefore, does not reflect differential incorporation of Env into virus particles. Env glycoproteins encoded by laboratory-adapted strains such as HIV-1NL4-3 are not representative of those viruses most commonly found in natural infection. HIV-1NL4-3 uses CXCR4 as a co-receptor whereas the majority of HIV-1 strains in people are CCR5-tropic [46]. Additionally, lab strains are generally more sensitive to antibody-mediated neutralization of cell-free virions than are primary isolates [47], [48]. Therefore, the effect of Nef on neutralization of HIV-1 bearing CCR5-tropic Envs from primary isolates was examined. Wild-type and Nef-negative HIV-1 particles, pseudotyped with EnvJRFL, were produced by transfecting Jurkat cells with env-defective HIV-1NL4-3 proviral DNA, together with the EnvJRFL expression plasmid. Surprisingly, in contrast to the experiments with EnvNL4-3, in the absence of nAb, Nef caused no increase in the infectivity of virions pseudotyped with EnvJRFL; particle normalization based on RT activity resulted in equal infectivity for the wild-type and Nef-negative virion stocks, obviating the need for AZT to normalize infectivity. As with EnvNL4-3, the sensitivity of EnvJRFL-pseudotyped virus to neutralization by 2F5 or 4E10, but not by 2G12, b12 or Z13e1, was decreased by Nef (Figure 3A). The same specific effect of Nef on sensitivity to nAb activity was observed with EnvSF162 (Figure 3B). HIV-1NL4-3, HIV-1JRFL, and HIV-1SF162 are all clade B strains, the subgroup most common in the USA and Europe. In other regions of the world, non-clade B viruses predominate. In sub-Saharan Africa, for example, where the prevalence of HIV-1 is highest [49], clade C is common. Env glycoproteins from two primary, clade C viruses [50] were therefore tested for effects of Nef on sensitivity to nAbs. As previously reported [50], the clade C Env glycoproteins were insensitive to neutralization by 2F5 (data not shown). Virions pseudotyped with either of the two clade C Envs were 10 to 50-fold less sensitive to neutralization by 4E10 in the presence of Nef than in the absence of Nef (Figure 3C). Neither clade C Env was neutralized by 2G12 (not shown). One of the two was neutralized by b12, but Nef did not change the sensitivity of HIV-1 to neutralization by this antibody. Overall, then, the specific effect of Nef on the sensitivity to 2F5 and 4E10 was observed with Env glycoproteins derived from different clades, with either co-receptor preference, and irrespective of virus adaptation to tissue culture. The effect of Nef on nAb activity was originally examined for the purpose of identifying virion modifications that correlate with the Nef-associated increase in virion infectivity. The different viral pseudotypes used in this study had different levels of intrinsic infectivity (Figure S2), and Nef had a highly variable effect on infectivity. This was greatest (30-fold) when virions bore HIV-1NL4-3 Env, intermediate (9-fold) for Env from HIV-1SF162, minimal (2 to 4-fold) for the two clade C Envs, and undetectable for EnvJRFL (Figure 4A). In contrast, the effect of Nef on sensitivity to neutralization by 2F5 and 4E10 was the same for EnvJRFL as it was for EnvNL4-3, and significantly greater than EnvNL4-3 for the clade C Envs (Figure 4B). Thus, the effect of Nef on nAb sensitivity does not correlate and therefore is not a consequence of the Nef-mediated increase of infectivity. Since the infectivity of HIV-1 pseudotyped with the EnvJRFL was not changed by Nef, using viruses bearing this envelope glycoprotein avoids the problem of dealing with unequal infectious titres. JRFL Env was therefore used for most of the subsequent experiments. Nef downregulates CD4. Therefore, in the absence of Nef, the higher levels of CD4 that result might interfere with the function of Env. To test whether the effect of Nef on neutralization depended on CD4 expression in producer cells, HIV-1NL4-3 was generated from Jurkat D1.1, a CD4 negative subclone of Jurkat cells, and from HSB-2 [51], another CD4-negative T cell line. In both cases, the effect of Nef on MPER neutralization by 2F5 was similar in magnitude to the effect observed with virions produced from CD4-positive Jurkat (Figure 5A and Figure 1). To further investigate the variability of this activity of Nef in different producer cell types, HIV-1NL4-3 pseudotyped with the EnvJRFL was also produced from a panel of different cell lines, including the T-cell lines MT4 and CEM-SS, the B-cell line DG75, and the adherent cells HEK293T and TE671. Nef altered MPER neutralization of viruses produced from all cell lines tested (Figures 5B and C). However, virus produced from HEK293T was only minimally responsive to this effect. Although small in magnitude, the effect of Nef on the IC50 of both 2F5 and 4E10 was significant (p<0.05 calculated by 2-tail Mann-Whitney test) and specific, since Nef did not alter susceptibility to neutralization by 2G12 (Figure S3). These results show that the effect of Nef on neutralization does not depend on CD4 expression in producer cells. Moreover, the magnitude of this effect may depend on other cell-type-specific factors. To test whether Nef alters the neutralization sensitivity of virus generated by cells naturally infected by HIV, the effect was also tested on viruses produced with PMBCs derived from three different donors. Virus was generated by infecting PBMCs with a modified version of NL4-3 where EnvJRFL replaces the NL4-3 sequence. Using all three donors, the Nef-defective viruses were at least 10-fold more sensitive to neutralization by 2F5 and 4E10 than wild-type viruses, while their sensitivity to b12 remained the same (Figure 6). Nef, therefore, affects the neutralization sensitivity of virus derived from primary cells. To determine if this new activity is conserved among Nef proteins encoded by different lentiviruses, virions were produced by transfecting Jurkat T cells with a nef-defective HIV-1HXB2 provirus, together with Nef expression plasmids from a lab-adapted, clade B virus (HIV-1LAI), a primary, clade C virus, and from SIVAGM. With virions bearing EnvHXB2, all three nef alleles enhanced virion infectivity (Figure S4A) and conferred 5 to 10-fold resistance to 2F5 (Figure 7A). When HIV-1NL4-3 particles were pseudotyped with EnvJRFL, none of the nef alleles increased virion infectivity (Figure S4B), but all decreased sensitivity to neutralization by 4E10, 5 to 15-fold (Figure 7B), mirroring the nAb result obtained using EnvHXB2. This indicates that the effect of Nef on neutralization is conserved among disparate nef alleles, and confirms the independence of this phenotype from the effect on virion infectivity. MoMLV Glycosylated-Gag (Glycogag) substitutes for Nef in promoting HIV-1 infectivity but differs from Nef in that it does not downregulate MHC-I or CD4 [20]. To determine if Glycogag decreases susceptibility to neutralization by 2F5 or 4E10, EnvJRFL-pseudotyped single-cycle Nef-positive and Nef-defective viruses (Figure 8) were produced in Jurkat T cells by co-transfection of provirus constructs with plasmids expressing MoMLV-Glycogag or an empty control vector. Glycogag conferred to HIV-1 a decrease in sensitivity to 2F5 and 4E10 of identical magnitude to that produced by Nef. Glycogag expression did not further decrease 2F5 and 4E10 sensitivity of wild-type HIV-1, indicating that the activities of Nef and glycogag on susceptibility to neutralization are redundant. The absolute infectivity of the EnvJRFL pseudotypes was unaffected by either protein (Figure S5A) confirming that the effect on neutralization is not linked to enhanced infectivity. Glycogag had a similar effect on the susceptibility to neutralization using HIV-1 bearing EnvNL4-3 (Figure S5B), which is also fully sensitive to the effects of Nef and Glycogag on infectivity (Figure S5C). Altogether, results show that the activity of Nef on neutralization is not a prerogative of the lentivirus protein. Nef is a pleiotropic factor able to perform many activities and, via distinct surfaces, to interact with a plethora of cellular proteins. To determine if the effect on neutralization sensitivity is linked to other activities, a panel of Nef mutants (Table 1) was screened for effects on antibody neutralization. We tested the Nef mutated on the SH3 binding domain (PP72,75AA) [52], the mutant unable to interact with adaptor protein complexes, (LL164,165AA) [53]–[55], mutants unable to interact with dynamin 2 and thioesterase (L112A and FD121,123AA) [56], Nef mutated on the putative cholesterol binding motif (LYYK/RSSL) [57], and finally the myristoylation defective Nef (GG2,3AA) [27]. All nef mutations were inserted into an env-defective HIV-1NL4-3 provirus to allow expression in cis together with the rest of the viral genome. Single cycle of replication virus was generated by transfecting Jurkat cells with the virus constructs and the EnvJRFL expression plasmid, and the effect of Nef on sensitivity of neutralization by 2F5 and 4E10 tested using TZM-bl reporter cells. Among all Nef variants, only the myristoylation mutant had a defective activity on neutralization sensitivity (Figure 9 and Figure S6). In contrast, all other mutants retained the ability to decrease virus sensitivity to both 2F5 and 4E10. Since mutations abrogating Nef capabilities to enhance infectivity, to recruit src kinases, and to downregulate CD4 and MHC-I did not abrogate the activity on neutralization, we conclude that the latter is genetically distinguishable from the others and that the effect of Nef on neutralization is a novel activity. Being Nef myristoylation important to mediate the correct interaction of the protein with the lipid environment, results suggest that its association with the cell membrane is crucial for the activity on neutralization. HIV-1 TMgp41 interacts with intracellular transport machinery via leucine- and tyrosine-based sorting signals in the 151 aa cytoplasmic tail [30]–[32]. To determine if the ability of Env to engage the vesicular transport machinery is required for the effect of Nef on neutralization sensitivity, a stop codon was engineered that allows translation of HIV-1JRFL TMgp41 up to the 7th aa of the cytoplasmic tail. As previously reported [58], absence of the cytoplasmic tail had little effect on virion infectivity which, like full-length EnvJRFL (Figure 3D), remained insensitive to Nef (not shown). Despite the lack of the cytoplasmic tail, Nef conferred 12-fold resistance to neutralization by 4E10, but not by 2G12 (Figure 10). The cytoplasmic tail of TMgp41 is therefore not required for the activity of Nef on neutralization. Given that Nef is a cytoplasmic protein, and that the cytoplasmic tail of TMgp41 is the only part of Env within the cytoplasm, this result indicates that the effect of Nef on neutralization does not involve direct contact between Nef and Env. A quantitative, virion immunoprecipitation assay was established to determine if Nef decreases the efficiency by which 2F5 and 4E10 bind HIV-1 virions. Protein G-coupled beads decorated with the nAb to be tested were incubated with suspensions of EnvJRFL–pseudotyped, Nef-positive and Nef-defective viruses that had been normalized by RT activity prior to incubation. Beads were washed to remove unbound virions and the amount of virus captured was quantified using a PCR-based, reverse transcriptase assay [59]. Preliminary experiments showed that magnetic beads, rather than porous sepharose beads, provide a much superior tool, because non-specific binding of virus particles (either in the absence of antibody or Env) is negligible (Figure 11). In contrast, the background produced by sepharose beads was 10 to 100-fold higher than that obtained with magnetic beads, critically reducing the sensitivity of the assay (data not shown). The efficiency of virion capture varied among nAbs (Figure 11). However, in all cases, the amount of virus captured was significantly higher than the background obtained with virions devoid of Env or in the absence of nAb. 2G12, b12, and Z13e1 captured a similar amount of Nef-positive and Nef-defective virions (Figure 11A). In contrast, up to four-fold more Nef-defective than Nef-positive virus was captured by 2F5 and 4E10 (Figure 11B). Since MoMLV Glycogag also decreased HIV-1 susceptibility to the nAbs (Figure 8), the effect of Glycogag on virion binding by 2F5 and 4E10 was also tested. The capture assay was repeated with nef-positive and nef-defective HIV-1, generated in the presence or absence of Glycogag. Glycogag specifically decreased the efficiency of Nef-defective virion capture by 2F5 and 4E10 (Figure 11C). It did not change the efficiency of capture of the Nef-positive virus, indicating that the activity of the two proteins was redundant. These results mirror the effect of Nef and Glycogag on sensitivity to neutralization by these antibodies, and provide direct evidence that these proteins specifically reduce the efficiency of HIV-1 virion binding to 2F5 and 4E10. Here we describe a novel function for lentiviral Nef: it renders the HIV-1 virion refractory to the broadly-neutralizing antibodies 2F5 and 4E10. This effect was extraordinarily specific, and not the result of a global decrease in neutralization sensitivity, since Nef had no effect on sensitivity to nine other, well-characterized Env neutralizing agents. This Nef activity targets a property that is conserved among diverse HIV-1 Envs, irrespective of clade, chemokine receptor preference, or adaptation to tissue culture. Divergent HIV-1 Nef proteins, as well as an SIVAGM Nef that shares <40% amino acid identity with HIV-1 Nef, all have this activity, suggesting that it reflects a core Nef function. Analyses of Nef mutants revealed that the activity on neutralization susceptibility is a novel and yet unreported activity, mechanistically distinct from other Nef activities. This same specific effect on HIV-1 sensitivity to neutralization by 2F5 and 4E10 was evident with MoMLV Glycogag, indicating that this activity is shared by an unrelated protein encoded by a gammaretrovirus. The TM glycoprotein of gammaretroviruses contains an MPER with clusters of aromatic residues like those in the HIV-1 MPER [60]. Additionally, there are reports that the gammaretrovirus MPER is targeted by potent neutralizing antibodies [60]–[62]. Taken together, these findings suggest that, like lentiviruses, gammaretroviruses encode a protein to protect from similar broadly-acting nAbs that target the MPER of TM. Though HIV-1 particles pseudotyped with Env glycoproteins from a range of disparate HIV-1 isolates were all equally sensitive to the effect of Nef on neutralization sensitivity, the same virion pseudotypes responded very differently from one another with respect to the enhancing effect of Nef on virion infectivity. In particular, virions bearing EnvJRFL were fully sensitive to the Nef effect on neutralization but totally unresponsive to the Nef effect on infectivity. Such large Env-dependent variation in the effect of Nef on infectivity has not previously been reported. Given the evidence that virions pseudotyped with virus glycoproteins driving entry to an endocytic compartment are not sensitive to the effect of Nef on infectivity [63], we hypothesize that different HIV-1 Envs could target cell entry of virus particles to different pathways, altering the requirement of Nef for optimal infectivity. Deciphering the primary sequence determinants for Env responsiveness may prove valuable for a better understanding of the mechanism by which Nef promotes virion infectivity. Though the neutralization experiments reported here were initiated to understand how Nef promotes virion infectivity, the lack of correlation indicates that the two phenotypes are independent. This conclusion is also supported by the evidence that several Nef mutants, which lack activity on virus infectivity, retain an unaltered ability to decrease sensitivity to neutralization. A screen of Nef mutants revealed that Nef myristoylation is required for the activity on neutralization, while mutations impairing the interaction of Nef with some well characterized cellular partners, such as src kinases, adaptor proteins and dynamin 2, had no fundamental effect on this activity. Accordingly, overexpression of dominant-negative dynamin 2 had no effect on the ability of Nef to increase HIV resistance to 2F5 and 4E10 (Figure S7). We therefore conclude that the effect on neutralization susceptibility is unrelated to other Nef activities. Deletion of the TMgp41 cytoplasmic tail, and the intracellular trafficking signals that it possesses [30]–[32], did not alter the effect of Nef on neutralization sensitivity. This activity, then, is unlikely to involve intracellular redistribution of Env resulting from direct interaction with Nef, or from indirect effects of Nef on cellular signaling components. Additionally, the activity cannot be ascribed to an effect of Nef on Env encapsidation since the level of virion-associated Env was unchanged by Nef. The effect of Nef remained the same in the absence of CD4 expression in producer cells excluding more subtle interference of Env by its cognate receptor. In response to sequential binding of gp120 to CD4 and chemokine receptors, gp41 undergoes a series of structural changes, first an extended conformation, followed by a fusogenic bundle of six α-helices called the trimer of hairpins. Previous reports proposed that 2F5 and 4E10 bind to the pre-hairpin intermediate with an extended conformation [64]. However, Nef had no effect on sensitivity to T20 (Figure 1A), a fusion inhibitor that binds to the prehairpin intermediate [43], or to fusion of cell-free virions with cells [26], [65]. Perhaps the most compelling evidence that the extended conformation is not necessary for binding by 2F5 and 4E10 is the experiment presented here showing that these antibodies capture virus particles in the absence of receptor engagement (Figure 11). These findings are in agreement with recent studies showing that the extent of 2F5 and 4E10 binding to cell-free virions correlates with neutralization [66], and causes gp120 shedding [67]. Thus, the data here demonstrate that Nef attenuates the interaction of MPER-specific nAbs with cell-free virions, rather than modulating structural changes that occur subsequent to encounter with cells. Based on recent structural studies, the MPER epitopes targeted by 2F5 and 4E10 are believed to be partially embedded within the lipid bilayer of the virion [68]–. The neutralizing activity of these antibodies is proposed to rely on their ability to interact with membranes [71]–[73], and is dependent on the long hydrophobic CDR3 loop [74], [75]. This might facilitate their interaction with the lipid bilayer and be instrumental for the ability to dock with the epitope, to extract hydrophobic residues from the lipid environment [64], [70] and ultimately contribute to the neutralizing activity [73], [76]. In contrast, neutralization with Z13e1 might not require an interaction with lipids [64], in line with evidence that the crucial residues of the epitope are located on the solvent exposed face of MPER [64]. Interestingly, several studies have reported a link between Nef and lipid biosynthesis and trafficking. Nef was reported to induce expression of genes involved in cholesterol biosynthesis [77], to reduce cholesterol efflux [78] , to enhance the raft-like character of virions via an increase in their cholesterol content [57] and/or a preferential incorporation of sphingomyelin [79], a phospholipid with a neutral head group. The efficiency of the interaction of 2F5 is favored by phospholipids with negatively charged head groups [80], and the association of the MPER with membranes is favored by the presence of sphingomyelin and cholesterol [81]. We therefore propose that Nef, by altering the lipid composition of the virion, alters the susceptibility to neutralization by 2F5 and 4E10, either by reducing the preliminary contact of the antibodies with the virus particle, and/or by increasing the strength of the association of the MPER with the viral membrane (Figure 12), which would make the epitopes less accessible to the antibodies. Supporting this hypothesis, our experiments revealed that the effect on neutralization is totally dependent on Nef myristoylation, which is essential for the localization of Nef into lipid rafts [82]–[84] and was found to be required for enhanced synthesis and efflux of cholesterol. We found that mutating the cholesterol binding motif, which had been initially linked to an increased cholesterol content of virus particles, does not abrogate the activity on neutralization. However, the role of such a binding motif, however remains unclear, in light of a more recent study [79] which failed to confirm its cholesterol binding function. We have observed variability of the magnitude of the Nef activity on neutralization when screening virus produced by different cell lines, observing that virus generated by HEK293T cells is only minimally responsive. Interestingly, it has been recently reported that the lipid composition of cellular membranes and viruses derived from different cell lines can vary significantly. The composition of the cell membrane isolated from HEK293T and from the T-cell line MT4, as well as the lipid pool of progeny viruses derived from these cell lines were found to vary significantly in their sphingomyelin content [85]. It is therefore plausible that cell-type specific variabilities of the membrane lipid composition modulate the responsiveness of the progeny virus to the effect of Nef on neutralization. The most potent monoclonal antibodies targeting Env have been cloned from bone marrow and B-cells of HIV-1 infected patients, and were instrumental in identifying crucial antibody specificities associated with protection. Within gp120, these antibodies were found to target the CD4 binding site (e.g. b12 and VRC01 [36], [86]), CD4 inducible epitopes (e.g. 17B [37] and E51 [38]), a carbohydrate dependent epitope (2G12 [39]) and a quaternary structure-dependent epitope comprising the V2 and V3 loops (e.g. PG16) [87]. Within gp41, antibodies targeting the membrane proximal external region (MPER) were found to be the most potent and broadly neutralizing, including the monoclonal antibodies 2F5 and 4E10 [40], [41] and Z13e1 [42]. The use of MPER as an antigen to induce a protective immune response in vivo has therefore been widely tested [88], [89]. However, MPER-specific neutralizing antibodies are only rarely found in HIV-1 infected subjects [90], [91] and different immunization strategies using the MPER have failed to induce significant neutralizing immunity [88]. This could be the result of limited accessibility of nAbs to such native epitopes which are located at the interface with the retroviral membrane. By favoring the interaction of MPER with the membrane Nef might contribute to hiding these crucial viral epitopes from the humoral immune response. Long-term nonprogression to AIDS has been reported in people infected with Nef-defective HIV-1 [1]–[3]. It would be of interest to determine if the strong, broadly neutralizing antibody responses observed in some of these individuals [92] are caused by MPER targeting antibodies. Wild type, env-defective and nef-defective HIV-1NL4-3 and HIV-1HXB2 provirus constructs have been previously described [34], [93]. Nef mutations GG2,3AA, PP72,75AA, L112A, FD121,123AA, LL164,165AA and LYYK202,206-208RSSL were introduced into env-defective HIV-1NL4-3 by site directed mutagenesis. HIV-1 pseudotypes were produced using pSV-JRFL Env, pCAGGS SF162 gp160, and the two subtype C env-encoding plasmids SVPC13-ZM109F-PB4, SVPC16-CAP45.2.00.G3 (all from NIH AIDS Research and Reference Reagent Program). Deletion of the Env cytoplasmic tail was achieved by introducing a stop codon by site directed mutagenesis in position 704 of the Env ORF in pSV-JRFL env. The nef ORF of HIV-1 LAI (subtype B), 97ZA012 (subtype C) and SIVagm were expressed in PBJ5 plasmid as already described [20]. Minimal active MoMLV glycogag, truncated at residue 189, was also expressed in PBJ5, as already described [20]. Human lymphoblastoid Jurkat D1.1 (ATCC), HSB-2 (NIH AIDS Research and Reference Reagent Program), Jurkat T (modified to express the large T antigen from SV40), MT4, DG75 HAD and CEM-SS cells, were grown in RPMI (Invitrogen). Human HEK 293T and TE671 cells, TZM-bl and GHOST indicator cells (NIH AIDS Research and Reference Reagent Program) were grown in DMEM (Invitrogen). Media were supplemented with 10% fetal calf serum (PAA Laboratories) and cell cultures were maintained at 37°C and 5% CO2. Soluble CD4, AZT, saquinavir, the monoclonal antibodies b12, 17B, 2G12, the goat polyclonal serum ARP401 and 4E10 were obtained from the National Institute for Biological Standards and Control (NIBSC). The peptide T20, and the antibodies E51, 2F5, and Z13e1 were obtained from the NIH AIDS Research and Reference Reagent Program. Dextran sulphate (MW 500,000) was purchased from USB. Virions capable of a single round of replication were produced by transfecting suspension-growing cells using electroporation and adherent cells using calcium phosphate and Fugene 6 (Roche), with env-deficient HIV-1 proviral DNA and vectors encoding retroviral env glycoproteins at a 4∶1 ratio. Virus pseudotypes were harvested 48h after transfection. Replication-competent HIV-1 was harvested 48 hours after having infected Jurkat T cells or PHA and IL-2 stimulated PBMC with VSV-G pseudotyped viruses. Virus-containing supernatants were clarified by low speed centrifugation, and filtered through 0.45 µm pore filters. Single cycle infectivities were determined in triplicate by challenging target cells with serially diluted viruses normalized based on their reverse transcriptase activity [59], [93]. HIV-1 infectivities were revealed by staining infected TZM-bl cells with X-Gal as described [20]. When replication-competent HIV-1 was used, to limit replication to a single cycle and prevent syncytia formation, saquinavir (1 µM) and dextran sulphate (20 µg/ml) were added 2 hour after infection. Reverse transcriptase (RT) activity in the supernatants was quantified using a Sybr green I-based real-time PCR enhanced RT assay (SGPER) that possesses both high sensitivity and an extraordinary dynamic range. The assay is a modified version of that described earlier [59]. Briefly, virions in cell-free supernatants were disrupted by adding an equal volume of SGPERT lysis buffer containing 0.25% Triton X-100, 50 mM KCl, 100 mM TrisHCl pH7.4, 0.4 U/µl RNase inhibitor (RiboLock, MBI Fermentas). Lysed virions were used for reverse transcription of MS2 RNA template (Roche) [94]. Quantification of reverse transcribed products was carried out in a CFX96 thermal cycler (Biorad) using Sybr-Green I, hotstart Taq and reaction buffer (Fermentas), and a MS2 primer set already described [94]. A standard curve was obtained using known concentrations (expressed in functional units) of recombinant HIV-1 RT (Ambion). Sensitivities of the functional env-pseudovirus or replication competent NL4-3 to neutralizing agents were assayed on TZM-bl or GHOST cells, seeded onto 96-well tissue culture plates a day prior to neutralization. Viruses were normalized based on RT activity and the inocula were adjusted to produce between 1% and 3% infection of the monolayer. To equalize the level of infectivity and obtain a similar amount of Nef-positive and Nef-negative viruses, AZT was added to the Nef-positive virus at the final concentration of 15 µM when replication competent HIV-1NL4-3 was used. Viruses were incubated with serially diluted neutralizing agents for 1 hour at room temperature. The complexes were added to indicator cells, incubated at 37°C for 2 hours, followed by two washes with PBS before being cultured in fresh complete DMEM. The cells were incubated at 37°C for a further 40 to 42 hours before staining for β-gal or prior to measurement of luciferase (TZM-bl) or before flow cytometry (GHOST). When replication-competent virus was used, protease inhibitor Saquinavir (1 µM) (NIBSC) and dextran sulphate (20 µM) (USB) was added after the 2 PBS washes to limit replication to a single cycle of infection and to prevent formation of syncytia. Neutralization was measured by calculating the residual infectivity of treated virus samples considering the infectivity of the untreated sample as 100%. Fitted sigmoidal curves and IC50 were obtained using Prism (Graphpad) with the least square variable slope method and using the dose-normalized response protocol. Neutralizations were performed independently three times with each combination of virus and antibody to be analyzed and data shown are the average with standard deviations. Statistical significance for all data sets was assessed by subjecting the derived IC50 values from the triplicate independent neutralizations to 2-tail Mann-Whitney test. Differences with p<0.05 were considered significant. All differences in neutralization sensitivity described in results were found to be significant based on this test. To examine the association of Env with HIV-1 particles, viral particles present in filtered culture supernatants were pelleted through sucrose cushions as described [95]. Pelletable material and cell lysates were analyzed by SDS-polyacrylamide gel electrophoresis (PAGE) and Western blotting, using a mouse monoclonal antibody to HIV-1 p55/p24, a rabbit antiserum to GP120 (both from NIBSC, Centre for AIDS Reagents) and the monoclonal anti-gp41 Chessie 8 (from NIH AIDS Research and Reference Reagent Program). Virus samples were also analyzed by ELISA. Briefly, lysates were serially diluted in carbonate buffer (pH 9.4) and coated overnight onto maxisorb 96 well plates (Nunc), blocked for 1 hour with 2.5% non fat dry milk, and probed with a guinea pig anti-gp120 serum (1∶40 for 1.5 hours) followed by a HRP-conjugated secondary antibody (Jackson, 1∶10,000 for 1.5 hours). The signal was developed with TMB substrate before being stopped with 2M H2SO4 and measured at 450nm. A standard curve was generated using serially diluted recombinant SF162 gp140. An immunoprecipitation assay was used to study virus capture by nAbs. 10 µl protein G magnetic beads (Dynabeads, Invitrogen) were resuspended in 500 µl DMEM containing 10% FCS. For each sample subjected to immunoprecipitation, 1 µg nAb and 10 µl of beads were incubated in complete medium at room temperature with rocking to allow maximum nAb binding to Protein G. The beads were then washed twice with complete medium. Virus supernatant (500 µl) was added to the beads and incubated for 1 hour 37°C, with rocking. Unbound viruses were removed by 3 washes in complete medium. The virus bound to magnetic beads was lysed with 10 µl SGPERT lysis buffer and incubated for 10 minutes before being diluted 10-fold with SGPERT dilution buffer. The diluted lysate was then centrifuged at 800g for 1 minute to sediment the beads and the pelleted beads immobilized on the tube by applying a magnet. 10 µl of the lysate supernatant was then removed and used in the SG-PERT assay to quantify the amounts of virus captured. Buffy-coats obtained from anonymous blood donors were provided by the Blood Transfusion Center of the Hematology Service of the University Hospital of Geneva by agreement with the Service, after approval of our project by Ethics Committee of the University Hospital of Geneva (Ref# 0704). Peripheral blood mononuclear cells (PBMCs) were isolated from buffy coats prepared from healthy, anonymous donors using Ficoll-Paque Plus (GE Healthcare).

10.1371/journal.ppat.1006304

Fatty acid DSF binds and allosterically activates histidine kinase RpfC of phytopathogenic bacterium Xanthomonas campestris pv. campestris to regulate quorum-sensing and virulence

As well as their importance to nutrition, fatty acids (FA) represent a unique group of quorum sensing chemicals that modulate the behavior of bacterial population in virulence. However, the way in which full-length, membrane-bound receptors biochemically detect FA remains unclear. Here, we provide genetic, enzymological and biophysical evidences to demonstrate that in the phytopathogenic bacterium Xanthomonas campestris pv. campestris, a medium-chain FA diffusible signal factor (DSF) binds directly to the N-terminal, 22 amino acid-length sensor region of a receptor histidine kinase (HK), RpfC. The binding event remarkably activates RpfC autokinase activity by causing an allosteric change associated with the dimerization and histidine phosphotransfer (DHp) and catalytic ATP-binding (CA) domains. Six residues were found essential for sensing DSF, especially those located in the region adjoining to the inner membrane of cells. Disrupting direct DSF-RpfC interaction caused deficiency in bacterial virulence and biofilm development. In addition, two amino acids within the juxtamembrane domain of RpfC, Leu172 and Ala178, are involved in the autoinhibition of the RpfC kinase activity. Replacements of them caused constitutive activation of RpfC-mediated signaling regardless of DSF stimulation. Therefore, our results revealed a biochemical mechanism whereby FA activates bacterial HK in an allosteric manner, which will assist in future studies on the specificity of FA-HK recognition during bacterial virulence regulation and cell-cell communication.

Besides roles in nutrition, lipids also function as important signals in the regulation of prokaryotic and eukaryotic cells. In bacteria, fatty acids are part of the language of cell-cell communication known as quorum sensing for a decade. However, how bacteria detect these signals and regulate virulence remains elusive. Here, we provide multiple evidences to show that a full-length receptor histidine kinase, RpfC, directly binds to a fatty acid-based signal factor using a short sensor region. This binding event stimulates RpfC autokinase activity by triggering conformational change in its catalytic region, which is critical in regulating bacterial quorum sensing and virulence. Our results confirm a long-outstanding assumption in cell signaling of phytobacteria, and provide a technical pipeline to analyze fatty acid-receptor interactions.

Quorum-sensing is a process that bacterial cells communicate with each other to elicit specific physiological responses, including virulence against hosts [1,2]. How single-celled bacteria detect and respond to population density is a fundamental question in studying quorum sensing. Previous studies have reported that a number of chemicals, such as acylated homoserine lactones, peptides, quinolones, and small molecular fatty acids (FA), were implicated in the “bacterial languages” that are usually recognized by bacterial sensor histidine kinases (HK) to elicit quorum sensing [1,3]. Typically, HK and its cognate response regulator (RR) constitute a two-component signal transduction system (TCS), the predominant detection-response mechanism in prokaryotic cells. The N-terminal input region of an HK detects specific stimuli, and an invariant histidine residue within its C-terminal dimerization and histidine phosphotransfer (DHp) domain is autophosphorylated. The HK then modulates the phosphorylation level of the cytoplasmic RR by its phosphotransferase or phosphatase activity. Eventually, the RR uses its C-terminal output domain to regulate gene expression or cellular behavior [4]. In the past three decades, the basic biochemical processes of protein phosphorylation and dephosphorylation during TCS regulation have been well documented, however, as the first event to trigger the cell-cell communication, only a few of ligand-HK interactions were experimentally investigated [5,6]. Therefore, how HK recognizes various signals, especially in non-model bacteria, remains incompletely studied [7]. The major difficulty in studying HK-ligand interactions is the fact that the majority of HK are membrane-bound proteins with various hydrophobic helices [8]. Usually HK are enzymatic inactive in solutions containing detergents. Traditional strategies that express soluble truncated HK by deleting input regions and transmembrane helices then prevent the investigation of ligand-HK interactions [9,10]. In addition, HK are quite difficult to be crystalized so that the high resolution, three-dimensional structure of full-length HK is usually unavailable. This impedes the understanding of the structural mechanism of ligand-HK interaction [5]. Furthermore, signals like FA are hydrophobic molecules whose dissolution require organic solvents, these properties made the HK-ligand interactions difficult to be measured by commonly used biophysical methods, such as surface plasmon resonance (SPR) and isothermal titration calorimetry (ITC) [11,12]. Therefore, multi-disciplinary approaches based on extensive genetic analysis are needed to investigate FA-HK relationships. In bacteria, diffusible signal factor (DSF) is a special family of signaling FA molecules. Various DSF-family members (such as DSF, BDSF, CDSF, SDSF, etc.) have been found to control quorum sensing and virulence in a number of bacteria, including plant pathogen Xanthomonas spp. and human pathogens Pseudomonas aeruginosa, Stenotrophomonas maltophilia, and Burkholderia cepacia [13,14]. These FAs take part in inter-species and inter-kingdom communications between bacteria and other organisms, including bacteria, fungi and host plants [15–21]. The first identified molecule of this family, DSF, was found in the Gram-negative bacterium Xanthomonas campestris pathovar (pv.) campestris, causal agent of black rot disease of cruciferous plants which encode approximately 52 HKs [22]. DSF is a medium-chain FA with a cis-11-methyl-dodecenoic acid structure [23]. Previous studies have revealed that extracellular DSF stimulates a TCS, RpfC-RpfG, to control bacterial virulence and quorum-sensing [24,25]. Of them, RpfC encodes a putative hybrid-type HK with multiple phosphorylation sites, while its cognate RR, RpfG, was the first HD-GYP domain-containing protein proved to have the phosphodiesterase activity to hydrolyze second messenger c-di-GMP into GMP [26]. At low bacterial cell density, RpfC binds to and represses the DSF synthase RpfF by a receiver (REC) domain, preventing production of DSF [27]. At high cell density, high concentration of extracellular DSF activates RpfC-RpfG system to degrade c-di-GMP, releasing the suppression of c-di-GMP on a global transcription factor Clp that controls the expression of multiple virulence factors [15,28,29]. However, because of the afore-mentioned technical difficulties in studying membrane-bound HK-FA interactions, whether DSF is the ligand of RpfC and biochemical mechanism of RpfC activation is unknown. Recently, two PAS-domain-containing HKs, RpfS and RpfR, were shown to bind DSF in X. campestris and BDSF in Burkholderia cepacia, respectively [30,31]. However, both RpfS and RpfR are soluble, cytoplasmic proteins without transmembrane helices, they are probably cytoplasmic receptors of intracellular DSF and unlikely cell-surface receptors to sense extracytoplasmic stimuli. Therefore, how the bacterial pathogen detects DSF dispersed in external environment and triggers cell-cell communication remains an opening question. In this work, we show that DSF directly binds to a short N-terminal sensor of RpfC, elevating levels of RpfC autophosphorylation. A group of amino acid residues within the region adjoining to membrane were indispensable to DSF-RpfC binding and autokinase activation. DSF-RpfC interaction resulted in allosteric change in the DHp and catalytic ATP-binding (CA) regions, which may promote in trans phosphorylation of RpfC. In addition, substitutions of two amino acids within the juxtamembrane domain of RpfC caused constitutive activation of the HK. Our data revealed the biochemical mechanism responsible for the interaction between HK and FA, and provided insight into bacterial signaling during cell-cell communication. RpfC belongs to a group of hybrid-type of HK with sensing mechanisms associated with membrane-spanning helices [8]. The putative secondary structure of RpfC has two characteristics different from the prototypical HKs (Fig 1A): Firstly, the signal input region of RpfC contains five hydrophobic TM helices and a putative 22-amino acid (aa)-length, periplasmic sensor at the most front end of its N-terminus. Secondly, there is a short juxtamembrane domain (16 aa-length), rather than a HAMP linker (about 50 aa-length), connects the input region to DHp-CA domains. In addition, RpfC also contains a C-terminal histidine phosphotransfer (HPt) domain and a REC domain (Fig 1A). The enzymatic activity of RpfC has never been investigated before. To biochemically confirm that RpfC is a HK, a truncated, soluble RpfC protein (RpfCΔinput) lacking the N-terminal input region (including sensor and TM domains) was obtained and purified. However, RpfCΔinput did not exhibit any detectable autokinase activity (Fig 1B), suggesting that the input region is critical for maintaining enzymatic activity. To address this question, we obtained a full-length RpfC protein (RpfCFL) with a C-terminal His6 epitope tag. Two membrane-embedded forms of RpfCFL, liposome and inverted membrane vesicle (IMV), were reconstructed and purified. As shown in Fig 1C and 1D, both forms of RpfCFL exhibited clear autokinase activity, making it possible to enzymatically investigate the mechanism of RpfC activation. To determine if DSF affects the enzymatic activity of RpfC, DSF was added to reaction mixtures containing the liposome or IMV forms of RpfCFL. As shown in Fig 1C and 1D, the level of RpfCFL-P phosphorylation approximately doubled compared with the control. Kinetic analyses of the IMV and liposome forms of RpfCFL showed an increase in the phosphorylation level of the IMV form at 30 s post DSF addition, whereas a similar increase was not detected until 2 min for the liposome form. This difference might be caused by variation in the phospholipid compositions of the IMV and liposome forms, which would affect autokinase activity. In addition, dose-response analysis of DSF on the activity of RpfC revealed that addition of 0.5 μM DSF was sufficient to elicit a detectable increase in the level of RpfCFL-P (Fig 1E and 1F). This concentration agrees with the previously reported minimal bioactive concentration of DSF (approximately 0.5 μM) that required for eliciting cell-cell communication [23]. Increasing the DSF concentration resulted in a logarithmic increase in the RpCFL-P level, and RpfCFL-P levels tapered off as they neared 20 μM (Fig 1F), suggesting that the system had reached saturation point. RpfC is a hybrid histidine kinase that contains additional HPt and REC domains (Fig 1A). To exclude the possibility that the elevation of RpfCFL-P levels was caused by a change in DSF-dependent phosphoryl transfers from the DHp domain to these domains, the conserved phosphorylation sites within the HPt and REC domains were independently replaced [RpfCH657A and RpfCD512V]. The IMV forms of the two recombinant RpfC proteins were used in the phosphorylation assay. As shown in S1 Fig, neither of the amino acid replacement affected the DSF-dependent elevation of RpfC autokinase activity. Taken together, these findings provide direct biochemical evidences to demonstrate a long-term supposition that RpfC is an HK whose autokinase activity can be activated by the ligand DSF. Because membrane-bound HK usually employ periplasmic sensors and TM helices to detect signals, mutagenesis was used to identify regions critical for detecting DSF. A series of in-frame rpfC deletion mutants in a ΔrpfF genetic background that lost the capability to synthesize endogenous DSF were constructed. These constructs include a mutant with a deletion of the putative short sensor region (rpfCΔsensor), four mutants with their TM regions deleted in pairs to maintain the general topology of the protein (rpfCΔTM1-2, rpfCΔTM2-3, rpfCΔTM3-4, rpfCΔTM4-5), and a mutant with the input regions (sensor and TM regions) completely deleted (rpfCΔinput). Western blotting analyses revealed that deletions of the TM2–TM3 or input region caused instability of RpfC, whereas deletion of the short sensor region resulted in increase of the cellular levels of RpfC protein (S2A Fig), indicating the presence of a negative autoregulatory loop mediated by the sensor. As deletion of rpfF completely eliminated the synthesis of endogenous DSF, these double mutants were used to determine whether the different input regions were functional in sensing exogenously added DSF. Since the DSF-RpfC regulated, extracellular protease (EXP) activity can be directly observed without staining during bacterial growth [32], it was selected to be measured as a representative phenotype of the DSF-RpfC regulation. As compared with the positive control (ΔrpfFΔrpfC-rpfC), EXP activities of all these mutants were severely decreased, most likely because they lost the ability to detect exogenous DSF stimulation (Fig 2A). To quantify the effect of DSF perception in these mutants, we used pHM2 vector to construct a biosensor (PengXcc-GUS) by fusing β-glucuronidase gene (GUS) to the promoter of engXcc (XC_0639), which encodes an extracellular endoglucanase specifically regulated by the RpfC-RpfG system [15,23]. After providing the biosensor in trans to these mutants, GUS activity assay revealed that the transcription level of engXcc was reduced to 13.9–18.6% of that of the control (Fig 2B). Similarly, when DSF was added to the bacterial cultures, the ability of these mutants to form biofilms and produce extracellular polysaccharides (EPS) was significantly decreased (Fig 2C and 2D), exhibiting deficiencies in sensing DSF. To further investigate the regulatory function of the input region of RpfC in virulence, the same deletions as described above were generated in a wild-type (WT) background. Phenotypic characterization revealed that bacterial virulence (Fig 3A and S3A Fig), EXP production (Fig 3C), biofilm formation (Fig 3D), and EPS production (Fig 3F) of these mutants were all significantly reduced compared with the control. Quantification of the PengXcc-GUS activity (provided in trans by a recombinant pHM2 vector in each strain) showed that engXcc expression levels in these mutants decreased to 6.2–7.9% of the WT level (Fig 3B). In addition, as rpfC negatively modulates DSF synthesis, to measure the DSF production of the mutants, each strain was spotted onto a NYG-milk plate in close proximity to a ΔrpfF mutant deficient in endogenous DSF production. As shown in Fig 3E, deletion of the sensor region did not release the suppression of DSF production, while deletion of the transmembrane regions (TM12-TM45) moderately decreased the inhibition of DSF synthesis by RpfC, since the rpfF mutant exhibited higher EXP activity to degrade milk. Deletion of the entire input domain (sensor and TM regions) resulted extensive secretion of EXP, suggesting that the RpfC-mediated inhibition of DSF synthesis is remarkably eliminated. Taken together, mutational analyses suggested that the sensor and TM regions of RpfC are all involved in DSF detection and regulation of bacterial virulence. Of these, the function of the RpfC sensor appears to be particularly important as its deletion resulted in an increase in the cellular amount of RpfC protein, and had no effect in suppression of DSF-regulated EXP production as the other regions of input domain. The following analysis therefore mainly focused on the possible interaction between the RpfC sensor and DSF. We proposed that the RpfC sensor contains the amino acids essential for sensing the DSF signal. Multiple-Alignment of the RpfC sensor sequences from orthologs of close-relative bacteria of X. campestris pv. campestris showed that 15th to 22nd amino acids are highly conserved in species belonging to the Xanthomonadaceae family (Fig 4A), indicating this region is critical in function. Thereafter, alanine-scanning mutagenesis was used to identify essential amino acids in recognizing DSF. A full-length rpfC sequence was first amplified and then inserted into the pHM1 vector. Next, 19 non-Ala codons, except the initial Met1 residue, within the sensor region were individually point mutated into the Ala codon. The two indigenous Ala (Ala16 and Ala21) were also individually replaced into Val (Fig 1A). These 21 recombinant vectors were then individually electroporated into a ΔrpfFΔrpfC double mutant strain containing a GUS reporter fused to the promoter region of engXcc on the bacterial chromosome. Western blotting analysis revealed that apart from RpfCK2A, whose cellular amount was relatively low, all recombinant RpfC proteins were stable (S2B Fig). An exogenous DSF stimulation assay was then used to compare the engXcc transcription levels of these mutants to that of the positive control strain containing a plasmid-borne rpfC. As shown in Fig 4B, in the absence of exogenous DSF, replacements of S3A, L8A, R11A, D17A and Q22A caused slightly but significant changes in PengXcc activity (P ≤ 0.05). In the presence of exogenous DSF, three replacements, D17A, S18A, and Q22A, caused significant reductions in the PengXcc activity (levels 22.9–32.9% of the WT level). The R15A and E19A substitutions resulted in significant but intermediate reductions in the activation of the PengXcc (39.0–66.1% of the WT level). In contrast, the S3A replacement caused a stable, significantly increase in the PengXcc activity (to 110.0% of that of the control, Fig 4B). The above result revealed that substitutions in the amino acids from Arg15 to Gln22 (except Ala16, His20 and Ala21) of RpfC sensor resulted in deficiencies in sensing DSF. Consequently, further amino acid substitution analyses were conducted, which includes construction of recombinant strains containing a plasmid-borne RpfCR15K, RpfCR15H (Fig 4C), RpfCD17E, RpfCD17N (Fig 4E), RpfCS18T, RpfCS18C (Fig 4I), RpfCE19D, RpfCE19Q (Fig 4F), RpfCQ22E, and RpfCQ22N (Fig 4H) in the genetic background of ΔrpfFΔrpfC double mutations. Under DSF stimulation, RpfCR15K strain has similar PengXcc activity to that of the control (110% level) in sensing DSF (Fig 4C), suggesting that the positively charged, polar residues with similar side chains (R or K) are important in this location. RpfCS18T substitution, which naturally occurred in several species belonging to the genus Xanthomonas (Fig 4A), significantly increased the PengXcc activity to the 139% level of the control (Fig 4I), implying that the hydroxyl oxygen in the side chain of Ser (S) or Thr (T) is essential to sense DSF. However, all the other substitutions resulted in significant decreases of the PengXcc activity, indicating that these amino acids in the RpfC sensor region are essential and evolutionarily fixed to these locations (Fig 4C–4I). In addition, recombinant strains containing RpfCA16G, RpfCA16D (Fig 4D), RpfCA21G and RpfCA21D (Fig 4G) were also constructed. PengXcc activity assay showed that RpfCA16G and RpfCA21G replacements did not impact the DSF perception (Fig 4D and 4G). If the nonpolar Ala in the two sites were replaced by the negatively charged, polar Asp that disrupts the native conformation of the sensor region, the PengXcc activities were significantly decreased to 78% and 52% of the control, respectively (Fig 4D and 4G). Since the RpfCS18T substitution increased the PengXcc activity under DSF treatment and is the naturally occurred variation (Fig 4A and 4I), we further constructed 16 recombinant strains in the genetic background of ΔrpfFΔrpfC double mutant. Each strain contains a replacement of Ser18 of RpfC to one of the other 16 amino acids besides afore-mentioned Ala, Thr, and Cys. Quantification of PengXcc activity showed that except the RpfCS18T replacement, all the other substitutions led to significantly decrease in the engXcc expression (Fig 4I), supporting the view that a Ser/Thr residue in this site is critical in sensing DSF. To genetically evaluate the biological roles of the identified amino acids in sensing DSF, phenotypes of the recombinant bacterial strains (in the background of ΔrpfFΔrpfC double mutant) were examined following addition of exogenous DSF. Compared with the rpfC complementation strain, the strain containing RpfCS3A replacement showed similar levels of EPS production, biofilm formation and EXP production (Fig 5A–5C), while strains containing RpfCR15A, RpfCD17A, RpfCS18A, RpfCE19A, and RpfCQ22A replacements exhibited significant decreases in EXP and EPS production, as well as in biofilm formation (Fig 5A–5C). Strains containing a RpfCS18T substitution had similar or increased levels as the control in EXP activity, EPS production and biofilm formation (Fig 5D, 5E and 5F). Collectively, these genetic analyses suggest that five amino acids of Arg15, Asp17, Ser18, Glu19, and Gln22 adjoining to the transmembrane helices of RpfC are essential in detecting DSF, which is in parallel to the fact that they are highly conserved in bacterial evolution. Besides DSF perception, the roles of these essential amino acids in controlling bacterial virulence were also analyzed. Recombinant bacterial strains were constructed in the ΔrpfC background, each contains a pHM1 vector to produce a RpfC derivate with an amino acid replacement. Western blotting analysis revealed that RpfC protein amounts of these strains were stable (S2J and S2K Fig). Plant inoculation assays showed that R15A, D17A, S18A, E19A and Q22A substitutions resulted in substantial attenuation in virulence against host cabbage B. oleraceae (Fig 5G and S3B Fig), while the S18T replacement did not affect bacterial virulence (Fig 5G and S3C Fig). In addition, the production of one of the major virulence factors of X. campestris, EPS, was also significantly decreased in all of the tested strains except of the one containing RpfCS18T replacement (Fig 5H and 5J). These data strongly suggested that deficiency in DSF binding and perception resulted in deficiencies in bacterial virulence. It is noticeable that S3A substitution also resulted in remarkably decrease in virulence, albeit it caused slight hypersensitivity in detecting DSF as afore-mentioned (Fig 5G and S3B Fig). This result suggests that besides DSF perception, S3A is involved in additional, unknown function in regulating virulence. To detect a possible direct DSF-RpfC interaction, microscale thermophoresis (MST) was used because this technique is superior in studying membrane-bound receptors that are embedded in liposomes or nanodiscs with diverse ligands such as fatty acids and metals [33,34]. As shown in Fig 6A, DSF bound to the RpfCFL liposome with a dissociation constant (Kd) of 0.82 ± 0.12 μM, which represents a stronger binding affinity than those of the RpfS-DSF and RpfR-DSF interactions [31,35]. However, DSF didn’t bind to the truncated RpfCΔsensor liposome (Fig 6B) and soluble RpfC protein without input region (RpfCΔinput, Fig 6C). Thermal shift assay (TSA) using differential scanning fluorimetry was also employed to measure the DSF-RpfC interaction. As shown in Fig 6E, during thermal denaturation, addition of DSF resulted in the significant increase of melting temperature (Tm) of RpfCFL liposome from 60.27°C to 64.27°C (5 μM DSF vs. 10 μM RpfC) or 66.27°C (10 μM DSF vs. 10 μM RpfC), strongly supporting a direct binding between DSF and RpfC. When the RpfCΔsensor liposome and soluble RpfCΔinput were applied in TSA, no thermal shift was detected after DSF stimulation (Fig 6F and 6G). These MST and TSA results suggest that DSF binds to the sensor region of RpfC. To directly detect the sensor-DSF interaction, the sensor peptide fused with a glutathione S-transferases (GST) tag was successfully obtained by a pGEX6P-1 expression system, and this sensor peptide was purified by on-column cleavage together with size exclusion chromatography to remove the GST tag. MST analysis revealed that DSF bound to the sensor peptide with a similar affinity of the full-length RpfC liposome (Kd = 0.14 ± 0.04 μM, Fig 6D). Since this sensor peptide only contains a tryptophan (Trp7), its autofluorescence is quite low and not applicable in TSA, circular dichroism spectra (CD) analysis was used to measure the effect of DSF stimulation on the secondary structure of sensor peptide. CD analysis showed that after addition of DSF, the secondary structure of the short peptide was remarkably changed. The ratio of strand was gradually elevated along with the increase of DSF concentration (Fig 6H), while DSF stimulation did not have recognizable impact on the secondary structure of GST (Fig 6I). We further measured the impacts of substitutions of identified essential amino acids on the binding affinity with DSF. RpfCFL with corresponding replacements of amino acids in the sensor region were purified, assembled into liposomes, and their interactions with DSF were measured by MST. As shown in S4 Fig, substitutions of RpfCR15A, RpfCD17A, RpfCS18A, RpfCE19A and RpfCQ22A completely eliminated the binding between DSF and RpfC liposomes. In contrast, the RpfCS3A replacement decreased the Kd value to 0.613 ± 0.34 μM (S4A Fig), indicating that the RpfCS3A substitution slightly enhanced the DSF-RpfC interaction. Collectively, the above results experimentally demonstrated that DSF directly binds to the sensor domain of HK RpfC. Since the structures of RpfC and its orthologs remain unclear, limited proteolysis together with shotgun mass spectrometry were used to assess the conformational changes of RpfC involved in detecting DSF. During the analysis, the non-hydrolyzable ATP analog adenosine 5′(β,γ-imido) triphosphate (AMP-PNP) was added as a mimic for nucleotide binding. As shown in Fig 7A and S5 Fig, following addition of DSF to the reaction mixture, the patterns and amounts of most of the degraded RpfCFL liposome fragments were similar to those of the control (DSF minus). However, DSF stimulation repeatedly caused a large accumulation of a protein fragments (30 kDa, including those in the S3A substitution, S5A Fig). Nano-LC-MS/MS analysis revealed that the band represented the DHp-CA region of RpfC (from 192nd to 474th aa.). Proteolysis of RpfCFL liposomes with D17A, S18A, E19A, and Q22A replacements revealed similar degradation footprints, regardless of the presence or absence of DSF (Fig 7B and S5 Fig). These results suggest that the binding of DSF to the RpfC sensor caused conformational changes in the HK associated with the DHp-CA region. To investigate whether substitutions of the identified essential amino acids influence the ability of the RpfC autokinase to react to DSF, phosphorylation levels of the liposomes of the RpfCFL derivatives were compared to that of the WT. As shown in Fig 7C, stimulation of the RpfCS3A substitution using a physiological concentration of DSF (0.75 μM) resulted in hypersensitivity in detecting DSF, with the RpfC-P level remarkably increased under stimulation. RpfC with R15A, D17A, S18A, and E19A replacements exhibited substantially decreased autokinase activities, regardless of the absence or presence of DSF. The Q22A replacement did not affect RpfC autophosphorylation in the absence of DSF, but decreased DSF sensing, as shown by the significant decrease in RpfC-P levels under DSF treatment (Fig 7C). In addition, the autophosphorylation level of RpfCS18T was increased under DSF stimulation, albeit that the increase was slightly lower than that of the WT RpfC (Fig 7C). These results suggest that the identified essential amino acids, especially those located in the region adjoining to membrane, play important roles in RpfC autophosphorylation and DSF perception. RpfC contains a 16 aa-length, short juxtamembrane domain between the transmembrane helices and DHp-CA domains (Fig 1A). The juxtamembrane domain is highly conserved among RpfC orthologs from close-relatives of X. campestris pv. campestris (Fig 8A), implying that it has an important role in activating the RpfC autokinase after ligand perception. Alanine-scanning mutagenesis was again used to analyze the function of this region. As shown in Fig 8B, under stimulation of near-saturated concentration of DSF, 13 amino acid substitutions (in the background of ΔrpfFΔrpfC-rpfC background) gave rise to significantly decrease of the capability to sense DSF that is quantified by PengXcc activities. In the absence of DSF, although 12 replacements also caused significant decrease in the background PengXcc activities, it is noticeable that two substitutions, RpfCL172A and RpfCA178D, exhibited significant elevation of PengXcc activities (Fig 8B). When a low concentration of DSF (1 μM) was applied in treatment, both strains with RpfCL172A or RpfCA178D substitution also show constitutive activation of RpfC (Fig 8C). Especially, the PengXcc activity of the strain containing RpfCA178D replacement in the absence of DSF treatment even increased to the similar level to that of the DSF stimulation (Fig 8C). The phenotypes of the bacterial strains were also characterized. Without DSF stimulation, recombinant bacterial strains containing RpfCL172A and RpfCA178D, which were constructed in the genetic background of ΔrpfCΔrpfF double mutation, produced more biofilm, EPS, and EXP towards the levels of the positive control strain (ΔrpfFΔrpfC-rpfC) under DSF stimulation (Fig 8D, 8E and 8G). For example, in the absence of DSF treatment, RpfCL172A and RpfCA178D replacements caused significantly increases of bacterial EPS production to 259% and 247% levels of that of the control strain, respectively, similar to the EPS amounts that were generated under DSF stimulation. Since RpfCL172A and RpfCA178D substitutions constitutively activated the RpfC-regulated processes regardless of the presence or absence of DSF, we reasoned that mutations in the codons of Leu172 or Ala178 suppress the phenotypic deficiencies caused by mutations in the codons of essential amino acids of the sensor region in detecting DSF. To challenge this, we selected the Ala178 site for further genetic epistatic analysis. Three double mutants in the background of ΔrpfFΔrpfC-rpfCA178D were constructed by point mutating the codon of Asp17, Ser18, or Gln22. PengXcc activity assay revealed that all of these double mutations (rpfCD17A-A178D, rpfCS18A-A178D and rpfCQ22A-A178D) significantly suppressed the deficiency in the engXcc expression that is caused by the point mutations in the codons of essential amino acids (Fig 8F), regardless of DSF stimulation. In addition, although both deletion of rpfF gene and point mutation of the codons of essential amino acids sensing DSF caused serious decrease in bacterial virulence, RpfCA178D substitution suppressed the deficiency by recovering bacterial virulence level toward that of the positive control (Fig 8H and S3D Fig). The above genetic analysis suggests that the Leu172 and Ala178 are involved in autoinhibition of RpfC kinase activity without DSF stimulation. In vitro autokinase assay then revealed that the autophosphorylation levels of the recombinant RpfCL172A and RpfCA178D proteins are remarkably higher than that of the WT RpfC in the absence of DSF treatment, exhibiting a constitutive activating state (Fig 8I). Collectively, these results support a view that the juxtamembrane domain of RpfC inhibits its autokinase activity when the concentration of DSF is low. However, DSF perception by the sensor region releases this inhibition to activate the RpfC autophosphorylation. During this process, Leu172 and Ala178 in the juxtamembrane domain play critical roles in inhibiting the RpfC activity. How a ligand interacts with HK is one of the fundamental questions in studying bacterial quorum sensing. Here, we provided enzymological, genetic and biophysical evidences to demonstrate that a HK of X. campestris, RpfC, is a bona fide membrane-bound receptor that directly binds a fatty acid signal, DSF. The results confirmed a long-held hypothesis regarding cell-cell communication in phytopathogenic bacteria [23,25,36]. DSF binds with high affinity to a 22-amino acid sensor region in the N-terminal of RpfC (Figs 1 and 6). The binding of DSF causes allosteric change associated with the DHp-CA domain of RpfC, which facilitates RpfC autophosphorylation (Fig 7). Systematic mutational investigation together with biochemical analysis identified six essential residues in the DSF-RpfC interaction (Figs 4 and 5). Of these, five amino acids located in the region adjoining to membrane are indispensable for DSF-RpfC binding (Figs 4 and 5), while the Ser3 of the RpfC sensor region is functionally unique, as replacement of this residue resulted in slight hypersensitivity in detecting DSF. In addition, two point mutations (rpfCL172A and rpfCA178D) in the coding sequence of RpfC juxtamembrane domain effectively suppressed the phenotypic deficiencies caused by mutations in the sensor coding region, which is due to the constitutive activation of RpfC autokinase. These results support a molecular model (Fig 9) that the juxtamembrane domain inhibits the autokinase activity of RpfC when the extracellular concentration of DSF is low. However, when the DSF concentration increases along with the rise of bacterial population, DSF binds to the sensor region of RpfC and activates the HK by releasing this inhibition. To our best knowledge, this work provides the first experimental evidence to support a direct membrane-bound HK-FA interaction during bacterial quorum sensing and regulation of virulence. In addition to their roles in nutrition, FA acts as important signaling molecules in both eukaryotes and prokaryotes. For example, in animals, free FA signals are detected by G-protein-coupled receptors (GPCR) or receptor tyrosine kinases (RTK) [37]. A recent structural study revealed that FA ligands of an human insulin secretion modulator, GPR40, bind to hydrophilic/positively charged residues in various docking sites that are formed by the characteristic seven-TM helices of GPCR [38]. In addition, complex FA, such as cholesterol, inhibits the autokinase activity of human RTK epidermal growth factor receptor (EGFR) by influencing membrane heterogeneity-mediated transmembrane signal transduction [39]. In the present work, MST and TSA analyses revealed that DSF binds RpfC liposome with a relatively strong affinity (Kd = 0.82 ± 0.12 μM, Fig 6A). This binding affinity is higher than those of the DSF-RpfR (1.37 μM, measured by ITC) and DSF-RpfS (1.40 μM, by ITC) interactions [31,35], while both RpfR and RpfS are cytosolic, soluble proteins without transmembrane helix. In vitro autokinase assay showed that 0.5 μM DSF was sufficient to activate RpfC autokinase (Fig 1). This concentration is very close to the reported minimum DSF concentration in eliciting cell-cell communication [23], and in the physiological range of extracellular DSF, which approximately ranged from 0.002 to 27.4 μM dependent on the growth stages of bacterial population [27,40,41]. Our data suggest that the N-terminal short sensor region of RpfC plays a fundamental role, if not exclusive, in directly binding DSF (Fig 6D and 6H). In the sensor region, five amino acids adjoining to membrane of the RpfC are highly conserved in the bacteria belonging to the Xanthomonadaceae family (Fig 4A). It is likely that these amino acids form a primary docking site for DSF. This hypothesis is supported by the following data: 1) Replacements of these residues completely dissociated the DSF-RpfC interaction (S4 Fig). 2) This region contains hydrophilic or charged residues, especially Ser18 which can be functionally substituted by a Thr residue and favors hydrogen bond formation, most possibly with the carboxyl group of DSF. 3) Replacement of these residues completely eliminated the DSF-triggered, conformational change associated with DHp-CA domain (Fig 7). 4) Amino acid replacements, especially of Ser18 and Gln22, decreased the level of autophosphorylation of RpfC in response to DSF (Fig 7). The above results suggest that the short sensor region of HK likely acts as a “hook” to catch DSF when it is diffusely passing through the periplasm. In a previous report, ITC analysis revealed that DSF binds to the PAS_4 domain of RpfS of X. campestris [31]. Collectively, it suggests that the binding sites of DSF to proteins are diverse and need to be investigated further. It is noteworthy that the function of Ser3 within the RpfC sensor appears unique. This residue is highly conserved in all sequenced Xanthomonas species, whereas in other closely related bacterial species, Ala, Asp, Asn and Lys, respectively, occupy this position (Fig 4A). In the presence of exogenous DSF, S3A replacement caused hypersensitivity of RpfC in DSF detection: the binding affinity increased by approximately 20% compared with WT RpfC (S4 Fig), and the autokinase activity were also elevated (Fig 7C). However, as with the other essential residues, S3A replacement also attenuated virulence (Fig 5G and S3B Fig), suggesting that the Ser3 plays an additional role in regulating virulence, or DSF-RpfC binding affinity is subtly optimized during evolution so that any abnormality is detrimental to bacterial fitness. The role of Ser3 in binding DSF remains unclear. In previous study on the autoinducer CAI-I and HK CqsS in Vibrio cholerae [5], it revealed the length of hydrocarbon chain of autoinducer is critical in the ligand-HK interaction. The length of the FA chain is also a critical parameter in the activity of DSF-family compounds [23], therefore, one possible function of Ser3 is to act as a “ruler” to help suitable FA molecules gain entry into the RpfC docking site. However, further evidence is needed to clarify this hypothesis. How a ligand activates a membrane-bound HK remains an important question. To date, several structural mechanisms for HK activation have been proposed [42], including scissor blade [43], piston-like change [44], or retortion triggered by ligand binding mechanisms [45]. In these models, the movement of CA domains is at the center of in trans phosphorylation of the homodimer of HK. Our data revealed that DSF stimulation resulted in conformational change associated with DHp-CA domain (Fig 7A). In addition, DSF-sensor interaction released the juxtamembrane domain-mediated autoinhibition on the RpfC kinase activity. Among them, Leu172 and Ala178 are two critical amino acids since replacement of them resulted in constitutive activation of RpfC. These results support a view that DSF acts as an allosteric activator of RpfC by releasing the autoinhibition of its juxtamembrane region. Further investigation is necessary to obtain the high resolution structure of the DSF-RpfC complex to dissect the structural mechanism of RpfC activation, which is also meaningful to study the specificity in sensing different DSF-family signals. All bacterial strains and recombinant vectors used in this work are listed in S1 Table. Xanthomonas campestris pv. campestris (Xcc) 8004-derived strains and wildtype strain (WT) grew at 28°C in NYG medium (tryptone 5 g L-1, yeast extract 3 g L-1, glycerol 20 g L-1, pH 7.0) or 210 medium (sucrose 5 g L-1, casein enzymatic hydrolysate 8 g L-1, yeast extract 4 g L-1, K2HPO4 3 g L-1, MgSO4·7H2O, 0.3 g L-1, pH 7.0). E. coli DH5α was used as the host for construction of all recombinant vectors. E. coli BL21(DE3) strain was used as the host for expressing recombinant proteins with pET30a vector (Novagen, USA). Appropriate antibiotics were added when needed as following concentrations: kanamycin (50 μg ml-1); spectinomycin (150 μg ml-1); ampicillin (100 μg ml-1) and rifamycin (25 μg ml-1). Xcc 8004 and E. coli electro-competent cells were prepared by extensively washing bacterial cells three times with ice-cold glycerol (10%). Transformation condition of both X. campestris pv. campestris and E. coli cells was set as 1.8 kV cm-1, 25 μF and 200 Ω and conducted in a Bio-Rad Pulser XCell (Bio-Rad, USA). HPLC purified diffusible signal factor (DSF, CAS No. 677354-23-3, purify > 90.0%) was purchased from Sigma Aldrich (USA) and used in different concentrations as indicated in different experiments. If not specially mentioned, general molecular biology techniques, including PCR, DNA ligation, enzyme restriction, western blotting, etc, were according to the protocols in Molecular Cloning[46]. All in-frame deletion (markerless) mutants of rpfC and double mutant of rpfC-rpfF were constructed using suicide vector pK18mobsacB[47] by a homologous, double cross-over method. Briefly, the 5' and 3' genomic sequences of a targeted region were amplified using the primers listed in S2 Table, and correct PCR products were ligated into suicide vector pK18mobsacB. The recombinant pK18mobsacB vector was electroporated into competent cells of Xcc 8004 to generate single-crossover mutants by selection on NYG plates containing kanamycin. Afterwards, single-crossover mutants were cultured in NYG medium (antibiotic-free) for 1–2 hours and then grew on NYG plates containing 10% sucrose to select second-round homologous cross-overs. Correction of candidate bacterial mutants (resistant to 10% sucrose but sensitive to kanamycin) was verified by PCR and subsequent sequencing. To genetically complement the ΔrpfC mutant, a full-length rpfC gene was amplified using primers listed in S2 Table, ligated into the broad-host vector pHM1 [48], and electroporated into E. coli DH5α to generate the recombinant vector. This vector was then extracted from E. coli DH5α and electroporated into the ΔrpfC or ΔrpfCΔrpfF mutant, in which transcription of full-length rpfC was under the control of a PlacZ promoter. Besides the initial residue Met1, the N-terminal sensor region of RpfC contains 21 residues, with two of them being Ala. To conduct alanine-scanning mutagenesis, full-length rpfC coding sequence was amplified by PCR and inserted into a pGEM T-easy vector (Promega, USA), and Easy Mutagenesis System (TransGen Biotech, China) was used to construct point mutations according to the manufactory’s manual. Coding sequences of 19 non-Ala residues were mutated into Ala, respectively, and Ala16 and Ala21 were mutated into Val, respectively. The point mutation was confirmed by sequencing. These inserts with corresponding point mutations were cut by restriction enzymes, purified, and ligated into broad host vector pHM1 [48]. Their expressions were under the control of a PlacZ promoter. Recombinant vectors were then electroporated into ΔrpfC or ΔrpfCΔrpfF mutants as needed. Primers used to create these mutants are listed in S2 Table. To construct a biosensor using expression level of engXcc, which is subject to the control of RpfC-RpfG, as a parameter to estimate DSF-RpfC interaction, 5′ promoter region of engXcc (254 bp) was amplified by PCR, transcriptionally fused with a gusE gene to create PengXcc-GUS reporter insert (with native gusE Shine-Dalgarno to drive protein translation). For different purposes, the reporter sequence was provided in trans or integrated into bacterial chromosome. For in trans complementation, this reporter sequence was cloned into a pHM2 vector (no promoter upstream multiple cloning site), which was then electroporated into ΔrpfC or ΔrpfCΔrpfF mutants for GUS expression analysis. For alanine-scanning mutagenesis, the PengXcc-GUS reporter sequence was integrated into the chromosomal locus of engXcc by homologous, double cross-over via a recombinant pK18mobsacB vector (pKengXcc-GUS). For GUS activity assay, bacterial strains were cultured and adjusted to OD600 = 0.1, then grew without DSF or with 10 μM DSF for about 9 hours. Cells were collected by centrifugation (12,000 g, 10 min at 4°C), and immediately frozen in liquid nitrogen. GUS extraction buffer (50 mM sodium phosphate [pH 7.0], 5 mM DTT, 1 mM EDTA [pH 8.0]) was added to resuspend the cells and then these bacterial cells were lysed by sonication. The mixture was centrifuged (12,000 g, 10 min at 4°C) and the supernatant was used for GUS activity assay. Levels of GUS expression were quantified by its activity using 4-methylumbelliferyl ß-D-glucuronide (4-MUG, purchased from Sigma Aldrich, USA) as a substrate. A standard curve was prepared by diluting the 4-MU stock solution. The fluorescence of samples and standard curve solutions were measured using an excitation wave-length of 360 nm and an emission wave-length of 460 nm. Protein concentrations of supernatants were measured using Coomassie brilliant blue G-250 Protein Assay (Bio-Rad, USA) with BSA as a standard. For each experiment, at least three independent repeats were conducted for calculating the parameters. Plant inoculation and virulence assay were conducted as previously described[49]. In brief, six-week-old cabbage cultivar Brassica oleraceae cv. Jingfeng 1 was used as host plants. WT strain of Xcc and sterile 10 mM MgCl2 were used as positive and negative controls, respectively. All bacterial strains were cultured overnight in NYG medium containing appropriate antibiotics. Cells were collected, washed by 10 mM MgCl2, and the concentrations were adjusted to OD600 = 0.4 before inoculating into plant leaves using sterile scissors. After inoculation, the plants were kept in a greenhouse at 25°C–30°C and relative humidity >80%. Lesion length was scored 10 days after inoculation, and virulence level was scored semi-quantitatively as follows: 0, no visible effect; 1, limited chlorosis around the cut site; 2, chlorosis extending from the cut site; 3, blackened leaf veins, death, and drying of tissue within the chlorotic area; 4, extensive vein blackening, death, and drying of tissue. Assay of extracellular polysaccharides production (EPS) was conducted according to previous study[50]. Bacterial strains were cultured at 28°C in NYG medium until OD600 = 0.4. If necessary, DSF with appropriate concentration was added and bacteria were cultured for 75–96 hours before EPS production measurement. Quantification of biofilm development was conducted by classic method of crystal violet staining and according to previous study [51]. Bacterial strains were grown at 28°C in NYG medium until OD600 = 1.0, and 200 μl culture were inoculated into a 96 well plate (Costar, USA), cultured for 12 hour before quantification. To test the effect of DSF on the formation of biofilm, bacterial strains were cultured and adjusted to OD600 = 0.1. Then, those bacteria strains with appropriate concentration of DSF were grown for about 9h at 28°C and 200 μl culture were inoculated into 96 well plate as mentioned above. Estimate of extracellular protease (EXP) were conducted on NYG-milk plate as described previously [52]. If needed, 3.5 μL DSF (30 μM or 10 μM) was added near the bacterial colony. C-terminal His6-tagged recombinant proteins were expressed by constructing corresponding recombinant pET30a (Novagen) vectors that were electroporated into E. coli BL21(DE3) strain (S1 Table). Primers used to generate these constructs are listed in S2 Table. His6-tagged proteins were expressed and purified using affinity chromatography with Ni-NTA agarose beads (Novagen, USA), according to manufacturer′s instructions. Purified proteins were concentrated using Centricon YM-10 columns (Millipore) and the elute buffer was changed into storage buffer for further use (50 mM Tris-HCl, pH 8.0, 0.5 mM EDTA, 50 mM NaCl and 5% glycerol). Preparation of inverted membrane vesicles (IMV) containing full-length RpfC was according to the protocol of our previous study with minor modification[53]. Briefly, after sonication, cell debris of E. coli BL21(DE3) was abandoned by 6,000 g and the membrane containing full-length RpfC in supernatant was collected by ultracentrifugation at 60,000 g at 4°C for 60 min. After ultracentrifugation, the membrane was washed in high-salt buffer (20 mM sodium phosphate, pH 7.0; 2 M KCl; 10% glycerol; 5 mM DTT; 1 mM PMSF) twice. Finally, the membranes were resuspended in 0.5 ml storage buffer (20 mM Tris-HCl, pH 7.5; 10% glycerol) for autokinase assays. Liposomes were reconstituted as described by previous study[54]. Briefly, IMV of RpfC was obtained as above-mentioned, and dissolved in suspension buffer (20 mM phosphate, 500 mM NaCl, 20 mM imidazole, pH 7.4) to an approximate concentration of 10 mg ml-1 for preparation of liposomes. 900 μl IMV suspension with 100 μl 10% n-Dodecyl-ß-D-maltoside (DDM) were mixed by “end-over-end” mixing for 45 min at 4°C. The supernatant was collected by centrifugation 50,000 g for 30 min and purified by Ni-affinity chromatography. The Ni-NTA beads (Novagen, USA) were pre-equilibrated with 5 volumes of binding buffer (20 mM phosphate, 500 mM NaCl, 20 mM imidazole, pH 7.4, 0.1% DDM). Then the solubilized IMV and pre-equilibrated beads were mixed and incubated at 4°C for 30 min. After the mixing, the supernant was removed and the deposition was washed with binding buffer until the absorbance at OD280 nm returned to base line. Finally, 100 μl of elution buffer (20 mM phosphate, 500 mM NaCl, 250 mM imidazole, pH 7.4, 0.1% DDM) was added to elute the purified RpfC-His6. For embedding RpfC by liposome, 10 mg liposomes (Avanti Polar Lipids) were dissolved into 1 ml buffer (50 mM Tris-HCl pH7.5, 10% glycerol, 0.47% Triton-100). Then purified RpfC-His6 in elution buffer was added. The mixture was stirred at 4°C for 45 min. The final ratio of phospholipids to protein was about 10:1 (w/w). Bio-Beads (beads: detergent = 10:1, Bio-Rad, USA) were added to remove the detergent and the solution was stirred gently at 4°C overnight. Residual detergent was removed completely by addition of Bio-Beads after further incubation for 2 hours. The Bio-Beads were pipetted off and liposomes containing RpfC-His6 were gathered by centrifugation with 200,000 g, 4°C for 30 min. The RpfC liposome was resuspended in final buffer (20 mM Tris-HCl, pH 7.5, 10% glycerol) and stored at 80°C until used. in vitro phosphorylation assay was conducted as described in our previous study [53]. For autophosphoylation, RpfC liposomes or IMV were incubated with 100 μM ATP containing 10 μCi [γ-32P]ATP (PerkinElmer, USA) in autophosphorylation buffer (50 mM Tris-HCl, pH 7.8, 2 mM DTT, 25 mM NaCl, 25 mM KCl, 5 mM MgCl2) for indicated time (28°C). If necessary, DSF was added into the mixture 20 min before addition of ATP. The reaction was stopped with 6 × SDS-PAGE loading buffer. The phosphorylated proteins were separated by 12% SDS-PAGE. After SDS-PAGE electrophoresis, gels were separated from back glass plate and placed in a Ziploc bag and exposed to a phosphor screen for 1 hour. The screen was scanned with a PhosphorImage system (Amersham Biosciences, USA) at 50 μM solution. If necessary, signal intensity was measured by Quantity One software (Bio-Rad). RpfC sensor fused with a GST tag was obtained and purified according to the GST Gene Fusion System Handbook (Amersham Biosciences) with GST Resin (TRANS). In order to acquire sensor peptide with GST tag cleaved off, on-column cleavage procedure was conducted. In brief, lysate of recombinant E.coli BL21 (DE3) strain was mixed with pre-equilibrated GST Resin with PBS buffer for ten minutes before loading into column. The column was washed by PBS buffer and resuspened with PreScission buffer (50 mM Tris-HCl, pH 8.5, 150 mM NaCl). Following the injection of 2 units PreScission Protease (GenScript), the column was sealed and placed on a rotator at 4°C. After 10 h of digestion, the flow fractions were collected, which contains preliminary sensor peptide with the GST tag being moved. If necessary, the sensor peptide was purified again by the GST Resin to get rid of uncleaved sensor-GST fusion protein. Eventually, purified sensor peptide was obtained by using size exclusion chromatography with column Superdex 75 10/300 GL (GE Healthcare), stored under -80°C before use. DSF was mixed with purified proteins or liposomes of RpfC to a final concentration of 0 μM, 5 μM or 10 μM respectively in the reaction buffer (50 mM Tris-HCl pH 7.8, 25 mM NaCl, 100 mM KCl). The protein concentration was 0.5–1 μg/μl. The TSA was performed by a Prometheus NT.48 nanoDSF device with a temperature gradient of 20–95°C, 1°C /min. Unfolding transition points were determined according to the changes of intrinsic tryptophan fluorescence at 330 nm, 350 nm. The ratio of fluorescence and the melting temperature (Tm) were calculated by the NT Melting Control software (NanoTemper Technologies). Binding reactions of RpfC to DSF was measured by microscale thermophoresis in a Monolith NT.Label Free (Nano Temper Technologies GMBH, Germany) instrument which detects changes in size, charge and conformation induced by binding. RpfC liposomes were collected with centrifugation of 200,000 g for 40 min and resuspended in MST buffer (50 mM Tris-HCl pH 7.8, 150 mM NaCl, 10 mM MgCl2, 0.05% Tween-20) to an approximate concentration of 0.1 μM. A range of concentration of DSF (range from 0.06 μM to 2 mM) in assay buffer (50 mM Tris-HCl pH 7.8, 150 mM NaCl, 10 mM MgCl2, 0.05% Tween-20, 5% methanol) was incubated with RpfC liposomes (1:1, v/v) for 10 minutes. The sample was loaded into the NT.Label Free standard capillaries and measured with 20% LED power and 40% MST power. Purified sensor protein was dissolved in reaction buffer (50 mM Tris-HCl pH 8.5, 150 mM NaCl, 0.1% Tween-20) to a final concentration as 8 μM. Dilute DSF from 0.0122 μM to 25 μM in buffer (50 mM Tris-HCl pH 8.5, 150 mM NaCl, 0.25‰ methanol). Different concentrations of DSF and sensor protein (1:1, v/v) were mixed and loaded into NT.Label Free standard capillaries. The label free MST assay was performed with 20% LED power and 40% MST power. KD Fit function of the Nano Temper Analysis Software Version 1.5.41 was used to fit curve and calculate the value of dissociation constant (Kd). The limited proteolysis experiments were performed at 0°C with 1.4 μg RpfC liposomes in a reaction buffer containing 50 mM Tris-HCl, pH 8.0, 100 mM NaCl, 2 mM DTT and 1.13 mM AMP-PNP. Trypsin was added to a final concentration of 0.018 μg μl-1 to degrade RpfC liposome. Aliquot was removed at indicated time and the reaction was stopped by 5 × SDS loading buffer (250 mM Tris-HCl pH 6.8, 10% (w/v) SDS, 0.5% (w/v) bromophenol blue, 50% (v/v) glycerol, 25 mM PMSF). Samples were separated by 12% SDS-PAGE gel and protein bands were detected by silver staining. The sequence of the different peptide fragments were determined by a nanoLC-MS/MS with Orbitrap Fusion system (Thermo scientific, USA). To determine if DSF has impact on RpfC sensor conformation change, CD analysis was carried out on a Chirascan CD Spectrometer (Applied Photophysics, UK), with 10 mm pathlength and 1 nm bandwidth. Sensor protein with GST tag cleaved off was dissolved in buffer (50 mM Tris-HCl pH 8.5, 150 mM NaCl) to 60 μM. Dilute DSF with Sensor protein to a series of concentration (0 μM, 100 μM and 500 μM). CD wavelength scans were collected between 200 nm-260 nm. The spectra data were analyzed on the http://dichroweb.cryst.bbk.ac.uk website with Contin-LL method [55].

10.1371/journal.pbio.1000294

Rare Variants Create Synthetic Genome-Wide Associations

Genome-wide association studies (GWAS) have now identified at least 2,000 common variants that appear associated with common diseases or related traits (http://www.genome.gov/gwastudies), hundreds of which have been convincingly replicated. It is generally thought that the associated markers reflect the effect of a nearby common (minor allele frequency >0.05) causal site, which is associated with the marker, leading to extensive resequencing efforts to find causal sites. We propose as an alternative explanation that variants much less common than the associated one may create “synthetic associations” by occurring, stochastically, more often in association with one of the alleles at the common site versus the other allele. Although synthetic associations are an obvious theoretical possibility, they have never been systematically explored as a possible explanation for GWAS findings. Here, we use simple computer simulations to show the conditions under which such synthetic associations will arise and how they may be recognized. We show that they are not only possible, but inevitable, and that under simple but reasonable genetic models, they are likely to account for or contribute to many of the recently identified signals reported in genome-wide association studies. We also illustrate the behavior of synthetic associations in real datasets by showing that rare causal mutations responsible for both hearing loss and sickle cell anemia create genome-wide significant synthetic associations, in the latter case extending over a 2.5-Mb interval encompassing scores of “blocks” of associated variants. In conclusion, uncommon or rare genetic variants can easily create synthetic associations that are credited to common variants, and this possibility requires careful consideration in the interpretation and follow up of GWAS signals.

It has long been assumed that common genetic variants of modest effect make an important contribution to common human diseases, such as most forms of cardiovascular disease, asthma, and neuropsychiatric disease. Genome-wide scans evaluating the role of common variation have now been completed for all common disease using technology that claims to capture greater than 90% of common variants in major human populations. Surprisingly, the proportion of variation explained by common variation appears to be very modest, and moreover, there are very few examples of the actual variant being identified. At the same time, rare variants have been found with very large effects. Now it is demonstrated in a simulation study that even those signals that have been detected for common variants could, in principle, come from the effect of rare ones. This has important implications for our understanding of the genetic architecture of human disease and in the design of future studies to detect causal genetic variants.

Efforts to fine map the causal variants responsible for genome-wide association studies (GWAS) signals have been largely predicated on the common disease common variant theory, postulating a common variant as the culprit for observed associations. This has led to extensive resequencing efforts that have been largely unsuccessful [1]–[5]. Here, we explore the possibility that part of the reason for this may be that the disease class causing an observed association may consist of multiple low-frequency variants across large regions of the genome—a phenomenon we call synthetic association. For convenience, these less common variants will be referred to here as “rare,” but we emphasize that we use this term loosely, only to refer to variants less common than those routinely studied in GWAS. The basic idea of how synthetic associations emerge in this model is illustrated in Figure 1, which shows how rare variants, by chance, can occur disproportionately in some parts of a gene genealogy. Any variant “higher up in the genealogy” that partitions those parts of the genealogy containing more disease variants than average will be identified as disease-associated. It is well appreciated that a noncausal variant will show association with a causal variant if the two are in strong linkage disequilibrium (LD). We use the previously introduced term synthetic association [6], however, to describe how such indirect association can occur between a common variant and at least one and possibly many rarer causal variants. Using the term synthetic as opposed to indirect emphasizes that the properties of the association signal are very different when the responsible variant or variants are much less frequent than the marker that carries the signal, as we detail below. To assess the tendency of rare disease-causing variants to create synthetic signals of association that are credited to single polymorphisms that are much more common in the population than the causal variants, we have simulated 10,000 haplotypes based on a coalescent model in a region either with or without recombination (Materials and Methods). We assumed that gene variants that influence disease have an allele frequency between 0.005 and 0.02, which is generally below the range of reliable detection (either by inclusion or indirect representation) using the genome-wide association platforms currently in use. We assumed a baseline probability of disease of φ for individuals with none of the rare genetic risk factors. The presence of at least one rare risk allele at the locus increased the probability of disease from φ to γ. We considered two values of φ (0.01, 0.1) and chose values of the penetrance γ such that the genotypic relative risk (GRR) of the rare causal variants varied incrementally between 2 and 6, where GRR is the ratio γ/φ. These values were chosen to explore the space around a GRR of 4, a threshold above which consistent linkage signals would be expected [7]. We simulated scenarios with one, three, five, seven, and nine rare causal variants. Across the conditions we have studied, not only is it possible to achieve genome-wide significance for common variants when one or more rare variants are the only contributors to disease, it is often the likely outcome (Figure 2). Overall, 30% of the simulations were able to detect an association with a common SNP at genome-wide significance (p<10−8). Three factors—GRR, sample size, and the number of rare causal variants—had a notable impact on power to detect an association with a common SNP. As expected, greater proportions of synthetic associations were created when GRR increased for the rare causal variants and when sample size increased. As the number of rare causal variants increased, the probability of creating a synthetic association did as well. One possible explanation for this increase due to increasing the number of rare causal variants is that adding more causal variants increases the size of the disease class, which is the proportion of haplotypes that carry one or more disease alleles [8]. The size of the disease class varied in the simulations both because the frequency of causal variants was allowed to vary, and because the disease class increases on average with the number of causal variants. To investigate the effect of the disease class on synthetic associations, we separated the results by size of disease class and found first that the larger the disease class the higher the chance of a significant synthetic association. We also find, however, that within a disease class size, the probability of significant synthetic associations decreases with the number of causal variants (Figure 3). Importantly, association with the strongest causal variant in individual simulations was more significant than with the strongest common synthetic association in 98% of the simulations, and for each combination of parameters, the proportion of simulations with genome-wide significant associations was always higher for the strongest causal variant than for synthetic associations when testing for association with individual variants. Of particular importance to note, except for the case of GRR = 2, all conditions considered here produced a nonnegligible proportion of simulations with significant common variants. It is also noteworthy that significant signals of association can be credited to common variants even when there is only a single rare causal site. A control simulation was run by testing the common variants from one genealogy against phenotypes generated by a separate genealogy with the same parameter settings and not a single test fell below genome-wide significance of 10−8 for all simulations. This shows that significant synthetic associations depend on the associations that occur within a single gene genealogy (or correlated ones in a recombination graph) and that sites undergoing free recombination cannot create genome-wide significant synthetic associations. Intuitively, it seems obvious that when rare variants are the cause of the associations, there should then be multiple common variants that carry significant independent associations. To evaluate this expectation, we took those genealogies that produced a genome-wide significant association and asked what the strongest association was when the top genome-wide significant association was first incorporated in the model. We found that almost 40% of genealogies with a genome-wide significant variant had secondary, independent associations that also achieved genome-wide significance. We also found that fewer than 10% of genealogies had no further significant associations (at α = 0.05). These results demonstrate a clear tendency of rare variants to create multiple independent signals of synthetic association. One essential question about synthetic associations is whether they are expected to be robust to the presence of recombination. Surprisingly, not only does recombination fail to eliminate synthetic associations, but low rates of recombination can enhance them compared with no recombination (Figure 2B). For example, for GRR = 4 and 9 risk alleles, and a sample size of 3,000 cases and 3,000 controls, we find the proportion of trees showing significance for zero recombination is 0.66. When we introduce a recombination rate of 5×10−5 (ten times the genome-wide average for 500 bp) between segments, however, we find that the proportion increases to 0.92. When recombination is increased further, the expected decline in the synthetic association is observed. Importantly, however, even at exceptionally high recombination through the region (5×10−4 between segments), we find that almost 30% of the simulations show a significant common variant, and recombination must increase to 5×10−3 to reduce the proportion to below 1%. Importantly, the simulations involving recombination prohibit evaluation of any common variant that has a rare causal site within the same segment. Thus the synthetic associations emerging in these simulations occur between sites that are separated by a minimum recombination distance of that between segments, which is 1×10−3 to 5×10−3. It is counterintuitive that recombination would increase synthetic associations since recombination reduces the average LD in a region. The observation can be explained, however, by the effect of recombination on the distribution of association amongst sites within a genomic region. Although the average LD declines as recombination increases, it is not known how higher moments behave and these moments can influence the proportion of pairs of sites that exceed some given threshold level of association. We tested this as the explanation for the capacity of recombination to enhance associations by directly evaluating the mean and the variance of the association between rare and common variants in a simplified simulation. We considered two regions separated by a specified recombination rate. We calculated the average pairwise association between rare and common variants and also the variance of the pairwise LD between rare and common variants in each simulation, and evaluated both these parameters as a function of recombination. We found that although the mean is nonincreasing, the variance first increases then decreases (Figure 4), suggesting that increases in recombination can “widen” the distribution of LD among sites sufficiently to increase the density in the tail and thereby create stronger synthetic associations. These patterns make clear that so long as a given genomic region has one or more rare variants that contribute to disease, these rare variants can generate synthetic associations that are observed in much more common polymorphisms. Under ideal conditions for such synthetic associations, they can be detected with sample sizes far smaller than those routinely used in GWAS. Under less ideal conditions (for example, higher prevalence attributable to environment or to other genetic factors outside of the locus being considered or lower penetrance for the local rare variants), the sample size must be larger. One essential quality of synthetic associations is that although they are often likely to be created when multiple rare variants exist in a region, there are certain conditions under which very little association will be detected even with very large sample sizes and large effects of the causal variants because causal alleles will segregate to opposite common alleles. In other words, no common variant will be able to partition the rare variants on a genealogy to create a large enough imbalance to create association. We also investigated trends in association with causal variants and found that even though our model specified that only derived alleles at causal sites are deleterious, more than a third of the most highly associated common SNPs showed a higher penetrance for the ancestral allele. This result follows observed patterns [9]. Another important trend is that if only rare variants are contributing to the disease class in a region, the risk allele frequency of the most significant synthetic association will tend toward the low end of the distribution of more common allele frequencies (median = 0.10), although over 20% of genome-wide significant synthetic associations had a risk allele frequency above 0.25 (Figure 5). Of course, this trend is noted when all common variants in a region are included, which is not the case with the available commercial genotyping chips, which have a greater probability of including more common variants. In this case, the skew towards lower-frequency variants would be less. We next attempted to determine the expected genomic distances over which rare variants could create synthetic associations. To do so, we simulated a 10-Mb region with a typical recombination rate (1 cM/Mb), nine rare causal variants, 2,000 cases and 2,000 controls, and GRR = 4. We then identified the most distal causal variant that was confirmed to actually contribute to the signal of synthetic association. We did this by finding the most distal variant that resulted in a minimum of a one-log drop in p-value when its effect was statistically removed (by incorporation as a covariate into the regression). We found that when a synthetic association reached genome-wide significance, the most distant causal variant that affected the significance of the synthetic association was closer than 2 Mb from a synthetic association in fewer than 13% of the simulations and at least 9 Mb away in 4% of the simulations. The median distance of the most distant causal variant was 5 Mb. A simulated Manhattan plot showing a 10-Mb region with average recombination and nine causal variants with GRR = 4 shows an example of a signature created by synthetic association (Figure 6). Finally, we evaluated the genomic pattern of synthetic associations using two real-world examples: hearing loss and sickle cell anemia. These two examples represent two possible extremes for synthetic associations. Sickle cell anemia is a serious Mendelian disease in which the body makes sickle-shaped red blood cells. The disease mostly affects subjects with African ancestry, and prevalence among African Americans in the United States is approximately 1 in 600 [10]. It is known to be caused by autosomal recessive mutations in HBB, and the frequency of the most common causal variant (Hb S allele) is ∼3.6% in Americans of African ancestry [11]. In comparison, hearing loss is a complex human disease, occurring in one per 1,000 newborns on average [12]. More than two dozen causal genes have been identified for autosomal recessive nonsyndromic hearing loss [13],[14], but mutations in the GJB2/GJB6 locus account for about half of the cases of European ancestry [12],[15]. Among hundreds of known causal mutations in the GJB2/GJB6 locus [14], the 35delG mutation in GJB2 is the most common, with an allele frequency of 1.25% in European Americans [16], but hundreds of other point mutations in GJB2 as well as a 342-kb deletion encompassing GJB6 also represent known causal variants [17],[18]. For sickle cell anemia, a total of 179 SNPs reached genome-wide significance (p<5×10−8), encompassing an ∼2.5-Mb region on chromosome 11p15.4 (from 3.59 Mb for rs12422109 to 5.98 Mb for rs997433). The region contains dozens of genes and dozens of visually discernable LD blocks in HapMap YRI population. The top association signal (rs7120391, p = 1.1×10−136) is 9 kb from OR51V1, which is very near the causal gene, HBB (Figure 7). Clearly, highly significant association signals can travel across multiple LD blocks to distant genomic regions. The three most significantly associated SNPs for hearing loss are all located at the GJB2/GJB6 locus on 13q12.1 (Figure 8), including rs870729 near GJB6 (p = 3.38×10−11, OR: 1.69), rs3751385 within GJB2 (p = 1.50×10−9, OR: 1.63), and rs7329467 within GJA3 (p = 6.87×10−8, OR: 1.68). The three SNPs have weak LD with each other (pairwise r2 values range from 0.02 to 0.62), but all of them are common variants. For example, rs870729 has a minor allele frequency (MAF) of 18.7% in controls and 28.0% in cases. To evaluate the independence of the association signals from the three SNPs, we tested association again by incorporating rs870729 in a logistic regression model, yet still found residual association for rs7329467 (p = 4.3×10−6), but not rs3751385 (p = 0.33), consistent with the expectations derived above for the behavior of synthetic associations. The locus has been extensively resequenced in numerous studies, and there is no common causal variant at the locus with ∼18.7% allele frequency similar to rs870729. Therefore, rare variants at the locus create multiple independent association signals captured by common tagging SNPs. These results show that a large proportion of genomic regions that harbor one or more rare variants that contribute to disease is likely to create “synthetic” signals of association [6]. If the region carries an excessively large number of causal variants, this expectation decreases, but for intermediate numbers of (causal) rare variants, detection of many such regions appears inevitable due largely to the fact that increasing the number of rare causal variants increases the size of the disease class in that region. Separately considering the number of causal variants and the proportion of alleles that are disease causing (the disease class) makes clear that the latter is the key driver of the ease of creating synthetic associations. The intuition for this is obvious. Even when the frequency of disease-causing variants is very low individually, as the disease class grows, collectively they come closer to the frequency of common variants, allowing the possibility of a strong signal to be generated for one of the common variants. This is only prohibited when the causal variants are so numerous as to be distributed roughly evenly through the genealogy (or if an even distribution appears by chance for smaller numbers of causal variants). In considering the likelihood of rare variants creating a large disease class, it is essential to appreciate that signals can combine in the face of considerable recombination. This makes clear that the “locus” associated with GWAS signals may be far larger than has often been assumed. We also note that the apparent size of the disease class is not a good guide as to the number of causal variants responsible. Even if the disease class is quite large, it is easily possible that it consists of only rare variants if there are a relatively large number of rare causal variants and these could be spread very broadly over genome regions stretching into the megabases. With respect to the size of the disease class, our simulations highlight the counterintuitive result that, under some genetic models, increases in the number of causal mutations at a locus can increase the probability of a synthetic association. Although our simulations only show that synthetic associations are likely to occur, coupling this demonstration with the available data does suggest that some of the reported associations are likely to be due to this effect, and many more may be enhanced by the signal of surrounding rare causal variants. First, despite considerable efforts, the vast majority of genome-wide associations have never been tracked to causal sites, even though many surrounding regions have been extensively resequenced [2]–[4]. If all of the responsible variants were common SNPs, one might expect that more clear evidence of causation would have been identified by now for a nontrivial number of common variants. Although this expectation is valid for common causal variants, because we know roughly where to look in the genome, this does not hold for synthetic associations due to rare variants, which may reside at a considerable distance from the associated common variants. Second, it is now known that rare variants contribute to common diseases, and that cases that carry the rare high-penetrant contributors to disease often have “typical” clinical presentations [19]–[21]. On balance, therefore, our results suggest that even though the apparent impact of common variants is only modest for many traits [6],[22],[23], this impact may have been systematically overestimated [24]. It is worth emphasizing that the alternative explanation provided here makes clear, testable predictions. As noted, in a model of synthetic associations, regions that show significant effects for common variants will often consistently show significant residual independent effects after the effect of the most important variant has been accounted for. Second, since rare variants are much more likely to be population specific, synthetic associations are expected to be inconsistent across population groups. In fact, a number of recent studies have confirmed differences in effect between populations [24]–[35]. Table 1 lists variants from these studies in which the point estimate for a follow-up study in a separate population fell outside the 95% confidence interval for the odds ratio of the original study. This includes 13 variants and odds ratios with confidence intervals for the population in which association was first discovered (12 European and one Japanese) and 20 odds ratios for subsequent tests in separate populations, consisting of eight tests in African Americans (seven not significant and one significant in the opposite direction), four tests in Japanese (one not significant and three significant in the same direction), four tests in Koreans (one not significant and three significant in the same direction), two tests in the Indian subcontinent (two not significant), one test in Europeans (not significant), and one in Chinese (not significant). Although it is possible that many of these differences are related to differences in LD (association) between markers and causal sites, genetic or environmental interactions, or simply genetic heterogeneity, it appears likely that many of these differences are due to multiple underlying rare variants that create different synthetic effects in the populations. There are also likely to be other diagnostics of synthetic associations observable in GWAS data. For example, one would expect distinctive extended haplotypes to be enriched in cases relative to controls in large regions surrounding GWAS signals that are synthetic (K. Wang, S. P. Dickson, C. A. Stolle, I. D. Krantz, D. B. Goldstein et al., unpublished data). Perhaps most importantly, the observation that association statistics are stronger for the causal sites in the vast majority of cases implies that in many cases, it should be possible to identify candidate causal sites using whole-genome sequence data surrounding GWAS signals and evaluate these for association. When the association is synthetic, association statistics would be expected to strengthen considerably when the correct causal sites are assayed. There are also practical implications related to finding the variants responsible for observed associations. Perhaps the most important of these is that targeted sequencing within a “block” of LD surrounding GWAS discoveries is often not expected to identify the causal sites. Because modest amounts of recombination can enhance synthetic associations, and because recombination must be exceptionally high to eliminate the possibility of genome-wide significant associations, one or more of the responsible causal sites could be a very considerable distance from the common variant showing a signal of association. This possibility is starkly illustrated by the sickle cell anemia example in which genome-wide significant synthetic associations span ∼2.5 Mb around the causal mutation, although heterosis may also influence this result. This possibility suggests that efforts to identify causal variants responsible for GWAS signals that concentrate on a region of high LD surrounding the implicated variant are not well motivated and are likely to miss many and perhaps most of any rare variants that contribute to synthetic associations (see, for example, [5]). The distance over which synthetic associations occur also offers an alternative explanation to the increasingly common observation of rare variants that occur within the vicinity of a GWAS signal but cannot explain that signal entirely. A simple explanation for such observations is that extending the sequencing to at least 4 Mb and ideally up to 10 Mb around the GWAS signal would pick up other rare variants. In some cases, identifying all the contributing rare variants may explain all of the original signal, whereas in other cases, there could be a combination of rare and common variants contributing. In addition, if synthetic associations are responsible for many of the observed signals, then sequencing in a small number of control samples (even over a much broader genomic region) is also unlikely to succeed. Under our model, the causal sites are both rare and relatively high-penetrant contributors to disease, and will therefore be unlikely to be detected in a small number of control samples. Finally, the focus of attention on genes that are near GWAS signals may be incomplete or misleading in that the actual causal sites may occur in many different genes surrounding the implicated common variant. It is also worth emphasizing that as few as one or two rare variants, at much lower frequency than the associated common SNP, can create a significant synthetic association. In such a case, sequencing a small number of cases that carry the “at risk” common variant might miss entirely the causal rare variants even if the correct genome region is resequenced. These considerations argue for caution in efforts to resequence around genome-wide associations and argue instead that genome-wide sequencing in carefully phenotyped cohorts might be a better use of resources. It has been suggested that rare high-penetrant variants would produce a signal inconsistent with those observed in many common traits in favor of models with thousands of common variants with marginal penetrance [36]. We have shown that multiple rare variants in a region are capable of acting over large distances to create associations in common variants similar to observed associations. A key point is that multiple rare causal variants may be causing the observed associations, therefore a single haplotype would be insufficient to explain such associations. Ultimately, the proportion of GWAS signals that is due to common versus rare variants is a question that can only be resolved empirically. Our analyses simply illustrate that in following up GWAS signals, the possibility of synthetic associations must be taken into account. If it were true that many signals were synthetic in nature, however, one interesting and potentially encouraging implication of these results is that some of the very modest associations emerging from genome-wide associations may in fact be pointers to rare variants of much larger effect that could be directly informative about disease pathophysiology or be sufficiently high penetrance to be of useful predictive value. For the primary simulation, two simulated haplotypes were randomly selected with replacement for each individual, and sufficient individuals were generated to simulate the desired number of cases and controls. Case/control status was designated based on the assigned risk, and equal numbers of cases and controls were selected for association testing. We tested all common variants in the genealogy for association with disease status, where common was defined by a minor allele frequency of 0.05 or greater. Thus we exclude any variant that is actually disease causing and focus on those that are generally represented directly or indirectly in the current genome-wide genotyping platforms [37]. Association tests were performed by comparing 1,000, 2,000, or 3,000 each of cases and controls, and we screened for common variants with p-values less than 10−8, a now-typical threshold for genome-wide significance [1]. We defined a single “simulation” as follows. A random gene genealogy was drawn with mutations distributed along the genealogy, and disease-causing mutations were assigned at random from those variants that were in the allowed frequency range. Then cases and controls were sampled as described, and the common variants screened for association. We then determined the proportion of such simulations that resulted in a genome-wide significant signal being credited to at least one of the common variants in the genealogy. Genealogical trees were simulated using GENOME with an effective population size of 10,000 and a mutation rate of 10−8 in a 100-kb region. When recombination was simulated, 200 fragments of 500 bp each were used with recombination occurring between each fragment [38]. Trees were drawn using Dendroscope [39]. p-Values were obtained using logistic regression on the case-control status under an additive model. Odds ratios (for the common variants) were estimated using the β term from the logistic regression. A second p-value reported for common variants was based on a logistic regression, with the most strongly associated common variant as a covariate in the model to assess residual association after discounting the strongest synthetic association. For both disease association studies, we performed a standard GWAS using Illumina HumanHap550 BeadChip with over ∼550,000 SNPs, which represent common tagging variants and do not include any of the disease-causing mutations for either condition. We carried out a standard association test on all markers on the chip passing default quality control measures (minor allele frequency >5%, Hardy-Weinberg equilibrium p-value >1×10−6, SNP call rate >95%), using the PLINK software [40]. For the sickle cell anemia GWAS, we compared 194 cases and 7,407 controls of inferred African ancestry via multidimensional scaling, with a genomic control inflation factor of 1.01. For hearing loss, we performed a GWAS on 418 cases and 6,892 control subjects, all of whom were of genetically inferred European ancestry via multidimensional scaling, with a genomic control inflation factor of 1.02.

10.1371/journal.ppat.1005990

Mechanistic Studies and Modeling Reveal the Origin of Differential Inhibition of Gag Polymorphic Viruses by HIV-1 Maturation Inhibitors

HIV-1 maturation inhibitors (MIs) disrupt the final step in the HIV-1 protease-mediated cleavage of the Gag polyprotein between capsid p24 capsid (CA) and spacer peptide 1 (SP1), leading to the production of infectious virus. BMS-955176 is a second generation MI with improved antiviral activity toward polymorphic Gag variants compared to a first generation MI bevirimat (BVM). The underlying mechanistic reasons for the differences in polymorphic coverage were studied using antiviral assays, an LC/MS assay that quantitatively characterizes CA/SP1 cleavage kinetics of virus like particles (VLPs) and a radiolabel binding assay to determine VLP/MI affinities and dissociation kinetics. Antiviral assay data indicates that BVM does not achieve 100% inhibition of certain polymorphs, even at saturating concentrations. This results in the breakthrough of infectious virus (partial antagonism) regardless of BVM concentration. Reduced maximal percent inhibition (MPI) values for BVM correlated with elevated EC50 values, while rates of HIV-1 protease cleavage at CA/SP1 correlated inversely with the ability of BVM to inhibit HIV-1 Gag polymorphic viruses: genotypes with more rapid CA/SP1 cleavage kinetics were less sensitive to BVM. In vitro inhibition of wild type VLP CA/SP1 cleavage by BVM was not maintained at longer cleavage times. BMS-955176 exhibited greatly improved MPI against polymorphic Gag viruses, binds to Gag polymorphs with higher affinity/longer dissociation half-lives and exhibits greater time-independent inhibition of CA/SP1 cleavage compared to BVM. Virological (MPI) and biochemical (CA/SP1 cleavage rates, MI-specific Gag affinities) data were used to create an integrated semi-quantitative model that quantifies CA/SP1 cleavage rates as a function of both MI and Gag polymorph. The model outputs are in accord with in vitro antiviral observations and correlate with observed in vivo MI efficacies. Overall, these findings may be useful to further understand antiviral profiles and clinical responses of MIs at a basic level, potentially facilitating further improvements to MI potency and coverage.

HIV-1 continues to be a serious health threat, with nearly 40 million infected individuals worldwide. Despite effective treatment options, issues with resistance and drug toxicities illustrate the need for new drugs with novel mechanisms. Maturation inhibitors (MIs) block a key protease cleavage within its target, preventing formation of infectious HIV-1 virus. A first generation MI, (bevirimat), failed in clinical studies due to lack of broad spectrum activity, a result of amino acid polymorphisms around the site of action. BMS-955176 (GSK3532795) is a second generation MI active against these polymorphisms, and is currently in a Phase 2b study. We used a combination of antiviral and novel biochemical approaches to understand the mechanism for these spectrum differences. We find that while bevirimat exhibits incomplete antiviral activity, even at saturating drug concentrations, BMS-955176 exhibits greater ability to maximally inhibit these viruses, in part due to higher affinity for its target. These data were integrated into a semi-quantitative kinetic model whose outputs are in accord with in vitro antiviral observations and correlate with observed in vivo MI efficacies and the results of recent crystal and cryo-electron tomography structures. Our findings offer insights into MI activity and mechanism and may prove useful to help guide development of new MIs, with potential applicability to other virus systems and inhibitors.

Currently there are more than 1.2 million individuals (age 13 years older) in the United States (CDC data)[1] and more than 35 million worldwide infected with HIV, with 39 million people already having died from the disease and 2.3 million new cases reported in 2013.[2] There are presently >35 FDA-approved HIV therapies or combinations of agents which can be categorized into different classes: NRTIs, NNTRIs, PIs, integrase and entry inhibitors, (the latter includes attachment and fusion inhibitors, along with CCR5 antagonists).[3, 4] However, co-morbidities associated with long-term use of antiretrovirals (ARVs)[4–6] and the continued development of resistance remains a problem. [7, 8] Thus, there is a continuing need for new HIV-1 drugs which lack cross-resistance to existing classes and have excellent long term safety profiles. HIV-1 maturation inhibitors (MIs) are a class of agents that may be effective in the treatment of HIV-1.[9–12] MIs disrupt the final step in the HIV-1 protease-mediated cleavage of the HIV-1 Gag polyprotein between capsid (CA) and spacer peptide 1 (SP1), a step which is responsible for a major conformational rearrangement of viral proteins within the virion that leads to the production of infectious virions.[13–15] The first generation HIV-1 maturation inhibitor, bevirimat (BVM), was halted in development[16] due to lack of clinical response in subjects whose viruses contained certain polymorphic Gag variants present in ~50% of the subtype B population, with such variations common among non-subtype B HIV-1 viruses.[17–27] Despite this result, BVM provided proof of concept (POC) in the clinic [28,29] that HIV-1 maturation inhibitors (MIs) per se might provide an effective alternative, should a next generation agent possess suitable pan-genotypic coverage.[30–32] BMS-955176 (GSK3532795) was developed as a second generation MI that possesses antiviral activity against viruses containing BVM-resistant Gag polymorphisms.[9, 19, 23, 33–40] It is currently in Phase 2b clinical trials.[41–43] However, an understanding of the mechanism for how BMS-955176 achieves this improved antiviral coverage has not been described. Such an understanding at the mechanistic level is of intrinsic interest, potentially providing further insights into the maturation process itself, and the biology and biochemistry of HIV-1 infection. Of clinical importance, such understanding may also be of value to help guide the development of newer MIs with further improvements to MI activity, genotypic coverage and spectrum. We took three approaches to address how BMS-955176 achieves these improvements to antiviral coverage. In the first, details of the antiviral dose-response profiles of BVM and BMS-955716 with respect to viruses containing various Gag polymorphs were studied. In a second approach, the mechanism of cleavage of capsid/spacer peptide 1 (CA/SP1) was evaluated using a novel LC/MS assay to quantitatively characterize the kinetics of cleavage HIV-1 Gag VLPs as a function of polymorph, while also determining the inhibitory effects of BVM and BMS-955176 in that system. Thirdly, the affinities and kinetics of dissociation of these MIs to these same Gag polymorphs in VLPs were measured using a radioligand binding assay. Results reported herein indicate that reduced BVM antiviral activities toward certain polymorphs (elevated EC50 values) were accompanied by incomplete (less than 100%) inhibition of antiviral activity, even at saturating BVM concentrations. Thus, depending on polymorph, BVM may be described as a partial antagonist. On the other hand, BMS-955176 exhibits a significantly greater ability to maximally inhibit these Gag polymorphs. Biochemical characterization indicates that improvements to polymorphic coverage (both lower EC50s and higher degrees of maximal antiviral inhibition) are a result of its higher affinity for its target (Gag), which was shown to primarily be a result of its slower rate of dissociation. The antiviral and biochemical data herein reported were integrated into a model that calculates rates of CA/SP1 cleavage as a function of MI concentration and Gag polymorph, predicting in vitro antiviral profiles and estimating in vivo efficacy. These findings offer new insights into MI activity and mechanism and may prove useful to understanding the pre-clinical and clinical responses of MIs at a mechanistic level, potentially facilitating further improvements to newer MIs. MT-2 cells were obtained from the NIH AIDS Research and Reference Reagent Program; 293T cells were obtained from the ATCC. Cell lines were sub-cultured twice a week in either RPMI 1640 (MT-2) or DMEM (293T) media (Gibco), supplemented with 10% heat inactivated fetal bovine serum (FBS, Gibco), and 100 units/mL penicillin with 100 μg/mL streptomycin (Gibco). The parent WT virus was generated at Bristol-Myers Squibb from a DNA clone of NL4-3 obtained from the NIH AIDS Research and Reference Reagent Program[44] and contains the Renilla luciferase marker in place of viral nef, and the substitution of Gag P373 for serine, the most common B subtype variation at that position among B subtype viruses (NLRepRlucP373S). NLRepRlucP373S (WT) was modified to contain changes in Gag (for example, V362I, V370A, A364V, ΔV370,[40] the latter three of which encode high level resistance to BVM.[33, 40, 45] The recombinant viral DNA was then used to generate virus stocks by transfecting 293T cells (Lipofectamine PLUS kit, Invitrogen). Titers of all stocks were determined in MT-2 cells, using luciferase as the endpoint (Dual-Luciferase Reporter Assay System, Promega).[40, 46] The TCID50/ml (tissue culture infectious dose) was calculated by the method of Spearman-Karber.[47] Compound susceptibilities of NLRepRlucP373S variants were examined using a multiple cycle infectivity assay as follows[40]: MT-2 cell pellets were infected with virus and re-suspended in cell culture medium. After a 1-hour pre-incubation at 37oC/CO2, cell-virus mixtures were added to a dose range of compound in 96-well plates at a final cell density of 10,000 cells per well. All compounds were tested at 1% final DMSO concentration. After 4–5 days of incubation at 37°C/5% CO2, virus yields were determined by Renilla luciferase activity (Dual-Luciferase Reporter Assay System, Promega) and the signals read using an Envision Multilabel Reader (PerkinElmer Product number: 2104). Maximal percent inhibition (MPI) values were calculated using the equation: MPI = (1- (signal from average at two highest drug concentrations/signal from no-drug control) * 100). MI susceptibilities were also determined using an assay format similar to that reported. [36, 40, 48] which restricts viral growth to one replication cycle as follows: In a first step. 10 μg of the proviral clone of NLRepRlucP373S variant (containing the appropriate Gag substitution) and 8 μg of plasmid pSV-A-MuLV-env (MuLV envelope gene under control of the SV40 promoter, NIH AIDS Research and Reference Reagent Program, Cat# 1065) were co-transfected (calcium phosphate, Invitrogen, K2780-01) into 293T cells (60–70% confluence, T75 flask). After overnight incubation at 37°C/5% CO2, the transfected cells were washed, trypsin treated, and re-suspended in fresh medium at a density of 5 x 105/mL. Cells were then distributed (100 μL/well) to 96 well plates that contained 100 μL of media with compound (compound was 3x serially diluted in DMSO, 1% final concentration of DMSO). In a second step, after 30 hours at 37°C/5% CO2, 100μL of supernatant (containing the newly produced virus) was transferred to a second 96 well plate to which fresh 293T cells (3x104/well) were added. Cultures were maintained for 2 days, after which cell-associated Renilla luciferase activity was measured upon the addition of Enduren (EnduRen Live Cell Substrate, Promega, catalog # E6485) and the signals read using an Envision Multilabel Reader (PerkinElmer Product number: 2104). MPI values were calculated as that compound concentration which inhibits 50% of the maximal signal (no-drug control) as described above. To demonstrate the late inhibitory phenotype of BMS-955176, the above single cycle assay was modified by the use of the HIV-1 envelope-deleted derivative pNLRepRlucP373 Δenv,[40] transfected with plasmid encoding HIV-1 LAI envelope (pLAIenv was constructed within BMS, contains the entire sequencing encoding LAI GP160 under control of the CMV promoter). LAI pseudotyped virus produced in a first step in the presence of inhibitor was added to MT-2 cells in the second step, instead of 293T cells as performed above. HIV-1 virus-like particles (VLPs) are non-infectious particles that are made through transfection of a partial HIV-1 genome and contain only the Gag structural protein. VLPs used in these experiments [35, 36, 40] did not contain HIV-1 genes other than gag, and were prepared as follows: a synthetic gene (GagOpt)[49–51], under the control of the CMV promoter in plasmid 1_pcDNAGagOpt, was constructed to encode full length HIV-1 LAI Gag, with codons optimized for expression in mammalian cells. Various GagOpt clones were used, containing the coding sequence of LAI Gag or variant Gag polymorphs, starting from the N-terminus of matrix (MA, amino acid position 1) and extending to the stop codon of p6. The VLPs were produced[52, 53] by transfection (Mirus Bio LLC, TransIT®-LT1, cat# MIR 2300) of 293T cells (70–80% confluency in a T175 flask) with 18 μg of the appropriate pGagOpt plasmid. After 2 days of incubation at 37°C, supernatants (containing secreted VLPs) were cleared from cell debris by filtration (0.45-μm filter, Millipore #SCHVU01RE). The VLP particles were then pelleted through a 20% sucrose cushion at 25,000 rpm in an SW28 rotor for 2 hours, re-suspended in PBS at a total protein concentration of about 1000 μg/mL and then stored at -80°C. Purified VLPs (~100 ng) were incubated at room temperature for 10–30 min in 10 μL of VLP buffer (50 mM MES pH 6.0, 100 mM NaCl, 2 mM EDTA and 2 mM DTT) supplemented with 0.06% Triton X-100 to remove the VLP lipid bilayer. Delipidated VLPs (~100 ng) were incubated with 3 μM MI (0.1% final DMSO) at 22°C for 2 hours, and then digested with HIV-1 protease by adding 1 μL of 2.7 μM of HIV protease (final concentration 0.27 μM, HIV-1 protease constructed to contain substitutions that limit auto-proteolysis: Q7K/L10I/I13V/L33I/S37N/R41K/ L63I/C67A/C95A)[54] One μL samples were taken at the indicated time points and digested with trypsin as follows: one μL of each HIV-1 protease digested sample was added to 24 μL of 50 mM ammonium carbonate (pH 8) containing 4 mM DTT. Samples were incubated at 60°C for 60 minutes, and then alkylated by the addition of 1 μL of 100 mM iodoacetamide. Samples were then kept in the dark for 30 minutes. Subsequently, 1 μL of 0.1 mg/mL reconstituted trypsin (Promega sequence grade modified trypsin, cat# 9PIV511) was added to each sample, and trypsin digestion was allowed to proceed at 37°C overnight. Reactions were stopped with 1 μL of formic acid, and peptides were analyzed by LC/MS. For MI inhibition studies MI (3 μM >500-fold antiviral EC50, 2 hour pre-incubation) MIs were first added to VLP to effect binding, after which time HIV protease was added to catalyze cleavage. Under these conditions the molar ratio of MI to Gag monomer is approximately 30-fold. Liquid chromatography/mass spectrometry (LC-MS) analysis was performed using a Waters nanoAcquity UPLC system interfaced with a Thermo Scientific LTQ XL Orbitrap mass spectrometer affording nanoflow-LC/accurate mass data. Data were acquired by positive ion electrospray ionization using a Michrom Bioresources, Inc. Advance CaptiveSpray ion source operated at 1.5 kV and a transfer tube lens set at 150°C. Data on unique tryptic peptides of interest was acquired by single ion monitoring (SIM) using a 5 amu window in profile mode at a resolution of 7500 @ ½ ht. NanoLC analysis was carried out using a Waters Symmetry C18 180 μm x 20 mm 5 um (PN-186003514) trap column and a Microm Magic C18AQ 0.1 x 150 mm (PN-CP3/61271/00) analytical column. Trapping was performed at 5 μl/minute for 2 minutes at the initial gradient composition prior to the analytical gradient. The mobile phase composition was water (MP-A) and acetonitrile (MP-B) with each containing 0.1% formic acid. The analytical gradient was as follows 5%B to 35% MP-B over 30 min (ramped to 70% MP-B followed by equilibration at 5% MP-B) at a flow rate of 500 nL/minute. The mobile phase composition was water (MP-A) and acetonitrile (MP-B) with each containing 0.1% formic acid. Three microliters injection volumes were used for each sample. Data was analyzed using Thermo Xcalibur Processing software 3.0.63 and Thermo Xcalibur Quanbrowser software 3.0.63. Areas were measured using the plus 2 charge state mono isotopic mass of the peptides of interest +/- 0.2 amu from the peak apex. The raw peak area for the SQ peptide (SLFGNDPSSQ, internal trypsin cleavage fragment at the C-terminal terminal end of Gag), was used as an internal control for normalization of the response for the peptides of interest. The data for the AM peptide (Gag SP1, generated by HIV-1 Pr cleavage at the N- and C-termini of SP1), the VM peptide (generated by HIV-1 Pr cleavage between SP1 and nucleocapsid, then cleavage by trypsin) and the VR peptide (cleavage by trypsin only, no internal cleavage by HIV-1 Pr) were normalized against the data from the SQ peptide. The percent of total = 100 x [AM/SQ / (AM/SQ + VM/SQ + VR/SQ)], where AM/SQ + VM/SQ + VR/SQ = the sum of all the peptide fragments encompassed within the two trypsin cleavage sites on either side of the SP1 peptide. Specific binding of MIs to VLPs were determined using a scintillation proximity (SPA) radiolabeled binding assay. VLPs (0.5 to 1.2 μg in PBS) were mixed with 100 μg of SPA beads (PBS suspension, PVT WGA SPA beads, PerkinElmer, cat # RPNQ0250) in 40 μL of total volume per well (96-well plate, Corning, white low binding, cat# 3600) After 1-hour incubation at room temperature, the volume was increased to 180 μL /well by the addition of binding buffer (100 mM Tris, pH 6.5, 2 mM EDTA, 0.03% Tween-20, 5 mM MgCl2). The final concentration of DMSO in the assay was 10% by volume. For determination of Kd values by a competition method, 20 nM of [3H]-BMS-977660 (a C:20 double bond reduced (tritiated) form of BMS-955176) [35, 40] was added to the VLP/bead mixtures, to which was added a serial dilution (0.04–3000 nM) of non-radiolabeled MI. After 4 hour-equilibration at room temperature, bound [3H]-BMS-977660 was measured using a Top Count plate reader (PerkinElmer). The data were fit to an equation for heterologous competition (GraphPad v 5.1). MI dissociation rates were measured by adding 30 nM [3H]–BMS-885221 (a C:20 double bond reduced form of BVM) or 20 nM [3H]–BMS-977660 to SPA bead/VLP (0.5 to 1.2 μg) complexes, allowing binding to reach equilibrium for 3 hours at room temperature. After this time a 40-fold molar excess of unlabeled competitor MI was added to effect irreversible displacement of the [3H] MI. Kinetics of disappearance of the bound 3H MI were monitored using a Microbeta2 plate reader (PerkinElmer) and the data fitted to a first order exponential equation (GraphPad Prism v5.1). Details of the model are shown in later in the Results section. In the presence of an MI, the rate of CA/SP1 cleavage, and thus the formation of mature virus, is derived below. Since the measured dissociation rate constants of the MIs (koff) are faster than the innate rates of CA/SP1 cleavage (k1) for the WT and polymorphic viruses, a rapid equilibrium assumption was employed to derive the observed rate constant (kclv,ob) to form mature virus (C). With this assumption, the association and dissociation rates of MI binding are the same. The concentration of total immature virus equals the sum of free immature virus (B) plus the MI bound immature virus (A) and is defined as Imtotal. Replacing [A] in Eq 2 from Eq 1: [Imtotal]=[MI][B]Kd+[B] (3) Rearranging Eq 3, the relationship between B and Imtotal [B]=Kd[MI]+Kd[Imtotal] (4) The formation of mature virus (C), and the depletion of total immature virus (Imtotal) have the same rate, thus d[C]dt=−d[Imtotal]dt=k1[B]+k2[A]=k1[B]+k2[MI]Kd[B] (5) Substituting [B] in Eq 5 with Eq 4: d[C]dt=−d[Imtotal]dt=(k1*Kd[MI]+Kd+k2*[MI][MI]+Kd)[Imtotal] (6) Integrating Eq 6, the solution of equation for the cleavage of CA/SP1, and thus formation of mature virus (C) is Ln[C][T]==(k1*Kd[MI]+Kd+k2*[MI][MI]+Kd)t (7) where t is the time, assuming at t = 0 there is no cleaved CA/SP1 or mature virus existing, and T is the total concentration virus. The observed rate constant (kclv,ob) to form capsid from CA/SP1, and thus mature virus, in the presence of MI is: kclv,ob=k1*Kd[MI]+Kd+k2*[MI][MI]+Kd (8) Previous reports indicated that a first generation MI, BVM, demonstrated poor antiviral activity both preclinically and in a POC study toward clinical isolates[28, 29] containing polymorphic substitutions in Gag around the site of its mechanism of action, i.e., at or near the HIV-1 protease-mediated cleavage site between capsid (p24) and spacer peptide 1[9, 19, 20, 23, 33, 45] These polymorphs include substitutions at Gag positions V362, Q369, V370 and T371.[19–21] BMS-955176 was identified as a clinical candidate with improved potency against viruses containing these polymorphic substitutions, low human serum binding and excellent PK properties [35–38, 40, 55] BMS-955176 (Fig 1) retains potent activity toward these polymorphic variants in vitro and was active in a Ph2a POC study[41–43] As shown in Table 1,[40] BMS-955176 is 5.4-fold more potent than BVM toward WT virus, and polymorphic viruses retain sensitivity to BMS-955176, with FC values (EC50/WT EC50) between 1- and 6.8-fold.[35, 36, 40] The protease inhibitor nelfinavir was used as a control, which exhibits similar antiviral characteristics towards all the polymorphic viruses. By comparison, BVM exhibits significantly reduced activity toward these variants (up to >1000 fold). For example, BMS-955176 retains activity toward variants with substitutions at Gag V370 by alanine or methionine (1.4- and 1.5-fold, respectively), and V362I (2.4-fold), as compared to 54-, 177- and 7.2-fold losses of sensitivity by BVM, respectively. In addition, BMS-955176 retains activity toward virus with V370A/ΔT371 and ΔV370 substitutions, both minor polymorphs in subtype B, but characteristic of subtype C isolates[24] (FCs of 3.5 and 6.8-fold, respectively). By comparison, BVM is >100-fold less active toward both V370A/ΔT371 and ΔV370-containing viruses. An early BMS compound in the series leading to the identification of BMS-955176 was BMS-1 (Fig 1),[37] with an antiviral profile similar to BVM. It was included in this study to determine if the results were able to be generalized beyond BVM and BMS-955176. BMS-955176 does not inhibit A364V[40], a resistance mutant selected for by BVM in vitro[33] and also reported in two HIV-1 subjects in a clinical trial with BVM.[56] Overall, these results indicate that BMS-955176 exhibits significantly improved in vitro antiviral activity toward polymorphic variations in Gag which result in reduced sensitivity to first generation MIs. With these results in hand we initiated virological and biochemical studies whose aim was to understand the mechanistic basis for the improved antiviral profile of BMS-955176. Earlier biochemical studies had noted that while BVM disrupts the final step of HIV-1 maturation, that of CA/SP1 processing, this disruption is not an absolute block: some mature CA is generated, even at high concentrations of the compound.[57] We considered it possible that partial biochemical inhibition might translate into partial inhibition in antiviral assays. This concept was evaluated by conducting detailed studies of the antiviral inhibition dose response curves of BVM toward less susceptible Gag polymorphs, focusing on the degree of inhibition at the highest BVM concentrations tested. Initial studies made use of a multiple cycle (MC) infectious virus assay using HIV-1 luciferase reporter viruses. In this format, a low viral input (multiplicity of infection typically 0.002–0.005) was used, and therefore multiple rounds of virus release and viral re-infection were required to achieve sufficient luciferase signal for detection at the assay endpoint (typically 4 days post infection). The results are shown in Fig 2 and Table 2. In this format, BVM inhibition of the control WT virus (percent maximal inhibition (MPI +/- Stdev) of 98.4 +/- 1.0) was similar to that achieved by NFV, an HIV protease inhibitor used as a positive control ((percent maximal inhibition (MPI) of 98 +/- 0.9 compared to 100%, respectively). By comparison, BVM inhibition of the polymorphic viruses V362I and V370A reproducibly achieved only partial inhibition (MPI values of 81.3 +/- 2.0 and 65.4 +/- 3.8, respectively). Improvement to the MPI of the V362I and V370A viruses was not achieved at higher BVM concentrations, as a plateau in inhibition at a concentration occurred at approximately 1 μM. Representative examples of dose response curves are shown in Fig 2 for BVM, BMS-1 and BMS-955176. The ΔV370 virus variant is resistant to BVM, exhibiting an MPI of 9.2+/- 6.3 (Table 2). Control experiments performed with 3 and 6 μM MI dissolved in 10% FBS media vs. PBS buffer for up to 4 days, found that upon subsequent evaluation the concentration of MI between 80–100% in the media, indicating no loss due to precipitation of MI under these conditions. Average recovery in the PBS condition was ~50%, indicating precipitation and binding to the walls of the tube (S1 Fig). This result indicates that the plateau in inhibition is not an artifact of limited MI solubility under the cell culture conditions. Secondly, as discussed later below, (behavior of BMS-955176 toward the ΔV370 virus), there is an obvious plateau in inhibition at 100 nM in a single cycle assay of approximately 50% of maximal, but in a multiple cycle assay the maximal percent inhibition (MPI) is higher (91.9%). Such pronounced plateaus were observed in other single cycle assays (see values in Table 2). If solubility were to be the limiting factor, both single and multiple cycle formats would be expected to provide similar plateaus. BMS-1, an early compound in the development of the structure activity relationship (SAR) leading to the identification of BMS-955176, shares structural similarity (Fig 1) to both BVM and BMS-955176 but differs from BVM by replacement of the C3 dimethylsuccinic acid by a benzoic acid.[37] Similar to BVM, the first generation BMS-1 (FC values similar to BVM, Table 1) exhibits a high MPI against wild type virus (97.0 +/- 0.5), but only partially inhibits V362I and V370A viruses (MPIs = 71.4 +/- 5.4 and 63.2 +/- 4.7, respectively) and does not inhibit ΔV370 containing virus (MPI = 3.7 +/- 2.9) (Fig 2B and Table 2). By comparison, BMS-955176 exhibits high MPI values of 100 +/- 0.4, 98.2 +/- 1.3, 93.8 +/- 1.8 and 91.9 +/- 4.4 for WT, V370A, V362I and the ΔV370 containing viruses, respectively (Fig 2C and Table 2). To further probe the phenomenon of incomplete inhibition of various polymorphic viruses, we employed a 2-step single cycle assay in which HIV-1 LAI pseudotyped virus is first released into the supernatant by co-transfection of NLRepRlucP373Δenv and pLAI envelope plasmids into 293T cells in the presence of MI.[40] Subsequently, the supernatant is harvested and used for infection of MT-2 cells in a second infection step. In a manner similar to the Magi assay[48, 58] a signal in the second infection step indicates that infectious virus had been produced in the transfection step. However, subsequent rounds of infection are prevented as virus produced in the second stage lacks an HIV-1 envelope, and is thus unable to infect MT-2 cells. An inhibitor which blocks the production of infectious virus in the transfection stage of the assay will score as inhibitory in the second stage of the assay. Since this assay monitors the events that have taken place in a single cycle of infection, we refer to this format as a single cycle assay, or SC assay. Control experiments established that when a late inhibitor such as nelfinavir is added at the transfection step, luciferase activity is inhibited in the infection stage (Fig 3A). However, when NFV is added only at the infection stage, luciferase production is not inhibited (Fig 3B). The HIV-1 attachment inhibitor, BMS-378806[59] is fully active (Fig 3B), as expected for an agent which inhibits early in the HIV-1 life cycle. The MIs BVM and BMS-955176 behaved similarly to nelfinavir, inhibiting luciferase production only when added in the first step of the assay, consistent with their late mechanism of action. As shown in Table 2, BVM exhibited an SC MPI value towards WT virus of 82.3 +/- 2.7, less than BMS-955176 in this more demanding format, while BVM barely inhibited the V370A virus (MPI of 19.0 +/- 3.5), a result which is qualitatively similar to that obtained using the MC format. The ΔV370 variant is resistant to BVM in this assay. In contrast, BMS-955176 exhibits SC MPI values of 93.0 +/- 2.5, 76.5 +/- 3.0 and 45.9 +/- 7.4 towards the WT, V370A and ΔV370 viruses, respectively (Table 2). Overall, MPI values in both the SC and MC formats follow the same trend, but SC MPI values are reproducibly lower, presumably due to the fact that viral challenge is higher in the transfection format vs. infection (MC, low multiplicity of infection = 0.005), and the absence of multiple cycles which inhibit breakthrough virus from within each preceding cycle. Antiviral dose response curves for inhibition of the ΔV370 virus by BMS-955176 are compared using the two formats (MC assay, Fig 4A and SC assay, Fig 4B). Fig 4C shows the differences in MPI values from Fig 4A and 4B, where the control NFV exhibits full inhibition in both formats. BMS-955176 inhibition of ΔV370 in the MC assay did not reach the 100% control value of NFV (Table 2, Fig 2C). The single cycle assay provides a wider dynamic range from which to understand the nature of the stable incompletely inhibited plateau, as compared to the multiple cycle assay (~100 nM BMS-955176 toward ΔV370 (Fig 2B)). To understand the partial antiviral inhibition results we considered the basic framework for the underlying mechanism of maturation inhibition, i.e. its capacity to block the last cleavage step during virion biogenesis, that of CA/SP1 cleavage by HIV-1 protease.[30, 31] As depicted in Fig 5, maturation inhibitor (MI) binds to the immature HIV Gag polyprotein in the vicinity of the cleavage site[60–62] to produce the MI-bound form (A), in which CA/SP1 is protected from HIV-1 protease cleavage. As reported, action of MIs on Gag VLPs requires that Gag be fully assembled in its quarternary state, [9, 63] in concordance with this we have observed that heat inactivation of VLPs abrogates specific MI binding. Binding is reversible,[35, 40] with association and dissociation rate constants defined as kon and koff respectively. The innate cleavage rate constant (k1) determines the efficiency of the irreversible conversion from immature virus (B) to mature virus (C). Based on the observed maximal percent inhibition (MPI) values from the cellular antiviral assays, we hypothesize that the MI-bound immature virus (A) can also be cleaved by HIV-protease, but at a reduced rate (k2, where k2 < k1), thus accounting for the production of virus, as a function of polymorph and MI, even at saturating concentrations of MI. In this model (derivation in Materials and Methods) k1 is specific for each Gag polymorphism, while k2 is a function of both MI and Gag polymorphism. Thus, mature virus C will be produced as a function of time in a manner dependent on the steady state concentrations of both the free immature form B and the MI-bound immature form A, and dependent on their respective rate constants, k1 and k2, for HIV-1 protease cleavage of CA/SP1. To challenge this scheme and model this process, we created appropriate biochemical assays needed to obtain the requisite protease CA/SP1 cleavage rate constants and MI affinities toward the WT and polymorphic variant viruses. An LC/MS analysis method to quantify the specific event inhibited by MIs was developed that measured the HIV-1-mediated cleavage of CA/SP1 (p25) to CA (p24) and SP1 through quantitation of a peptides released after subsequent trypsin cleavage (Fig 6A).[35] This method entails exposure of HIV-1 Gag virus-like particles to HIV-1 protease in vitro in the absence or presence of MIs, followed by trypsin cleavage of the resulting HIV-1 protease-mediated products. The starting parental material (no cleavage at either SP1/NC or CA/SP1) is referred to as peptide VR, by virtue of the N- and C-terminal amino acids of the peptide produced by trypsin cleavage of Gag (Fig 6B). Cleavage by HIV-1 protease between SP1 and NC at site H1, and subsequent cleavage by trypsin, produces intermediate peptide VM (Fig 6B). The N-terminal valine of VM is derived from trypsin cleavage while the C-terminal methionine is derived from HIV-1 protease cleavage. Lastly, HIV-1 protease cleavage at CA/SP1 at HIV-1 protease site H2 produces peptide AM (Fig 6B, peptide nomenclature as further described in Materials and Methods, peptide AM = Gag peptide SP1). This method is suitable for monitoring the 3 species simultaneously, allowing for measurement of the kinetics of cleavage at both CA/SP1 and NC/SP1 (representative experiments for wt and A364V are shown in Fig 7C and 7D, respectively). In the example of Fig 7C (wt), an average of two independent experiments at an HIV-1 protease concentration of 270 nM, the parent peptide VR disappears first due to rapid cleavage between SP1 and NC, as has been reported.[13–15] Rapid disappearance of VR is accompanied by the release of intermediate VM (VM is essentially the surrogate for p25), which decreases as AM (SP1) is formed. AM (SP1) peptide appears slowly, gradually increasing with time, but its formation remains incomplete at the last time point (240 minutes) under this set of conditions. By comparison, for A364V (Fig 7D) the disappearance of parent VR is similarly rapid vs. wt, while the appearance of intermediate VM, and product AM (SP1) are faster than wt. Measured rate constants (kclv,ob at 270 nM HIV-1 protease) for the cleavage of CA/SP1 by HIV-1 protease at CA/SP1 from WT and BVM-resistant polymorphic VLPs are shown in Table 3. For comparison to the cleavage rate data, multiple cycle antiviral sensitivities from Table 1 are also shown in this table. As might be expected, absolute cleavage rates were a function of the HIV-1 protease concentration; they were linear over the range of 67–540 nM HIV-1 protease (S2 Fig), indicating no loss of proteolytic activity within this time window, as expected for use of protease specifically engineered to not undergo autoproteolysis.[54] CA/SP1 cleavage of WT VLPs was the slowest, while VLPs containing V370A and V362I were cleaved approximately 3-fold faster than WT (Table 3). The subtype C-like surrogate polymorphic VLPs, V370A/ΔT371 and ΔV370, were cleaved 2.2- and 2.7-fold faster than WT. By comparison, A364V, the completely BVM and BMS-955176-resistant variant, [9] was cleaved ~10-fold faster than WT, as reported.[64] A set of representative AM peptide (SP1) appearance curves is shown in S3 Fig: the order of appearance of SP1 product is A364V > V370A, V362I, ΔV370 > V370A/ΔT371 > WT, which is a similar, but in inverse order, to the antiviral sensitivities of these polymorphic viruses to BVM (Tables 1 and 2). HIV-1 protease specifically designed to inhibit auto-proteolysis was used,[54] as initial experiments of wt HXB2 HIV-1 protease produced unsatisfactory results in terms of non-linearity of cleavage with time. As can be seen in S2 Fig, there is linearity of cleavage for wt for concentrations of protease up to 540 nM, an indication of no loss of proteolytic activity, with the kinetic data reported in Table 3 performed using 270nM protease. There was non-linearity for A364V cleavage at 270nM protease at longer time points, thus the rate constant data for A364V was derived from within the linear range only. A sub-analysis of the rates in the linear range over multiple concentrations of protease indicated that the relative 2nd order rate constant for A364V (S2 Fig) is 9-fold faster than wt, in agreement with the 1st order constant, and indicating that the first order rate constant accurately captures this information. The relative rate of cleavage of A364V (9.7-fold) is in accord with a value previously published (7.6-fold).[64] BVM and BMS-955176 were evaluated for their abilities to inhibit CA/SP1 cleavage of the Gag polyprotein using the LC/MS analysis method. Preliminary experiments established that MI binding to VLPs reached equilibrium within 2 hours, so incubations with MI were maintained, prior to adding protease. As shown in Fig 8A (left panel), 3 μM BVM or BMS-955176 inhibit the production of final product AM (SP1) from wt VLPs. In addition, inhibition of cleavage data by the MIs are not due to non-linear rates of cleavage, due artifactually from loss of proteolytic activity, but rather, are due to innate differences in cleavage rates (see above, protease engineered to limit autoproteolysis and cleavage rates calculated from within the linear range). However, BVM inhibition of WT CA/SP1 cleavage was not maintained throughout the entire time course of the cleavage experiment, as it dropped from 39% inhibition at 2 hours to 1% inhibition at 4 hours (Fig 8A). On the other hand BMS-955176 exhibited sustained inhibition over the 4 hour period with WT VLPs. This persistence of in vitro CA/SP1 cleavage inhibition trended with the antiviral cell culture MPI values (Table 2). For example, the sustained inhibition of cleavage of WT CA/SP1 by BMS-955176 correlates to its single cycle MPI value of 93% (100% for multiple cycle MPI) towards WT virus in cell culture, whereas the loss in in vitro inhibition of CA/SP1 cleavage at longer time points by BVM towards WT correlated to its single cycle MPI of 82% (98% for multiple cycle MPI). VLPs containing the ΔV370 polymorphism were also evaluated in this assay. BMS-955176 inhibited ΔV370 cleavage to a degree similar to BVM inhibition of WT at the earliest time point (30 minutes) and did exhibit time-dependence, but the loss of inhibition was slow, with ΔV370 cleavage still remaining partly inhibited (13%) at the 4 hour time point. BVM was not inhibitory at any time point toward ΔV370 containing VLPs. Again, the time-dependent inhibition in this assay correlates to lower MPI values in cell culture with BMS-955176/ΔV370 values of 46% for its single cycle MPI (92% for multiple cycle MPI) and BVM/ΔV370 values of -26% (SC MPI) and 9% (MC MPI). Interestingly, while BVM did not inhibit cleavage of A364V, BMS-955176 reproducibly exhibited a small degree (~10%) of inhibition at the earliest time point, but was not inhibitory by 2 hours (Fig 8B). Multiple cycle MPI values for both compounds against A364V containing virus were near zero. Thus, lower MPI values are correlated to both reduced antiviral potency (elevated MC EC50s) and a time-dependent loss of in vitro inhibition of CA/SP1 cleavage. Conversely, higher antiviral MPI values are correlated with greater antiviral potency (lower MC EC50s) and correlated with persistence of in vitro inhibition of CA/SP1 cleavage over time. To complete the data required to model MI inhibition of CA/SP1 cleavage (Fig 5) as a function of MI and Gag polymorph, specific binding affinities of BVM, BMS-1 and BMS-955176 toward VLPs containing Gag polymorphs were determined through the use of a competitive radioligand binding assay (Table 4).[40] Examples of competition displacement assay results are provided in S4 Fig, including BMS-955176/A364V. BMS-955176 affinity for WT Gag VLPs was 3.2 nM, with slightly lower affinity for V362I (4.3 nM), and reduced affinity (2- and 10-fold) for V370A and ΔV370 VLPs, respectively. By comparison, BVM affinity toward WT was 5.4 nM, which was reduced 2.9-, 9.1- and 48-fold toward V362I, V370A and ΔV370, respectively. BMS-1 affinities were 3-5x reduced, as compared with BVM. The binding of BMS-955176 toward A364V was measurable (Kd 98 +/- 13 nM), but severely attenuated. At a concentration of 3 uM, BVM only partly inhibited [3H]-BMS-977660 (BMS-‘176*). Total radiolabel binding to A364V was low; a reliable Kd could only be determined for BMS-955176. An adaptation of the binding assay was used to measure the kinetics of MI dissociation, as has been described for the determination of kinetics of dissociation of [3H] HIV integrase inhibitors from HIV-1 integrase.[65] Pre-formed MI/VLP complexes were treated with a large molar excess of a competitor MI, and the kinetics of dissociation followed over time. Representative examples of dissociation data are shown in S5 Fig, with data tabulated in Table 5. Dissociation half-lives for BVM and BMS-955176 were determined using their related C20:C29 double bond reduced tritiated derivatives, [3H] BMS-885221 ([3H BVM*) and [3H] BMS-977660 ([3H ‘176*), whose antiviral profiles are identical to BVM and BMS-955176, respectively. Dissociation half-lives for [3H]-BVM* and [3H]-176* were similar for WT VLP (41 vs. 51 minutes, Table 5). However, [3H]-BVM* dissociates rapidly from V370A and ΔV370 VLPs (≤ 3 minutes) and with an intermediate rate from V362I. By contrast, [3H]-176* displays similar dissociation kinetics towards all four VLPs. Dissociation of [3H]-176* from A364V was difficult to measure due to the low value of specific binding: T1/2 was rapid (< 2 minutes). Rates of dissociation of BVM from V370A, V362I and ΔV370 were >12, 2.0 and >19-fold faster, respectively, compared to BMS-955176 (Table 5). This is similar in magnitude to the decreased affinities of BVM for these VLPs (9.1-, 3.5- and 48-fold, respectively, as compiled in Table 4). The antiviral potencies in cell culture toward the viruses with these polymorphs share the same trend as the affinity and off rate data: when compared with BVM, BMS-955176 binds with higher affinity and dissociates more slowly from the polymorphic VLPs, a result which is qualitatively correlated with its improved ability to inhibit replication of the cognate polymorphic viruses (Table 1). Interestingly, while BMS-955176 affinity (Table 4) and dissociation rates (Table 5) for WT, V362I and V370A are correlated (similar Kd values, similar dissociation half-lives), thus indicating that affinity is mainly driven by dissociation rates, affinity of BMS-955176 toward ΔV370 is reduced 10-fold as compared to WT (Table 4), though the dissociation rate is reduced by only 1.1-fold. This may indicate that reduced BMS-955176 affinity toward ΔV370 is due to a slower rate of association or possibly more complex multi-step binding kinetics, as has been observed for HIV-1 integrase strand transfer inhibitors.[66] This slower association rate implies a less pre-organized binding site, hindering the association of the ligand to its binding site, a point later addressed in the Discussion section. Biochemical studies of rate constants of polymorphic cleavages (Table 3), on one hand, and the binding affinities of MIs (Table 4), on the other hand, indicate that there is a qualitative relationship of each to the efficacy of a given MI to inhibit viral replication of a given polymorphic virus. From cellular assays, a plateau in inhibition (MPI values of <100%) suggests an escape mechanism that appears to contribute to reduced efficacy of a given MI towards different polymorphic viruses. Here, a model integrates both biochemical and cellular data to provide a more quantitative estimation of MI inhibition of CA/SP1 cleavage, and thus formation of mature viruses in vivo. The model (detail in Materials and Methods) has two terms which describe the observed rate of cleavage (kclv,ob) at CA/SP1 by HIV-1 protease in the presence of MIs. The first term describes the cleavage of the immature virus in the unbound state (B) (Fig 5). This term incorporates the innate cleavage rate constant k1 for different polymorphs and the concentration of the MI and its affinity (Kd) for that polymorph. This is straightforward, and in accord with a simplified model (referred to here as model 1) in which only unbound state (B) is subject to protease cleavage. However, the observation of incomplete inhibition in antiviral assays (Table 2, MPI <100%) and the time-dependent loss in inhibition in in vitro cleavage assays (Fig 8) points to some degree of escape from inhibition as a function of both MI and polymorph, despite saturating concentrations of MI (Table 2, MPI <100%). To explicitly account for these observations, a second term was incorporated into the model for the rate of cleavage of immature virus in the MI-bound state (A), which is cleaved with rate constant k2, unique for each MI and polymorph. Essentially, the addition of this second term puts a cap on the maximal degree of inhibition of cleavage (referred to here as Model 2). Model 2a uses MPI values from the MC antiviral assay, with the concept that viral escape in the multiple cycle infection assay is a more realistic representation of clinical HIV replication (multiple cycle), as compared to the situation in the SC assay (Model 2b). An example of this approach is provided by BVM inhibition of the V370A virus where the antiviral MC MPI is 65.4%. This indicates that the fractional degree of cleavage arising from the bound form (A) is ((100–65.4)/100) = 0.345 of that arising from the unbound form (B). Thus the rate constant k2 = 0.345*k1. Model 1 and 2a reductions in the rates of cleavage of WT, V370A, V362I and ΔV370 by BMS-955176 and BVM are depicted in Fig 9 as log10 reductions in cleavage rates vs. the uninhibited (no MI) control. The upper panels (Fig 9A and 9B) were modeled with only the first term included (biochemical data only, model 1), while the lower panel was modeled with both terms included (biochemical and MC antiviral MPI data, model 2a). A key result for model 1 is that based solely on biochemical data, its estimation is in rough alignment with the antiviral results for these variants. A key result of model 2a (lower panel) is that there is a plateau in the degree of inhibition that depends on MI and polymorph, a direct result of the inclusion of the antiviral MPI data, bringing model 2a into closer alignment with the antiviral MPI data. A quantitative comparison between BMS-955176 and BVM can be made from the modeling approaches at a selected MI concentration, for example at 300 nM MI (reductions spanning the entire range of concentrations are plotted in Fig 9, and tabulated in S1 and S2 Tables). At this concentration, BMS-955176 log10 reductions in WT virus (from the MC MPI data alone), log10 reductions in WT VLP cleavage rates (from model 1) and log10 reductions in WT VLP cleavage rates from model 2a are < -2.00 log10, -1.98 log10 and -1.96 log10 (Table 6), respectively. For BVM, these values are -1.80 log10, -1.75 log10 and -1.48 log10, respectively. Differences among the methods for the two MIs are somewhat larger for V362I (BMS-955176: -1.20/-1.85/-1.12 vs. BVM: -0.73/-1.30/-0.64), and V370A (BMS-955176: -1.73/-1.67/-1.41 vs. BVM: -0.46/-0.85/0.36, respectively). Log10 calculated reductions by BMS-955176 toward ΔV370 (BMS-955176: -1.09/-1.00/-0.76) are lower, as compared to wt and V370A, in accord with clinical results for subtype C viruses,[43] while values for BVM indicate essentially no antiviral activity (-0.04/-0.33/-0.02) toward the ΔV370 polymorph. Model 1 wt reductions for BVM (-1.75), are similar to that from the antiviral MPI data (-1.80). Model 1 V362I, V370A and ΔV370 predictions are somewhat larger vs. antiviral data. By comparison, the inclusion of MC antiviral MPI data (model 2a) results in lower predicted log10 reductions for wt and polymorphs, in line with the antiviral data for BVM.[18, 68] Table 6 also contains calculated log10 reductions for wt, V362I, V370A and ΔV370, using a modification of model 2 in which MPI values are taken from the SC assay (model 2b, no term for k2). In these cases, model 2b gives similar results to those taken directly from log10 viral reductions calculated directly from the SC antiviral MPI values (as to be expected given the weight of the SC MPI-derived term in the equation which dominates the response over that of the biochemical-only model 1), and under-predicts the clinical responses. Model 2a time courses for the appearance of cleavage product SP1 peptide by HIV-1 protease for WT, V362I, V370A and ΔV370 VLP, and inhibition profiles by 300 nM BVM or BMS-955176, are shown in Fig 10. As further detailed in S1 Table (model 1, no MPI data included), 300 nM BMS-955176 reduces the rate of cleavage of WT, V362I, V370A and ΔV370 by 95, 71, 47 and 10-fold, respectively. BVM is effective at reducing the rate of WT cleavage (57-fold), less effective toward V362I (20-fold), much less effective toward V370A (7.1-fold), and ineffective toward ΔV370 (2.2-fold). Model 2a (S2 Table) indicates that 300 nM BMS-955176 reduces the rate of cleavage of WT, V362I, V370A and ΔV370 by 91-, 13-, 25- and 6-fold, respectively. By comparison, model 2a indicates that while BVM is effective at reducing the rate of WT cleavage (30-fold), it is far less effective toward V362I (4.4-fold) and ineffective toward V370A (2.3-fold) and ΔV370 (1.1-fold). Another way to visualize the results is to compare antiviral dose-response curves to those generated from the models across all concentrations. Plateaus in antiviral inhibition are apparent, particularly for BVM and BMS-1 toward polymorphic variants, as noted in Table 2. This is shown in Fig 11, which displays the antiviral dose-responses (MC assay) for combinations of BVM, BMS-1 and BMS-955176 toward wt, V362I, V370A and ΔV370 viruses, compared to the calculated values from models 1 and 2a (exception: the combination of BMS-1 with ΔV370 was not performed). The results illustrate that the antiviral data is in better alignment with model 2a compared to model 1. The data also highlight that binding per se is insufficient to result in a complete antiviral effect (MPI values of 100%) for several combinations of MIs and polymorphic variants, for example for BVM and BMS-1 toward V362I, V370A and ΔV370, despite binding, (Fig 11 and Table 4). Clinical viral load reduction (VLR) data from BMS-955176[41–43, 70] and BVM clinical trials[18, 28, 68] are shown in Table 6. Clinical VLR reductions were compared to reductions in rates of CA/SP1 cleavage using the different models at a concentration of 300 nM (fold reduction values relative to no MI added to each particular virus). This concentration of MI was chosen for the comparison for two reasons. First, BVM trough concentrations of >20 μg/mL were associated with the best clinical responses[18, 28, 68] and based on a BVM antiviral serum shift of 130-fold[35, 36, 40] the implied free concentration of 20 μg/ml BVM is 263 nM. Similarly, the clinical response of BMS-955176 in a 10 day Ph2a study reached a plateau at C24 exposures between 713 and 1289 nM[67] (mean = 1521 nM), implying a mean free concentration (based on a reported free fraction of 0.14)[40] of 213 nM. Thus, modeling was compared at 300 nM for both MIs. The maximal median decline for subjects having a WT genotype at 40 mg QD dosing by BMS-955176 in a Ph2a POC 10 day monotherapy study was (-1.75) log10 (Table 6).[41] This value is slightly less than both the model 1 and 2a values (~-2 log10), and less than the value directly calculated from the MPI in the MC assay (< -2.00 log10). With respect to subjects harboring Gag polymorphisms (Gag amino acids 362, 364, 370, 371) at a dose of 40 mg BMS-955176, a comparison can be made to V370A, with V370A acting as a kind of surrogate for such polymorphisms (there is currently no available data breaking out patient responses to individual polymorphic viruses). Model 1 values for polymorphs V370A and V362I (-1.67 and -1.85 log10, respectively), or the values directly calculated from the MPI in the MC assay (-1.73 and -1.20 log10), are similar to the clinical response of BMS-955176 reported for polymorphs (-1.71 log10), while the projected values from model 2a for V370A and V362I, incorporating the MC MPI data (-1.41 and -1.12 log10, respectively), are somewhat lower than reported for subjects with these polymorphic genotypes. The model 2b V370A value (-0.61 log10), using single cycle MPI data, greatly underestimates the clinical result for subjects harboring polymorphic viruses, thus suggesting that SC MPI values are likely too stringent, leading to an under-estimation of clinical responses (Table 6). For BVM, the mean decline for those subjects achieving trough concentrations of >20 μg/mL[28, 29] with a WT genotype at 250–400 mg QD in a Ph2 14 day monotherapy POC study was -1.26 log10.[18] This value is lower than both the WT model 1 calculated decline (-1.75 log10) or the value directly calculated from the MPI in the MC assay (-1.80 log10). The model 2a value (-1.48 log10) is closer to the clinical data. A mean -0.21 log10 decline was noted in subjects harboring Gag polymorphisms (Gag amino acids 369, 370, 371) at doses of 250–400 mg BVM,[68] which may be compared to the V370A and V362I polymorphic viruses used in this study. The model 1 BVM declines (-0.85 and -1.30 log10) over-predict the clinical response, while the projected declines calculated from the MPI values (solely from the MC assay) for these two variants (-0.46 and -0.73 log10) or model 2a (-0.36 and -0.64 log10), respectively, are in closer alignment for these types of polymorphic patient viruses (-0.21 log10). The calculated reductions in CA/SP1 cleavage rates for WT and polymorphic viruses V362I and V370A at 300 nM MI (Table 6) are compared in Fig 12. Overall, of the models, model 2a, incorporating MC assay MPI values, provides a better correspondence to both antiviral dose response curves (Fig 11) and clinical viral load reductions (Fig 12). An early MI failed in the clinic due to inability to inhibit ~50% of viruses containing polymorphic variation in Gag near the site of MI action. The 2nd generation MI, BMS-955176, is active toward these viruses. In this study we sought to understand the mechanistic origin for the improved antiviral activity of BMS-955176, and to model this behavior as a function of Gag polymorph cleavage rates, MI affinity and MI concentration, with consideration as to how this information relates structurally to MI binding. Such an approach may have utility in interpreting pre-clinical antiviral results and clinical data on MI action, and may also be helpful in the discovery of MIs with further improvements to potency and spectrum. The higher affinity of BMS-955176 toward Gag polymorphs appears to be a predominant driver for better antiviral activity toward Gag polymorphs (both lower EC50 values as well as higher MPI values). Similarly, higher BMS-955176 affinity is apparently an important driver for the superior performance in in vitro cleavage assays. BMS-955176 inhibition is maintained against WT at all time points (4 hours), while BVM inhibition is lost over time. Consistent with the overall relationship, in a case where BMS-955176 has a phenotype of partial time-dependent inhibition (in vitro cleavage for ΔV370, Fig 8B), this was correlated to an elevated FC in antiviral assays (Table 1, 6.8-fold) and a reduced MPI value in the MC assay as compared to WT (93 vs. 100%). One of the features of the inter-relationships of the data is that Kd values are linearly correlated with both MPI values and FC antiviral EC50 values (R2 0.96), with greater affinity providing higher MPI values and lower FC. Antiviral and biochemical data were integrated into a model for calculating the reduction in the rate of cleavage of CA/SP1 by a given MI/polymorphic combination. Modeled reductions in rates of CA/SP1 cleavage by BMS-955176 and BVM were compared to antiviral data in cell culture and viral load reductions observed clinically with these MIs using several models, the most relevant model being one which incorporates both biochemical MI affinities for its Gag target, innate cleavage rates for the viruses and values for MPI from multiple cycle anti-viral data (Model 2a). At a dose of 40 mg QD BMS-955176 in a 10 day monotherapy POC study, the maximal median viral load declines for subjects having WT or polymorphic genotypes were -1.75 log10 and -1.71 log10, respectively, in alignment to values calculated from model 2a (wt: -1.96 log10, V370A: -1.41 log10, V362I: -1.12 log10). Similarly, at doses of 250–400 mg QD BVM in a Ph2a 14 day monotherapy study of subtype B patients, the mean viral load declines for subjects having a WT or polymorphic genotype were -1.26 log10 and -0.21 log10, respectively, in the range of values calculated from model 2a (wt: -1.48 log10, V370A: -0.36 log10). These studies determined that in vitro inhibition of HIV-1 replication by early generation MIs BVM and BMS-1 does not always reach 100%. It should also be noted that for one polymorphic variant (ΔV370) BMS-955176 also does not always reach 100% inhibition as well; albeit to a significantly reduced degree. This observation with respect to early generation MIs was observed across polymorphs, and was correlated with a reduction in antiviral potency (increased fold change EC50 values) by a particular MI toward the particular virus containing that Gag polymorph. For example, in a multiple cycle assay, BVM maximally inhibits the replication of HIV-1 Gag V370A by 65.4%, and, in a single cycle assay, by 19%, exhibiting a 54-fold change in its multiple cycle EC50. These observations suggest that, depending on polymorph and MI, this phenomenon is analogous to one of partial antagonism. In seeking the mechanistic origins of this behavior we initially considered a simplified model for MI inhibition of CA/SP1 cleavage of viral particles in which cleavage only takes place in that fraction of particles not bound to the MI. Thus, CA/SP1 cleavage should continue apace on the MI-unbound particles at a rate determined by the steady state fraction of unbound MI. Because of this, model 1 places no upper limit on the degree of maximal inhibition: at saturating MI concentrations the fractional amount of unbound Gag will approach zero, and thus complete inhibition is to be expected. However, the antiviral phenotype of incomplete inhibition in cell culture at saturating BVM concentrations argues against this simple model, thus suggesting the need for a modification to the model to explicitly include a term which ultimately places an upper value on the degree of maximal inhibition. For this purpose we made use of the MC MPI values, which we interpret as a direct functional readout of viral escape from MI action in cell culture. Parameterizing the biochemical-only model (model 1) required determination of the appropriate biochemical values for the innate rates of HIV cleavage and the affinities of MIs toward assembled Gag virus-like particles. These measurements were made by developing two assays. In the first, we made use of an LC/MS-based assay that directly measures CA/SP1 cleavage vs. time, thus providing rate constants for this process as a function of polymorph. These results showed that Gag polymorphic variants that are less susceptible to inhibition of replication by early generation MIs BVM and BMS-1 (Table 1) are cleaved 2.7–9.7-fold more rapidly than the WT (Table 3) and they correspondingly exhibit the most pronounced incomplete inhibition profiles (MPIs <100%) in antiviral assays (Table 2 and Fig 11). In a qualitative sense, poorer antiviral coverage of these polymorphs appears to be in part a consequence of poorer MI affinity for Gag, but also is a reflection of a lack of ability of BVM and BMS-1 to fully inhibit when bound, i.e., consistent with the proposed pathway in which cleavage occurs despite MI binding (k2-mediated, Fig 5). This results in what is in essence partial antiviral antagonism, as a function of MI and polymorph, which cannot be overcome by merely increasing MI concentration. V362I is more sensitive to BVM inhibition vs. V370A. Though superficially posing a challenge to a model in which efficacy of inhibition of CA/SP1 cleavage is entirely a function of cleavage rate, this is not the case for model 2a, where terms 1 and 2 of equation (see model for inhibition) also incorporate the Kd value for the binding of the MI. In this case BVM affinity for V370A is 9.1-fold poorer than wt, while BVM affinity for V362I is 2.9-fold reduced. This 3-fold higher affinity for V362I contributes, in part, to allowing BVM to maintain, albeit incompletely and right shifted, activity toward V362I, while losing activity toward V370A. Mechanistically, what structural model might explain the result of escape from inhibition, despite binding? The following proposed model is based on a number of reported observations. First, NMR studies indicated that the superstructure around CA-SP1 in the region of MI binding (SP1) is in dynamic equilibrium between a random coil and an alpha helix.[71] In support of this dynamic equilibrium model, small changes to buffer and detergent alter the helicity of the SP1 region [72] while point mutations predicted to reduce helicity destroy particle production. [73–75] Earlier cryo-electron tomography work on immature particles found that the extension of SP1 from the C-terminal region of CA could be fitted as a six-helix bundle, leading to a proposal that cleavage at CA-SP1 acts as a molecular switch, facilitating the final conformational changes required for capsid rearrangement and core condensation. [76, 77] A deeper structural understanding is now at hand with the report of a cryo electron tomography structure of the immature assembled Gag lattice at 3.9 angstrom resolution and a crystal structure reported at at 3.2 angstrom resolution [61, 62] The structures indicate that the CA-SP1 cleavage site is hidden within this 6-helix bundle, and protected from cleavage due to inaccessibility, a structural explanation for why cleavage at this site is the slowest of the Gag cleavages.[15] MI binding is suggested to rigidify the structure and likely shifts the equilibrium of the superstructure in favor of the 6-helix structure, thus reducing the propensity for unraveling and presentation of the cleavage site. This is in accord with a report that BVM binding increases the stiffness of immature virions.[78] The formation of a more ordered helical state as a consequence of MI binding in this region, shown by cross linking studies of BVM analogs at sequences overlapping or proximal to the CA-SP1 cleavage site, is also consistent with previous biochemical data on the effect of bevirimat on Gag processing, and with genetic data from resistance mutations.[60] The results reported in this study are in alignment with these structural results and proposal for the role of polymorphic or MI resistance changes which increase cleavage site presentation. As compared to wt, the more rapid innate rates of CA/SP1 cleavage of certain polymorphs are therefore explainable as a reflection of a decrease in the stability or equilibrium concentration of the bundle, i.e., the inherently greater degree of disorder in the cleavage region allows for the presentation of the protease recognition site in its extended conformation a greater proportion of the time. The modeled biochemical and viral data, which showed improved inhibition of in vitro cleavage and higher maximal antiviral inhibition by BMS-955176 are consistent with a global explanation for the broader antiviral coverage of BMS-955176 vs. BVM: the increased affinity of BMS-955176 for its binding site increases the concentration and perhaps structural integrity of the quarternary structure of the assembled 6-helix bundle This thereby decreases dynamic fraying of the structure which would otherwise lead to protease cleavage. With respect to the observed phenotype of partial antagonism by certain MI/polymorph combinations, the data suggests that binding in and of itself is not always sufficient to induce changes in the local geometry needed to completely prevent protease recognition of CA/SP1 and thereby completely block cleavage. This may be the case for V362I vs. V370A. While these two polymorphs are cleaved with similar rates (Table 3), they exhibit differing MPI values depending on MI. The wt MI-bound Gag structure is likely innately more ordered to begin with, while polymorphic variants, with greater innate flexibility and reduced local order, retain some bias in this direction, despite MI binding (Fig 5, pathway 2), rendering them partly susceptible to cleavage even in the MI-bound state. This suggests that depending on the particular effects induced on the local conformation by a given polymorphic change and the particular binding poise of an MI, the consequences of that binding may be only partially transmitted to the key conformation changes that are meaningful for antiviral activity, that of maintaining reduced access of the CA/SP1 cleavage site to protease. Thus, biochemically one observes a time dependence to the in vitro cleavage inhibition, while in antiviral assays, less than maximal antiviral inhibition. This is an escape mechanism. At the structural level, the ability of the 2nd generation MI BMS-955176 to induce greater protection from cleavage of polymorphs is possibly due to additional binding contacts within the Gag structure, reflected in its higher binding affinity, but a detailed explanation must await MI bound structures. Such binding presumably contributes to a greater stabilization of that local conformation (presumably increased helicity of SP1) which renders the system less sensitive to protease recognition/cleavage. Further, while the generality of the conclusion that faster innate rates of polymorphic cleavage are a reflection of greater flexibility and accessibility of the CA/SP1 site to protease recognition and cleavage seems therefore to be sound, further studies are needed to understand the structural details of MI binding, in particular to shed light on those cases where saturable binding is still not maximally productive (partial antagonism). Given the similar dissociative off rates of BMS-955176 toward wt and ΔV370 VLPs, but the higher affinity toward wt, the calculated rate of association toward the ΔV370 variant is implied to be ~9-fold slower than wt (from consideration of a simple 1 step binding model (kon = koff x Kd). This slower on rate may reflect a more unstructured MI-unbound structure in the vicinity of the ΔV370 MI binding site (as compared to, for example, V370A, with a calculated relative on rate similar to wt). From the published structure, position 370 is at the end of the 6-helix bundle, so potentially deletions in this region introduce unzipping and greater local disorder, with such a disordered state obscuring the trajectory of MI binding, and thereby inducing an entropic penalty to binding. While further work is clearly needed to more fully understand the relationship of modeled to antiviral and clinical results, the approach described herein to understand MI activity and mechanism should prove useful to potentially facilitate further improvements to MI potency and coverage.

10.1371/journal.pgen.1005618

Allelic Variation of Cytochrome P450s Drives Resistance to Bednet Insecticides in a Major Malaria Vector

Scale up of Long Lasting Insecticide Nets (LLINs) has massively contributed to reduce malaria mortality across Africa. However, resistance to pyrethroid insecticides in malaria vectors threatens its continued effectiveness. Deciphering the detailed molecular basis of such resistance and designing diagnostic tools is critical to implement suitable resistance management strategies. Here, we demonstrated that allelic variation in two cytochrome P450 genes is the most important driver of pyrethroid resistance in the major African malaria vector Anopheles funestus and detected key mutations controlling this resistance. An Africa-wide polymorphism analysis of the duplicated genes CYP6P9a and CYP6P9b revealed that both genes are directionally selected with alleles segregating according to resistance phenotypes. Modelling and docking simulations predicted that resistant alleles were better metabolizers of pyrethroids than susceptible alleles. Metabolism assays performed with recombinant enzymes of various alleles confirmed that alleles from resistant mosquitoes had significantly higher activities toward pyrethroids. Additionally, transgenic expression in Drosophila showed that flies expressing resistant alleles of both genes were significantly more resistant to pyrethroids compared with those expressing the susceptible alleles, indicating that allelic variation is the key resistance mechanism. Furthermore, site-directed mutagenesis and functional analyses demonstrated that three amino acid changes (Val109Ile, Asp335Glu and Asn384Ser) from the resistant allele of CYP6P9b were key pyrethroid resistance mutations inducing high metabolic efficiency. The detection of these first DNA markers of metabolic resistance to pyrethroids allows the design of DNA-based diagnostic tools to detect and track resistance associated with bednets scale up, which will improve the design of evidence-based resistance management strategies.

Scale up of Long Lasting Insecticide Nets has massively reduced malaria mortality across Africa. However, resistance to pyrethroid insecticides in malaria vectors threatens its continued effectiveness. Here, we established that allelic variation in two CYP450s is the most important driver of pyrethroid resistance in the major African vector Anopheles funestus and detected key mutations controlling this resistance. The duplicated P450s CYP6P9a and CYP6P9b are directionally selected across Africa with alleles segregating according to resistance phenotypes. Alleles from resistant mosquitoes present significantly higher metabolic activities toward pyrethroids compared with alleles from susceptible mosquitoes. Furthermore, transgenic flies over-expressing resistant alleles of both genes were significantly more resistant to pyrethroids. Three amino acid changes from the resistant CYP6P9b allele are the key pyrethroid resistance mutations which induce high metabolic efficiency. The detection of these first DNA markers of metabolic resistance to pyrethroids allows the design of diagnostic tools to detect and track resistance.

Despite the recent decrease in malaria mortality (47%) [1], the disease remains a serious public health burden in the tropical world, with 584,000 deaths globally in 2013, of which 90% occurred in WHO African region, and mostly in children under the age of 5. Malaria control relies heavily on the use of insecticide-impregnated LLINs and indoor residual spraying (IRS) [2]. Unfortunately, resistance to insecticides, especially pyrethroids (the only class approved by WHO for LLINs [3]), in major malaria vectors such as and An. funestus [4–6] and An. gambiae [7, 8] is threatening to derail these intervention tools [9]. An. funestus is widely, geographically distributed across Sub-Saharan Africa [10], and it has high vectorial capacity in some places surpassing even that of An. gambiae [11]. It reaches maximal abundance in the dry season when the density of An. gambiae and An. arabiensis have declined, thereby extending the period of malaria transmission [12]. Cases of resistance to pyrethroid, carbamate and organochlorine insecticides are increasingly reported in An. funestus populations across Africa [4, 5, 13–15]. It is imperative to design and implement suitable resistance management strategies to limit the impact of such resistance (WHO, 2012). One prerequisite is the development of appropriate diagnostic tools to facilitate the monitoring of insecticide resistance at an early stage, in order to inform control programs of the best course of action to take. However, the design of DNA-based diagnostic tools requires a thorough understanding of the molecular basis of the resistance. To date, efforts to characterise mechanisms of resistance in malaria vectors have implicated knockdown resistance (kdr) and metabolic resistance through elevated expression of resistance genes, especially cytochrome P450s [6, 8] as the two major mechanisms conferring resistance to pyrethroids. No kdr mutation has been reported in the voltage-gated sodium channel of An. funestus [2]; pyrethroid resistance is therefore mainly metabolic. However, despite the numerous reports of implications of over-expressed P450s in pyrethroid resistance, the detailed molecular mechanisms through which they confer pyrethroid resistance in mosquitoes remain largely uncharacterised. It also remains unclear whether mechanisms other than P450 over-expression are also involved in resistance, for example, allelic variation with changes in the coding sequences through mutations of key amino acid residues or cis and/or trans mutations which could impact gene regulation [16]. Recent observation of polymorphism variations for the two most important pyrethroid resistance genes, CYP6P9a and CYP6P9b, in the malaria vector An. funestus [17, 18] suggests that this mosquito species is an excellent candidate to assess the impact of allelic variations of resistance genes on pyrethroid resistance. Recently, for the glutathione S-transferase gene, GSTe2, in An. funestus, a single point mutation (Leu119Phe) was detected and established to confer DDT resistance [15]. If such causative mutation(s) could be identified for pyrethroid resistance mediated by P450 genes, it will facilitate the design of DNA-based diagnostic tools to easily detect and track such resistance in field populations. Knowledge of insect P450s’ preferential sites of metabolism of insecticides can facilitate the design of highly specific synergist inhibitors against the detoxification genes. For example, in insects and other organisms the 4´ spot of the phenoxybenzyl ring have been shown to be the preferential sites of hydroxylation [19, 20], which is followed by hydrolysis to generate intermediates including alcohols and acids which could easily be conjugated. On the other hand fluorogenic probes are increasingly used to identify inhibitors of P450s, and as diagnostic compounds to establish the degree of binding of insecticide substrates to P450s [21, 22]. In this study, sequence characterisation of CYP6P9a and CYP6P9b across Africa detected important polymorphism variations. In silico predictions, in vitro and in vivo functional characterisation tools were then applied to demonstrate that allelic variation in CYP6P9a and CYP6P9b is the key molecular change through which An. funestus mosquitoes acquire high resistance to pyrethroid insecticides. Furthermore, site-directed mutagenesis coupled with in vitro functional characterisation of the mutant recombinant proteins detected three major amino acid changes responsible for high pyrethroid metabolising efficiency of CYP6P9b from resistant populations of An. funestus s.s., compared to susceptible allele. This will allow the design of DNA-based diagnostic assay to detect and track this resistance in field populations in Africa. Having established that the allelic variations observed between CYP6P9a and CYP6P9b haplotypes induced significant differences in their 3D structures and possibly their ability to interact and metabolise pyrethroid insecticides, we next validated these observations through in vitro and in vivo characterisation experiments. Recombinant proteins of the various CYP6P9a and CYP6P9b alleles were produced with optimal expression between 36–56 hours as described previously [4]. No significant differences were observed in P450 content and cytochrome P450 reductase activity between the recombinant proteins expressed from the various alleles of CYP6P9a and CYP6P9b (hereafter named resistant alleles) compared to proteins obtained from the FANG alleles (hereafter referred to as susceptible alleles) (S6A and S6B Fig). Thus, reductase content may not induce any allelic variation in metabolic activities. The extent of the role played by allelic variation of these P450s in pyrethroid resistance was determined in order to establish whether it is alone sufficient to confer resistance in vivo, even more than gene over-expression. For this purpose, the most resistant (MAL) and the susceptible alleles (FANG) of each gene were over-expressed in vivo through a transgenic expression using the GAL4/UAS system, and their ability to confer resistance compared using contact bioassays with permethrin and deltamethrin. qRT-PCR confirmed that both CYP6P9a and CYP6P9b were expressed only in the transgenic F1 progenies from the crosses between the Actin5C-GAL4 driver line and the different UAS-CYP6P9 lines, and not expressed in the control flies (S9A and S9B Fig) (see S1 Text). Control flies are generated by crossing flies with the same background as the experimental group (but devoid of the UAS element and the candidate gene), with the driver (Actin5C-GAL4) lines to generate Actin5C-GAL4-null lines without candidate P450s. Detecting insecticide resistance at early stage is one of the prerequisite for the design and implementation of effective insecticide resistance management strategies. DNA-based diagnostic tools are essential for this purpose, but the diagnostics require a thorough understanding of the molecular basis of insecticide resistance. In this study, we dissected the molecular basis of a monooxygenase-mediated resistance to pyrethroid in one of the major African malaria vectors An. funestus demonstrating that (i) allelic variation in cytochrome P450 genes is a key molecular mechanism conferring pyrethroid resistance in field mosquitoes, and (ii) detecting key amino acid changes in resistant P450 alleles responsible for pyrethroids-metabolising efficiency. The finding of these resistance markers could help in the design of DNA-based diagnostic tools that will facilitate the detection and tracking of such resistance markers in the field. This study has provided several evidences supporting the key role of allelic variation in conferring pyrethroid resistance. Detection of DNA-based markers of metabolic resistance has so far proved challenging because of the redundancy of genes involved but also because of the multitude of mechanisms through which these mechanisms can operate. Here, using site-directed mutagenesis and comparative recombinant enzyme characterisation we established, for the first time in our knowledge in a mosquito species, that three amino acid replacements in the P450 CYP6P9b are responsible for high metabolic activity toward pyrethroids as their replacement in MALCYP6P9b with variants from the susceptible FANGCYP6P9b correlated with significant loss of catalytic activity and metabolic efficiency towards permethrin. We suggest that the three residues (Val109, Asn384 and Asp335) in the protein from the resistant alleles of CYP6P9b contribute in unique ways to create ensemble complementarity groups on the P450 active sites, a potent pharmacophore which recognise key features on the pyrethroid substrates, allowing optimal inter-molecular interactions and enhancing affinity and/or catalysis. Impact of amino acid changes in metabolic activity of CYP450 genes have been described previously in relation to metabolisms of other chemicals but not against pyrethroids. Indeed, the Glu318Asp substitution in human CYP1A2 has been shown to increase the Kcat of O-dealkylation of 7-ethoxycoumarin 13-fold [37] (see S1 Text). As these are the first markers of metabolic resistance to pyrethroids involving P450s in An. funestus and/or even other mosquito species, these mutations should facilitate the design of DNA-based diagnostic tests for tracking resistance in the field. However, unlike the case of DDT-resistance gene GSTe2 (Leu119Phe substitution) in An. funestus, the CYP6P9a and CYP6P9b are duplicated genes with high sequence similarity which will require particular attention in the design of diagnostic assays. The detection of these causative mutations in CYP6P9b also set a pace for studies of this kind to be carried out in other insect vectors for which allelic variation in key resistance genes could also play an important role. This study presents a detailed dissection of the genetic and molecular basis of metabolic resistance to insecticide in a major malaria vector demonstrating that allelic variation of key P450 genes is responsible for the resistance to pyrethroids, the bed nets insecticides. Key amino acid changes between resistant and susceptible alleles specifically, three residues, Val109, Asp335 and Asn384 in the resistant CYP6P9b alleles accounted for these metabolic differences. The finding and characterisation of these resistance markers paves a way to design a DNA-based diagnostic test that can allow tracking of these resistant alleles across Africa, enabling the design of evidenced-based resistance management strategies to mitigate the impact of this resistance on the success of ongoing and future control interventions. To assess whether allelic variation in CYP6P9a and CYP6P9b was the most important factor conferring pyrethroid resistance even more than gene over-expression, a resistant (Malawi) and a susceptible (FANG) alleles for each gene were independently expressed in D. melanogaster, using the GAL4-UAS system. The hypothesis being that if sequence variation in these two genes is the key determinant of pyrethroid resistance, then over-expression of resistant alleles of both genes will confer more resistance to pyrethroids than observed by the over-expression of the susceptible alleles. To detect the key amino acid changes conferring high pyrethroid-metabolising efficiency in resistant CYP6P9a and CYP6P9b alleles, a hypothesis-driven approach was used to select candidate amino acids from resistant alleles which were mutated into variants existing in the susceptible alleles, FANGCYP6P9a and FANGCYP6P9b. The hypothesis being that if the target residue is linked with high pyrethroid metabolising-activity in the resistant allele then a significant loss of activity will be obtained when mutated into a variant from the susceptible allele. For CYP6P9a the substitution aimed at were Ser320Tyr, Phe431Ser, Gln301His and a double mutant Leu63Phe_Lys66Gln. For CYP6P9b the following replacements were effected: Val109Ile, Asp335Glu, Asn384Ser and Pro401Ala. For both genes, the first amino acid in each case represents residue from the resistant allele and the second amino acids being variants present in the susceptible allele. Amplification was achieved using mutagenic primers (S8 Table) in a whole plasmid primer extension PCR followed by self-ligation at transformation step [62, 63]. The PCR was carried out using plasmidic pJET1.2::ompA+2-MALCYP6P9a and pJET1.2::ompA+2-MALCYP6P9b respectively as templates, with Phusion HotStart II High-Fidelity DNA Polymerase (Thermo SCIENTIFIC, USA). In a reaction mix containing 1X Buffer (with 7.5mM MgCl2), 80μM dNTP mixes, 0.3μM each of forward and reverse mutagenic primers, 2-4ng plasmidic DNA template, 0.04U/μl of Phusion HotStart Taq and sterile water was added to give 50μl. The reaction was started by preheating the mixture to 98°C for 10 minutes; and then followed by 35 cycles each of 94°C for 30 seconds; annealing (at 65°C) for 30 seconds and extension at 72°C for 2.5 minutes. This is then followed with final extension at 72°C for 10 minutes, and 4°C (hold). The PCR product was incubated for 1 hr with 2μl of 1X FastDigest Buffer and 1μl of DpnI (Thermo SCIENTIFIC), at 37°C to digest the Dam+-methylated parental template [62]. 4μl of the digest was then transformed into E. coli DH5α. Positive colonies were mini-prepped and sequenced on both strands to confirm presence of desired mutations. Mutagenic plasmids of MALCYP6P9b were successfully digested with NdeI and XbaI restriction enzymes, gel extracted with QIAquick Gel Extraction Kit (QIAGEN) and ligated into pCWOri+ already linearized with the same restriction enzymes, to construct the mutagenic cassettes pB13::ompA+2-MALCYP6P9b; mutagenic plasmids of MALCYP6P9a could not be cut out from the pJET1.2 construct and consequently the P450 insert could not be introduced into the expression vector pCWOri+. The mutant plasmids of MALCYP6P9b were then co-transformed together with AgCPR into JM109 and functional membranes expressed as described previously. Efforts to introduce mutations into the constructs pB13::ompA+2-CYP6P9a and pB13::ompA+2-CYP6P9b directly were not successful due possibly to the larger size of pCWori+ plasmid and as such amplification of mutants from MALCYP6P9a plasmids were not successful. Recombinant MALCYP6P9b membranes expressed from mutant plasmids were used alongside the wild type recombinant MALCYP6P9b (wrMALCYP6P9b) to screen for metabolic activities against Type I and Type II pyrethroids (permethrin and deltamethrin, respectively), in order to determine the functional impact of each amino acid changed on metabolism of insecticides. In addition, fluorescent probes were also screened to establish differences in the O-dealkylation activity of the mutants compared with the wild type enzyme. All in vitro and kinetic analyses were conducted as described above. However, as different mutants expressed with varying concentrations all assays were conducted within linear range, with the amount of enzyme and time which produced highest activities. The DNA sequences reported in this paper have been deposited in the GenBank database (accession numbers: GenBank KR866022-KR866069).

10.1371/journal.pgen.1000477

Designing Genome-Wide Association Studies: Sample Size, Power, Imputation, and the Choice of Genotyping Chip

Genome-wide association studies are revolutionizing the search for the genes underlying human complex diseases. The main decisions to be made at the design stage of these studies are the choice of the commercial genotyping chip to be used and the numbers of case and control samples to be genotyped. The most common method of comparing different chips is using a measure of coverage, but this fails to properly account for the effects of sample size, the genetic model of the disease, and linkage disequilibrium between SNPs. In this paper, we argue that the statistical power to detect a causative variant should be the major criterion in study design. Because of the complicated pattern of linkage disequilibrium (LD) in the human genome, power cannot be calculated analytically and must instead be assessed by simulation. We describe in detail a method of simulating case-control samples at a set of linked SNPs that replicates the patterns of LD in human populations, and we used it to assess power for a comprehensive set of available genotyping chips. Our results allow us to compare the performance of the chips to detect variants with different effect sizes and allele frequencies, look at how power changes with sample size in different populations or when using multi-marker tags and genotype imputation approaches, and how performance compares to a hypothetical chip that contains every SNP in HapMap. A main conclusion of this study is that marked differences in genome coverage may not translate into appreciable differences in power and that, when taking budgetary considerations into account, the most powerful design may not always correspond to the chip with the highest coverage. We also show that genotype imputation can be used to boost the power of many chips up to the level obtained from a hypothetical “complete” chip containing all the SNPs in HapMap. Our results have been encapsulated into an R software package that allows users to design future association studies and our methods provide a framework with which new chip sets can be evaluated.

Genome-wide association studies are a powerful and now widely-used method for finding genetic variants that increase the risk of developing particular diseases. These studies are complex and must be planned carefully in order to maximize the probability of finding novel associations. The main design choices to be made relate to sample sizes and choice of commercially available genotyping chip and are often constrained by cost, which can currently be as much as several million dollars. No comprehensive comparisons of chips based on their power for different sample sizes or for fixed study cost are currently available. We describe in detail a method for simulating large genome-wide association samples that accounts for the complex correlations between SNPs due to LD, and we used this method to assess the power of current genotyping chips. Our results highlight the differences between the chips under a range of plausible scenarios, and we demonstrate how our results can be used to design a study with a budget constraint. We also show how genotype imputation can be used to boost the power of each chip and that this method decreases the differences between the chips. Our simulation method and software for comparing power are being made available so that future association studies can be designed in a principled fashion.

The International HapMap project [1],[2] documented the strong correlations between alleles at polymorphic loci in close physical proximity along human chromosomes. As a consequence it is necessary to genotype only a subset of loci to capture much of the common variation in the genome. Combined with recent technological innovations this observation has made the concept of genome-wide association (GWA) studies a reality [3],[4]. Over the few last years these studies have been very successful in uncovering new disease genes for many different complex diseases [5]. Well over 300 such loci have already been published and many more studies are currently being planned. In the design of such studies two fundamental decisions have to be made: which loci to genotype, and in how many individuals. Both decisions have practical constraints. For example it is currently not possible to assay all known variation in the human genome at a reasonable cost and choices must be made between a set of commercially available genotyping chips. Similarly, sample sizes are often limited by the number of well characterized clinical samples. Therefore, ultimately, the researcher and funding bodies must ask how to use the financial and practical resources available in order to best further the understanding of the genetics of the disease or trait of interest. A primary consideration should be the power of the study: the probability of detecting a variant assumed to be causal. In comparing chips for GWA studies it has been common to ask what proportion of SNPs not directly genotyped are “captured” or “tagged” by the chip, i.e. are well predicted, via LD, by a SNP, or combination of SNPs, on the chip. To do so it is necessary to define the level of prediction required, or equivalently to set a threshold for the required level of correlation. Although arbitrary, this has often been set at 0.8 [6],[7],[8]. The resulting proportion of SNPs captured at this level is often referred to as the coverage of the chip. Having specified the threshold it is possible to estimate the coverage of a particular chip from HapMap data, although we note that some care is required to account for SNPs not in HapMap [8]. Here we focus instead on the power of particular chips to detect causal variants of different effect sizes, and the way in which this varies with study size and/or study cost and when using genotype imputation methods. Although coverage is straightforward to estimate, power is a complicated function of the set of SNPs on the chip, effect size, and sample size, and can only be assessed by simulation. It turns out that differences in coverage between chips are often not reflected in substantial differences in power and that the use of genotype imputation further reduces these differences. Study power is routinely used throughout science in experimental design and we argue that it should be the primary consideration in designing GWAs. This approach was used in settling several design questions in the Wellcome Trust Case Control Consortium [5]. Our results have been encapsulated in a user-friendly R package that allows the power of different chip and sample size combinations to be assessed given a total budget for the study. Knowledge of study power is also invaluable when analysing data from a study. Assessment of whether positive results at a particular significance level are “real” or due to chance requires knowledge of power [5], and the practical decision of how far down the list of potential associations one should go in replication studies should be informed by power considerations. Other comparisons of chips have been carried out but have either focussed exclusively on estimating coverage [8], have been limited in scope of which chips have been evaluated [9] or have used analytical calculations that do not properly take into account the complex LD structure of the human genome [10],[11] or failed to assess the impact of imputation correctly [11]. A recent paper [12] has used chip data to assess the performance of the chips but the small sample size (N = 359) means that these results cannot be used to assess power of new study designs of more realistic sizes. In addition, the simulations of quantitative phenotypes used the Signal to Noise Ratio (SNR) to measure effect size of the causal SNP which is non-standard and difficult to interpret. For binary traits, simulations assumed a disease prevalence of 25%, a relative risk of 3 and a sample size of only 75 cases and 75 controls. These parameter settings are not realistic for genome-wide association studies or useful when designing new studies. Study power depends on assumptions about the underlying disease model, in addition to effect sizes and sample sizes. When the true causative SNP is not on the genotyping chip there will typically be several SNPs on the chip which are correlated with it. One or more of these could give a signal of significant association and hence allow detection of the locus. The LD structure of the human genome is sufficiently complicated that this effect cannot be captured analytically. It must be assessed via simulation studies. Nonetheless, there is one very simple situation for which analytical calculation is possible and helpful: that of the simplest disease model in which only a single SNP, correlated with the causal variant, is genotyped. For a design with the same number of cases and controls, under the disease model in which disease risk changes multiplicatively with the number of copies of the risk allele carried by an individual (this model is often referred to as the additive model because risk increases additively on the log scale), there is a known analytical relationship [13]: (1)where χ2 is the chi-squared test statistic, the number of cases and controls, γ the effect size, p the allele frequency of the risk variant and γ2 is the correlation between the marker and causal SNP. Although the real problem is much more complicated than this setting, Equation 1 does provide some useful intuition. Firstly, when the relative effect size is large () the correlation between the marker and causal SNP may only need to be weak (r2≪0.8) for the association to be detected (the expected test statistic is big). Equally, if the relative effect size is small () then even strong or complete association (0.8<r2≤1) may not generate sufficient power to reject the null hypothesis of no association. Assessment by simulation of the power of a particular chip requires simulation of large sets of case and control samples which mimic the LD patterns in human populations (see Figure S1 for an example). The approach we use, implemented in a software package called HAPGEN is conceptually simple and is illustrated in Figure 1. We have previously used this approach to compare different analysis methods and has been briefly describe before [14]. In this paper we provide full details of the approach and these are given in the Methods section. Informally, the required samples are built up from the known haplotypes in HapMap. Consider first the simulation of control samples in a region of the genome. A particular control individual is simulated by separately simulating its two haplotypes in the region. Each of these haplotypes is made up as mosaics of the known haplotypes in HapMap, with the mechanism for constructing these mosaic haplotypes based on population genetics theory. Fine-scale estimates of recombination rates are used to calculate the probability of breaks in the mosaic pattern as one moves along the region. For a given SNP assumed to be causal under a particular disease model and effect sizes, it is straightforward to calculate the genotype frequencies in cases at that SNP. Case samples are simulated separately by first simulating the genotype of the case at the causative SNP and then working outwards in each direction to construct the haplotypes carrying the alleles simulated at the causative SNP. Loosely, this process will result in oversampling of HapMap chromosomes which carry the risk allele, with the effect dropping off as one moves away (in genetic distance) from the causative locus (see Figures S2, S3, and S4 for examples). We apply the method here to an assessment of the power of different chips, but we note that there are many other settings which require simulations of large case-control samples. These include comparisons of analysis methods [14] and tagging approaches [15], assessments of parameter estimates, and design questions for follow-on studies such as resequencing and fine mapping of associated regions. We assessed the power of commercially available chips via simulation. Each simulation assumed a particular SNP in HapMap was causative, with a given effect size and used the HAPGEN package to simulate case and control samples of different sizes. In the simulated data we then restrict attention to the genotypes at only the SNPs on the chip in question and ask whether analysis of these would yield a significant result for any of the SNPs on the chip. An estimate of power is obtained by repeating the simulation over a large number of putative disease SNPs across the genome and using the proportion of simulations in which we find a significant test statistic. For definiteness, in the results presented below we simulate data under the additive disease model, and in analysis of the data consider each SNP separately and apply the so-called trend, or Cochran-Armitage test [16], a chi-squared test with one degree of freedom. We fix a significance level of 5×10−7, and vary the number of cases and controls in the simulated study. There are various other versions of these assumptions which could be made. We explicitly look at one set of multi-marker tests below and also carry-out a limited set of simulations to assess the impact of genotype imputation. There are two somewhat different perspectives that could be adopted regarding genome-wide association studies. One is to regard the GWAS as a self-contained experiment in its own right with the statistical inference being a formal hypothesis test of the null hypothesis of no association. From this perspective, the goal at the conclusion of the GWAS is to decide whether particular SNPs are, or are not, associated with the phenotype of interest. But this is not what happens in practice. There is a strong consensus in the field that the results of association studies should not be relied upon without additional (statistically significant) evidence from analyses in independent replication samples [17] , and many major journals have policies which preclude publication of GWAS studies by themselves, without such replication evidence. Common practice is thus to regard the GWAS as an experiment to highlight SNPs of interest, and then to take as many as possible of the interesting SNPs into replication studies. We adopt the second perspective throughout this paper, and our power calculations are for the probability that for each of the genotyping chips considered, there will be SNPs reaching a prespecified, low, p-value, under specific assumptions about the underlying genetic effects. Given current practice, we believe the right quantity to calculate would be the probability, for the respective chips, effect sizes, and sample sizes, that the experiment would give rise to SNPs showing enough signal to be taken forward for replication. This is (inevitably) ill-posed, so we focus instead on a surrogate for it, namely the probability that at least one SNP will have a p-value below a very stringent threshold. In this context there is nothing special about the choice of p-value threshold, and it is now well understood, for example from meta-analyses, that SNPs well down any ranked list of hits from the GWAS associations can still be genuine associations. For definiteness, we focus throughout on the threshold of p<5×10−7). This is deliberately set so that false positive rates will be low – for example, most SNPs with trend test p-values passing this threshold in GWAS studies, including all of those in the WTCCC experiment, have had associations confirmed in replication studies (see [18] and the NHGRI Catalog of Published Genome-Wide Association Studies at http://www.genome.gov/GWAStudies/). Choice of a different p-value threshold changes the numerical value of the power we calculate, but does not affect the relative performance of the chips, or the relative effect of sample size (data not shown). If one were to adopt the first of the two perspectives on a GWAS study, namely that it is a formal statistical hypothesis test in its own right, then power comparisons become more complicated, at least under a frequentist statistical perspective: for a given nominal per-SNP significance level, the overall GWAS experiment will have somewhat different false positive rates for the different commercial chips, because they have different SNP sets, or when some SNP genotypes are imputed, depending on the number of imputed SNPs, for the same reason. Actually, even for a fixed chip, overall false positive rates will differ depending on the population in which the GWAS is conducted, because of differing patterns of LD between the SNPs on the chip (and hence different effective numbers of independent tests). We do not pursue this approach here, principally because it does not reflect the way GWAS experiments are typically used in practice: regardless of the genotyping chip used, whether or not genotype imputation is employed, and the population studied, researchers tend to focus on the most significant SNPs after the GWAS and try to confirm that they are real in replication studies. In addition, as noted above, overall GWAS false positive rates are low, for any of the commercial chips, at the very low per-SNP significance level we consider. Nonetheless, in what follows, readers should be aware that we are comparing power, defined here as the probability that at least one SNP reaches a fixed p-value threshold under specific assumptions about design and effect sizes, across settings in which these very low false positive rates will differ between chips (and across populations). In calculating power, as thus defined, we simulate data under the assumption that a particular allele is causal and then look to see whether any SNPs on the respective genotyping chip, within a large region around the causal SNP attain the specified significance level. In ignoring the SNPs on the chip elsewhere in the genome, this approximation will underestimate the probability of there being a SNP meeting the significance threshold, but at the very low threshold, the probability of there being a SNP elsewhere in the genome meeting the threshold is extremely small, so that effect of this approximation will be minimal and our power calculations based on only on SNPs within the 1Mb region containing the causal SNP will be very close to the true values. We simulated putative disease loci at SNPs in phase II of the HapMap within twenty-two one megabase regions on each of the autosomes, a total of nearly 50,000 SNPs, which together are typical for the genome in terms of SNP coverage and recombination rates (see Figure S5 and Text S1 for details). We investigated the power afforded by seven different genotyping chips: the 100 k, 500 k and 6.0 chips from Affymetrix (www.affymetrix.com) and the 300 k, 610 k, 650 k and 1 M chips from Illumina (www.illumina.com). These chips sets differ in the way in which the SNPs are chosen and the total number of SNPs assayed. As technology develops and genotyping chips become denser it is a natural question to ask how much power would be gained by genotyping additional SNPs or by using genotyping imputation methods [14]. To facilitate such comparisons we evaluated the performance of a hypothetical chip that contains all the SNPs in HapMap to act as a point of reference in our results. The performance of this ‘complete’ chip is shown as a solid black line in all of the figures showing power. Since the simulations we carry out only use HapMap SNPs as causal SNPs this analysis approximates the scenario in which we have a chip which types all possible SNP variation. We return below to consideration of results for studies in the Yoruban population. Focussing now on the power curves in the top row of Figure 2 several features are evident. The first is the profound effect of sample size. Effect sizes of 1.5 or smaller might be typical of what would now be expected for most variants affecting susceptibility to common human diseases [5]. For effect sizes at the top of this range (1.3–1.5) very large studies (say 2,000–3,000 cases and the same number of controls) are needed to have reasonable power, while for smaller effect sizes even studies of 5000 cases and 5000 controls have very little power. This ties in with growing empirical evidence. For example, for Crohn's disease, the WTCCC study, of 2000 cases and 3000 controls found 9 loci with p<5×10−7, whereas several smaller studies published around the same time each found only one or two of the loci, with little overlap across these smaller studies, consistent with each having modest power for the larger set of loci. Further, recent meta-analyses of 4,539 cases for type 2 diabetes and 3,230 for Crohn's disease have been needed to discover further loci with estimated effect sizes in the range 1.1–1.2. Even for a disease not previously studied by GWA, studies with fewer than 2000 cases and 2000 controls will have low power, except in special circumstances, for example if there are loci with larger effect sizes than has been typical across many other diseases. A second general feature of the power curves for Caucasian studies in Figure 2 is that aside from the Affymetrix 100 K chip (which is no longer available), there are not major differences in power across the other seven chips. For Caucasian samples the chips are typically ordered (with decreasing power): Illumina 1M, Illumina 650 k, Illumina 610 k, Affymetrix 6.0, Illumina 300 k, Affymetrix 500 k, but the absolute difference in power between the best and worst of these chips is often no more than around 10%. Put another way, for effect sizes in the range 1.3–1.5, a study with the Affymetrix 500 K chip would have the same power as one with the Illumina 1 M chip if its sample size were larger by 10–20%, with smaller increases in sample sizes giving studies with other chips the same power. Further, in Caucasian studies, power for all chips other than the Affymetrix 100 K chip is quite close to the best which could be obtained, namely by directly genotyping the causative SNP. Equation 1 makes clear the dependence of power on the frequency of the risk allele. The results in Figure 2 are averaged over putative causative SNPs with a risk allele frequency (RAF) in the range 5–95%. Figure 3 shows that this hides quite different behaviour depending on whether the putative disease SNP is rare or common, and that the conclusions in the preceding subsection apply principally for common causative SNPs. The Figure shows a substantial difference in power for common and rare alleles with the same effect size and that power is minimal for the rare alleles when the effect size is small. These results refer to single-SNP analyses. While there are definitely more powerful analysis methods for rare alleles [14], this is not a major factor in the loss of power, and neither is the incomplete coverage of the SNPs on the commercially available chips: even using a sample size of 3000 cases and controls and genotyping the causal locus directly (black line) is unlikely to lead to a test statistic which will reach the small levels of significance thought appropriate for GWAS. There is an open question as to whether rarer causal alleles might have larger effect sizes than common causal alleles. If this were though plausible, then in assessing power overall for a particular chip, one could focus in Figure 3 on particular ranges of effect sizes for common causative alleles and a different range of effect sizes for rarer causative alleles. It is becoming clear that many loci harbouring common alleles affecting common diseases will have effect sizes in the range 1.1–1.2, and our simulations demonstrate that there is almost no power to detect these in studies of the size currently underway. As has already been shown empirically [19],[20] these loci can be found by meta-analyses and follow up in larger samples of GWA findings. Slightly larger relative risks do become detectable in large samples. For example the power to detect an effect of size 1.3 jumps from almost zero with 1000 cases and 1000 controls to over 50% in a study three times the size. Figure 3 also demonstrates that chip sets differ in the power they offer to detect associations at different frequencies. Most noticeably, when averaged over common alleles the Illumina 300 k chip set offers more power than Affymetrix 500 k. For rare alleles, the opposite is true with the Affymetrix 500 k chip having more power than the Illumina 610 k chip. This is most likely due to the way in which the Illumina SNP sets have been designed to specifically tag the common variation present in the HapMap panels. Immediately apparent is how close, for studies in Caucasian populations, the genotyping chips track the power afforded by the ideal “Complete chip” in a given study design and disease model. Figure 3 illustrates that the potential benefits of increasing SNP density on the chips or from using imputation [14] are greatest for low frequency SNPs. When focusing on common alleles, the potential benefits are greatest for the Affymetrix 100 k and 500 k chips and the Illumina 300 k chip and we show this when specifically consider imputation below (see Table 1). However, a clear consequence of these results is that for any of the chips in current use, increasing sample size is likely to have a bigger effect on power than increasing SNP density. A striking feature of Figures 2 and 3 is that substantial differences in coverage between different chips do not translate into big differences in power. Put another way, coverage is often a poor surrogate for power. As an example, the coverage in the CEU HapMap population (r2≥0.8) provided by the Affymetrix 500 k and Illumina 610 k chips are 65% and 87% respectively, a difference of 22%. On the other hand, the difference in power e.g. for relative risk 1.5 and 1500 cases and controls, is only 7% (66% and 73% respectively). In one sense this shouldn't be surprising. Coverage is measured to a hard threshold: so if SNP has r2 of 0.85 to its best proxy on one chip and 0.75 to its best proxy on another chip, it will be counted as “covered” by one chip but not by the other, whereas the difference in power is small. Coverage statistics also do not depend on study size or disease model. Figure 4 illustrates the differences in correlation structure for two chips. For each HapMap SNP we found it's best “tag” (the SNP on the chip with which it has the highest r2) and generated a histogram of these maximized r2 values. To recover coverage we simply count the proportion of SNPs for which the best tag r2 is ≥0.8, coloured red in the bottom row of figure 4. In this sense, informally, it is useful to think of coverage as assuming that there is power one for every “tagged” SNP and no power for every other SNP. This is of course false, in ways which help to explain why coverage differences do not translate into power differences. When a SNP is common and the effect size is moderate or large, there will still be good power to detect it even if the best SNP on the chip only has r2 = 0.5 or less. At the other extreme, for rare SNPs, unless the effect size is very large, power would be low even if the SNP had a perfect proxy on the chip. Thus even if these SNPs were well covered by one chip and completely missed by another they would not contribute to a difference in power between the chips because both chips would have power close to zero for them. The top row of Figure 4 shows the average power for SNPs in each LD bin. For the Affymetrix 500 K chip, there is a greater contribution to power from the sets of SNPs which are not well “covered”, than for the Illumina chip, and hance a smaller difference in power than in coverage. For several reasons it is of interest to study the power of commercially available chips in different populations. Firstly the Illumina 100 k, 300 k and 610 k chips are aimed at capturing variation in the CEU population, whereas the Affymetrix 500 k chip is not designed with a specific population in mind. Furthermore the Illimina 650 k chip has a subset of SNPs targeted at capturing variation in the HapMap YRI (Yoruba, Africa) population. LD will not extend as far in the YRI collection [1] as in the CEU, reducing the coverage of a given set of SNPs. Figures 2 and 3 show the results of power calculation using the distribution of diversity in both the HapMap CEU and YRI populations. The results show that the increased ancestral recombination leads to a loss of power and coverage across all chips for a range of study designs. The difference between the power available from commercial genotyping chips and that achievable by exhaustively assaying all SNPs shows that increasing marker density may yield a better return than a similar approach in non-African populations. The Illumina 650 k chip, with the YRI fill-in illustrates these potential benefits, showing a marked increase in power over the 610 k. However the performance of the Illumina 300 k chip, designed using the CEU HapMap, falls below the Affymetrix 500 k when genetic diversity is modelled on the YRI HapMap panel. It is not yet clear how closely patterns of diversity and LD in other African populations mimic those in the Yoruba, and hence to what extent the power results will translate to studies in other populations. One general point is that the Illumina 650 k chip was designed specifically to capture common Yoruban variation, so one might expect power for this chip to decrease in other African populations, for which it is not specifically designed. On the other hand, the Affymetrix 500 k chip was not designed using this data, so there would be not a systematic effect changing power estimates for other African populations. As a consequence, differences in power between the Illumina 650 k chip and Affymetrix chips may well be smaller in other African populations. Multi-marker methods, which use combinations of SNPs, have been suggested as an efficient way to increase both coverage and power [15]. Figures S6 and S7 show the results of simulations that implement the multi-marker tests. In these figures the dotted lines, which represent coverage, are higher for all chips in comparison to single marker approaches (Figures 2 and 3) consistent with previous observations. We find that multi-marker approaches also increase statistical power to detect disease loci, but that the increase is modest relative to coverage, and the broad conclusions above are not much affected. Interestingly, when comparing across genotyping platforms, we find for example that the Affymetrix 500 k chip gains more by combining SNPs than the Illumina 300 k chip. Genotype imputation methods [21],[14] are now being widely used in the analysis of genome-wide association studies [5] and meta-analysis of such studies [22],[20]. These methods can be thought of as a more sophisticated version of Multi Marker tests but are relatively much more computationally demanding. We carried out an evaluation of the boost in power that can be gained by imputation using the program IMPUTE [14]. For our simulations with a sample size of 2000 cases and 2000 controls and a relative of the causal SNP or 1.3 we ran IMPUTE on the genotype data from each of the chips under study using the CEU HapMap as the basis for imputation. We then carried out a test of association at all the imputed SNPs in addition to the SNPs on each chip. We used our program SNPTEST to carry out tests of association at imputed SNPs to properly account for the uncertainty that can occur at such SNPs[14]. The results of the simulations are shown in Table 1 and shows that the use of IMPUTE provides a noticeable boost in power over testing just the SNPs on each chip or using Multi Marker tests (as defined in [15]). This agrees with our previous results [14]. It is also very noticeable that imputation reduces the differences in power between the chips and that the use of imputation produces a level of power that is almost as high as our hypothetical ‘complete’ chip. We also note that the boost in power is more substantial than that estimated in another recent study [11]. A close look at the details of this other study shows that the only imputed SNPs used were those (a) which had real genotype data from one of the other chips, and (b) the imputed and real data at the SNP agreed with an r2>0.8. So for example, for the Affy 500 k chip only genotypes at 427,838 imputed SNPs were used, rather than all those available from HapMap (approximately 2.5 milion SNPs), as normal practice when carrying out imputation. Using such a filter clearly creates a bias towards imputed SNPs that are almost perfect tags for SNPs on the chip so it is not surprising that this study shows such small increases in power when using imputation. One option open to researchers who would like to increase power in the context of limited case series is just to increase the control collection. This strategy might include using cases for one disease as extra controls for another (assuming suitably different disease aetiologies and similar population history). We investigated the utility of such an approach by performing simulations with 1000 cases and an increasing number of control (Figure S8). Although the gains are not as strong as increasing both the case and control sample sizes (Figure 2), the ability to reject the null hypothesis of no association increase considerably with the size of the control panel. For example, adding an extra 2000 controls to a case-control study with sample size 1000–1000 increases power to detect an effect of 1.5 typically by 20%. Subject to care in their use, the growing availability of genotyped sets of controls promises to make this a possibility worth investigating for many studies. The results of our simulations can be used to assess the power of a range of possible designs for a given budget and have been encapsulated in a user friendly R package for this purpose (see Software section). Table 2 shows the study size and power that can be achieved on a budget of $2,000,000 for each of the chips assuming the disease causing allele of has a relative risk of 1.5, a risk allele frequency of at least 0.05 and that a p-value threshold of 5×10−7 is used to define power. Since the different chips vary in their prices and their per sample processing costs we obtained quotes from service providers for the various chips and averaged them (see Text S1). The prices were based on quotes for 4000 chips and quotes were converted to US dollars using current exchange rates where necessary. We obtained 5 different quotes for the Affymetrix chips and 6 different quotes for the Illumina chips. The results show that in this scenario the Illumina 300 k chip produces the most powerful design (82.1%) primarily due to its relatively cheap price compared to the other chips. Using the same sample size (2653 cases and controls) the ‘Complete’ chip has a power of 88.1%. It is also notable that the power of thie Illumina 300 k chip is nearly 17% greater than the power that can be achieved by the Illumina1 M chip (63.5%) which has approximately 3 times the SNP density. These result further illustrate the deficiencies in using coverage as a measure of chip performance as sample size is not factored into the calculation. Although these results are interesting we advise against using them directly in the design of a new study. There were noticeable variations in the quotes we obtained from the service providers and prices are likely to change through time. We encourage new studies to re-calculate power of various designs based on a set of up to date and competitive prices and to take into account the general effect that genotype imputation can have on these power estimates. Because of the complexity of human LD patterns, many questions of interest cannot be addressed analytically. We have described in detail our simulation method, HAPGEN, for generating large samples of case and control data at every HapMap SNP, which mimic the patterns of diversity and LD present in the HapMap data. The software can simulate case data under a single causal disease SNP model for specified genotypic relative risks. We have used the method here to assess the power of various commercially available genotyping chips for case-control genome-wide association studies, but note that it could be utilised to assess other design questions, in the evaluation of analytical methods, and in considering follow-on studies such as resequencing and fine-mapping. In Caucasian populations the differences in power afforded by current-generation genotyping chips are not large, and the power of these chips is close to that of an optimal chip which always directly genotyped the causal SNP. Listed in order of decreasing power for the CEU population, averaged over all potential disease SNPs with RAF ≥5%, the chips we considered were: Illumina 1M, Illumina 650 k, Illumina 610 k, Affymetrix 6.0, Illumina 300 k, Affymetrix 500 k and Affymetrix 100 k. In line with our previous work we have shown that imputation can boost the power of each chip substantially and that the resulting power will approach that which could be obtained by a hypothetical ‘complete’ chip that types all the SNPs in HapMap. One limitation of the approach we (and others [9],[10],[12],[11]) have used is that the causal SNP is assumed to be one of those SNPs in the HapMap panel and this will not always be true. Other studies [1] have shown that the majority of SNPs not in HapMap will be highly correlated with the SNPs that are in HapMap and this is especially true for the more common SNPs. This means there is a slight bias in our power results for each chip and for the use of imputation but we do not expect it to be large. A consequence of this point is that the power we estimate for the ‘complete’ chip approximates the power we might obtain if we had a chip which typed all the SNPs that exist in the human genome. A main conclusion from our analysis is that study size is a crucial determinant of the power to detect a causal variant. Increasing study size typically has a larger effect on power than increasing the number or coverage of SNPs on the chip, at least amongst chips currently available. Even for effect sizes at the larger end of those estimated to date for common human diseases (RRs of 1.3–1.5) quite large sample sizes, at least 2000 cases and 2000 controls and ideally more, are needed to give good power to detect the causal variant. When case numbers are limited, there are still non-trivial gains in power available from increasing just the number of controls. Care is needed in assessing the appropriateness of a set of controls, but as larger sets of control genotypes are made publicly available this strategy has considerable appeal, whatever the number of available cases. SNPs with smaller effect sizes are unlikely to be detected even in studies of the sizes currently undertaken, but as has been shown empirically for several diseases, these can be found by meta-analyses which combine different GWAs, or by follow-up in large samples of SNPs which look promising in the original GWA but fail to meet the low levels of significance thought appropriate for GWAS. When the causal SNP is rare (MAF<10%), all chips have low power unless its effect is large and sample sizes are large. This conclusion would hold even if the chip directly genotyped the causal SNP. The relative ordering of different chips, on the basis of power, also changes in this context. As would be expected, power is also lower for all chips for samples which match the patterns of LD seen in the Yoruba HapMap sample, and again the relative ordering of chips changes in this setting. It is not yet clear how well the results for the Yoruba would extend to other African populations. An often-quoted metric in assessing chips is the coverage of each chip: an estimate of the proportion of SNPs which have r2>0.8 with at least one SNP on the chip. Although relatively simple to calculate (and even simpler to miscalculate), not least because it does not depend on study size, our results show that coverage can be a poor surrogate for power, and that relatively large differences between chips in coverage do not translate to large differences in power. The sets of SNPs on Illumina chips are chosen in part to maximize particular criteria, such as coverage, for certain populations, typically those in HapMap. One difficulty of analyses such as those in this paper is that these resources are also the natural ones with which to assess properties of the chips. Thus when Illumina chips “tuned” to one population (say the 610 K chip for CEU) are used in other populations, power might be systematically lower than the levels assessed here. In contrast, SNP sets of Affymetrix chips are chosen largely in a non-population specific way. While power is likely to vary in populations other than those we have considered here, there is not the same systematic effect which would lead to a decrease in power. A quantitative assessment of this phenomena will be possible when dense genotype data is available for other populations, such HapMap Phase 3. We have assumed here that accurate genotypes are available for all SNPs on each chip. In practice some SNPs on each chip will fail QC tests and not be available for analyses. As a consequence, our study will overestimate power, though this effect is unlikely to be large. We are only able to use SNPs in HapMap as potential disease SNPs. These may not be systematically representative of all potential disease SNPs. HapMap SNPs have systematically higher MAFs than do arbitrary SNPs [2], but for SNPs within a particular range of MAF, it seems unlikely that their LD properties will differ systematically, so, for example, we would expect our results for common SNPs to extend beyond those in HapMap. We have focussed on the most common GWA design, namely of a single-stage study, and the simplest disease model. The flexibility of the simulation approach allows many other practical aspects of study design to be incorporated into power calculations. These include more complex disease models, two-stage strategies (the starting point for our work was a comparison of power for one- and two-stage designs in the context of the WTCCC study [5]), genotyping errors, QC filters, misidentification of cases as controls and simple types of population structure. The HAPGEN software also provides a useful tool for the development and comparison of more sophisticated multi-marker approaches to detecting disease association (e.g. imputation [14]). We therefore believe that simulations are an essential tool in the design of association studies by allowing a focus on study power and an assessment of the affect on power of following a given study design. We hope that this method will continue to find use and can be extended to new catalogs of genetic variation such as the 1000 Genomes Project http://www.1000genomes.org/. As in other areas of science, power seems a central consideration in study design and choice of genotyping chip. But other issues may also play a role. These include coverage of particular genes, or genomic regions of interest; the utility of GWA data for directing downstream studies such as resequencing and fine mapping; data quality for particular chips; and the extent to which a chip reliably assays other forms of genetic variation such as copy number polymorphisms. Adding data to existing studies is straightforward if the same chip is used, but the success of imputation methods, in particular in meta-analyses [19],[20] means that this is not essential. In general, Affymetrix chips have more redundancy than do Illumina chips, in the sense of containing sets of SNPs which are correlated with each other. The immediate consequence of this is lower coverage and lower power for the same number of SNPs, but there can be advantages to this redundancy: loss of a particular SNP to QC filters may not be as costly; and signals of association are likely to include more SNPs, thus making them easier to distinguish from genotyping artefacts. Ultimately power can only be calculated under an alternative model. Thus on a practical level the optimal choice of assays and sample sizes will actually depend on the researcher's belief regarding the unknown distribution of effect sizes and models relating genotype and phenotype. In particular we show that one might adopt different strategies depending on the expected frequency of disease causing variant, the effect size and even the population from which cases and controls are sampled (Figure 3). In the continuing search to better understand the genetic basis of common human diseases, numerous study designs can be adopted which may involve combining data sets, imputing missing SNPs [14], distilling signals of association over multiple experimental stages, and so on. In this complex setting study power will remain a central criterion in study design, and the kinds of approaches developed here will continue to allow informed decision making by experimenters. We adopt the model introduced by [23] (denoted LS from now on), who described a new model for linkage disequilibrium, which enjoys many of the advantages of coalescent-based methods (e.g. it directly relates LD patterns to the underlying recombination rate) while remaining computationally tractable for huge genomic regions, up to entire chromosomes. Their model relates the distribution of sampled haplotypes to the underlying recombination rate, by exploiting the identity (2)where h1,…,hn denote the n sampled haplotypes, and ρ denotes the recombination parameter (which may be a vector of parameters if the recombination rate is allowed to vary along the region). This identity expresses the unknown probability distribution on the left as a product of conditional distributions on the right. LS substitute an approximation for these conditional distributions into the right hand side of (3), to obtain an approximation to the distribution of the haplotypes h given ρ (3)If h1,…,hn are n sampled haplotypes typed at S bi-allelic loci (SNPs) LS modelled the distribution of the first haplotype as independent of ρ, i.e. all 2S possible haplotypes are equally likely, so . For the conditional distribution of hk+1 given h1,…,hk, LS modelled hk+1 as an imperfect mosaic of h1,…,hk through the use of a Hidden Markov Model (HMM). That is, at each SNP, hk+1 is a (possibly imperfect) copy of one of h1,…,hk at that position where where the transition rates between the hidden copying states are parameterized in terms of the underlying recombination rate. The transition rates are different for each of the conditional distributions in such a way so as to mimic the property that as we condition on an increasingly larger number of haplotypes we expect to see fewer novel recombinant haplotypes. A parameterisation for the mutation rate (or emission probabilities of the HMM) is used that has similar properties (see [23] for more details). The simulation of a new set of haplotypes for control and case individuals is proceeds using the following algorithm. 1. Pick a locus from the set of markers in the real dataset as the disease locus. The disease locus is chosen at random from all those loci with a minor allele frequency (MAF) within some specified range [l,u]. We use to denote the disease locus, a and A to denote the major and minor alleles at the disease locus and use p denote the sample minor allele frequency at this locus. 2. For a given disease model simulate the alleles at the disease locus of the new individual conditional upon case-control status. At the disease locus we use a general genotype model in which the frequencies of the genotypes aa, Aa and AA in control individuals are given by (1−p)2, 2p(1−p) and p2 respectively. This assumes that the control individuals are so-called population controls (as used by the WTCCC study [5]) rather than individuals who have been selected to specifically not have the disease. For case individuals the genotype frequencies are determined by specification of the two relative risks (4)where denotes the probability that an individual is a case conditional upon having genotype g. Under this model (5)where γ = (1−p)2+2αp(1−p)+βp2. As an example, if p = 0.1, α = 2 and β = 4 the control and case genotype frequencies are (0.81, 0.18, 0.01) and (0.67, 0.30, 0.03) respectively. Assuming we have a set of k known haplotypes, the generation of a case (control) starts by simulating a genotype g using the case (control) genotype frequencies. This simulated genotype specifies the alleles on the the two haplotypes of the new individual at the disease locus. For example, if g = Aa then hk+1,d = 1 and hk+2,d = 0. 3. This step involves the simulation of two new haplotypes for the individual conditional upon the alleles simulated at the disease locus in Step 2 and conditional upon the fine-scale recombination map across the region. This involves simulating the rest of hk+1 and hk+2. We only describe the generation of sites right flanking of the disease locus as the generation of the left flanking markers is virtually identical. Also the simulation of hk+2 follows directly from our description of how the rest of hk+1 is simulated. Let Xj be the hidden state of the HMM that denotes which haplotype hk+1 copies at site j (so that ). This state variable is initialized at the disease locus as follows (6)The value of , as with LS, is Watterson's point estimate (Watterson, 1975) (7)Simulation of the hidden state of the HMM then proceeds using the following transition rule (8)where zj is the physical distance between markers j and j+1 (assumed known); and , where is the effective (diploid) population size, and cj is the average rate of crossover per unit physical distance, per meiosis, between sites j and j+1 (so that cjzj is the genetic distance between sites j and j+1). This transition matrix captures the idea that, if sites j and j+1 are a small genetic distance apart (i.e. cjzj is small) then they are highly likely to copy the same chromosome (i.e. Xj+1 = Xj). To mimic the effects of mutation the copying process may be imperfect: with probability k/(k+θ) the copy is exact, while with probability θ/(k+θ) a mutation will be applied to the copied haplotype. Specifically, 4. Return to step 2 to generate another individual or terminate. Illustrations of the HAPGEN method in practice and details of the testing the method against coalecent simulations are given in Text S1. We used release 21 of the HapMap data for which phased haplotypes are available in NCBI b35 coordinates. The SNPs that occur on each genotyping chip were obtained from the websites of Affymetrix and Illumina respectively. Some of the SNPs in these sets do not occur in the HapMap phased haplotype data due to QC measures applied to the raw genotype data. For the Affymetrix 6.0 and Illumina 1 M chips 90.8% and 88.1% of the SNPs on these chips respectively are in this release in HapMap. This will have the effect of making our estimates of power slight underestimates of the true power. We simulated data for twenty-two one megabase regions chosen at random, one from each autosome. To ensure that the regions used to approximate genome-wide power were representative of the genome at large we their SNP density. Figure S5 plots the distribution of inter-SNP distances within the 22 analysis regions and across the whole genome for three of the genotyping chips analyzed. The close match between the distribution, both on the physical scale and in terms of genetic distance suggests that our results are insensitive to the regions we chose to simulate, and can be used to make comparisons of genotyping chips genome-wide. We used data from the HapMap project Phase II to estimate coverage. Single marker coverage was defined to be the proportion of all variation (with minor allele frequency greater than 5%) in r2 with a SNP on the genotyping chip above 0.8. Using this definition we achieved very similar estimates to previous studies which used the whole genome (we use twenty two representative megabases). Multi-marker coverage was calculated by an aggressive search of all 2-SNP and 3 SNP haplotypes within 250kb of the SNP being tagged [6]. The SNP was tagged if any of these multi-marker tags had r2 above 0.8, the rule defining the haplotype was also stored and added to the list of multi-marker tests. Single marker tests (Cochran-Armitage test) were performed at each SNP on the genotyping chip where information were simulated from the relevant HapMap panel. Multi-marker tests of association were performed in an identical fashion with the marker being formed by the multi-marker haplotypes known to tag HapMap variation. To avoid over estimation of power, multi-marker tags chosen to tag the current putative disease SNP in the simulations were excluded from the test set. Tests at imputed SNPs took account of the uncertainty in genotypes through a missing data likelihood as described in [14]. The HAPGEN software is freely available for academic use from the website http://www.stats.ox.ac.uk/̃marchini/software/gwas/gwas.html. In addition, the results of the power calculations for the 7 commercially available genotyping chips have been included in an R package called GWASpower available from http://www.stats.ox.ac.uk/̃marchini/#software. This package allows the user to determine the most powerful study design for a given budget. As new commercial genotyping chips become available we will update the package to include results of new chips. The package works by fitting a Generalised Linear Model to the results of the simulation study and using the model fit to predict the power for a given number of cases and controls.

10.1371/journal.pgen.1005985

Intermittent Stem Cell Cycling Balances Self-Renewal and Senescence of the C. elegans Germ Line

Self-renewing organs often experience a decline in function in the course of aging. It is unclear whether chronological age or external factors control this decline, or whether it is driven by stem cell self-renewal—for example, because cycling cells exhaust their replicative capacity and become senescent. Here we assay the relationship between stem cell cycling and senescence in the Caenorhabditis elegans reproductive system, defining this senescence as the progressive decline in “reproductive capacity,” i.e. in the number of progeny that can be produced until cessation of reproduction. We show that stem cell cycling diminishes remaining reproductive capacity, at least in part through the DNA damage response. Paradoxically, gonads kept under conditions that preclude reproduction keep cycling and producing cells that undergo apoptosis or are laid as unfertilized gametes, thus squandering reproductive capacity. We show that continued activity is in fact beneficial inasmuch as gonads that are active when reproduction is initiated have more sustained early progeny production. Intriguingly, continued cycling is intermittent—gonads switch between active and dormant states—and in all likelihood stochastic. Other organs face tradeoffs whereby stem cell cycling has the beneficial effect of providing freshly-differentiated cells and the detrimental effect of increasing the likelihood of cancer or senescence; stochastic stem cell cycling may allow for a subset of cells to preserve proliferative potential in old age, which may implement a strategy to deal with uncertainty as to the total amount of proliferation to be undergone over an organism’s lifespan.

Stem cell cycling is expected to be beneficial because it helps delay aging, by ensuring organ self-renewal. Yet stem cell cycling is best used sparingly: cycling likely causes mutation accumulation—increasing the likelihood of cancer—and may eventually cause stem cells to senesce and thus stop contributing to organ self renewal. It is unknown how self-renewing organs make tradeoffs between benefits and drawbacks of stem cell cycling. Here we use the C. elegans reproductive system as a model organ. We characterize benefits and drawbacks of stem cell cycling—which are keeping worms primed for reproduction, and reducing the number of future progeny worms may bear, respectively. We show that, under specific conditions of reproductive inactivity, stem cells switch back and forth between active and dormant states; the timing of these switches, whose genetic control we start delineating, appears random. This randomness may help explain why populations of aging, reproductively-inactive worms experience an increase in the variability of their reproductive capacity. Stochastic stem cell cycling may underlie tradeoffs between self-renewal and senescence in other organs.

An important goal of aging research is not just to extend lifespan—which in C. elegans can be simply achieved by a pause in developmental and reproductive activities in the “dauer” state [1]—but to do so in a way that increases “healthspan” without diminishing organ activity. To this end, it is critical to understand whether aging is driven by organ activity or whether it is a simple function of chronological age [2]. The C. elegans gonad provides a powerful model system to address this question. Previous studies have identified mechanisms by which the “reproductive lifespan”—the period of adulthood over which C. elegans hermaphrodites can bear progeny—can be extended (e.g. [3]). But this extension does not increase the brood size, which is in fact substantially reduced. This suggests a tradeoff between reproductive lifespan and brood size, compatible with reproductive senescence being driven by reproductive activity (see also [4]). That reproductive senescence is driven by reproductive activity is however contradicted by a report that aging individuals lose “reproductive capacity”—the maximum brood size an individual is capable of producing from a given point in time until cessation of reproduction—as a function of chronological age rather than reproductive activity [2]. Here we resolve this apparent contradiction by showing that the loss in reproductive capacity—a phenomenon we refer to as “reproductive senescence” because it mimics the loss of function in other self-renewing organs—is driven chiefly not by increasing chronological age, but by activity of the gonad, and in particular by germline stem cell cycling. To ask whether reproductive senescence is a simple function of chronological age, or whether it is driven by reproductive activity itself, it is useful to manipulate that reproductive activity (e.g. [2]). There are two naturally-occurring C. elegans sexes: males and hermaphrodites. Hermaphrodites can either self-fertilize (abbreviated as “self” below) with the ~300 stored self-sperm they produce during development, or be cross-fertilized with male sperm transferred during mating, which allows brood sizes of up to 1,200 [5]. Brood size of mated hermaphrodites is limited by senescence of the reproductive system, which ultimately stops producing fertilizable oocytes [3]. Reproductive activity can be modulated in a physiological way by controlled mating of hermaphrodites that are feminized—i.e. turned into “females”—by mutation of genes such as fog-1 or fog-2 [6–8]. These females do not produce self-sperm, but form an otherwise fully-developed reproductive system in which oocyte maturation and growth by cytoplasmic streaming are substantially reduced [9,10]. Females can bear progeny only after mating with males, whose sperm trigger oocyte maturation and fertilization. Virgin fog-2 females were shown to undergo reproductive senescence at roughly the same rate as reproductively-active hermaphrodites [2] despite the fact that the ovulation rate of fog-2 females is much reduced [9], suggesting a time-intrinsic senescence mechanism. But the mitotic zone of the germ line in feminized worms was subsequently shown to possess M-phase cells [11], suggesting that stem cells keep actively cycling even in virgin females. Reproductive senescence could therefore be driven by activity rather than being a function of chronological age. Here, we use genetic, environmental, and pharmacological manipulations to assay the relationship between gonad activity and senescence. We establish a causal relationship between the two. We further characterize germ cell cycle behavior on a gonad by gonad basis, using a new technique we developed, and find intermittent activity. Our results strongly suggest that switching between active and inactive states is stochastic. To begin identifying causes of reproductive senescence, we followed a two-fold approach. First, we characterized loss of reproductive capacity over time in various genetic backgrounds known to differ in proximal gonad activity. Second, we asked whether high gonad activity early in life diminishes remaining reproductive capacity. To modulate proximal gonad activity, we selected fog-1 and fog-2 females described above, which have low ovulation rates [9,10]. We used for comparison inx-22; fog-2 females and spe-8 hermaphrodites—both of which are also sterile (unless mated); loss of inx-22 results in precocious oocyte maturation in feminized gonads even in the absence of the sperm signal [9,10,12], while loss of spe-8 preserves stimulation of oocyte maturation by self-sperm that are incapable of fertilizing the oocytes [13,14]. Virgin spe-8 and inx-22; fog-2 females ovulate at a rate close to wild-type (1.7/h, 0.9/h, and 2.2/h, respectively [9,14]), substantially higher than that for fog-2 (0.2/h; n = 31), which is in turn significantly higher than that for fog-1 (0.1/h; n = 35; p < 0.04). We mated virgins at either day 0, 1, 2, 3, 7 or 10 of adulthood and assayed total brood size (Fig 1A). We found that spe-8 and inx-22; fog-2 undergo faster reproductive senescence than either fog-1 or fog-2. A two-way analysis of variance considering age of mating and genotype identified a significant main effect of genotype on reproductive capacity, as well as a significant interaction effect (S1A Table). For example, at day 2 of adulthood (S1 Fig) post-hoc analysis showed all pairwise differences to be significant except for the inx-22; fog-2 / spe-8 pair (S1B Table). Thus, across the genotypes that we studied, the ranking of reproductive senescence rates from fastest to slowest is: inx-22; fog-2 = spe-8 > fog-2 > fog-1. Therefore, gonad activity as measured by oocyte production correlates positively with the rate of reproductive senescence. To ascertain whether early gonad activity diminishes remaining reproductive capacity, we compared the reproductive capacities of three groups of fog-2 females at day 7 of adulthood (Fig 1B). The first group was mated at the onset of adulthood, which causes more active germ cell cycling (see below for detailed cell cycle analysis). To verify that females in the first group did not run out of sperm late in life, a second group was mated both at the onset of adulthood and again at day 5, which did not lead to a significant increase in brood size (S1C Table; see also [2]). The third group was only mated at day 5 and had more than 3 times as many progeny after mating as the first or second group did over the same period (S1C Table). Therefore, consistent with [4], increased germline activity caused by mating is associated with hastened reproductive senescence. We next asked whether strong inhibition of germline activity using conditions likely to be encountered in the wild also led to a delay in reproductive senescence. We starved fog-2 females from the last larval stage (L4) for two days (Fig 2A). Despite being starved and experiencing a ~30-fold drop in germline mitotic index (S2A Table), females progressed from L4 to the adult stage and produced fully-formed oocytes. We did not observe shrinking of the germ line or the reproductive diapause identified in wild-type by [15] (Fig 2B), likely because the absence of sperm in females prevents the redirection of resources to slowly-growing embryos proposed by [16]; consistent with this, starvation for 2 days markedly reduced the number of apoptotic cells identified in females using a feminized ced-1::gfp reporter strain, from 18.3 per gonadal arm to just 1.9 (S2B Table and S2 Fig; note that this is in contrast to starvation of hermaphrodites over a 6-hour period [17]). We then returned the females to food for 24 h, mated them with males, and assayed their brood sizes. Compared to continuously-fed controls of the same age, brood size was increased by almost 2-fold (S2C Table and Fig 2C). This increase occurred even as apoptosis is restored to normal levels when females are returned to food (S2D Table and S2 Fig). Starved worms thus retain a higher percentage of their reproductive capacity during the starvation period compared to controls (Fig 2D). While starvation has pleiotropic effects such as increased stress resistance [18], these results are also compatible with the idea that germline activity—measured either by germ cell cycling or oocyte production—drives reproductive senescence. To ask directly if reproductive senescence is driven by germline activity, and to distinguish between the influence of cell cycling distally and oocyte maturation proximally, we next performed cell cycle inhibition experiments (Fig 3A). We fed the small molecule hydroxyurea (HU), a specific inhibitor of DNA synthesis (germ cells are the only mitotically active cells in adults). We found that HU treatment substantially reduces the incidence of M-phase cells within 12 h (from 3.2 to 0.35; n = 18 and 17 respectively) and eliminates it after 1 day (n = 20). We tested whether germ cell cycle arrest was accompanied by a reduction in ovulation rate by exposing virgin females to HU for 24 h at day 1 of adulthood and counting the number of oocytes laid during this period. We found that for both fog-2, which has a low ovulation rate, and inx-22; fog-2, which has a much higher ovulation rate, there was no effect of HU treatment on ovulation (S3A Table and Fig 3B). We compared reproductive capacities of females that prior to mating received a 24 h treatment with HU or a control treatment without HU. HU treatment increased fog-2 reproductive capacity by 27% (S3B Table and Fig 3C) and increased inx-22; fog-2 reproductive capacity by 50% (S3B Table and Fig 3C). Cell-cycle arrested females thus retain a higher percentage of their reproductive capacity compared to controls (Fig 3D). HU treatment has a number of effects other than cell cycle arrest. While these effects are detrimental [19,20] and are expected if anything to hasten senescence, the possibility remained that HU had a hormetic effect. To test this idea we asked whether HU treatment increased lifespan or thermotolerance, but found that it in fact decreased them (S3C Table and S3A and S3B Fig). While this does not formally exclude the possibility of a hormetic effect on reproductive lifespan, it shows that HU does not have a global beneficial effect on worm health. Overall, the increase in brood size that results from the HU treatment is thus remarkable. To confirm that delayed reproductive senescence induced by HU is not an off-target effect, we also performed a cell cycle inhibition experiment using the selective cyclin-dependent kinase inhibitor (CDKI) Roscovitine. Roscovitine has previously been used to reversibly inhibit mitosis in starfish and sea urchin embryos [21]; we found that it effectively reduces the mitotic index in worm germ lines (S3D Table and Fig 3E). We found that the reproductive capacity of females treated with Roscovitine for 24 h prior to mating was over 2-fold larger than that of females that received a DMSO control treatment (S3E Table and Fig 3F). While HU- and Roscovitine-treatment brood sizes cannot be compared directly, because treatment with DMSO—used as a Roscovitine solvent—decreases brood size [22], the qualitative effects of HU and Roscovitine treatments are the same. The role that we uncovered for cell cycling in driving reproductive senescence, combined with the role of DNA damage in driving senescence in other systems [23], led us to wonder if cell cycling could curtail reproductive output at least in part through the DNA damage response (DDR). To test whether cell cycling leads to increased DDR, we focused on the single stranded DNA binding Replication Protein A (RPA-1 in worms). While RPA-1 is activated in response to multiple forms of DNA damage [24–26], it plays in particular a role in the repair of induced double strand breaks (DSBs) associated with meiotic crossovers. This role leads to presence of RPA-1 foci in the pachytene stage of meiosis [27,28]. Defects in chromosome segregation in the mitotic zone do not lead to increased prevalence of foci in the mitotic zone itself, but instead lead to accumulation in late pachytene of meiosis-induced DSBs that are not resolved [29]. We quantified RPA-1::YFP foci in pachytene (zones 5 and 6 defined by [28]), in young rpa-1::yfp hermaphrodites taken at day 1 of adulthood, and in rpa-1::yfp hermaphrodites taken at day 4 of adulthood that had been selfed or that had been mated at day 0 of adulthood. We found over ~5-fold more foci per nucleus in the mated group than in either of the selfed groups (S4A Table and Fig 4A). Consistent with the pattern of increased cycling being associated with increased RPA-1 foci, selfed hermaphrodites had more foci at day 4 than at day 1 (S4A Table). Overall, although it remains to be established how increased cycling may lead to chromosome segregation defects or other defects in the mitotic zone, these results strongly suggest that this increased germ cell cycling leads to increased proximal germline DDR. Our results show that gonad activity strongly contributes to reproductive senescence, mostly as a result of germ cell cycling. We thus hypothesized that, at any given point in the reproductive life of a worm, remaining reproductive capacity is inversely related to the total amount of germ cell cycling that has occurred up to that point. To test this hypothesis, we decided to ask whether a dose-dependent relationship exists between average cell cycle rate and the rate of reproductive senescence by using genetic or environmental mutations that modulate cell cycling. To identify suitable genetic manipulations, we compared fog-1 and fog-2 females with inx-22; fog-2 and spe-8. Since the former two strains undergo slower reproductive senescence than the latter (see above; Fig 1A) and have lower proximal gonad activity, we surmised that they might also have slower distal cycling. We also assayed for changes in cell cycle when oogenesis rates are high due to the presence of self or male sperm, or low due to depletion of self sperm. We first detail our cell cycle analysis, and subsequently compare results to reproductive senescence data. We first attempted to determine average cell cycle speeds. To this end we carried out pulse-chase experiments using bacterial food labeled with 5-Ethynyl-2’-deoxyuridine (EdU), which is incorporated by cells in S phase. Although virtually all young, selfed wild-type mitotic zones contained EdU-positive cells following a 30-minute pulse (consistent with previous reports; [32,33]), many mitotic zones from virgin females contained no labeled cells. If this lack of S-phase labeling was due to all cells in a given gonad being found by chance in G1-, G2-, or M-phase, given that M-phase length is about 10% that of G1, G2, and M combined (e.g. [33,34]), and mitotic zones contain ~260 cells (S5E Table), one would expect to find on average in such mitotic zones 10% of 260 cells, i.e. 26 cells, in M-phase. Since neither we nor others [35] have observed mitotic zones with such a large number of M-phase cells, lack of any S-phase cell in a mitotic zone indicates that the mitotic zone as a whole is in a dormant state. The two most likely explanations for the presence of unlabeled mitotic zones were thus either that the unlabeled worms had not ingested the EdU-labeled bacteria, or that the mitotic zones were dormant at the time of the pulse. To explore these possibilities, we switched to a continuous EdU labeling assay. In wild-type gonads at day 1 of adulthood, virtually all mitotic zones had at least one labeled cell within the first time point we assayed, with only 5% remaining unlabeled (1 h labeling; see example images in Fig 5A and detailed graphs in Fig 5B). The proportion of unlabeled mitotic zones at 1 h was substantially higher for fog-1 (55%) and for fog-2 (17%) than for wild-type; it took in excess of 6 h for labeled cells to have appeared in all feminized mitotic zones (Fig 5A and 5B). To test whether labeling delays could be due to pauses in feeding, we fed virgin worms bacterial food mixed with fluorescent beads. At the very first time point we assayed (1 h), 100% of wild-type, fog-1, and fog-2 worms had ingested the fluorescent beads (see example in S4A Fig). This strongly suggests that different labeling rates are due to bona fide differences in mitotic zone cycling, with a substantial proportion of fog-1 and fog-2 gonads being in a dormant state in which germ cells are not progressing through the cell cycle. Continuous EdU labeling results show the existence of at least two states in which reproductively-inactive gonads can reside—an actively-cycling state, and a dormant state—and suggest that stochastic switching occurs between these states. Specifically, that different cycling states exist within the population of feminized, reproductively-inactive gonads can be inferred from the large differences in labeling times for e.g. fog-2 gonads (the same reasoning applies to fog-1 gonads): some gonads label almost immediately, showing that they are actively progressing through S-phase at the time of label application, while others take up to 8 h and thus do not have cells actively progressing through S-phase during these 8 h. If fog-1 and fog-2 simply cycled continuously, in the same way as wild-type but with uniform slowing down of all cell cycle phases, the distribution of cell cycle phase indices would be unchanged and the number of gonads with at least one cell progressing through S-phase would be the same as in wild-type at any given time—and the continuous labeling time courses would therefore appear identical for all genotypes. Stochasticity can be inferred from the fact that EdU labeling of reproductively-inactive gonads happens immediately for some gonads but takes up to 8 h for others: this shows independent behavior of gonads with identical genotypes, for which stochastic switching between active and dormant states is the most parsimonious explanation. To begin confirming the existence of such states by independent means, we quantified the coefficient of variation (CV) in mitotic index. The CV is defined as the ratio of standard deviation to mean and thus provides a unit-free measure of noise. The fog-1 and fog-2 CVs were 2.3- and 1.7-fold higher, respectively, than the wild-type CV, differences that are significant at a 95% confidence level (S5A and S5B Table). These differences show a larger spread among the population of the mitotic index measured at a given point in time, as expected if there is a mixture of gonads with little or no mitosis (the dormant subpopulation) and gonads cycling as normal (the active subpopulation). This supports the idea that feminized gonads switch back and forth between active and dormant states. This idea is further tested by independent means below. We asked whether the dormant state we identified was an artificial byproduct of mutations in fog-1 or fog-2, or whether it was a state naturally occupied by sperm-deprived gonads. We first performed continuous EdU labeling of selfed wild-type hermaphrodites taken at day 3 of adulthood, by which time reproductive activity has started declining; full labeling took 4 h (S4B Fig), twice as long as for hermaphrodites at day 1 (which are fully active). Mitotic zones from mated hermaphrodites assayed at day 3 of adulthood, however, labeled at the same speed as selfed worms at day 1 (S4B Fig). We next performed continuous EdU labeling using inx-22; fog-2 or spe-8 gonads, which do not produce embryos but retain sustained oocyte maturation. The fractions of dormant inx-22; fog-2 and spe-8 mitotic zones were closer to that of wild-type (Fig 5B) than to that of fog-1 or fog-2. Similarly, inx-22; fog-2 and spe-8 times to labeling of all mitotic zones were closer to that of wild-type than to that of fog-1 or fog-2 (Fig 5B). The differences in median labeling times were significant for all genotype pairs (p < 7.8E-3, Wilcoxon test with Bonferroni correction for 15 tests; see Methods), except for pairs taken within the group formed by wild-type, spe-8, and inx-22; fog-2, which all behave similarly (p > 0.05). Differences in fractions of dormant gonads matched differences in mitotic index CV: the CV for fog-1 was significantly higher than that for fog-2, which in turn was significantly higher than the CV for wild-type, inx-22; fog-2 or spe-8 (S5B Table). Finally, we asked whether the behavior we observed was particular to C. elegans, or whether it was shared with other nematode species. We chose C. remanei [36], for which genetic manipulation is unnecessary for feminization because it is a male/female species, but which undergoes reproductive senescence much like C. elegans females (S4C Fig). We found that the fraction of dormant mitotic zones was 27% for virgin C. remanei females, but that mating decreased that fraction to 5% (S4D Fig). Therefore, the mitotic zones are dormant at a substantially higher frequency in reproductively-inactive gonads than in reproductively-active gonads, both in C. elegans and in the related nematode species C. remanei. Having established that less active gonads experience periods of cell cycle dormancy, we returned to our pulse-chase dataset. To quantify cell cycle rates of active mitotic zones, and to verify that mitotic zones switch back to the dormant state from the active state, we developed a method that, instead of relying on population averages, computes cell cycle progression of individual mitotic zones. We segmented individual cells in confocal stacks of mitotic zones that were fixed and stained for DNA and EdU, and fit the data to computational simulations of germ cell cycling (see S1 Text). Overall progression of mitotic zones through the cell cycle is computed as a phase that defines a position on a circle, with one full revolution corresponding to a complete cell cycle (Fig 6A). A mitotic zone that labels during the EdU pulse but becomes dormant during the chase would stop progressing around the circle (Fig 6B). We validated our technique using wild-type young adult mitotic zones, for which a full revolution took ~5.5 h to complete (Fig 6C); this estimate of cell cycle length is consistent with previous findings [33,37]. We then compared initial cell cycle progression during the EdU chase between hermaphrodites and virgin females. Initial progression rates (i.e. rates of non-dormant mitotic zones) of wild-type, fog-1, fog-2, inx-22; fog-2 or spe-8 were largely similar (S5C Table). We estimated average cycling rates as the product of the fraction of active mitotic zones and their initial progression rates. Rankings are: wild type (0.17 cycles / h) ≃ spe-8 (0.20 cycles / h) ≃ inx-22; fog-2 (0.18 cycles / h) > fog-2 (0.13 cycles / h) > fog-1 (0.08 cycles / h), which are equivalent to the rankings for reproductive senescence rates. We do not know the molecular basis for the difference in average fog-1 and fog-2 cell cycle rates—slower oocyte maturation in fog-1 (S4E Fig) could increase mitotic zone dormancy—but in any case this difference provides a useful tool to investigate the effects of cell cycling activity. To ask if mitotic zones switch from the active to the dormant state, similar to switching from the dormant to the active state shown by continuous EdU labeling experiments, we further assayed the cell cycle progression of mitotic zones pulsed with EdU. Individual mitotic zones of young adult hermaphrodites pulsed with EdU showed little dispersion in their overall cell cycle progression, even after 6 h of chase—covering over a full cycle length (as shown by visual inspection of phase plots and by quantification of inter-gonad synchrony depicted as wedge width in Fig 6C). This shows that reproductively-active gonads have mitotic zones that progress through the cell cycle in a highly-similar fashion. By contrast, individual mitotic zones from virgin, feminized gonads showed substantially larger dispersion along the set of possible phases as shown by increased wedge width in Fig 6C, 6 h time point, for virgin fog-1 and virgin fog-2 (compare to wild-type day 1 or to mated wild-type day 3). Dispersion was even stronger in wild-type hermaphrodites at day 3—to the point that cell cycle progression appeared largely randomized. Given that initial progression rates are highly similar across the genotypes and conditions that we tested, the simplest interpretation of these results is that the time at which labeled mitotic zones return to the dormant state is stochastically distributed. To test whether dispersion in the amounts of cell cycle progression after labeling was due to reduced reproductive activity, we measured cell cycle activity in mated young females, and in hermaphrodites mated at day 1 of adulthood and assayed at day 3. In both cases, the increase in reproductive activity caused by mating was accompanied by a switch of cell cycle behavior to that of young adult hermaphrodites (Fig 6C) and by a reduction in the mitotic index CV (S5B Table), showing that occupancy of the dormant state is regulated by reproductive activity. Therefore, gonads with reduced reproductive activity switch back from the actively-cycling state to the dormant state in a way that is in all likelihood stochastic. A prediction from a model whereby feminized gonads switch back and forth stochastically between active and dormant states, and whereby gonad activity leads to loss of reproductive capacity, is that there may be an increase in variability of remaining reproductive capacity as the population ages (Fig 7A). This is because stochasticity of switching may lead to mitotic zones spending different total amounts of time in the active state. We thus tested whether inter-individual variability in remaining reproductive capacity does increase with age. To this end, we computed the coefficients of variation (CVs) of brood size in fog-1 populations mated at different times (Fig 7B). The CV is defined as the ratio of standard deviation to mean and thus provides a unit-free measure of noise. When mated at the onset of adulthood (day 0), the brood size CV was 0.12 (Fig 7C and S6 Table). By contrast, the brood size CV for mating on day 2 of adulthood was 0.33, i.e. ~3 fold higher. To distinguish between a simple increase in variability associated with aging (as is observed for many phenotypes; [38]) and an increase in variability driven by stochastic cell cycling, we considered brood sizes computed from day 2 of adulthood onwards, for worms mated at the onset of adulthood. The CV was 0.13 (Fig 7C and S6 Table), virtually identical to the CV of brood sizes scored from day 0 and significantly lower than the CV for worms mated on day 2 of adulthood. Therefore, the age-dependent increase in variability of remaining reproductive capacity for reproductively-inactive females is stronger than the increase that occurs for reproductively-active worms. We verified using a simple simulation that, given a suitable probability distribution of times spent in the dormant and in the active state, a population of females losing the same reproductive capacity as fog-1 over a period of two days can indeed experience a ~3-fold increase in the CV in remaining reproductive capacity (see S1 Text). Overall, these results support the idea that stochastic bursts of cycling drive reproductive senescence. To start querying organismal regulation of the dormant state, we first asked whether the state of a mitotic zone in one gonadal arm relates to that of the mitotic zone in the sister arm from the same worm. At 1 h of continuous EdU labeling, we found 92% agreement in EdU status (defined by presence or absence of at least one labeled cell; n = 36 pairs) between pairs of mitotic zones from the same worm; this substantial synchrony in gonadal arm states suggests that dormancy may be regulated at the organismal level, which could perhaps occur through neuronal control of TGF-beta signaling [39] or insulin signaling [40,41]. Since population density affects a number of aspects of worm physiology [42–44], we next asked whether that density also affects mitotic zone dormancy. Starting from the L2 stage, we kept populations of virgin fog-1 females either singled (low density) or as group of 70 individuals (high density) on 35 mm plates, and mated them at day 4 of adulthood to assay reproductive capacity. We found that females that had been kept at low density had a significantly reduced brood size compared to those that had been kept at high density (Fig 8A and S7A Table). We then assayed mitotic zone dormancy at day 1 of adulthood, and found a significant effect of population density on that dormancy (Fig 8B and S7B Table): 23% of singled female mitotic zones were dormant vs 46% for the higher-density population. Females thus undergo faster reproductive senescence and cycle more actively when they are singled. To ask whether hermaphrodite dormancy is also density-dependent, we performed the same experiment on selfed hermaphrodites assayed at day 3 of adulthood. We observed a similar effect as for females, with 9% dormancy for singled hermaphrodites vs 27% for the higher-density population (S7C Table). As a first step in elucidating the mechanism underlying this population-density dependence, we repeated the experiment using the daf-22(m130) mutation, which abrogates dauer pheromone synthesis [45], and found that density dependence of mitotic zone dormancy was lost (42% vs 47% for high and low densities, respectively; Fig 8B and S7D Table). Mitotic zone dormancy is therefore regulated by population density, most likely through a mechanism that involves dauer pheromone. Finally, we asked whether intermittent cycling is a behavior that is always associated with reduced overall reproductive activity, or whether it is specific to a state in which worms are well fed but deprived of sperm. Caloric restriction strongly reduces the rate of reproduction as well as total brood size [2,46]. In our hands, hermaphrodites kept in liquid culture with 1x1010 bacteria/mL (“high” concentration) had a similar brood size to hermaphrodites kept on solid medium under standard conditions (294 vs. 306; n = 11 and 12, respectively; p > 0.4), but brood size dropped more than two-fold (138; n = 11; p < 1.5E-6) at 1x109 bacteria/mL (“medium” concentration). Further, [46] showed a ~4-fold drop between 1x109 bacteria/mL and 1x108 bacteria/mL (“low” concentration). At day 1 of adulthood, we transferred wild-type adult hermaphrodites to low, medium or high E. coli concentrations for 24 h, and subsequently performed 1 h continuous EdU labeling. In contrast to mitotic zones from females or older hermaphrodites, those from food-restricted young hermaphrodites were all active (n = 20 each), even at the lowest food concentration. This suggests that intermittent cycling is a specific response to sperm deprivation. If stem cell cycling drives reproductive senescence, why would gonads that are not reproductively active maintain cycling and thus hasten their demise? We hypothesized that gonads that are in the active state are poised for reproduction, and thus respond quickly to a favorable change in environmental conditions. To test this hypothesis we characterized the dynamics of reproduction initiation after mating aged females. After mating of inx-22; fog-2 females at day 3 of adulthood, the rate of viable progeny production increased sharply and remained sustained over the first ~18 h (Fig 9). By contrast, mated fog-2 females of the same age experienced a transient increase followed by a trough in progeny production between 9 h and 13 h after mating; this trough was more marked and longer-lasting for fog-1 females (Fig 9; note that the maximal rates were lower for inx-22; fog-2 than for fog-1 or fog-2, which is expected given the faster reproductive senescence of inx-22; fog-2). The trough likely resulted from a drop in the numbers of developing oocytes in diplotene or diakinesis, which was not experienced by inx-22; fog-2 (S8A Table and S6A Fig), and appeared to be independent of apoptosis: initial progeny production occurred at the same rate in apoptosis-deficient strains carrying a ced-3 mutation (S8B Table and S6B Fig), and apoptosis levels decreased within the first 8 h after mating (S8C Table and S6C Fig) and thus likely did not account for a diminished supply of oocytes. In summary, mutant populations that maintain a higher proportion of dormant gonads are less capable of sustaining a high rate of progeny production shortly after mating. Overall, our results are consistent with dormant gonads needing time to make the switch to a state with maximal reproduction rate. Diminished initial reproductive activity of gonads that have been in a dormant state might stem from a detrimental impact of prolonged meiotic arrest on gamete quality or on the gonad region that houses them [47], or from a delay in returning to the active state. In any case, there is strong selection against reproduction delays [5,48]. Avoiding delays provides a powerful rationale against gonads staying fully dormant until reproduction becomes possible. Our findings significantly extend understanding of C. elegans reproductive senescence. We showed, using direct experimental manipulation of the cell cycle as well as using comparisons of senescence rates in strains that differ in gonad activity, that activity of the gonad is largely responsible for the progression of reproductive senescence—in particular as a result of germ cell cycling. In the course of this study we discovered that germ cell cycling can be intermittent, which as discussed further below suggests that worm populations may employ an active strategy to manage the progression of reproductive senescence. These results are in contrast to the idea that this progression is solely the result of passive damage accumulation as a function of elapsed time. What mechanisms tie reproductive senescence to germ cell cycling? Mutations that delay senescence may provide an entry point to start addressing this question. But although a number of mutations have been uncovered that increase reproductive lifespan, most do not result in an increase in total brood size and many in fact result in a decrease in brood size—i.e. the reproductive system is less active for a longer period of time, with reduced total output (e.g. [3]). Given our observations that gonad activity drives reproductive senescence, the primary effect of many of these mutations may thus simply be to alter the rate of reproduction, rather than to alter the relationship between reproductive senescence and gonad activity. By contrast, the DDR reduction of function mutant hus-1(op241) is to the best of our knowledge the first worm mutant reported to have a total reproductive capacity greater than wild-type. What link between cell cycling and reproductive senescence does this mutant suggest? It is possible that an intact DDR is required for full-speed germ cell cycling, and that hus-1(op241) germ cells thus cycle more slowly, leading to delayed reproductive senescence; given that hus-1(op241) does not display slower reproduction as might be expected from slower germ cell cycling, we do not favor this possibility. Another possibility is instead that the DDR, whose intensity according to RPA-1 foci correlates positively with the amount of past cell cycling, could mediate the effects of germ cell cycling on reproductive senescence—through molecular pathways that will require further study. An interesting additional problem will be identifying the downsides of increased reproductive output in hus-1(op241) mutants. We speculate that late-born progeny are more prone to mutation accrual or genomic rearrangements; further study will be required to characterize the fascinating tradeoff between increased brood size and decreased genome quality in progeny. We also note that germ cell cycling has been shown by others to shorten lifespan, in addition to reducing reproductive capacity as we report here: mating acts via DAF-12 –a key mediator of lifespan control by the reproductive system [49]–to cause somatic “shrinking” and death, while cell cycle inhibition protects from this effect [50]. It will prove interesting to ask whether the reproductive senescence and somatic lifespan effects of germ cell cycling are enacted by overlapping mechanisms. The role of germ cell cycling in driving reproductive senescence makes intermittent cycling particularly intriguing. Much of this intermittent cycling phenomenon remains to be characterized. The lack of suitable live imaging techniques makes it necessary to infer the existence of dormant and active states and the rules for transition between these states, and certainly leaves open the possibility that, in the future, more fine-grained studies will identify a broader array of sub-states and more complex transitions between them. The existence of at least two states, an active and a dormant one, is strongly supported by EdU continuous labeling studies that show that some virgin females mitotic zones label immediately when placed on EdU while some do not. This is compatible with germ cells being capable of stopping replication within S-phase or significantly lengthening S-phase [51,52]. Further, analysis of mitotic index CV shows greater variability between mitotic zones of virgin female populations than between those of mated female or young hermaphroditic populations. This is compatible with the existence of at least two cell cycle states—dormant and active—but there could be a continuum of intermediary states, whose practical relevance may depend on the amount of time they perdure. Until cell cycle states are more finely defined at the molecular level, control of transitions between states can only be addressed in qualitative terms. The spread in virgin female mitotic zone cell cycle progression during EdU pulse-chase experiments (Fig 6C) suggests that the transition between dormant and active states has a stochastic aspect (although we cannot formally exclude that individuals possess strongly intrinsic differences in switching rates); this is also suggested by the larger reproductive capacity CV in virgin females mated at day 2 than in those mated at day 0, which likely stems from a random distribution of time between days 0 and 2 spent in the active state that reduces reproductive capacity. We note however that biological phenomena are often described as “stochastic” until a deterministic underlying is identified; for example, control of the lysis-lysogeny decision of the lambda phage was described as stochastic but cell volume turned out to be a strong predictor of cell fate [53]. The switch between actively-cycling and dormant states could thus be downstream of an unknown highly-deterministic mechanism rather than being controlled by molecular noise. But in any case stochastic switching between active and dormant states provides a fitting and parsimonious model for our current data. The molecular controls of intermittent cycling remain to be fully established. It may seem surprising that fog-1 and fog-2, both known for their role in germline sex determination, have different cell cycle dynamics: fog-1 undergoes slower average cell cycling that does fog-2. This could be consistent with the previously-established, dose-dependent role of FOG-1 in promoting proliferation [54]. But a complication in comparing fog-1 and fog-2 exists in that virgin fog-1 ovulates at a slower rate than virgin fog-2; it is possible that a slower loss of cells to oogenesis leads to slower homeostatic mitotic zone cycling, rather than there being a direct role for fog-1 in controlling cell cycling. In line with this idea, it has most interestingly been shown recently that oocyte accumulation inhibits germ cell proliferation [41]. Lastly, fog-1 and fog-2 differ in the foraging behaviors [55], which could impact germ cell cycling. Overall, further experiments will be required to address the different behavior of fog-1 and fog-2. In addition to the fog-1 and fog-2 genes better known for their role in germline sex determination, our results and those from another recent study [41] implicate a number of genes and pathways in control of proliferation in virgin female mitotic zones. We have identified a role played by daf-22, likely through the dauer pheromone pathway, while the daf-18 and insulin pathways were implicated by [41] (in a way that did not address dormant and active states but that is compatible with our results when considering average cell cycle speed). Further experiments will be required to derive a more comprehensive understanding of the integration of these molecular controls. What is the relevance of intermittent cycling? First, we address whether it is a general response to a reduced need for germ cell production. A germline proliferation stop occurs quickly in females after deprivation of food [51,52]. In cases where a reduced average proliferation rate is warranted, different possibilities would be for mitotic zones to be reduced in size, for them to cycle at a slower but steady rate, or for them to go through periods of dormancy. We have not observed substantial changes in mitotic zone size. We did observe changes in cell cycle length across developmental stages: we recently reported that germ cells cycle ~60% more slowly in young adult hermaphrodites than in L4 hermaphrodites, in a way that does not involve intermittent cycling but rather S-phase lengthening [34]; we have also shown in the present study that caloric restriction that reduces the rate of reproduction over 2-fold also does not result in intermittent cycling. Therefore, intermittent cycling is a response that so far appears specific to well-fed females and hermaphrodites with a diminished sperm supply. Next, we address whether intermittent cycling may occur in the wild in a fashion that is relevant to “fitness.” We focused largely on genetically-feminized C. elegans hermaphrodites in the present study, because C. elegans has been more thoroughly studied than other species and can be investigated with better established genetic tools. Although C. elegans females do not occur naturally, we showed that the same intermittent cycling behavior occurs in C. remanei, which is a gonochoristic (i.e. male/female) species, and in older hermaphrodites. What is the relevance of germ cell behavior in older, sperm-depleted hermaphrodites? Two important questions to address this relevance are 1) how strong a selective pressure applies to hermaphrodites around the end of reproduction and 2) whether hermaphrodite mating is of much relevance in a mostly selfing species. We discuss these two questions in turn. Regarding whether reproduction of older hermaphrodites is under active selection, there is a general expectation that late-life reproduction might only be under weak selection ([56]; although see e.g. [57] for recent developments). Interestingly however, hermaphrodites show mating behavior that is specific to old age. Specifically, older hermaphrodites release a volatile pheromone that attracts males [58], which is regulated by the CEH-18 sperm sensing pathway [59]. In addition, older hermaphrodites mate more rapidly with males and eject sperm with reduced frequency than young hermaphrodites [60]. If evolutionary pressures were such that behavior of older hermaphrodites was of little relevance to fitness, one would expect that mutation accumulation would have caused these old-age specific behaviors to be lost [see e.g. 61]. It may of course be that these behaviors are just a remnant of a not-too-distant gonochoristic past. But in that case, the intermittent cycling behavior we observe would likely also be preserved from the previous gonochoristic state and thus amenable to study in C. elegans—even if it has lost its direct purpose. Overall, hermaphrodite mating that occurs under conditions of intermittent cycling is likely of relevance to fitness and is thus under active selection, or was in a suitably-recent past. Regarding the general in-the-wild relevance of outcrossing (irrespective of hermaphrodite age at the time of mating), although C. elegans males have only been isolated at a very low frequency in the wild [62,63] heterozygosity data show that outcrossing does occur in the wild—crucially, at a higher frequency than expected from mere spontaneous generation of males through meiotic non-disjunction of the X chromosome [reviewed by 61]. Males thus appear to be actively maintained in the wild. Consistent with this, experiments in the laboratory have identified conditions likely to be encountered in the wild, such as increased mutational pressure or exposure to changing environments, that lead to maintenance of males at a high frequency [61,64]. Lower than expected observed frequencies in the wild may be due to outbreeding depression [61], and male frequency may thus be actively maintained at an equilibrium frequency by counteracting evolutionary forces. Taken in combination with the fact that older hermaphrodites are more likely than young hermaphrodites to have cross progeny with males, this strengthens the idea that intermittent cycling in older hermaphrodites is a behavior that is of relevance in the wild and actively selected for, even if that selection is weaker than the selection for early hermaphroditic reproduction. If future reproduction of intermittently-cycling hermaphrodites is relevant to in-the-wild conditions and is specifically selected for, how would intermittency be beneficial? We speculate that uncertainty in the time at which reproduction will become possible could perhaps underlie a bet-hedging strategy. Bet-hedging is a well-established strategy followed by unicellular and multicellular organisms to avoid or spread risks in the face of uncertain environmental conditions [65,66], and phenotypic heterogeneity is a mechanism by which bet hedging can be implemented [67]. Specifically, a model is conceivable whereby in unfavorable conditions C. elegans populations hedge their bets by maintaining individuals that are primed for reproduction at the cost of faster senescence, and individuals whose gonads are dormant and that thus senesce more slowly, helping preserve reproductive capacity of the population over time. Stochasticity in the behavior of individual gonads would then have two roles. First, it would allow individuals to modulate the average rate of germ cell cycling. Second, and more interestingly, stochasticity is a parsimonious mechanism to develop a broad distribution of effective reproductive senescence rates in the population, without those rates being pre-assigned to each individual. Individuals that “take a chance” by cycling allow the population to quickly initiate reproduction if conditions become favorable shortly after they cycle, while individuals that stay in a dormant state avoid the population going extinct if conditions only become favorable after an extended period of time (during which individuals that frequently cycled exhausted their reproductive capacity). To our knowledge, such bet hedging would be the first to directly control senescence of a self-renewing organ, and the first to be implemented by integration of stochastic switching between different states. An interesting problem for future studies would be to ask whether bet-hedging strategies are used by other self-renewing organs in which stem cell cycling contributes to maintenance of tissue function, but increases the probability of a cell becoming senescent to avoid cancer [68]. Finally, we note that there may be a number of alternative reasons for intermittent cycling. Intermittent cycling may be an unselected-for, side-effect behavior that derives from the particular structure of the yet-to-be-characterized gene network that regulates it, while the increase in phenotype variability in old age could be explained by phenotypic drift [69]. Rather than implementing a bet hedging strategy, intermittent cycling could provide a convenient means for cells to alternate between different metabolic states, or could be a way of maintaining the same distribution of cell cycle phases in the mitotic zone irrespective of the average cycling rates—so that mitotic zones can assume an optimal distribution quickly when reproduction initiates. Future studies will be required to explore these ideas. Strains used were Bristol N2, JK3743: fog-1(q785) I [11], JK1078: fog-2(q71) V [6], XM1012: inx-22(tm1661) I; fog-2(q71) V [9], EM464: C. remanei [36], BA784: spe-8(hc50) I [13], CNQ41: fog-1(q785) I; ced-1::gfp(bcIs39) V (obtained by a cross of MD701 [70] and JK3743), CNQ44: ced-3(n717) IV; fog-2(q71) V (obtained by a cross of MT1522 [71] and JK1078), WS4581: rpa-1::yfp(opIs263) [28] and WS2277: hus-1(op241) I [31]. Strains were maintained as described [72] using E. coli HB101 as a food source. Worms were staged by picking at the L4 stage as identified by visual inspection of vulva shape. Unless otherwise specified, 50 worms were kept per 60 mm plate prior to mating. Mating at “day 0” of adulthood refers to exposure to males for 24 h starting from L4—so that females are exposed to males at the earliest time at which they are sexually mature. More generally, mating at day n of adulthood refers to exposure to males from L4 + n * 24 h to L4 + (n+1) * 24 h. For mating and cell cycle experiments, a sharp increase in cycling intermittency was noted in wild-type selfed hermaphrodites around day 3 adulthood. Therefore, when using day 3 hermaphrodites, special care was taken to use worms at precisely L4 + 72 h. Female mating was performed on 35 mm-diameter plates (CC7672-3340, USA Scientific, Ocala, FL), at a density of 1 female and 3 young adult males per plate. For C. remanei matings, females were continuously exposed to males, which were refreshed every three days; without this continuous exposure, female reproductive capacity is not exhausted (progeny count with 24 h male exposure was 288, n = 20, compared with 734, n = 20, with continuous exposure). For brood size scoring, mothers were passaged every day to 35 mm fresh plates, until the end of reproductive activity. Plates on which embryos had been laid were incubated for ~2 days to let progeny hatch and grow in size; progeny were counted before they reached the adult stage. Mothers that crawled off agar were censored from the analysis. To analyze the early response to mating in fine detail, females isolated 3 days after L4 were each plated with 3 males and any progeny removed from the plate every ~3 h for a total of 21 h. After another 24 h, transferred progeny were scored for viability. For cell cycle inhibition experiments, hydroxyurea (HU) or the CDK inhibitor Roscotivine (H8627 and R7772, respectively Sigma-Aldrich, St. Louis, MO) were freshly prepared and added to NGM-agar (cooled to 55°C prior to dispensing) at a final concentration of 40 mM for HU [73] or 50 μM for Roscovitine. Roscovitine was diluted in DMSO so that the final dilution of the solvent in NGM was 1:1000; control plates were also supplemented with DMSO at 1:1000. The following day, bacteria were UV-killed by placing open seeded plates in a Spectrolinker XL-1500 (Spectroline, NY) for 5 min, and then transferred onto HU, Roscovitine or control plates. Immediately after, 1-day adult virgin females were transferred to HU or control plates for 24 h. They were then returned to regular plates and mated as above. For starvation experiments, females were picked at the late L4 stage, rinsed 5 times in M9, and starved in complete S-medium [74]. Following starvation, females were transferred to seeded plates, on which they were kept for 24 h prior to mating. For the thermotolerance assay, worms were transferred to HU or control plates at day 1 of adulthood, kept on these plates for 24 h, returned to regular plates, and shifted to 35°C. Every 2 h until all worms had died, plates were removed individually from the incubator and the number of live and dead worms recorded. For the lifespan assay, worms were initially treated as for the thermotolerance assay but without temperature upshift. A count was made of live and dead worms every day until all worms had died. Worms that had desiccated on the side of plates or died due to an exploding vulva were censored from the time of death. Worms were kept on NGM plates until day 1 of adulthood, at which point they were transferred to 500 μl S-medium containing serial dilutions of HB101 at final concentrations of 108 (low), 109 (medium) or 1010 (high) cells / mL in 24-well tissue culture plates rocked on a nutator. At day 2 of adulthood worms were pulsed for 1 h with EdU resuspended in water and delivered at a final concentration of 0.4 mM and processed as detailed below. To label cells with the thymidine analogue EdU (C10337, Life Technologies, Grand Island, NY), worms were fed EdU-labeled E. coli. To prepare labeled E. coli, strain MG1693 was grown in minimal medium supplemented with glucose [75] and 75 μM EdU for pulse-chase experiments and 7.5 μM for continuous labeling experiments. When required, Fluoresbrite fluorescent microspheres (19507–5, Polysciences, Warrington, PA) were added to bacteria prior to seeding, at a 1:100 dilution. Immediately following seeding, plates were stored at 4°C. Plates were warmed to 20°C prior to use. Worms were kept for 0.5–8 h on EdU-labeled bacteria in the dark, returned to non-labeled bacteria for experiments that required a chase, and were fixed and processed as described [76] using 0.1 μg/ml DAPI to label DNA and 1:200 anti-PH3 antibody (9706, Cell Signaling, Beverly, MA) to label M-phase cells, and imaged at ~0.3 μm z intervals with LSM 710 or 780 confocal microscopes (Carl Zeiss MicroImaging, Oberkochen, Germany), using a 63x objective. EdU continuous labeling image data were assayed manually for the presence of at least one EdU-positive cell in individual mitotic zones. EdU pulse chase data were analyzed as described in S1 Text. For whole germ line imaging, for counts of apoptotic cells detected with the CED-1::GFP reporter or foci detected with the RPA-1::YFP reporter, and for counts of cells in diplotene or diakinesis extruded germ lines were fixed and stained with DAPI, 0.16 μM Alexa 594-conjugated Phalloidin (A12381, Life Technologies), and 1:1000 anti-GFP (for CED-1::GFP and RPA-1::YFP; ab5450, Abcam, Cambridge, MA). Image stacks were acquired at 0.3–0.6 μm z intervals using a 63x objective. In some instances, several panels were imaged for each gonadal arm that were subsequently stitched [77]. RPA-1 foci were analyzed in the proximal meiotic region (from the beginning of zone 5 and into 6, as described [78]). The dataset was blinded by a user who copied renamed files from all datasets into a single directory. A second user scored each of the renamed image files by selecting 20 random cells in zones 5 and 6 using the DNA channel of the image and recording RPA-1::YFP foci within each of these cells using Parismi [52]. The Wilcoxon rank sum test as implemented by the R project or Matlab was used to test for significance of differences (two-tailed test), unless otherwise stated. A generalization to interval-censored data of the Wilcoxon test implemented by the “interval” R package [79] was used to analyze labeling times in continuous EdU labeling experiments. Bootstrapping was performed using the “boot.ci” function of the “boot” R package [80]. The log-rank test was used for analysis of survival data.

10.1371/journal.pcbi.1003421

A Multi-Scale Model of Hepcidin Promoter Regulation Reveals Factors Controlling Systemic Iron Homeostasis

Systemic iron homeostasis involves a negative feedback circuit in which the expression level of the peptide hormone hepcidin depends on and controls the iron blood levels. Hepcidin expression is regulated by the BMP6/SMAD and IL6/STAT signaling cascades. Deregulation of either pathway causes iron-related diseases such as hemochromatosis or anemia of inflammation. We quantitatively analyzed how BMP6 and IL6 control hepcidin expression. Transcription factor (TF) phosphorylation and reporter gene expression were measured under co-stimulation conditions, and the promoter was perturbed by mutagenesis. Using mathematical modeling, we systematically analyzed potential mechanisms of cooperative and competitive promoter regulation by the transcription factors, and experimentally validated the model predictions. Our results reveal that hepcidin cross-regulation primarily occurs by combinatorial transcription factor binding to the promoter, whereas signaling crosstalk is insignificant. We find that the presence of two BMP-responsive elements enhances the steepness of the promoter response towards the iron-sensing BMP signaling axis, which promotes iron homeostasis in vivo. IL6 co-stimulation reduces the promoter sensitivity towards the BMP signal, because the SMAD and STAT transcription factors compete for recruiting RNA polymerase to the transcription start site. This may explain why inflammatory signals disturb iron homeostasis in anemia of inflammation. Taken together, our results reveal why the iron homeostasis circuit is sensitive to perturbations implicated in disease.

The nutritional iron uptake is tightly regulated because the body has limited capacity of iron excretion. Mammals maintain iron homeostasis by a negative feedback loop, in which the peptide hepcidin senses the iron blood level and controls iron resorption. Molecular perturbations in the homeostasis loop lead to iron-related diseases such as hemochromatosis or anemia of inflammation. Quantitative studies are required to understand the dynamics of the iron homeostasis circuitry in health and disease. We investigated how the biological activity of hepcidin is regulated by combining experiments with mathematical modeling. We present a multi-scale model that describes the signaling network and the gene promoter controlling hepcidin expression. Possible scenarios of hepcidin regulation were systematically tested against experimental data, and interpreted using a network model of iron metabolism in vivo. The analysis showed that the presence of multiple redundant regulatory elements in the hepcidin gene promoter facilitates homeostasis, because changes in iron blood levels are sensed with high sensitivity. We further suggest that inflammatory signals establish molecular competition at the hepcidin promoter, thereby reducing its iron sensitivity and leading to a loss of homeostasis in anemia of inflammation. We conclude that quantitative insights into hepcidin expression regulation explain features of systemic iron homeostasis.

Hepcidin is a humoral polypeptide that plays a central role in systemic iron homeostasis (reviewed in [1]). One main function of hepcidin is to maintain constant levels of iron circulating in the blood despite imbalances in external iron availability: Iron overload in the blood stimulates hepcidin transcription in hepatocytes. Hepcidin in turn blocks intestinal iron uptake and macrophage iron release into the blood by binding to the iron exporter ferroportin and triggering its degradation. Thus, hepcidin is part of a negative feedback circuit that stabilizes the iron blood concentration. Negative feedback is known to play a key role for the robustness and homeostasis of biochemical networks (reviewed in [2]). Biochemical negative feedbacks have been shown to compensate for various perturbations including stochasticity in gene expression [3], [4], mutations [5] and pharmacological inhibition [6], [7]. In systemic iron homeostasis, iron diet content affects the iron concentration in the blood [8], [9], suggesting that the hepcidin feedback loop only partially compensates for perturbations. Nevertheless, genetic perturbations of iron-dependent hepcidin regulation result in hemochromatosis, a common hereditary disease [10]. At the molecular level, hereditary hemochromatosis (HH) is caused by inappropriately low hepcidin expression and an inability to compensate changes in iron blood levels. HH patients are characterized by chronic iron overload that causes organ damage such as liver fibrosis. Hepcidin expression is primarily controlled at the transcriptional level. Information about iron blood levels is transduced from the hepatocyte cell membrane to the nucleus by the bone morphogenetic protein (BMP) signalling pathway [11], [12]: Increasing iron blood concentrations are sensed by Hfe and Tfr2, two transmembrane proteins mutated in HH [13], [14]. The signal is transmitted by the BMP co-receptor HJV and BMP receptor 1 to trigger the phosphorylation and nuclear translocation of SMAD1/5/8 transcription factors (referred to as SMADs hereafter). BMP6 is regulated by hepatic iron levels and plays a critical role in this process [15], [16]. The hepcidin promoter contains two BMP-responsive elements (BRE1 and BRE2) that are recognized by the SMADs (sometimes also abbreviated as BMP-RE1 and BMP-RE2) [17]–[19]. Mutations in the BRE1 promoter element and in the BMP signalling pathway are associated with HH [20]. Hepcidin expression is also regulated by inflammatory cytokines and hypoxia [21], [22]. Inflammatory cytokines such as IL6 activate the STAT3 signaling pathway in hepatocytes. Phosphorylated STAT3 transcription factors (TFs) are directly recruited to a STAT binding site (STATBS) in the hepcidin promoter, thereby enhancing hepcidin expression and reducing iron blood levels [23], [24]. Chronic inflammation causes an iron-related disorder, known as anemia of inflammation (AI), because the persistent lack of iron availability blocks erythropoiesis [10]. This indicates that the integration of BMP and IL6 signals at the level of hepcidin expression plays a key role in systemic iron homeostasis. Combinatorial gene regulation by binding of multiple different transcription factors to the same promoter is a recurrent aspect of transcription networks. Thermodynamic modeling employs methods from statistical thermodynamics to describe combinatorial binding of transcription factors and RNA polymerase (RNAP) to the promoter, depending on the protein concentrations and binding energies. [25]–[32]. The modeling framework focusses on transcription initiation and is based on the assumption that gene activity is determined by RNAP recruitment to the transcription start site (TSS). Thermodynamic modeling has been shown to accommodate various modes of signal integration on a promoter [25], [26], [29], some of which have been confirmed experimentally for bacterial and yeast promoters [33]–[36]. More recently, the framework was extended to aspects of eukaryotic gene regulation, including nucleosome positioning effects [37], [38]. In this work, we combined experimental measurements and thermodynamic modeling to quantitatively analyze how the iron-sensing BMP and inflammatory IL6 pathways coordinately control hepcidin expression. Systemic iron homeostasis is maintained by an auto-regulatory negative feedback loop that involves the transcriptional induction of hepcidin in the liver: elevated iron levels in the blood induce hepcidin expression by activating the BMP signaling pathway in the liver. Released hepcidin in turn induces the degradation of intestinal iron transporters, thereby lowering the blood iron level. We analyzed a conceptual mathematical model of this circuitry to understand how systemic iron homeostasis can be maintained despite imbalances in iron availability and consumption (Supplemental Text S1). The model suggests that the regulatory loop most potently balances iron blood levels if hepcidin expression responds in a steep, nonlinear manner to alterations in iron blood levels. Robust homeostasis further requires that the hepcidin promoter is able to sense and to respond to a broad range of iron blood concentrations. We therefore reasoned that data of dose-dependent hepcidin promoter regulation, and its modulation by inflammatory IL6 stimulation, could provide valuable insights into the regulation of iron homeostasis. Hepatic cell culture systems do not directly respond to stimulation with extracellular iron [39]. BMP6 is involved in hepatic iron-sensing in vivo, and is commonly used as an external stimulus to characterize how hepcidin expression responds to changes in iron blood levels in vitro [15], [16]. We performed co-stimulation experiments with BMP6 and IL6 in human hepatoma (HuH7) cells, and measured the activity of a luciferase reporter gene driven by the hepcidin promoter 24 h after stimulation. The hepcidin expression response of HuH7 cells was characterized in previous studies, and it was concluded that HuH7 cells reflect known features of hepcidin expression in vivo (see Discussion). Reporter gene assays were performed in HuH7 cells transiently expressing luciferase constructs under control of wildtype (WT) or mutant hepcidin promoters (Fig. 1B and C). The WT promoter spans 3 kb upstream of the transcription start site, and contains a proximal STAT-binding site (STATBS), a nearby BMP-responsive element (BRE1), and a distal BMP-responsive element (BRE2). We and others previously showed that these sequence motifs are necessary and sufficient for responsiveness towards BMP and IL6 stimulation (see Discussion). In each of the three promoter mutants, one of these transcription factor binding sites is non-functional: The BRE1m and BRE2m constructs exhibit point mutations in the corresponding BMP-responsive elements, while STATdel is characterized by a complete deletion of the STAT-binding site. For simplicity, we will generally refer to transcription binding site mutations even when discussing the deletion of the STATBS. The raw luciferase activity reads of at least four biological replicates were processed (see Methods) and the merged data is shown in Fig. 1C (using the same arbitrary concentration units for all heatmaps). BMP6 and IL6 mono-stimulation both increased the luciferase activities of the WT promoter construct. Maximal IL6 stimulation enhanced luciferase activities by 20-fold, while a 450-fold increase was observed upon maximal BMP stimulation. The IL6 response saturated within the concentration range used, because IL6 concentrations beyond 6 ng/ml hardly increased expression any further. Thus, BMP6 increased expression much more efficiently than IL6 in terms of maximal possible inducibility. Co-stimulation with IL6 and BMP further enhanced the luciferase activity of the WT promoter compared to mono-stimulation. We confirmed by qPCR measurements that this co-stimulation response quantitatively reflects the expression of endogenous hepcidin mRNA (Supplemental Protocol S1, Supplemental Fig. S1). As expected, the STATdel promoter fails to respond to IL6 stimulation, but is still sensitive to BMP treatment. The luciferase activities of the BRE1m construct are much lower than WT for both basal and induced conditions (Fig. 1C). However, the BRE1m promoter is qualitatively similar to the WT promoter in terms of stimulus inducibility, because maximal BMP stimulation induces a large increase in luciferase activity, while the maximal IL6 dose has a much lesser impact. The co-stimulation response of the BRE2m construct is qualitatively different from that of the WT and BRE1m promoters. The BRE2m promoter resembles a coincidence detector (‘logical AND gate’): Mono-stimulation with either BMP or IL6 raises luciferase activity to intermediate levels only, and co-stimulation with both ligands is required for maximal expression (Fig. 1C). This suggests that BRE1 and BRE2 fulfill very distinct functions in hepcidin regulation: The BRE2m promoter shows only a slight reduction in basal luciferase activity compared to the WT promoter. This reduction in basal activity did not reach statistical significance (paired t-test), and was much less pronounced than the effect of a BRE1 mutation. The promoter responsiveness to IL6 mono-stimulation was significantly reduced by the BRE1 mutation but not by the BRE2 mutation (paired t-test). Thus, BRE2 has lesser impact on the IL6-induced fold-change in hepcidin expression than BRE1. The most prominent feature of the BRE2m promoter is the reduced ability to respond to BMP stimuli compared to the WT promoter (0.001<p<0.0025, paired t-test): maximal BMP stimulation enhances luciferase activity by only 80-fold in BRE2m, when compared to 400-fold in WT and BRE1m. The loss of BRE1 hardly affects the promoter responsiveness to maximal BMP doses, although it has some impact at intermediate BMP doses. Taken together, these data raise the interesting question of how two transcription factor binding sites with very similar sequence can show qualitatively distinct behavior in hepcidin expression regulation. The coordinated regulation of hepcidin expression by BMP and IL6 may involve crosstalk at the level of signal transduction. We analyzed signaling pathway interactions by measuring the phosphorylation of STAT3 and SMAD1/5/8 after stimulation with BMP and/or IL6 using quantitative immunoblotting. Cells were stimulated for 12 h after starvation, and phosphorylated SMAD1/5/8 and STAT3 were detected in the cell lysate using phospho-specific antibodies (Supplemental Fig. S2). The signals were quantified by densitometry and the results of biological duplicates were merged (see Methods). SMAD1/5/8 but not STAT3 showed basal phosphorylation in unstimulated cells which is consistent the role of BRE1 in controlling basal luciferase activities (Fig. 1C). As expected, stimulation with BMP or IL6 alone resulted in dose-dependent increases of SMAD1/5/8 and STAT3 phosphorylation, respectively (Fig. 1D). Co-stimulation with a saturating dose of IL6 appeared to slightly reduce BMP-induced SMAD1/5/8 phosphorylation (Fig. 1D, bottom), but the effect is not statistically significant (paired t-test). High doses of BMP had significant inhibitory effects on IL6-mediated STAT3 phosphorylation (p<0.001, paired t-test). However, the effect was moderate and typically resulted in a less-than two-fold reduction in phospho-STAT3 levels (Fig. 1D, top). Thus, while co-stimulation with BMP and IL6 enhances hepcidin expression relative to mono-stimulation, a slight cross-inhibition is observed at the level of transcription factor phosphorylation. We integrated these measurements at different levels into a mathematical model to quantitatively understand the determinants of hepcidin expression regulation. Our model describes luciferase activity as a function of the extracellular IL6 and BMP concentrations, and consists of two modules: The signaling module describes SMAD and STAT TF phosphorylation in response to BMP and IL6 stimulation, while the promoter module characterizes combinatorial TF binding to the promoter and gene expression. The kinetic parameters of the model as well as the regulatory details at the promoter level were unknown. We therefore estimated the parameters by model fitting to TF phosphorylation and luciferase activity data (Figs. 1C and D), and systematically compared the ability of different promoter variants to fit the data in an unbiased modeling approach. The signaling module assumes that the dose-response curves of IL6-induced STAT phosphorylation and BMP-induced SMAD phosphorylation are of sigmoidal shape, as suggested by previous studies [4], [40]. Dose-response curves of TF phosphorylation are therefore represented using sigmoidal Hill equations (Supplemental Text S3). Inhibitory crosstalk between SMAD and STAT proteins was modeled by assuming that the phosphorylation degree of one TF affects the Hill equation parameters describing the other pathway. The thermodynamic promoter model describes luciferase activity as a function of the phosphorylated TF concentrations, and assumes that TFs affect the occupancy of the promoter: If the pSMAD and pSTAT concentrations are zero, the promoter will either be completely empty or RNA Polymerase II (RNAP) may bind to the transcription start site (TSS) at a basal level (Fig. 2A, bottom row). For increasing TF concentrations the promoter will be occupied by pSMAD, pSTAT or RNAP, or a combination of these, giving rise to multiple promoter states. The presence of three specific binding TF sites and RNAP binding to the TSS yields to 24 = 16 promoter states (Fig. 2A). In which of these states the promoter exists depends in a complex manner on the concentration of phosphorylated TFs, their binding affinity for DNA and may also be altered by TF/TF or TF/RNAP interactions on the promoter. Equations that describe the probability of each promoter state as a function of protein concentrations and binding affinities can be derived based on principles of statistical thermodynamics (Supplemental Text S2, Supplemental Protocol S2). The promoter states in the model are linked to gene expression by assuming that RNAP-bound promoters are transcriptionally active, while promoters devoid of RNAP binding are silent (Fig. 2A). A transcription initiation rate can thus be calculated from the probabilities of the promoter states (Supplemental Text S2, Supplemental Protocol S2). We neglected gene regulation at the levels of transcription elongation and post-transcriptional processing in our model. The experimentally measured luciferase activity is therefore assumed to be proportional to the simulated initiation rate. All variants of the promoter module comprised well-known aspects of promoter regulation such as pSTAT/pSMAD binding to the promoter and RNAP recruitment by TFs (grey arrows in Fig. 2B). We additionally allowed cooperative TF binding, implying that TFs may mutually enhance their recruitment to the promoter (red arrows in Fig. 2B). Given three possible cooperative TF interactions alone or in combination, we considered 8 promoter module variants in total (Fig. 2B). Model fitting was done by minimizing the χ2 metric which allows for larger difference between model and experiment if the experimental error is large (see Methods). Model variants with different numbers of parameters were compared with respect to goodness-of-fit to the training data using the Akaike information criterion (see Methods). Based on these measures, the data were best explained by a promoter module containing a single cooperative interaction among SMAD and STAT TFs bound to BRE1 and STATBS, respectively (model 4 in Fig. 2B). More complex models also containing the cooperative interaction between BRE1 and STATBS (models 6—8 in Fig. 2B) did not fit the data better than the selected model. Model variants lacking the BRE1-STATBS cooperativity (models 1—3 and 5 in Fig. 2B) fitted the experimental data less well than the selected model from a quantitative point of view. Moreover, they qualitatively failed to explain why IL6 mono-stimulation induces a lesser fold-change in the BRE1m construct when compared to WT (Supplemental Fig. S3, Fig. 1C, [19]). The best-fit model (Figs. 2C and D) comprises 19 kinetic parameters and describes the data with an accuracy close to experimental measurement noise (χ2 = 124, N = 80). Nine model parameters enter the sigmoidal Hill functions which describe the stimulus-induced TF phosphorylation, and signaling crosstalk between transcription factors. The remaining 10 parameters enter the promoter module, and describe the TF affinity for DNA binding sites, the TF interaction strength with RNAP, and the cooperativity of TF binding to DNA. The nature of the model parameters is described in detail in Supplemental Protocol S2, and a list of best-fit parameter values can be found in Supplemental Table S1. Not all model parameters could be unequivocally identified based on the experimental data, implying that multiple parameter sets yield a comparable fit to the training data (Supplemental Fig. S4). This non-identifiability of parameters gives rise to uncertainties in the model predictions. We therefore performed all model analyses for many measurement-compliant parameter sets (with χ2<135), not only for the best-fit solution (see Methods). The model predictions in Fig. 3 were therefore formulated as a range corresponding to the simulation runs with the highest and lowest predicted luciferase activity. In most cases, reliable model predictions were possible despite non-identifiability of individual parameter values. In our model, hepcidin expression was mostly determined by the dynamics of TF binding to the promoter, while inhibitory signaling crosstalk played only a minor role: Elimination of signaling crosstalk did not significantly change the simulated luciferase activities, and this conclusion held true for all measurement-compliant parameter sets (Supplemental Fig. S5). Moreover, we compared the ability of model variants with and without signaling crosstalk to fit the luciferase activity data (Fig. 1C). The fit of the crosstalk-less model to the training data was significantly better as judged by the Akaike information criterion (Supplemental Fig. S5). We therefore focused our model validation on regulation events at the promoter level, and neglected the relatively weak signaling crosstalk effects. We sought to verify our model by an independent set of experiments not used for model calibration. The formulation of model predictions was focused on double mutant promoters which simultaneously lack two TF binding sites (Fig. 3A). One central promoter mechanism predicted by the model is the cooperative interaction between pSMAD and pSTAT TFs, bound to BRE1 and STATBS, respectively. The double mutant promoter lacking functional BRE1 and STAT elements (BRE1mSTATdel promoter, Fig. 3A) was employed to independently confirm the cooperativity effect. The BRE1STATdel promoter shows a ∼30-fold reduced expression relative to WT upon stimulation with 2 ng/ml IL6 (Fig. 3B, red bar). The corresponding STATBS and BRE1 single deletions reduce expression by ∼3 and ∼30-fold, respectively (Fig. 3B, blue and green bars). Thus, the BRE1m and BRE1mSTATdel promoters exhibit similar expression levels. This is consistent with BRE1-STATBS cooperativity, because the single BRE1m single deletion already eliminates the cooperativity effect, and thereby a large part of the STATBS contribution to expression. In contrast, a transcription model lacking the cooperativity would predict an independent contribution of both sites, implying that the expression reduction in the BRE1mSTATdel promoter equals the product of the single deletion fold-changes (Fig. 3B, top). Thus, the double mutant data qualitatively supports the model with cooperativity between BRE1 and STATBS, also for higher doses of IL6 (Fig. 3B). BRE1 and BRE2 play different roles in hepcidin expression with respect to basal expression, BMP inducibility and co-stimulation response. One difference between the two sites is the above-mentioned cooperative interaction with the STATBS that is specific to BRE1. How do the BMP-responsive elements differ beyond this interaction? The model predicted that BRE1 has higher affinity for phosphorylated SMAD than BRE2, explaining why BRE1 plays a predominant role under basal conditions. Upon sufficiently strong BMP stimulation both sites are predicted to activate RNAP with comparable efficiency. In conclusion, the model suggested that BRE1 and BRE2 should behave similarly in the absence of cooperative promoter interactions. This prediction can be tested by co-stimulation of BRE2mSTATdel and BRE1mSTATdel promoters which solely contain BRE1 and BRE2, respectively (Fig. 3A). The experimental data was in good qualitative agreement with model predictions: Both mutants showed very similar co-stimulation heatmaps and primarily responded to BMP stimulation (Fig. 3C, bottom row). Maximal luciferase activity at high BMP levels was comparable for both constructs, indicating that BRE1 and BRE2 indeed drive RNAP activation with similar efficiency (Figs. 3C and D). Basal activity was approximately 10-fold higher in the BRE2STATdel promoter, suggesting that the isolated BRE1 has indeed a higher pSMAD affinity than BRE2 (Fig. 3D). Quantitative model predictions for the BRE2mSTATdel and BRE1mSTATdel heatmaps were only possible up to a certain range of absolute luciferase activities owing to non-identifiability of model parameters (Fig. 3C, top and middle row). The experimentally observed luciferase activities were within the predicted range (Fig. 3C). The model suggested that saturating RNAP binding to the TSS limits hepcidin expression upon strong stimulation, whereas signaling pathway saturation plays only a minor role. Saturation implies that single deletions promoters should exhibit expression levels relatively similar to WT, because the remaining TF binding sites maintain near-complete occupancy of the TSS with RNAP. This buffering effect should be abrogated upon a combined TF binding site deletion, leading to the prediction that hepcidin double mutant promoters exhibit very low expression compared to the corresponding single mutants. We confirmed this prediction by measuring the activity of the BRE1mBRE2m double mutant promoter (Fig. 3E): The double deletion of BRE1 and BRE2 reduces the luciferase activity by up to ∼1000-fold compared to WT (red bars). The corresponding single deletions typically reduce expression by ∼10-fold or less, and are thus much more similar to the WT promoter (blue and green bars). This behavior indicates TSS saturation, because the fold-change by a double deletion equals the product of the corresponding single deletion fold-changes in thermodynamic transcription models without saturation. The validity of our model is further supported by the quantitative agreement of the models' predictions with the data: the buffering of single deletions is more pronounced for strong (co-) stimulation (Fig. 3E), where promoter saturation is particularly prominent. We find similar, albeit less pronounced, buffering effects between STATBS and BRE2 (Supplemental Fig. S6) and conclude that saturating RNAP binding is an important aspect of promoter behavior. Promoter saturation also explains why the BRE1 single mutation strongly reduces expression at intermediate BMP concentrations, while having lesser impact at basal and strong stimulation conditions (Fig. 3E, blue bars, grey corridor). The BRE1 deletion has a strong effect on expression at intermediate BMP doses, because the high-affinity BRE1 is already fully occupied by pSMAD. Higher BMP concentrations alleviate the impact of BRE1 deletions, because low-affinity pSMAD binding to the BRE2 saturates the promoter, thereby buffering the loss of BRE1. We investigated in silico how the simultaneous presence of two BREs affects the behavior of the hepcidin promoter, and found that this promoter design enhances the steepness of the BMP dose-response curve: A certain increase in phospho-SMAD levels induces a larger expression fold-change in the WT promoter when compared to BRE1m and BRE2m promoters (Fig. 4A). The hepcidin promoter only contains a single STAT binding site as opposed to two BREs. Accordingly, the model predicts that the hepcidin promoter specifically responds with high sensitivity towards BMP stimulation, and is much less sensitive towards IL6 stimulation (Fig. 4B). The initial stimulation experiments used for model training (Fig. 1C) were based on three BMP and IL6 concentrations, and therefore did not allow for conclusions concerning the steepness of the BMP and IL6 response. To verify the model prediction by an independent set of experiments, we performed detailed dose-response measurements with multiple doses of IL6 and BMP, respectively (Fig. 4C). These data confirm that the BMP response is much steeper than the IL6 response, and thereby validate the model. Systemic iron homeostasis is maintained by an auto-regulatory negative feedback loop in which hepcidin expression levels depend on and control the circulating iron levels in the blood (Fig. 4D). A high BMP sensitivity of the promoter may allow the iron-BMP signaling axis to sense minor changes in iron blood levels, and to maintain systemic iron homeostasis. We simulated iron homeostasis in the living animal using an extended model with feedback (Fig. 4D). Iron blood levels were described by the species Feb, whose levels are controlled by influx and efflux reactions. The iron influx rate is proportional to the intestinal iron concentration (species Fei) which reflects the dietary iron content. Iron blood levels control the activity of the BMP signaling pathway, and thus hepcidin expression. Negative feedback regulation was considered in the model by assuming that the iron influx is negatively influenced by hepcidin (Supplemental Text S4). Hepcidin expression regulation by BMP and IL6 in the model was described using the previously derived best-fit model (Figs. 2C and D). The remaining parameters of the model describe the iron influx and efflux, hepcidin degradation, and the strength of hepcidin-mediated feedback on the iron influx. Homeostasis was analyzed by assessing how the iron blood level in the model (Feb) responds to a change dietary iron content (Fei), efficient homeostasis implying that a given fold-change in Fei elicits a much lesser fold-change in Feb. The key assumption we made was that hepcidin-mediated feedback regulation has a very strong impact on the iron influx. For the limit of strong feedback, it can be shown analytically that the degree of iron homeostasis is solely determined by the steepness of the hepcidin promoter response, and independent of the remaining model parameters (Supplemental Text S1). The range of intestinal iron concentrations for which homeostasis is observed is determined by the range of BMP concentrations that can be sensed by the hepcidin promoter (Supplemental Text S1). The model thus allowed us to quantitatively analyze how the hepcidin promoter architecture affects systemic iron homeostasis, although the remaining model parameters were unknown. The simulations in Fig. 4E show that a model with the WT hepcidin promoter efficiently maintains systemic iron homeostasis, as the iron blood levels remain essentially constant over a broad range of intestinal iron concentrations. Models with BRE1m and BRE2m promoters perform less well, as the perturbation-response curves are steeper and homeostasis is restricted to a narrower range of influx rates (Fig. 4E, green and red curves). This suggests that the simultaneous presence of two BMP-responsive elements in the promoter indeed optimizes the performance of the systemic iron homeostasis loop. One important question is why IL6 stimulation reduces iron blood levels and induces anemia of inflammation even though the homeostasis loop should effectively buffer IL6-induced perturbations in hepcidin expression. Simulations of the extended feedback model show strongly diminished iron levels and a loss of homeostasis if high IL6 levels are assumed (Fig. 4E, blue dashed line). This effect can be understood by considering the BMP dose-response curve of the best-fit promoter model for varying IL6 concentrations (Fig. 4F): increasing IL6 levels reduce the sensitivity of the BMP dose-response curve due to (partial) saturation of the TSS with RNAP. Moreover, significant changes in hepcidin expression are restricted to a narrower range of phospho-SMAD levels. The hepcidin promoter therefore responds less efficiently to changes in the iron/BMP signaling in the presence of IL6. This impairs the iron sensing capability of the hepcidin promoter in vivo, and leads to a breakdown of feedback homeostasis. Iron blood levels are chronically elevated in HH, in most cases due to inactivating mutations in the iron-sensing BMP signaling axis. One unexplored question is why HH is commonly associated with inactivating mutations in the SMAD signaling pathway, while mutations in the BRE1 promoter element are rare and BRE2 mutations have not been identified yet [20]. The iron homeostasis model predicts that a BRE1 deletion affects the iron blood levels more strongly than the BRE2 deletion in the WT homeostasis range (compare green and red lines in Fig. 4E, respectively). The more critical role of BRE1 may explain why only BRE1 mutations have been associated with HH, and can be explained by its higher phospho-SMAD affinity when compared to BRE2. The model further predicts strong buffering of BRE1 and BRE2 single deletions: Single mutations in either site have much weaker effects than a complete feedback ablation by a BRE1mBRE2m double mutation (light blue line in Fig. 4E). The very strong effect of a BRE double deletion in cultured cells (Fig. 3E) is thus predicted to be preserved in vivo. These simulations may explain why BMP signaling pathway mutations that simultaneously inactivate expression regulation via BRE1 and BRE2 are by far the most common cause of HH. Taken together, the steepness of the hepcidin promoter response appears to be a key parameter controlling how well the systemic iron homeostasis loop compensates for fluctuations in iron diet content. The fine-tuned expression of hepcidin plays a central role in systemic iron homeostasis, and is deregulated in two major clinical settings, HH and anemia of inflammation. Here, we comprehensively characterized the gene regulatory function of the hepcidin promoter using systematic promoter mutagenesis and co-stimulation experiments. We employed a multi-scale modeling approach capturing signaling and gene expression, and discriminated various promoter regulation scenarios. This approach complements existing strategies linking signaling and gene expression events [41], [42], and may be extended in future studies to model global transcription patterns based on mRNA expression profiles, TF binding information and mRNA half-life data. Gene expression may be a gradual or a binary event at the single cell level [43]. Our mathematical model assumes that TF phosphorylation and reporter gene expression measurements in a cellular ensemble reflect the behavior of single cells, and thus presumes a gradual mode of hepcidin expression. Experimental studies indicate that BMP-induced target gene expression is indeed a gradual event at the single-cell level [4]. In any case, the population-based model reflects physiologically relevant aspects of hepcidin expression, because systemic hepcidin levels in vivo are governed by expression in an ensemble of hepatocytes. The architecture of the hepcidin promoter was characterized in detail in previous studies, and BRE1, BRE2 as well as STATBS were identified as central cis-regulatory elements mediating BMP and IL6 responsiveness [19], [24], [44]. Our results confirm the central role of the STATBS, as STATBS deleted promoters cannot be induced by IL6 stimulation. The critical role of BRE1 and BRE2 for the BMP responsiveness was shown by reporter gene assays with truncated versions of the promoter, and is also supported by the high sequence conservation of these elements [17]–[19], [44], [45]. In a search for additional BREs, we performed a linker scanning analysis of the hepcidin promoter, systematically replacing short nucleotide stretches along the promoter sequence (unpublished data). This analysis revealed no additional BMP target motifs. Accordingly, we observe that the BRE1mBRE2m double mutant shows near complete ablation of basal expression and BMP responsiveness (Fig. 3C). However, stimulation with very high doses of BMP appears to slightly enhance expression from the BRE1mBRE2m promoter (Fig. 3C). We suggest that highly active BMP receptors may weakly phosphorylate SMAD2/3 TFs which are part of the TGFβ signaling pathway, thereby activating the previously described TGFβ-responsive elements in the hepcidin promoter [46]. Our model suggests that BRE2 enhances transcription with similar or higher efficiency than BRE1 (Fig. 3, Supplemental Fig. S4), thus raising the question of how a sequence element located as far as 2 kb away from the TSS enhances transcription with high efficiency. The 2 kb distance from the TSS corresponds to a length of 680 nm along the strand; this number exceeds the length of the mediator complex (∼40 nm) which links RNA polymerase to transcription factors [47]. Thus, BRE2-mediated transcription initiation likely involves DNA looping. Our current model neglects the details of DNA looping, but could be extended by existing quantitative modeling approaches taking into account the thermodynamics of DNA bending [48], [49]. Such a detailed promoter model should also consider that the BRE2 element of the hepcidin promoter is flanked by bZIP, HNF4alpha/COUP binding sites in the immediate neighborhood [17]. The single deletion of the bZIP or HNF4alpha/COUP sites reduces the BMP responsiveness of the hepcidin promoter. This indicates that a complex of multiple transcription factors cooperates at the BRE2 site to recruit RNAP which may explain the apparently high efficiency of BRE2 in driving transcription. Another prediction of our model is that the strong impact of the BRE1 in the basal state is due to its high affinity for phospho-SMAD binding when compared to BRE2 (Supplemental Fig. S4, Fig. 3E). Both BREs are characterized by the same sequence motif (GGCGCC), suggesting that epigenetic differences in the chromatin state may be responsible for the apparently different affinity of BRE1 and BRE2. Taken together, the present model is likely to be a simplified representation of the real events at the promoter. Future studies are required to model individual events which are currently lumped into overall interaction energies. Different modes of signal integration may be realized in transcriptional regulation. Two stimuli may control expression in an additive or multiplicative way. We systematically compared the co-stimulation response of the hepcidin promoter with the corresponding mono-stimulation responses (Supplemental Fig. S7). The fold-change in expression upon co-stimulation is generally less than the product of the mono-stimulation fold-changes in the WT promoter (Supplemental Fig. S7; Fig. 1C), as also supported by previous studies in HuH7 and Hep3B cells [50], [51]. Analytical studies in the Supplemental Text S1 show that this sub-multiplicative signal integration explains why IL6 co-stimulation leads to a breakdown of systemic iron homeostasis (Fig. 4E): Homeostasis is lost, because IL6 reduces the BMP-induced fold-change in expression, thereby reducing the efficiency of negative feedback regulation. Our model suggests that the less-than multiplicative behavior of the WT promoter arises from saturating RNAP binding to the TSS. The saturation effect is less pronounced in single mutant promoters, explaining why these exhibit near-multiplicative signaling integration (fold-change over basal in response to co-stimulation equals the product of the mono-stimulation fold-changes) (Supplemental Fig. S7). Interestingly, the model predicts and experiments support that pSMAD and pSTAT may also drive hepcidin expression in a synergistic, more-than multiplicative manner due to the presence of the cooperative BRE1/STATBS interaction if promoter saturation effects are negligible: We changed the basal pSMAD level in the model, and observed that a certain basal BMP signaling activity is required for optimal responsiveness of the promoter towards IL6 mono-stimulation (Supplemental Fig. S7). This model prediction is supported by experiments in HuH7 cells showing that SMAD4 siRNA lowers the IL6 inducibility of the hepcidin promoter ([46]; unpublished observation), and by data in hepatocyte-specific SMAD4 knockout mice [52]. We conclude that the hepcidin promoter shows high plasticity in the integration of BMP and IL6 signals, depending on the strength of basal and induced signaling. In this study, we used in vitro measurements in HuH7 cells to parameterize an in vivo model of systemic iron homeostasis, thereby assuming that HuH7 cells quantitatively reflect hepcidin regulation in vivo. HuH7 cells have been widely used in the field of iron metabolism to study hepcidin regulation, and have been to behave very similarly to other hepatoma cells (HepG2, Hep3B) and primary hepatocytes [17], [19], [23], [44], [46], [50], [53]. While, to the best of our knowledge, there are no reports of quantitative comparisons on hepcidin regulation in vivo and in vitro, there is abundant evidence that the in vivo data that qualitatively mirrors the results obtained in vitro. For instance, Wang et al. reported that SMAD4 is essential for hepcidin activation both in mice and in primary hepatocytes [52], while Pietrangelo et al. showed in both models that STAT3 is a key transcription factor for IL-6 activation of hepcidin gene expression [54]. These results can be consistently reproduced in HuH7 cells, suggesting that the molecular mechanisms involved in these signaling pathways are preserved in this cell line. The hepcidin promoter contains two BMP-responsive elements as opposed to a single STAT binding site, raising the question of why such a promoter design may be advantageous for the regulation of systemic iron homeostasis. We find that the presence of two BREs ensures that hepcidin expression is very sensitive towards changes in the iron-sensing BMP pathway (Fig. 4B–D). This makes the negative auto-regulation loop more nonlinear, thereby promoting systemic iron homeostasis (Fig. 4E). Systems biology studies at the organismal level are required to confirm that our simple model of iron homeostasis faithfully predicts the dynamics of iron metabolism in vivo. The mathematical model describing hepcidin expression (used to generate Figs. 2C, 2D, 3B–3E, 4A–4C and 4F) consists of two modules: (i) the signaling module which describes the phosphorylation of SMAD and STAT transcription factors as a function of the BMP and IL6 concentrations. (ii) the promoter module uses the concentrations of phospho-SMAD/STAT (described by module i) as inputs, and computes the hepcidin expression level as an output. In the signaling module, we described the dose-response behavior of SMAD/STAT phosphorylation at steady state using sigmoidal Hill equations (Supplemental Text S3). Potential signaling crosstalk was considered by assuming that the phosphorylation of one TF modulates the maximal phosphorylation of the other TF (Supplemental Text S3). The promoter module was described based on the thermodynamic model derived in Supplemental Text S2. In Supplemental Protocol S2, we describe how this thermodynamic model was applied to hepcidin expression regulation, combined with the signaling model and we also provide a detailed description of all model parameters. The hepcidin expression model was embedded into an ODE model to describe systemic iron homeostasis by hepcidin-mediated negative feedback (Figs. 4D and E). A detailed description of this ODE model can be found in Supplemental Text S4. The models were fitted to the experimental data by minimizing χ2 metric, given by χ2 = (Mi-sj⋅Di)2/σi2. Mi, Di and σi are the simulated value, the measured value and the experimental error, respectively. The fitted scaling factor si accommodates that the model is formulated in absolute concentration units, while signaling and luciferase activities could only be measured in arbitrary units. The simulated phospho-SMAD/STAT concentrations were fitted to the immunoblotting data (Fig. 2C), while the simulated transcription initiation rate was fitted to the luciferase measurements (Fig. 2D), i.e., gene expression was assumed to be at steady state (see text). Parameter optimization was done using a deterministic trust region optimizer in Matlab. In order to circumvent local minima, we repeatedly fitted the model starting from 80.000 quasi randomly distributed positions in the space of allowed parameter ranges. The optimization apparently converged to a global optimum, because ∼45% of the fitting runs yielded χ2 values close to the minimum all 80.000 runs. We fitted model topologies of different complexity by eliminating certain reaction steps (Fig. 2B, Supplemental Protocol S2), and compared the ability of model variants to fit the data. Models of different complexities were compared based on their goodness-of-fit to the training data set using the Akaike Information criterion (AIC = χ2+2k; k…number of model parameters) and the likelihood ratio test (Supplemental Protocol S2). Both statistical measures favored model topology 4 in Fig. 2B. Parameter identifiability was analysed using the strategy proposed by Hengl et al. [55]: The parameter vectors of the top 45% fitting results from quasi-randomly distributed starting parameter sets (see above) had a similar goodness-of-fit (χ2<135), and were analysed with respect to parameter ranges and parameter correlations (Supplemental Fig. S4). The robustness of model predictions was estimated by repeatedly simulating predictions for the top 45% of the model solutions. The simulations predicting the highest and lowest values are given as a prediction range in Fig. 3. The human hepatocarcinoma HuH7 cell line was cultured in Dulbecco's Modified Eagle's Medium (DMEM, high glucose; Invitrogen) supplemented with 10% heat inactivated low-endotoxin fetal bovine serum (FBS, Invitrogen), 1% penicillin, 1% streptomycin and 1 mM Sodium Pyruvate. Cell cultures were maintained in a 5% CO2 atmosphere at 37°C. Generation of the luciferase reporter construct containing a 2762-bp fragment of the human hepcidin promoter (WT) and derivate constructs with mutations in BMP-responsive element 1(position −84/−79; BRE1m), BMP-RE2 (position −2255/−2250; BRE2m), STAT binding site (position −72/−64; STATdel), and in BMP-RE1 and BMP-RE2 (BRE1mBRE2m), have been previously described [19], [44]. In this study we generated two additional reporter constructs that combined mutations in BMP-RE1 or BMP-RE2 with the deletion of the STAT binding site (constructs BRE1STATdel and BRE2STATdel, respectively). HuH7 cells (1.5×105 per well) were seeded onto six-well plates. The next day, 500 ng of pGL3 reporter vectors containing the hepcidin promoter constructs were transfected, together with 10 ng of a control plasmid containing the Renilla gene under the control of the CMV promoter. Plasmid transfections were performed using Lipofectamine 2000 (Invitrogen) according to manufacturer's instructions, and medium was replaced by FBS-free medium. Twenty-four hours after transfection, cells were treated with human BMP-6 (50 ng/ml, 24 h) and/or IL-6 (2 ng/ml, 24 h). Cells were harvested in Passive Lysis Buffer (Promega) for measurement of luciferase activity and cellular extracts were analyzed using the Dual-Luciferase-Reporter assay system (Promega) and a Centro LB 960 luminometer (Berthold Technologies). HUH7 cells (1.5×105 per well) were seeded onto 6-well plates and the day after the culture medium was exchanged to FBS-free medium. After 12 hours the cells were treated with increasing doses of BMP-6 (60; 200; 800 ng/mL; R&D Systems) and/or IL-6 (2; 4; 6; 25 ng/mL; R&D Systems) for 12 hours and then harvested for protein analysis. Cells were washed twice in ice cold Dulbecco's phosphate-buffered saline (PBS). Cell pellets were lysed in ice-cold NET buffer (1% Triton X-100 (v/v), 50 mM Tris-HCl pH 7.4, 150 mM NaCl, 5 mM EDTA, 20 mM NaF, 1 mM Na3VO4) supplemented with 1× Complete Mini Protease Inhibitor Mixture (Complete Mini,Roche Applied Science). The protein concentration was measured using the BCA (bicinchoninic acid) Protein Assay (Pierce). Protein lysates (15 µg) were subjected to 10% SDS-PAGE and transferred to a nitrocellulose membrane (Whatman) for protein immunodetection using rabbit anti-phospho SMAD 1/5/8 (Cell Signaling #9511), mouse anti-phospho STAT3 (Cell Signaling, #9138) and mouse anti-actin (Sigma Aldrich, A2228). Blots were then incubated with horseradish peroxidase conjugated anti-mouse or anti-rabbit secondary antibodies (Sigma Aldrich) and then subjected to chemiluminescence (Amersham Biosciences, ECL Plus). For the densitometric analysis the resulting bands were digitalized and quantified using the NIH Image J software (rsb.info.nih.gov/nih-image/) Luciferase Reporter Assays were performed in at least four biological replicates. Reporter gene expression was monitored using firefly luciferase, and co-transfection with Renilla luciferase allowed for correction with respect to cell number and transfection efficacy. The relative light units of firefly luminescence of each experimental condition were divided by the relative light units of the corresponding Renilla luminescence. Each replicate measurement series was normalized by the median over all data points of that series to correct for slight differences in absolute luciferase signals between replicate experiments. Experimental errors were estimated by calculating standard deviations over all replicates. Errors in the fold changes of luciferase expression (Fig. 3B and E) were estimated using a Monte-Carlo approach: Random realizations were drawn from normal distributions with mean and standard deviation equal to those of the measured luciferase expression data. Fold-changes were calculated for 103 pairs of realizations, and the fold-change error was evaluated by calculating the standard deviation of the resulting probability distributions. Signaling crosstalk was monitored by immunoblotting against phosphorylated SMAD and STAT in two biological replicates (Supplemental Fig. S4). Bands were quantified by densitometry, and duplicate measurements were merged by multiplying one of the duplicate series with a fitted scaling factor to correct for differences in arbitrary units between gels. Figs. 1D and 2C show mean and standard deviation of merged duplicate experiments. Some experimental errors estimated from scaling were unreasonably small; therefore a minimal experimental error was assumed, based on typical variability in Western Blot measurements (relative error of 5% plus an absolute error value).

10.1371/journal.pntd.0006630

Field evaluation of a 0.005% fipronil bait, orally administered to Rhombomys opimus, for control of fleas (Siphonaptera: Pulicidae) and phlebotomine sand flies (Diptera: Psychodidae) in the Central Asian Republic of Kazakhstan

Plague (Yersinia pestis) and zoonotic cutaneous leishmaniasis (Leishmania major) are two rodent-associated diseases which are vectored by fleas and phlebotomine sand flies, respectively. In Central Asia, the great gerbil (Rhombomys opimus) serves as the primary reservoir for both diseases in most natural foci. The systemic insecticide fipronil has been previously shown to be highly effective in controlling fleas and sand flies. However, the impact of a fipronil-based rodent bait, on flea and sand fly abundance, has never been reported in Central Asia. A field trial was conducted in southeastern Kazakhstan to evaluate the efficacy of a 0.005% fipronil bait, applied to gerbil burrows for oral uptake, in reducing Xenopsylla spp. flea and Phlebotomus spp. sand fly abundance. All active gerbil burrows within the treated area were presented with ~120 g of 0.005% fipronil grain bait twice during late spring/early summer (June 16, June 21). In total, 120 occupied and 14 visited gerbil colonies were surveyed and treated, and the resulting application rate was minimal (~0.006 mg fipronil/m2). The bait resulted in 100% reduction in Xenopsylla spp. flea abundance at 80-days post-treatment. Gravid sand flies were reduced ~72% and 100% during treatment and at week-3 post-treatment, respectively. However, noticeable sand fly reduction did not occur after week-3 and results suggest environmental factors also influenced abundance significantly. In conclusion, fipronil bait, applied in southeastern Kazakhstan, has the potential to reduce or potentially eliminate Xenopsylla spp. fleas if applied at least every 80-days, but may need to be applied at higher frequency to significantly reduce the oviposition rate of Phlebotomus spp. sand flies. Fipronil-based bait may provide a means of controlling blood-feeding vectors, subsequently reducing disease risk, in Central Asia and other affected regions globally.

Plague and cutaneous leishmaniasis are two diseases transmissible to humans vectored by fleas and sand flies, respectively. Although the diseases are vectored by two different insect types, the primary reservoir host for both diseases in desert foci in Central Asia is the great gerbil. Therefore, a promising strategy for controlling both vectors, and subsequently both diseases, is to target the host. A field study was conducted in southeastern Kazakhstan to evaluate the use of a rodent bait containing an insecticide (0.005% fipronil), applied to active gerbil burrows, in reducing field-caught flea and sand fly abundance. Results suggested that the fleas infesting gerbils could be reduced to zero for at least 80-days after completing treatment. The number of reproductive female sand flies were also reduced to zero 3-weeks after completing treatment, but environmental factors also influenced abundance. Additionally, because the bait contained a low insecticide concentration, the application rate remained low, posing reduced risk to non-target animal species. Our approach incorporates detailed evaluation of the use of a systemic insecticide in controlling fleas infesting gerbils and sand flies caught in light traps. Hence, this study provides an explicit means of evaluating the use of a new vector control approach in reducing two distinct insect vectors parasitizing the same host.

Vector-borne diseases transmissible to humans were responsible for more human disease and death than all other causes combined between the 17th and 20th centuries [1] and since the 1970s a global reemergence of several vector borne diseases has occurred [2]. Vector-borne diseases occur most frequently in areas of extreme poverty [3], and cost-effective measures, which consider socio-economic and environmental risk factors are warranted. Vector-borne diseases are prevalent in the Central Asian Republic of Kazakhstan, with plague (Yersinia pestis) and cutaneous leishmaniasis (Leishmania major) being two of concern [4,5]. Plague is a fatal, rodent-associated, flea-borne pathogen found throughout Asia, Africa and the Americas. Between 2010–2015, 584 plague-related deaths were reported globally [6]. While plague-induced mortality is far less common than during the pandemics of previous centuries [7], the disease is still regarded as “emerging” and changes in land-use have increased the probability of interaction between host species and humans [8]. Additionally, plague can wreak havoc on native wildlife, such as in North America, where plague outbreaks damage efforts to re-introduce endangered black-footed ferrets (Mustela nigripes) [9] which are dependent on plague-susceptible black-tailed prairie dogs (Cynomis ludovicianus) for food and habitat [10]. The last plague outbreak to occur in central Asia occurred in the Qinghai province of China in 2009 in which 12 people tested positive for plague with 3 people dying [11]. Central Asian plague outbreaks have been reduced in recent years with 17 cases and 8 deaths reported 2010–2015 [12]. However, desert plague is still a focus in central Asia because of the increasing risk of plague outbreaks from factors such as anthropogenic influence [13] and climate change [14]. Cutaneous leishmaniasis (CL) is a sand fly-borne neglected tropical disease. Although not lethal, it results in ulcerations on the skin which can lead to severe disability and lifelong scars, often resulting in severe social prejudice [15]. It is by far the most common form of Leishmania with ~700000–1200000 new cases being reported annually [16]. Although primarily a disease of the poor, as of 2009 it was estimated that there were ~3000–5000 cases of CL among U.S. military personnel [17]. As a result, the Deployed Warfighter Protection Research Program (DWFP) has invested in developing new pesticides for sand fly control [18]. The disease is difficult to control in part because aspects of sand fly ecology remain largely unknown [19]. In Central Asia, zoonotic CL is largely rodent associated [20]. Less than 100 cases were reported in Kazakhstan in 2015 [21] and of 333 leishmaniasis cases reported in Kazakhstan between 1996–2006, 332 of them were zoonotic CL [22]. The Great gerbil (Rhombomys opimus) is a colonial rodent that is considered the primary reservoir of plague [23] and zoonotic CL [24] in Central Asia and Kazakhstan in particular. In most instances, human plague infestation occurs because of epizootics among wild rodent populations, while the sources of infection are linked most closely with fleas and less often with contact with wild animals [25]. Great gerbils live in family groups that inhabit and defend burrow systems (gerbil colonies) [26]. These gerbil colonies are extensive, typically ranging from 15–40 m in diameter [27], but the diameter of a single colony can exceed 50 m with the sizes of colonies dependent on the nature of the soil and vegetation cover [28]. These complex and usually well-marked structures have a pronounced ecological center and periphery, with up to several hundred burrow entrances [29]. The total length of underground passages is on average 300–400 m but will occasionally exceed 1 km [30]. The depth of burrow systems generally averages 2–3 m [30]. The burrows of great gerbils play an important role in desert ecosystems of Central Asia and Kazakhstan because many animal species are associated specifically with the burrows of these rodents [31]. The number, position and size of gerbil colonies do not change over time, but the occupancy of these gerbil colonies can fluctuate greatly, and disease abundance fluctuates in response [32]. The rodent is regarded as an enzootic host for plague, in that the infections have been reported in high seroprevalence, but mass mortalities of gerbils are not often reported [23]. The vectors of plague and CL in the system involving great gerbils are Xenopsylla spp. fleas [32], particularly X. gerbili minax [33,34], and Phlebotomus spp. phlebotomine sand flies [35,36], respectively. Currently, indoor residual spraying (IRS) is one of the primary measures for control of endophagic flying vectors such as mosquitoes and phlebotomine sand flies [37]. Because the success of this type of control is dependent on vectors being endophilic [37], IRS may not be a logical means of control in desert-type areas where sand flies inhabit burrows. Additionally, recent research suggests that current program-initiated IRS control may not adequately reduce sand fly abundance in areas where they are believed to be primarily endophagic and endophilic [38]. A common technique for controlling ectoparasite vectors such as ticks and fleas in burrow systems is to dust rodent burrows with insecticides such as deltamethrin [39,40] or permethrin [41], and this is the preferred method of controlling arthropod pests such as fleas and sand flies inhabiting great gerbil burrow systems in Kazakhstan [25]. Dusting is reported to be costly and have negative impacts on non-target species [42]. Fleas and sand flies have both previously been reported to be resistant to insecticides used in IRS and insecticide dusting campaigns [43,44], because of the large amount of insecticide which is required to perform these applications. Systemic insecticides (endectocides) could provide an additional means of controlling disease vectors which blood feed on desert rodent species, by directly targeting the host with a bait containing a reduced insecticide concentration. The phenylpyrazole, fipronil, is a broad spectrum insecticide which disrupts the central nervous system of insects [45]. The whole blood half-life of a single oral dose (4 mg/kg) administered to rats is 6.2–8.3 days [46] and it is liposoluble which results in prolonged insecticidal effect in organisms [47]. Approximately 45–75% of fipronil may be excreted in rodent feces and 5–25% in the urine [46]. The excretion of fipronil in rodent feces may be beneficial in larval control. Fipronil has been shown to be highly efficacious against Phlebotomus argentipes sand fly adults and larvae when administered to lesser bandicoot rats (Bandicota bengalensis) and roof rats (Rattus rattus) [48]. It has also proved effective against P. papatasi feeding on Meriones shawi under laboratory and field conditions in Tunisia [49]. Additionally, pen and field studies suggest that phlebotomine sand flies and Anopheles spp. mosquitoes blood feeding on fipronil-treated cattle (Bos taurus, Bos indicus) can be reduced significantly [50–52]. Fipronil administered to rodents has also been highly efficacious in reducing on-host ectoparasites such as Ixodes spp. ticks [53,54] and Oropsylla spp. [55] and Xenopsylla spp. [56] fleas. Recently, a rodent grain bait (0.005% fipronil) applied to the openings of black-tailed prairie dog burrows, at a rate of 0.096 mg fipronil/m2, resulted in >95% reduction in on-host fleas infesting prairie dogs for a minimum of 52 days post-treatment application [55]. The potential for a fipronil-based bait to successfully target fleas and sand flies feeding on rodents has been indicated by these previous experiments. However, the impact of a rodent grain bait (0.005% fipronil) on Xenopyslla spp. flea and Phlebotomus spp. sand fly abundance, has never been reported in a field trial in southeastern Kazakhstan or anywhere else in Central Asia. The objectives of this study were to 1) apply 0.005% fipronil bait to active great gerbil burrows in southeastern Kazakhstan and 2) monitor Xenopyslla spp. flea and Phlebotomus spp. sand fly abundance. The great gerbil is the primary reservoir of plague and zoonotic CL in Central Asia and therefore serves as an abundant blood meal source for fleas [39] and phlebotomine sand flies [24]. A reduction in flea and sand fly abundance could potentially lead to a reduction in risk of human plague and CL transmission. The results of this study will help determine the efficacy of a fipronil bait in controlling fleas and phlebotomine sand flies when targeting a host upon which both blood feed. All activities involving animals for this study were reviewed and approved by the Institute of Animal Care and Use (IACUC) at Genesis Laboratories and followed the Animal Welfare Act and Genesis Laboratories Animal Care and Use Guidelines (Approval Date March 18, 2016). Additional approval for animal use was granted by the Animal Care and Use Review Office (ACURO) of the US Army Medical Research and Material Command (Approval for Protocol No. CBMS-FY15-010.03 Dated: April 29, 2016). The authors whose affiliations were with the M. Aikimbaev’s Kazakh Science Centre for Quarantine of Zoonotic Diseases were the only individuals who trapped gerbils in the field. Animals regarded as pest species (such as the great gerbil) do not require permits for trapping within the territory surveyed in southeastern Kazakhstan. The study was conducted in southeastern Kazakhstan (N 44.93621 E 76.02039), ~200 km north of Almaty city and 24 km northwest of Bakanas (June 1-September 3, 2016). The specific timing of study events are available in S1 Table. The Ili River, a major source of moisture in the region [57], was located ~16 km west. The study area was northern subzone desert, composed of flood plains and dunes, and consisting mainly of sandy and clay soils, with saxaul (Haloxylon aphyllum and H. persicum) being the primary vegetation. Two test areas were selected, one where 0.005% fipronil bait was applied to all active burrows (Treatment); and one where all active gerbil burrows remained untreated (Control). The boundaries of the test areas were separated by >400 m. The “Treatment area” was comprised of 1) a ~78750 m2 “Treatment plot”, from which flea and sand fly sampling would be performed, and 2) a “Treatment buffer zone”, extending ~200–600 m from the Treatment plot perimeter, which was established to account for gerbil movement, which has been estimated to be <200 m for >95% of individuals [58]. The “Control area” consisted only of a ~78750 m2 “Control plot” and no buffer zone was needed because no bait was applied. The main criterion for plot selection was the presence of >20 great gerbil colonies within each test plot. Trap sites were selected within the Treatment plot (n = 20) and Control plot (n = 20). GPS coordinates (Garmin Etrex 30, Olathe, KS, USA) were taken for all trap sites and the corners of the Treatment plot, Control plot and Treatment buffer zone (Fig 1). Environmental data (temperature, humidity, precipitation, wind speed) were collected from the nearest accessible monitoring station in Bakanas [59]. A gerbil colony census was conducted to determine the total number of gerbil colonies within the study areas, and their individual occupancy status (occupied, visited, empty). “Occupied” burrow systems were defined as those occupied by a family group. “Visited” colonies showed signs of activity, such as food storage, but were not occupied by a family group [27]. “Empty” colonies were those which had been completely deserted. All gerbil colonies were mapped using a handheld GPS (Garmin Etrex 30, Olathe, KS, USA). Prior to application of fipronil bait, field-collected sand flies and fleas were used to estimate baseline abundance for 2 weeks (June 1-June 15). At treatment Day-0 (June 16) ~120 g of fipronil bait was applied <1 m from each active gerbil burrow of each occupied and visited gerbil colony within the boundaries of the treatment area. Fipronil bait was effectively applied to ~771908 m2. A second analogous treatment application was performed June 21. Vector collection continued to be conducted as was done during the pre-treatment study period (June 16-June 21). Flea and sand fly collection continued to be conducted as was done during the previous study periods (Post-Treatment 1: June 22-July 29). Gerbil trapping, and flea collection were performed again at Study Days 79–80 (Post-Treatment 2: September 3–4) because prior studies have indicated systemic insecticides to suppress flea abundance by up to ~94% at 2-months and up to ~88.5% at 3-months post-treatment [69]. During the treatment and post-treatment periods (June 16-September 4), visual observations were performed 4x/week within the treatment area to monitor for any unexpected abnormal animal behavior, primary and secondary non-target mortality, or other negative signs that could be associated with fipronil bait application. The treatment area was traversed on foot and live animals (gerbils, foxes, etc.) were observed from a distance (>50 m) using binoculars. Any change in animal appearance was recorded. All animal activities performed during this study followed Animal Welfare Act regulations and were approved by the Genesis Laboratories, Inc. Institutional Animal Care and Use Committee (USDA Animal Welfare Act, 9 CFR Parts 1–3) and the Animal Care and Use Office of the US Army Medical Research and Material Command. Gerbil trapping was performed by researchers employed by the M. Aikimbaev’s Kazakh Science Centre for Quarantine of Zoonotic Diseases (Almaty, Kazakhstan), who had permission to perform the live-trapping. Infested great gerbils were defined as individuals having a minimum of one Xenopsylla spp. flea. The efficacy of fipronil bait was estimated by calculating flea and sand fly indices. The flea index was defined as the mean number of Xenopsylla spp. fleas collected per captured gerbil per plot during a single sampling period (Pre-treatment, Treatment, Post-treatment). The sand fly index was defined as the mean number of non-gravid female and gravid female Phlebotomus spp. sand flies collected per trap-night per plot during a single sampling period. Trap-night is defined as the average number of sand flies collected per trap per night of trapping. Male sand flies do not blood feed, feeding exclusively on sugar from plants [70] and hence do not transmit CL. Therefore, they were not used to calculate efficacy. Most female sand flies are gonotrophically concordant [71], requiring one blood meal for each batch of eggs produced, and reduction in gravid female sand flies could indicate a decrease in the rate of reproduction due to the ingestion of a blood meal acquired from a fipronil treated host. Pre-treatment, treatment, and post-treatment mean sand fly and flea indices were used to calculate efficacy of rodent grain bait (0.005% fipronil) in reducing vector abundance. For fleas, efficacy was determined by comparing the Pre-treatment flea index with the flea indices calculated during Treatment, Post-treatment 1, and Post-Treatment 2. For sand flies, abundance fluctuated dramatically during each time point and therefore efficacy was determined weekly during the post-treatment period. the Pre-treatment sand fly index was compared with indices calculated during Treatment, Post-Treatment Week-1 (PTW-1), PTW-2, PTW-3, PTW-4, PTW-5, and PTW-6. Efficacy of fipronil in reducing flea and sand fly indices was adjusted for potential vector reduction within the control plot using an equation described by [72]. Nonparametric statistical methods were used to determine significant differences in relative abundance (p<0.05). Differences in vector abundance occurring during pre-treatment, treatment, and post-treatment were compared between and within the study plots using a Wilcoxon Rank Sum test. Differences in nightly changes (+) in vector abundance in treatment and control were evaluated using a sign-test, to determine whether treatment might influence the tendency for vector abundance in increase and decrease during each trapping period. Environmental conditions in Bakanas were reported June-September 2016 (S2 Table). July was the warmest month (mean: 25.7°C) and September the coldest (mean: 19.7°C). Humidity was highly variable ranging from 6–100%. The greatest precipitation occurred in June (49.6 mm) decreasing exponentially in the months to follow, the least precipitation occurring in September (5.1 mm). The highest wind speed reported was in August (36 km/h) followed by June (32 km/h), the averages being 12.7 and 10.9 km/h, respectively. The occupancy status of individual gerbil colonies was obtained from 156 colonies within the treatment area (buffer = 127 + inner plot = 29) and from 31 colonies within the control plot. Gerbils occupied the majority of the burrow systems surveyed within the treatment (77%) and control areas (68%). Approximately 9% of the total colonies surveyed were “visited” with the remainder being “empty”. During each bait application, a total of 100 kg bait (~5 g fipronil) was used to treat active gerbil burrows of 120 occupied and 14 visited great gerbil colonies within the ~771908 m2 treatment area (Fig 4). Considering the treatment plot area (~771908 m2) and number of colonies treated (n = 134), this amounted to 129.5 mg bait/m2 and 0.006 mg fipronil/m2 being applied during this study (Table 1). No dead or moribund gerbils or non-target animals were observed within or around the treatment area. Additionally, no abnormal animal behavior was observed during the study. However, gerbils within the treated area appeared to have healthier skin and fur (reduced mange, sores) than those within the control area. The results of this study suggest that fipronil bait, applied twice during a 5-day treatment period, at a rate of ~0.006 mg fipronil/m2, 1) may be of reduced risk to great gerbils and non-target wildlife, 2) can significantly reduce or eliminate Xenopsylla spp. fleas for at least 80 days post-bait application, and 3) shows inconsistent efficacy against female Phlebotomus spp. sand flies, suggesting that different methodology such as more frequent applications may need to be implemented. To our knowledge, this is the only field trial, evaluating the use of a fipronil-based rodent bait in controlling fleas and sand flies, to be conducted in Central Asia. No Xenopsylla spp. fleas were collected within the treatment plot after bait application, suggesting that 100% control can be maintained for a minimum of 80 days post-treatment when applied in mid-June. Additionally, fleas were collected from the openings of active gerbil burrows, using handpump aspirators, five times during post-treatment between July 1 and July 29 within the Treatment and Control plot. No X. gerbili minax were collected from burrows within the Treatment plot while multiple (n = 212) were collected within the Control plot. The lack of baseline data for burrow fleas prohibited data analyses. However, the complete absence of X. gerbili minax from the sampled burrow entrances is worth noting, given the 0% X. gerbili minax infestation rate amongst Treatment plot gerbils relative to the 100% infestation rate in Control plot gerbils observed throughout the study. Considering the inability of fleas to survive outside of the burrow systems [22] our results would suggest that seasonal fipronil bait application could potentially remove Xenopsylla spp. fleas from plague-endemic areas. We should note that a single flea, of a different genus (Coptopsylla lamellifer) [81], a moderate plague vector [82], was collected within the treatment plot during September, suggesting that a second autumn treatment might be beneficial. In the future, studies should be designed to establish a baseline for burrow fleas during the pre-treatment period to better estimate the efficacy of the bait. Although treatment against fleas was successful, the precise rate of decline in efficacy of the bait was not calculable because 100% efficacy was still being achieved at test termination. [69] saw up to 90-day efficacy of imidacloprid-based bait in reducing fleas infesting ground squirrels and [48] found fipronil to be superior to imidacloprid for sand fly control. [55] reported fipronil efficacy of >90% against fleas infesting prairie dogs for at least 52-days post-initial application. Knowing the length of time required to achieve significant flea reduction would be highly beneficial to managers and might further suggest that two treatments performed in spring or autumn might be sufficient to reduce flea infestations. Fipronil efficacy against gravid sand flies was >70% during the treatment period and up to 100% during PTW-3, with reduced efficacy observed afterwards. Although gravid females were reduced markedly during these periods, 1) their relative abundance was low when compared with general sand fly abundance, 2) sand fly abundance differed significantly during the pre-treatment, treatment, and post-treatment periods within each test plot, and 3) fluctuations in relative abundance did not differ significantly between plots. Field populations of sand flies are sensitive to climatic variables such as temperature, strong winds and heavy rain [83]. The reduced efficacy during PTW-2 and PTW-3 may have been a result of the increase in windspeed. Several researchers have studied the movement of Phlebotomus spp. sand flies and have suggested that although they typically fly short distances, they may occasionally move distances more than 1–2 km [64,83,84]. If P. mongolensis is capable of flying distances >1 km with wind assistance, then it would suggest that reinvasion of the plots by gravid females may have occurred. If future studies indicate that this a persistent issue that needs to be addressed, future studies may consider treating more frequently and/or increasing the size of the buffer zone to better account for sand fly movement. However, the logistics of conducting a study within such a large area will need to be carefully considered. Sand fly abundance during PTW-5 and PTW-6 decreased markedly within both plots, possibly in response to increase in precipitation. Although not a subject of this paper, future studies might focus on explicitly evaluating the flight potential of sand flies and the influence of various climatic variables on sand fly abundance in southeastern Kazakhstan. Several additional ecological factors may have been responsible for the shorter duration and inconsistency of sand fly efficacy. Fleas are ectoparasites that rely heavily on the host they infest and cannot survive outside of the burrow systems [32]. Sand flies are less host-dependent in that only adult female sand flies blood feed and all other life stages (eggs, larvae, pupae) develop in organic matter. Male sand flies do not blood feed and instead feed exclusively on sugars from plants [70] and therefore were not exposed to the bait. Given development from egg-adult can range from 4-weeks to several months dependent on temperature [76], many developing sand flies may have not been exposed to fipronil-treated gerbils during the peak in fipronil blood concentration when efficacy is at its highest, and many sand flies likely emerged as the fipronil concentration in blood declined. This might explain why reduced sand fly efficacy was seen after PTW-3 and would suggest that monthly fipronil application might be a better means of controlling phlebotomine sand flies under these field conditions. Gravid females prefer to oviposit their eggs in areas containing organic material [85]. Previous researchers have shown fipronil excreted in animal feces to be highly efficacious in reducing laboratory reared phlebotomine sand fly larvae [48–50,86]. However, while many researchers have studied the potential oviposition sites of sand flies, there is a deficit of information regarding natural oviposition sites [87] and attempts to collect immature sand flies from the field have proven difficult and unproductive. For example, a researcher in Sudan processed ~2500 kg of soil to recover only a single sand fly larva [88]. Therefore, while it is speculated that sand flies may oviposit near animal feces, we cannot say with certainty that sand fly larvae were exposed to great gerbil feces. While this fell outside of the scope of our study, investigations of the natural oviposition locations of phlebotomine sand flies should be continue to be pursued, as this knowledge would greatly improve sand fly control programs. Managers must consider differences in ecology of fleas and phlebotomine sand flies and establish a clear distinction between the methodology for controlling these respective vectors of plague and CL. A concern associated with field application of insecticides is possible resistance of the target vector. The cat flea (Ctenocephalides felis), is resistant to multiple insecticides [89], but [90] reported no resistance in strains of cat fleas exposed to topical fipronil. We are confident that great gerbils consumed all of the fipronil bait immediately, without storing it in food chambers. We base this assertion on the fact that these large gerbils do not possess cheek pouches like hamsters [91]. The still images and videos taken with trail cameras further support the belief that gerbils consumed the bait. If great gerbils consumed all bait within a single feeding this would lessen the risk of potential insecticidal resistance. However, during future field applications, surveillance of the flea population in treated areas should be conducted to monitor possible insecticide resistance [92]. Occupancy status of gerbil burrows fluctuates rapidly [32] and it seems intuitive that fipronil-based treatment would be most appropriate when occupancy is high and thus vector abundance and disease transmission are subsequently high. A recent agent-based entomological model suggested that fipronil treatment had greater potential to reduce sand flies when applied during periods when the vector population is high and thus more individuals are acquiring blood meals while fipronil efficacy is at its peak [40]. We observed 100% reduction in Xenopsylla spp. flea abundance when treatment was applied twice in early summer when the gerbil colony occupancy rate was ~77%. This level of efficacy was sustained until autumn. Given the rate of gerbil colony occupancy can differ in spring and autumn, it would be beneficial to perform another treatment in autumn and compare the efficacy with that of treatment performed in spring/summer, as [93] found that the effects of dusting prairie dog burrows on flea abundance were longer lasting in autumn when compared to spring. Taking into consideration the fact that socio-economic factors could limit the number of feasible program-initiated fipronil treatment applications, comparing treatments performed in spring and autumn could help to determine the optimal timing in which treatment should be applied. [94] suggested that satellite imagery can be used to estimate gerbil colony occupancy, with only 2% error when compared with direct observation, with the potential to predict plague outbreaks. If occupancy can be correlated with fipronil bait efficacy, it is then possible that satellite imagery, in conjunction with recent data sets regarding vector and host abundance, could also be used to determine the timing of treatment application. As mentioned previously, no great gerbil or non-target wildlife fatalities were observed over the course of the experiment. Unlike other methods such as IRS or burrow dusting, the fipronil-based grain bait is applied orally to hosts, targeting blood feeding vectors exclusively, which reduces environmental exposure, minimizing contact with non-target wildlife or insects [52]. The gerbils within the treatment area appeared healthier than those within the control area (cleaner skin, fur), which could have possibly been a biproduct of bait application which reduced flea infestation from 100–0%. As previous researchers have already mentioned [55,69,95] we cannot be certain no non-target collateral occurred given observations were only performed above ground during the day. However, given the low fipronil concentration (0.005%) and an application rate of ~0.006 mg fipronil/m2, which was nearly 16x lower than what was used by [55], we can strongly suspect that negative non-target effects were reduced. The acute oral LD50 of fipronil in rodents is estimated to be 97 mg/kg body weight [46]. At this rate, a gerbil, weighing ~169–275 g [96], would need to consume ~>325–535 g bait in one feeding to attain the acute oral LD50, a feat which would be improbable. Fipronil is more toxic to fish and at least three gamebird species (Alectoris rufa, Phasianus colchicus, Colinus virginianus) [97]. Exposure of fish to this treatment is highly improbable in the desert, and fipronil binds to soil and has low solubility in water suggesting reduced risk to aquatic organisms under field conditions [97]. The most vulnerable bird species to be studied is the northern bobwhite quail (C. virginianus), which is native to the United States, with an LD50 of 11.3 mg/kg. Bobwhite quail weighing on average between 140–170 g [98] would need to consume ~>30–40 g of bait in one feeding. This is less than what a mammal would have to eat, and less than what quail consume on average per day (~20 g) [99]. The most common bird species observed within the study plots, and the one with the closest association with the rodent burrows was Oenanthe isabellina, a small insectivorous passerine [100]. Fipronil has been shown to be less toxic to passerines such as Spizella pusilla (LD50: 1120 mg/kg) and Taeniopygia guttata (LD50: 310 mg/kg) [97]. It is encouraging that no bird mortalities were observed during the study, but also not surprising given the low concentration of fipronil in the bait (0.005%). To put it another way, while 100 kg bait was applied to the 771908 m2 plot during each application, the total amount of fipronil applied during each application was marginal (~5 g). In contrast, dusting with permethrin, the preferred control method in Kazakhstan, does not target the host explicitly and is effective if applied at rates of up to 2 g permethrin per individual burrow [41]. Previous largescale field trials conducted in Wyoming, during which fipronil was applied as a spray to 33% of two 347-ha treatment areas at a rate of ~0.4 mg/m2, determined the risk of fipronil to birds and non-target insect species to be far lower than that of alternative insecticidal compounds because it could be applied at rates of 100-200x less than alternative insecticides [101]. This is termed the “reduced exposure” approach, and strongly supports the argument that a 0.005% fipronil bait applied at a rate >65x lower than that of these Wyoming field trials should have reduced non-target effects. This does not however imply that non-target organisms should not be monitored, and more research should be conducted to determine the risk of this proposed treatment to various vertebrate species, and to investigate measures to further reduce non-target risk. While efficacy against Xenopsylla spp. fleas exceeded requirements outlined by the EPA [73] caveats need to be addressed. Gerbils were collected 3x during Pre-Treatment (June 1-June 15), 1x during Treatment (June 16-June 21), 9x during Post-Treatment 1 (June 22-July 29), and 2x during Post-Treatment 2 (September 1–4). The trapping effort might be considered inconsistent and was largely a byproduct of trapping success and the availability of the individuals who were approved to perform the trapping. The flea numbers were also relatively low during the pre-treatment period. However, flea infestation was at 100% during pre-treatment and was reduced to 0% following treatment application. In contrast, flea infestation remained a constant 100% within the control plot throughout all study periods. While we are confident in our results, we proceed with caution, and recommend that future studies incorporate a more uniform trapping methodology and ensure that flea indices during the baseline period meet pre-determined criteria. We also note that while abundant fipronil laboratory and field data are available for other rodents such as roof rats [48], lesser-bandicoots [48], jirds [49], and black-tailed prairie dogs [55], laboratory-based fipronil studies involving great gerbils have not been conducted. While not a scope of this project, laboratory-based studies would be a useful addition to this field work to determine the fipronil half-life in the blood, feces and urine of great gerbils, as well as quantify its efficacy against both vectors. This information could aid in the design of future field studies and could be used to estimate parameters in predictive simulation modelling. The benefits of oral fipronil treatment may not be limited to fleas and sand flies under the current conditions. It was observed during the experiments that tick genera (Hyalomma, Rhipicephalus, Haemaphysalis, and Ornithodoros) within the treatment area declined markedly during the experiment. Previous studies have indicated that fipronil is efficacious against Ixodes scapularis [53,54] which transmit Borrelia burgdorferi, the causative agent of Lyme disease. Future studies conducted in this region would benefit from evaluating the ability of fipronil-based treatment to reduce tick abundance, which, if determined to be efficacious, would increase the versatility of the bait, allowing for the potential to control several tick, flea and sand fly-borne diseases. The efficacy of fipronil bait against fleas infesting great gerbils in southeastern Kazkhstan, and black-tailed prairie dogs in northern Colorado [55], suggest that fipronil bait application could be effective in a number of unique biotopes. The great gerbil has a wide distribution in Asia, also being found in Afghanistan, Iran, Turkmenistan, Kyrgyzstan, Pakistan, Tajikistan, Uzbekistan, Mongolia, and China [102]. Environmental and geographic conditions within these countries inevitably differ from southeastern Kazakhstan. Therefore, we cannot say with absolute certainty that the results of this study would be applicable to all biotopes in Central Asia. However, we argue that these results indicate field trials in other plague and CL-endemic Central Asian countries are warranted. Our study suggests that fipronil-based bait may serve as a promising new tool for reducing disease vectors in southeastern Kazakhstan, and its potential should be further investigated. X. gerbil minax fleas were reduced by 100% after fipronil bait was applied twice (June 16, June 21), but efficacy against gravid female phlebotomine sand flies was inconsistent. In addition to the potential efficacy against fleas, our study suggests that low-concentration fipronil-based treatment may present reduced risk for both gerbils and non-target species in the area. While efficacy against fleas was significant, modifications in bait application timing and frequency will be needed to adequately reduce sand fly abundance. In conclusion, fipronil-based bait, applied in southeastern Kazakhstan, has the potential to reduce or potentially eliminate Xenopsylla spp. fleas if applied every 80-days, and could reduce gravid Phlebotomus spp. sand flies if applied more frequently. However, the study would benefit from a larger number of gerbils being sampled and more uniform trapping methodology. This form of treatment would likely be best applied in small areas where threats to humans or endangered species are of concern. Future studies should aim to monitor the vector populations over a longer period and should consider applying bait during more than one season, to determine the longevity of fipronil-based grain bait efficacy and determine the optimum timing and frequency of application. Additionally, future studies should consider the possibility of applying multiple treatments over multiple years to evaluate efficacy, insecticide resistance, and risk to non-target organisms. Fipronil-based bait may provide a means of controlling blood-feeding disease vectors in Central Asia and other regions globally.

10.1371/journal.pntd.0004073

Syndromic Approach to Arboviral Diagnostics for Global Travelers as a Basis for Infectious Disease Surveillance

Arboviruses have overlapping geographical distributions and can cause symptoms that coincide with more common infections. Therefore, arbovirus infections are often neglected by travel diagnostics. Here, we assessed the potential of syndrome-based approaches for diagnosis and surveillance of neglected arboviral diseases in returning travelers. To map the patients high at risk of missed clinical arboviral infections we compared the quantity of all arboviral diagnostic requests by physicians in the Netherlands, from 2009 through 2013, with a literature-based assessment of the travelers’ likely exposure to an arbovirus. 2153 patients, with travel and clinical history were evaluated. The diagnostic assay for dengue virus (DENV) was the most commonly requested (86%). Of travelers returning from Southeast Asia with symptoms compatible with chikungunya virus (CHIKV), only 55% were tested. For travelers in Europe, arbovirus diagnostics were rarely requested. Over all, diagnostics for most arboviruses were requested only on severe clinical presentation. Travel destination and syndrome were used inconsistently for triage of diagnostics, likely resulting in vast under-diagnosis of arboviral infections of public health significance. This study shows the need for more awareness among physicians and standardization of syndromic diagnostic algorithms.

Physicians attending travelers with particular symptoms often neglect those infections that are transmitted by arthropods like ticks and mosquitoes (arboviruses) or don’t test for the appropriate arboviruses. This is because arboviruses cause symptoms that are similar to more common infections and because there is a geographical overlap in the arbovirus infections that people have a large chance of being infected with. We compared the amount of times that Dutch physicians had patients tested for arboviral infections with the likelihood that Dutch travelers would be exposed to particular arboviruses. Whereas research and diagnostics generally focus on only one virus, our study was uniquely comprehensive and systematic in that it analyzed multiple viruses simultaneously on the basis of a unique national database. The research shows that the current viruses travelers are tested for is incomplete and likely many more people carry these kinds of diseases than are diagnosed. As these diseases pose potential public health threats, physicians should be more aware of the diseases that travelers could be infected with, and protocols are needed regarding what infectious diseases physicians should check for when patients present with particular symptoms.

Globalization has resulted in a steep increase in travel and trade.[1, 2] In recent decades it has contributed to the spread of diseases that traditionally emerged only regionally but now threaten populations across the globe, stressing the need for global health surveillance.[1, 2] Among these emerging threats, arboviruses form a unique group, with a large public health impact in endemic countries, a tendency to expand their geographical distribution through trade and travelers, and colonize previously unaffected areas. Due to their vector-borne and often zoonotic nature, they require targeted surveillance and control schemes. This requirement is particularly relevant when evaluating symptoms of illness in travelers. Of all those returning from developing, tropical, or subtropical countries, 8% require medical care on return.[3] For those returning from Africa and Southeast Asia, fever is the most common reason for seeking medical care; for travelers returning from the Caribbean and South America, rash is the most common reason. Around 50% of the cases remain undiagnosed in clinics focused on travel medicine, and this percentage is likely higher in less specialized clinics.[3] The traveler’s personal physician is therefore an important link in ongoing arbovirus surveillance in travelers and the gate-keepers of disease detection. Correct diagnosis of arbovirus infections in travelers is challenging. Arboviruses have overlapping geographical distributions and cause symptoms that coincide with more common infections.[4] If general practitioners consider an arbovirus infection in their differential diagnosis, they commonly test for the best known arboviral disease, Dengue virus (DENV). Laboratory diagnostics for travelers are largely based on serologic testing, since viremia is short-lived and has often already dropped to undetectable levels when severe symptoms appear and diagnostics are performed.[5, 6] The use of serologic results for arbovirus diagnosis and surveillance requires careful evaluation due to cross-reactivity of antibodies to related viruses.[7] Also, several vaccines, notably for Yellow fever, Tick-borne encephalitis and Japanese encephalitis, can cause false-positive serological tests.[7] For these reasons, arbovirus illness is under-diagnosed, as evidenced by studies of unexplained illness in returned travelers.[8–10] A potential solution would be the development of syndromic arboviral disease detection methods that cover the most common arboviruses and simultaneously provide surveillance information.[11] Here we aimed to assess the potential added value of syndrome-based approaches for diagnosis and surveillance of neglected arboviral diseases in returning Dutch travelers. To map the patients high at risk of missed clinical arboviral infections in returned Dutch travelers, we compared the quantity and quality of all arboviral diagnostic requests by Dutch physicians, from 2009 through 2013, with a previously extensive literature-based assessment of travelers’ likely infection with an arbovirus.[4] The overlapping syndromes and geography, based on and updated from that review are depicted in Fig 1. For retrospective patient analysis, a database was created by integrating data from the two arbovirus diagnostic reference centers in the Netherlands: Erasmus Medical Centre in Rotterdam and The National Institute for Public Health and the Environment in Bilthoven. Previously, we described trends of DENV diagnostics in the Netherlands from 2000–2010.[12] The current study included almost all arbovirus diagnostic requests from Dutch physicians from 2009 through 2013 in the Netherlands. In the case of DENV not all data was included because 10% of the DENV diagnostics were performed outside the arbovirus reference centers and were not included in this dataset. For syndromic analysis, only entries were included where travel and clinical history were provided. To define the syndromes, entries in the database were reviewed by a consultant microbiologist, and infectious disease clinicians assigned them to syndrome categories (Table 1). Due to the laboratory-specific variety in diagnostic methods used, we classified each patient’s test results according to the validated methods and cut-offs for the pertinent laboratory. Results were classified as positive for a disease if the patient had (1) a positive PCR result with <40 cycles, (2) an IgM above an individual laboratory-determined cut-off, or (3) a minimum fourfold increase in IgG titers between two consecutive samples. For DENV patients, (4) a positive non-structural protein 1(NSI) antigen-capture test was among the criteria.[6] The likelihood of infections by arboviruses other than DENV was based on a previously published article in which we developed syndromic diagnostic algorithms based on data from an exhaustive review of the literature addressing geographic distribution and prevalence of arboviruses by syndrome.[12] Optimal diagnostic algorithms using a combination of clinical syndromes and geographical distribution presented were updated and used as a basis for our current analysis (Fig 1). In short, criteria used to prioritize arboviruses for the diagnostic algorithm were: a) circulation in urbanized areas, due to the use of humans as reservoir hosts, or the presence of reservoir hosts colonizing urban areas, b) known endemic disease, c) tourist activity in the area, d) high rate of exposure in resident population, and e) recorded cases of infections in travelers.[4] These diagnostic algorithms were used in the current article to identify gaps that may occur with a physician-indexed single-virus approach. Travel data for Dutch travelers was based on the year 2011. They were extracted from a commercial database “ContinuVakantieOnderzoek” (CVO) created for trend analysis in the tourism industry. Its data are collected and converted into national numbers every three months by interviewing individuals in about 15,000 Dutch households on their travel destinations, activities, lodging, transport, and booking method.[13] Using data from 2011 provided a representative distribution of Dutch travel behavior from 2009–2011. Only slight country specific fluctuations were reported.[13] The analysis was performed in STATA.[14] Pearson's chi-square test was used to assess for equality of proportions. Multivariable logistic regression models (Table 2) reporting odds ratios were used with a 95% confidence level.[14] Heatmaps were generated using the additional R package “stats”[15] and based on pair-wise correlation between rows and columns. This research was conducted in accordance with the Dutch law on medical research (WMO), article 1. In compliance with Dutch Law and medical ethical guidelines, no personal identifiers were included and no informed consent was required for use of data in this study. Over the five year study period 8126 patients were tested for arboviral diseases in the Netherlands. Of the patients, 44% presented to larger hospitals or specialized travel clinics. All other patients were seen at smaller hospitals or local clinics. Molecular tests comprised 1.3% of diagnostic tests performed. Larger hospitals and specialized travel/tropical clinics tested on average for 1.7 viruses per patient compared to 1.2 in smaller hospitals and local clinics. The patient male to female ratio was 1.04. Vaccination history was recorded on the diagnostic request for only 14 patients (<1%). Of all patients, 2153 (26%) had information on travel history and clinical history and were thus included for further syndrome and travel-based analysis. Of these, 23% had provided a second serum sample needed for determination of a potential IgG titer increase. With a median of 7 days, the average number of days elapsed between onset of symptoms and first sampling was 17.5 (95%CI 14.0–20.3). This number is based on the 317 patients with clinical and travel history for whom this chronological information was recorded. Elapsed time did not differ between patients seen at specialized hospitals/clinics and those visiting smaller hospitals/clinics. We analyzed the travel data of Dutch travelers in 2011 to determine the range and importance of arbovirus tests needed to cover the differential diagnosis for travelers with illness after return from the various destinations. In 2011, approximately 84% of Dutch travelers traveling abroad stayed within Europe. Western Asia (predominantly Turkey) was the most popular non-European destination, with nearly one million Dutch vacations booked annually (Fig 2).[13] The most diagnostic requests (35%) by far, however, were for travelers returning from destinations in South and Southeast Asia, while only 3% of all travelers had this region as their destination. The number of diagnostic requests by travel region and the proportion of positive test results (Fig 2) show that DENV testing was by far the most commonly requested (86%), yielding the highest absolute number of cases (Fig 2). When comparing the numbers of requests and proportions of positives by region of travel, substantial differences were observed: diagnostic requests for ill travelers returning from sub-Saharan Africa were frequent but not often positive, whereas ill travelers returning from popular arbovirus-endemic regions in Central and Western Asia were rarely tested. A low number of patients who had traveled within Europe were tested. DENV was tested (N = 41) almost as often as tick-borne encephalitis virus (TBEV) (N = 57), for which exposure is far more likely. Of note, two of these European travelers tested DENV-positive. One was a tourist returning from Croatia, who tested DENV-IgM-positive and borderline NS1-positive. The other tourist had taken a five-day trip to Southern France and was DENV IgM- and NS1-positive 10 days after return. However, 14 days previous to onset of symptoms, this traveler had been in Thailand before traveling on to France. Another virus considered endemic to Europe is Sindbis virus (SINV), for which diagnostics are not readily available in the Netherlands. Nor are they available for oropouche virus (OROV), endemic to South America. To assess the potential use of diagnostic requests for syndrome surveillance by region, we analyzed the symptoms recorded for each patient returning from a particular travel destination. Nearly all patients (86%) reported fever, followed by arthralgia/arthritis (22%) and enteric symptoms (14%). Information divided per travel region showed regional variation in symptoms recorded (Fig 3). For all regions, fever was the most reported symptom. Proportionally, neurological symptoms were more often reported for travelers returning from a European destination than for travelers from other regions. Arthralgia-arthritis was recorded more frequently for travelers returning from Oceania, with rash being most recorded for Southern Africa compared to other regions. Three heatmaps were created to visualize per continent (Africa, Asia and the Americas) the correlation between the physicians’ diagnostic requests and the literature-based syndromic algorithms (Fig 1). In the heatmaps, diagnostic requests are grouped based on the clinically important arboviral diseases per region within each continent (Figs 4–6). For most regions, Dutch physicians requested DENV diagnostics for 100% of the travelers who had recorded symptoms corresponding to DENV infection (fever, rash and joint pain). For some regions, a lower percentage of such patients was tested, i.e. Northern Africa (67%) (Fig 4), Western Asia (57%) (Fig 5) and Central America (38%) (Fig 6). In all regions, CHIKV testing was less frequently requested than DENV testing, even though the infections overlap in geographical distribution and range of symptoms to a great extent. On average, 45% of patients with febrile symptoms, rash and/or arthralgia after travel to CHIKV-risk areas in Asia were not tested for CHIKV. Patients with symptoms suggesting West Nile Virus (WNV), Japanese encephalitis (JEV), Rift Valley fever virus (RVFV) and TBEV were tested infrequently (0 to 25%) and only in association with neurological symptoms. Diagnostics on all other viruses presented in Figs 4–6 were minimally requested. We analyzed the association between symptoms recorded and test outcomes for DENV and CHIKV requests in Dutch travelers (Table 2). Patients with rash, hemorrhagic symptoms and fever had an increased odds of testing positive for DENV, but respiratory symptoms decreased the odds of being DENV-positive (OR 0·5). Positive test outcomes for CHIKV were associated with arthralgia combined with rash. Both DENV and CHIKV were positively associated with travel history to Southeast Asia. Here we assessed the extent of missed arboviral infections in travelers by a retrospective database analysis of all arboviral diagnostic requests in the Netherlands, from 2009–2013, in comparison with a literature-based assessment of arbovirus exposure while traveling (Fig 1). We found clear evidence for patient groups high at risk of being under-diagnosed for arboviral disease when evaluated by syndrome and by region. While DENV diagnostics are routinely requested, other relevant arboviruses are neglected, in particular CHIKV. Arthralgia, for example, is not only associated with DENV infections but also with many arboviruses, including CHIKV, as we found when calculating odds ratios within the current Dutch data.[4] Nevertheless, less than 55% of patients with symptoms compatible with CHIKV infection were tested (Fig 4a–4c). Interestingly, hemorrhagic symptoms and rash have a much higher odds ratio than arthralgia-arthritis for diagnosing DENV. Although arthralgia is an important symptom in dengue patients, rash and fever are often more pronounced.[16] In the case CHIKV arthralgia-arthritis is more pronounced and is known to have a higher predictive value for distinguishing CHIKV from DENV in endemic settings.[16, 17] Additionally, CHIKV is less well known by physicians in non-endemic countries so might be only considered if DENV diagnostics are negative. The analysis of diagnostic requests by region showed a bias toward the more well-known arboviral risk areas such as Southeast Asia (Figs 2 and 3). For travelers within Europe, arbovirus diagnostics are rarely requested, despite high incidence rates of TBEV reported across Europe and continuing circulation of WNV in parts of Europe popular with Dutch tourists.[8, 18] This is a general trend also seen in previous reports on travel associated infection presenting in Europe.[19] Housing type and location during travel is an import risk factor for exposure to specific vectors,[20, 21] and outdoor camping is popular among travelers in Europe.[13] The number of CHIKV and DENV requests within Europe was almost equivalent to the number of TBEV and WNV test requests, while only a small number of CHIKV and DENV have been reported.[22–24] The low number of TBEV and WNV requests may reflect a lack of physician awareness of European arboviruses and their risk to travelers; it may also reflect financial restrictions or limited time.[18] Our analysis showed that physicians were more likely to extend the diagnostic panel for patients with more severe or very specific symptoms. For instance, diagnostics for WNV and Western equine encephalitis virus (WEEV) were usually requested only for patients with neurological complaints, even though fever is the most common clinical presentation in >90% of WNV and WEEV patients.[25] Similarly, RVFV diagnostic requests were limited to patients with hemorrhagic symptoms (HS) and neurological symptoms (NS), although these severe symptoms occur in less than 1% of cases, and most patients present only with febrile symptoms (Fig 4a–4c).[26] This bias toward severe symptoms was likewise reflected by the finding that patients referred to large hospitals and travel clinics were more extensively evaluated than those visiting small hospitals and local clinics. Reasons for this difference were not assessed in our study but are likely related to the fact that 1) general practitioners often omit arbovirus diagnostics, due in part to budgetary constraints; 2) they may lack knowledge on arboviral disease, and 3) may believe that an arbovirus diagnosis is unlikely to influence their treatment decisions, particularly if symptoms are mild. However, even mild arbovirus infections can eventually cause severe or chronic symptoms like arthralgia and, in any case, they pose a potential risk to health workers. Lack of proper diagnosis may lead to unnecessary complications or extensive later testing of patients. A possible solution to this problem is diagnostic centers providing syndromic and region based diagnostic packages for travelers as presented by the algorithms here.[4] These can be continuously updated in collaboration with specialized physicians and Public Health professionals. This will relieve the general physicians from keeping up to date on such a complex and continuously changing area. At the same time physicians are provided with a complete diagnostic selection and data are more suitable for use in surveillance. Our results show a large variation in the timing of first diagnostic sampling. In our study, 50% of travelers contacted a healthcare provider during the first week of illness. This means that 50% did not, and viremic patients may introduce viruses into a region, when appropriate vectors are available, [23, 27] or pose a risk for nosocomial infection.[28, 29] Only 1.3% of all diagnostic tests performed were molecular, while 50% of patients fell within the range advised for molecular testing. The timeframe for molecular and serological diagnostics overlap to a great extent. Within the first days of illness, however, serology has a low sensitivity.[6] A number of the DENV cases may have been secondary, tertiary or quaternary infections. This reduces the sensitivity of serological detection by IgM in non-primary infections significantly.[6] Many patients are therefore probably missed due to lack of molecular testing within this timeframe. To use diagnostic data for syndromic surveillance, a two-tiered approach could be employed. First, samples collected after three days of illness onset would provide syndromic information by multiplex serologic testing. Second, if testing showed increased circulation of a target virus, confirmation and genomic surveillance would follow in patients suspected to harbor that virus sampled within seven days of illness. There are a number of limitations to this study. Nearly all patients tested for arboviral diseases in the five-year-period in the Netherlands were included. This group, however, only consists of patients that seek medical attention after travel and that are suspected of an arboviral infection by a clinician. Asymptomatic patients and patients where clinicians did not consider an arboviral disease are missed. Almost all patients lack vaccination history. Patients with recent yellow fever vaccinations could cause positive false positive serological tests.[6] Lack in reporting vaccination history and the resulting possible flavivirus cross-reactivity due to vaccination are known problems when using flavivirus serological diagnostic data.[9] Both diagnostic centers had extensively validated tests internally with yellow fever vaccines and changed diagnostic cut-offs provided by manufacturer to compensate if possible. However, false positive tests due to vaccination cannot be excluded. Infectious disease diagnostics and surveillance of travelers is primarily focused on those cases or diagnostic outcomes selected and reported by physicians.[30–33] Although this approach provides essential information, many patients remain undiagnosed, and re-evaluation of the selected pathogens has been advised.[10, 31, 34] However, much knowledge on probable arbovirus exposure of travelers is based on information originating from the destination country, which may have limited surveillance and diagnostic capabilities. In some of these countries, large-scale surveillance projects using a more syndromic approach to infectious diseases have shown extensive under-diagnosis and under-recognition of the importance of many arbovirus diseases as a cause of common syndromes.[35, 36] This underlines the need, in the Netherlands and other affluent countries, for more systematic syndrome-based diagnosis and surveillance in travelers to these regions. It demonstrates the added value of using routine travel information to support national and international surveillance programs. For such surveillance, capturing only a fraction of all cases may still provide reliable information on disease trends and possibly local outbreaks, provided the selection is systematic.[9] It is also important in terms of preparedness for emerging infectious diseases. A physician’s diagnostic requests for returned travelers can play a key role in infectious disease surveillance. However, while travel destination and syndrome could be used for triage and diagnostics, such use is inconsistent. We found clear evidence of patient groups at risk of under-diagnosis of arboviral disease when evaluated by syndrome and by region. Based on a comparison between all arboviral diagnostic requests by physicians in the Netherlands between 2009 and 2013 with a literature-based assessment of the likely exposure of the patients to an arbovirus, we showed that while dengue virus diagnostics are routinely requested, other relevant arboviruses such as chikungunya virus are neglected, even if travelers present with relevant symptoms and return from countries where the viruses are endemic. We also showed that for travelers to European destinations, arbovirus diagnostics were rarely requested and that for almost all arboviruses and travel destinations, diagnostics were requested only when patients presented with severe symptoms. Whether the low number of requests and overemphasis of physicians on patients presenting with severe symptoms reflects a lack of physician awareness of arboviruses and their risk to travelers, financial restrictions or limited time, it points at possible gaps in preparedness. Our paper shows that in order to limit the amount of missed clinical arboviral infections, and to increase the level of awareness of arboviral infections of public health significance, physicians should rely on diagnostics and surveillance with a syndromic approach and matching laboratory methods.

10.1371/journal.pntd.0000444

Community Management of Endemic Scabies in Remote Aboriginal Communities of Northern Australia: Low Treatment Uptake and High Ongoing Acquisition

Scabies and skin infections are endemic in many Australian Aboriginal communities. There is limited evidence for effective models of scabies treatment in high prevalence settings. We aimed to assess the level of treatment uptake amongst clinically diagnosed scabies cases and amongst their household contacts. In addition, we aimed to determine the likelihood of scabies acquisition within these households over the 4 weeks following treatment provision. We conducted an observational study of households in two scabies-endemic Aboriginal communities in northern Australia in which a community-based skin health program was operating. Permethrin treatment was provided for all householders upon identification of scabies within a household during home visit. Households were visited the following day to assess treatment uptake and at 2 and 4 weeks to assess scabies acquisition among susceptible individuals. All 40 households in which a child with scabies was identified agreed to participate in the study. Very low levels of treatment uptake were reported among household contacts of these children (193/440, 44%). Household contacts who themselves had scabies were more likely to use the treatment than those contacts who did not have scabies (OR 2.4, 95%CI 1.1, 5.4), whilst males (OR 0.6, 95%CI 0.42, 0.95) and individuals from high-scabies-burden households (OR 0.2, 95%CI 0.08, 0.77) were less likely to use the treatment. Among 185 susceptible individuals, there were 17 confirmed or probable new diagnoses of scabies recorded in the subsequent 4 weeks (9.2%). The odds of remaining scabies-free was almost 6 times greater among individuals belonging to a household where all people reported treatment uptake (OR 5.9, 95%CI 1.3, 27.2, p = 0.02). There is an urgent need for a more practical and feasible treatment for community management of endemic scabies. The effectiveness and sustainability of the current scabies program was compromised by poor treatment uptake by household contacts of infested children and high ongoing disease transmission.

Like many impoverished areas around the world, Aboriginal communities in Australia experience an unacceptably high burden of scabies, skin infections, and secondary complications. Young children are most at risk. Our study investigated scabies in a remote setting with very high rates of skin disease, a high level of household overcrowding, and limited infrastructure for sanitation and preventive health measures. We assessed uptake of scabies treatment and scabies acquisition following provision of treatment by a community-based skin program. In a household where scabies was present, we found that treatment with topical permethrin cream of all close contacts can significantly reduce a susceptible individual's risk of infection. Our findings also demonstrate the challenges of achieving a high level of treatment participation, with limited permethrin use observed among household contacts. This suggests an urgent need for a more practical treatment option. International efforts to reduce childhood morbidity and mortality have demonstrated the efficacy of numerous child health interventions but have also highlighted the deficits in their delivery and implementation. Experiences like this, where the effectiveness of a coordinated local program delivering an efficacious intervention is hampered by poor treatment uptake and ongoing transmission, are an important and timely message for researchers, program managers, and policy-makers.

Skin infections are a significant cause of morbidity in disadvantaged settings around the world [1],[2]. Scabies and pyoderma are endemic in many Aboriginal communities in northern Australia. These conditions cause local morbidity and contribute substantially to clinic workload and costs [3]. Moreover, the primary bacterial pathogen underlying most pyoderma in these communities is Group A Streptococcus (GAS), which can cause a myriad of debilitating secondary complications [4], including acute nephritis. Recent evidence also suggests a link between GAS skin infection and rheumatic fever, and consequently rheumatic heart disease [5]. Post-streptococcal disease rates in Aboriginal Australians are among the highest in the world [6]. Scabies is thought to underlie the majority of bacterial skin infections in these communities [7], thus controlling scabies is critical to improving skin health and reducing the burden of GAS secondary complications. Scabies is a disease that often accompanies poverty, with high prevalence being consistently associated with crowded living conditions [8]–[10]. Previous experience suggests community-based mass-treatment approaches are likely to be the most effective for control of scabies and skin sores in remote Aboriginal communities [1], [11]–[13]. Similar models are used to control endemic parasitic and other diseases around the world [14]–[16]. Substantial reductions in the prevalence of scabies in endemic settings have previously been described with mass community treatment using oral ivermectin [17] and topical permethrin [18],[19]. However, sustainability has been difficult to achieve [2],[12], particularly where there is high mobility between communities and households. A high level of community participation is critical to the success of community-based programs such as this. To maximise treatment uptake during both mass distribution and routine screening, the treatment must be acceptable and feasible in the setting to which it will be applied [20]. A skin health program incorporating mass annual distribution of permethrin cream, routine scabies screening and treatment at the clinic and in homes, has been operating in the remote Aboriginal communities of the East Arnhem region of the Northern Territory since 2004. Over this time, scabies prevalence amongst children in the region remained unchanged at 13% [21]. Due to the ongoing scabies burden in these communities we aimed to assess levels of treatment uptake among households in which one or more members was seen with scabies during routine screening. We also sought to investigate acquisition of scabies among household members during the month after the initial visit. The study was conducted in two Aboriginal communities participating in the East Arnhem Regional Healthy Skin Program in remote northern Australia. A review of presentations to the community health clinics in 2006 revealed that 63% of infants had been diagnosed with scabies and 69% with skin sores in the first year of life [3]. The skin program included annual mass community treatment for scabies (a designated “Healthy Skin Day”), combined with ongoing screening, treatment and follow-up of children aged <15 years in each community. On Healthy Skin Day, all individuals were encouraged to apply topical permethrin 5% (Lyclear), which was distributed directly to all households in the community. Residents were advised to apply the cream all over the body and leave it on overnight or for a period of at least 8 hours. Healthy Skin Day was held within the month prior to the commencement of data collection for this study. The East Arnhem program included routine skin assessments of children under 15 years throughout the year, which aimed to identify and treat reinfestation arising after the mass community treatment. These were conducted at the community clinic, during school screening programs and on home visits. Where a child with scabies was identified, permethrin cream was provided for all household members. Locally-relevant resources were produced and used by local workers to teach people about scabies and skin health during the assessments. The program team included doctors, nurses, dermatologists, Aboriginal health workers and locally employed community workers. A Darwin-based team visited regularly to support the local community workers. The program had been operating in these communities for 2.5 years when this nested study commenced. The nested study population was households where one or more children had been diagnosed with scabies during routine home visits. Data were collected between December 2006 and June 2007. Informed written consent was given by all participating households. This study was approved by the Human Research Ethics Committee of the Northern Territory Department of Health and Menzies School of Health Research (Approval number 04/11). A member of the Darwin-based team visited each community every two weeks during the 6-month study period to assist the community workers in recruitment and follow-up of households. Children with scabies were identified during home visits for routine skin screening undertaken as part of the existing skin health program (Day 0). The first child from a given household seen with scabies became the index child for that household. In line with normal skin program practice, the mother/carer of this child was provided with permethrin cream for all household members, together with standardised verbal instructions for cream use. Local guidelines for scabies treatment indicate that all household members should be treated, regardless of individual scabies status. Household members included the index case and all individuals considered by the index child's mother/carer to be living in the house at that time, as occurs during normal skin health program procedures. All household members other than the index case are referred to as the index cases' household contacts. Instructions were given for the index child and all other household members to use the cream that night. As noted, provision of treatment was according to normal skin health program procedures, and was not dependent on participation in this study. The date of birth, sex, scabies and skin sore status of each household member was recorded. Household members with skin sores were referred to the clinic for treatment. Recommendations regarding environmental measures to eradicate scabies from the household were also provided, which included washing all clothing and linen. Participation in these activities was not necessary for an individual or household to be considered to have participated fully in treatment. Evidence indicates that the role of fomites in transmission of scabies is minimal [22],[23]. However we consider these environmental recommendations to be positive public health messages, particularly in a highly overcrowded setting such as this. On Day 1 we revisited the home and asked the mother/carer which household members had used the treatment. Treatment uptake was defined at the individual and household level. For an individual, we accepted either self-report or report from the primary carer that the household member had used the cream. Complete household treatment uptake occurred if all household members were reported to have met the definition for individual treatment uptake. If household treatment participation was incomplete, the mother/carer was asked if she could say why household members had not used the cream. Subsequent home visits were undertaken on Days 14 and 28 to screen household members for scabies and skin sores. Additional permethrin cream was provided as required at these visits. Utilisation of cream provided after Day 1 was not assessed. Scabies was diagnosed clinically by specifically-trained staff, using accepted criteria and based on the nature and distribution of characteristic scabies lesions [24],[25]. This is the standard and widely accepted approach to scabies diagnosis in an endemic setting such as this [17],[26],[27]. Skin sores were also diagnosed clinically, based on the presence of any crusted, purulent or dry sores. Where a household member was not present during the home visit, we asked the primary carer in the household to report whether the individual had scabies and/or sores. Findings are described according to whether a trained Healthy Skin Worker (HSW) screened the individual or if skin status relied on family member report. When measuring scabies acquisition (defined as an individual who was scabies-free at Day 0 and had a clinical diagnosis of scabies within 28 days) we classified incident cases as either confirmed or probable. “Confirmed” scabies acquisition required an individual to be seen by a HSW both at baseline (with a clinical diagnosis as scabies-free) and again at a follow-up (clinical diagnosis of scabies) Report of scabies acquisition by a family member was considered “probable”. These were individuals who were: a) clinically diagnosed as scabies-free at baseline by a HSW and then reported to have scabies by a family member at the follow-up visit or; b) were not seen by a HSW at baseline but were reported to be scabies-free by a family member and were subsequently diagnosed as scabies by a HSW at follow-up or were reported to have scabies by a family member at the follow-up visit. Data were analysed using Intercooled Stata 9 (Stata Corporation, College Station, Texas, USA). Comparisons of continuous variables across two levels of a binary variable were conducted using the Mann Whitney U test. For comparison of proportions between two binary variables, relative risks and 95% confidence intervals (CI) were calculated. Chi square test or Fisher's Exact test were used to test the significance of associations where appropriate. To assess the relative contribution of independent (explanatory) variables to individual treatment uptake, logistic regression analysis was implemented using the method of marginal models estimated using Generalised Estimating Equations (GEE) with information sandwich estimates of variance. GEE was used due to the potential for clustering among households. An exchangeable correlational structure was used. These analyses were performed in Stata using the xtgee command. In exploring differences between susceptibles who did and did not acquire scabies over the one-month follow-up period, GEE was also employed given the potential for clustering by household. Since subgroups contained small numbers of individuals, there was insufficient power to test all explanatory variables by multivariate GEE model. Therefore, simple (univariate) GEE was performed for each independent variable of interest. This provided a measure of the significance of the association between acquiring scabies and each independent variable while controlling for the potential influence of household clustering. Forty households participated, involving 596 individuals (40 index children, 556 household contacts). Median household size was 15.5 persons (IQR 12, 20). At baseline (Day 0), a median of 23.6% (IQR 11.0, 44.7) of individuals in each household had scabies as determined by either HSW screening or family member report (Table 1). All households in which a child with scabies was identified during normal skin program activities agreed to participate in the study. Our study is the first to investigate levels and determinants of treatment uptake in a scabies endemic setting. Treatment uptake among index children was over 70%, suggesting that it is possible to achieve high levels of individual treatment use. However, treatment uptake among the household contacts of these children was poor and there were very few households in which all members were reported to have used the treatment. We are likely to have achieved a higher level of uptake in this study sample than would otherwise have occurred, given that householders were aware we would return the following day. Routine Healthy Skin program procedures do not include this home visit the day after treatment provision. In addition, self-reported treatment uptake is likely to be subject to reporting bias, which may also have resulted in an overestimation of treatment uptake. Therefore, even under circumstances of increased motivation to use the treatment and potential over-reporting of uptake, observed rates of treatment use were still very poor amongst household contacts. This low participation in treatment has important implications for the likelihood that sustained reductions in scabies burden can be achieved with the community-based model employed here. Our findings support current recommendations for universal treatment of close contacts where scabies is present, with the odds of scabies acquisition being greatest among young children and individuals in households with incomplete treatment uptake. Even in an endemic setting characterised by high mobility between households, universal treatment among family members in households where scabies was present significantly reduced the likelihood of acquisition among susceptible individuals. This is critical to protect young children, who are most at risk of infection. However, we have also demonstrated that the current approaches to achieving universal treatment are neither feasible nor effective in this setting. Although there was an established community-based control program in the two communities in which this study was conducted, we observed a very high level of secondary transmission. While the risk of secondary transmission remained unacceptably high in both communities, it was five times higher in Community A than Community B. The former had both a comparatively higher household burden of scabies and lower levels of treatment uptake amongst household contacts, which helps to explain the differential. Overall, reported treatment uptake was poorest in settings with a high scabies burden. This was observed at both the community level and at the household level. As noted, Community A had a higher median household scabies burden compared to Community B, and significantly less treatment uptake. In addition, irrespective of community of residence, an individual from a household with a high scabies burden was less likely to use the treatment than an individual from a household with a low scabies burden. The depressing reality for many households in these communities may well be that scabies has become part of life. At the individual level, treatment failure, or failure of others to adequately treat, creates an ongoing cycle of scabies transmission within both the household and the community. This may well contribute to the likelihood that those households with high scabies burden become increasingly less likely to treat. A number of community-based initiatives have documented a sustained reduction in scabies burden in endemic settings using topical permethrin. In Panama, directly observed permethrin treatment of all inhabitants of an island community, together with ongoing surveillance and treatment of new cases, resulted in a significant reduction in scabies prevalence [18]. Similar results have previously been reported in Australian Aboriginal communities without the requirement of directly observed treatment [12],[13]. That the achievements of previous programs have not been observed here may reflect community-specific factors, such as social or environmental characteristics, which may not be generalisable to other communities or regions. Indeed, even in the current study we identified marked differences in treatment participation and household scabies burden between the two participating communities. Furthermore, given the low levels of treatment participation observed here, it seems unlikely that complete participation would have been achieved during the mass community treatment event that took place in the weeks prior to our study. If a substantial and rapid reduction in community scabies prevalence could be achieved through complete participation in mass treatment, the subsequent management of new cases at the household level may be a more feasible and effective approach. However, the high level of movement within and between households and communities would still present significant challenges to maintaining a low prevalence of scabies. The difficulty in achieving a sustained reduction in scabies burden in these settings has previously been recognised [2],[12],[18]. In Australia and many other countries, topical preparations such as permethrin cream are the only approved treatment for community management of scabies. The practicality of topical treatment for the community management of endemic scabies has been questioned [9],[25]. Indeed, several characteristics of the treatment and setting give reason to doubt that a sustained reduction in scabies prevalence can be achieved with this approach. Environmental factors make total-body topical treatment impractical. These factors include the large number of people in each house, high heat and humidity, limited opportunities for privacy to apply the cream [1], and poor infrastructure for washing it off [28]. When complete treatment uptake does occur, there is a strong likelihood of rapid reinfestation due to the high prevalence of scabies, overcrowding and frequent movement between households and communities. A realistic consequence of this is low motivation to repeat the treatment process. The likelihood of individual participation in treatment for the community management of endemic parasitic infection has been linked to expected personal benefit and expected personal cost [29]. It seems likely that the low level of uptake observed here, particularly where there is a high burden of scabies, is symptomatic of low motivation to participate in a treatment regime that is onerous (high time and inconvenience) and has been seen to have limited effectiveness (low personal benefit). This is supported by the reasons cited for non-participation in treatment, the most common of which was that using the treatment wasn't a priority. Many also reported the inconvenience and unpleasantness of the treatment to be a key barrier to use. Of additional concern is the potential for the development of drug resistance [30],[31] when such long-running community disease control programs achieve only limited participation and disease reduction. Concerns regarding mite resistance to permethrin have recently been described in a number of Aboriginal communities in northern Australia [32],[33]. Thus it is possible that even if greater levels of treatment participation could be achieved, resistance to this treatment may undermine any potential impact on disease burden. These findings demonstrate an urgent need for a more suitable treatment for scabies to reduce the burden in endemic settings. Oral ivermectin has demonstrated success in the community management of endemic scabies [17],[26],[27]. For example, in the Solomon Islands, mass ivermectin administration to all residents of a small number of islands, combined with treatment of any individuals subsequently entering or returning to the communities, reduced scabies prevalence from 25% to <5% after four months, and was subsequently sustained at <1% [17]. Ivermectin is also widely used in community control of other parasitic infestations. Since 1987, it has been used extensively in community-based mass treatment programs in Africa and Latin America to control endemic onchocerciasis [34]. The Global Programme to Eliminate Lymphatic Filariasis recently reported that between 2000 and 2007, 149 million treatments of ivermectin had been given through community-based mass drug administration in 12 African countries [35]. Ivermectin is also effective against other parasitic infestations that can occur in high-scabies burden settings, such as strongyloidiasis [26], which is endemic in many Australian Aboriginal communities [36]. Ivermectin is not currently approved for the mass community management of scabies in Australia. Notwithstanding, a growing body of literature indicates it is safe and effective when used in mass drug treatment programs and is ideally suited for use in the community as it is a single oral dose that is easily administered. In 2003, it was estimated that 6 million people worldwide had taken ivermectin for various parasitic infestations with no serious drug-related adverse events reported [37]. Contributing to the safety profile is the accumulation of non-event data among pregnant women [38]–[40]. While ivermectin presents a viable alternative for the management of scabies, especially where compliance with topical treatment is improbable or impractical [34],[37], the provision of a more practical treatment alone is unlikely to completely resolve the low treatment participation seen here. Many households cited reasons for non-participation that may not be readily resolved simply with a more practical treatment. For example, there is an enduring perception that individuals without scabies do not need to be treated. Factors that have been identified as important for treatment participation and program success in mass ivermectin administration include strong community education and awareness-raising, engaging well-trained health workers who are trusted and respected by the community, and community involvement and ownership in the program [41]–[43]. It will be critical to consider such factors in the implementation of any initiative to reduce the burden of scabies in these communities, and to work together with communities to better understand household and individual barriers to program participation. Our study has several limitations that are inherent to research in this setting. With large and transient family groups, the concept of a household unit is tenuous and follow-up of individuals is compromised. While complicating data collection, this also further highlights the challenge of infectious disease management at the household level. In our study, almost half of the susceptible individuals were lost to follow up. We must therefore consider the possibility that those individuals who were followed up represent a biased sample of the population at risk. Those who were lost to follow up were significantly less likely to belong to a household in which all members used the treatment, and less likely to use the treatment themselves. This suggests that we have missed a segment of the susceptible population that may indeed be at a higher risk of acquiring scabies, and we may therefore have underestimated acquisition. Where a household member was unwilling or unavailable to participate in skin screening, a family member reported scabies status for that individual. Ideally, all household members would have been available for skin screening. The presence of itching and visible skin lesions in a household where scabies is known to be present has been reported to be highly specific [11]. Notwithstanding, here we have distinguished between scabies cases diagnosed by a trained worker (confirmed case) and those reported by a family member (probable case) in order to mitigate any potential impact of this on the validity of our data. There is limited evidence for the effective, long-term management of scabies in high prevalence areas [44]. Achieving a significant reduction in infectious disease burden in an endemic setting requires a high level of treatment coverage among those exposed to the disease. Here we have demonstrated a lower likelihood of scabies acquisition when all close contacts participate in treatment. This emphasises the importance of all close contacts of scabies cases being treated, whether symptomatic or not. This is recommended in all settings where scabies is present. However, achieving this high level of participation in treatment can pose a considerable challenge. Our findings indicate that the long-term management of scabies with topical permethrin has not been effective, as evidenced by low rates of treatment uptake and ongoing transmission. These results support the notion that, in controlling infectious diseases such as scabies at the community level, it is critical that the treatment be appropriate, acceptable and feasible in the setting to which it will be applied. It is also fundamental to recognise that the eradication of scabies and other infectious diseases in these settings cannot be achieved by treatment alone. To realise a significant and sustained reduction in disease burden requires that the underlying environmental and social conditions that promote such poor health be addressed.

10.1371/journal.pcbi.1001059

A Computational and Experimental Study of the Regulatory Mechanisms of the Complement System

The complement system is key to innate immunity and its activation is necessary for the clearance of bacteria and apoptotic cells. However, insufficient or excessive complement activation will lead to immune-related diseases. It is so far unknown how the complement activity is up- or down- regulated and what the associated pathophysiological mechanisms are. To quantitatively understand the modulatory mechanisms of the complement system, we built a computational model involving the enhancement and suppression mechanisms that regulate complement activity. Our model consists of a large system of Ordinary Differential Equations (ODEs) accompanied by a dynamic Bayesian network as a probabilistic approximation of the ODE dynamics. Applying Bayesian inference techniques, this approximation was used to perform parameter estimation and sensitivity analysis. Our combined computational and experimental study showed that the antimicrobial response is sensitive to changes in pH and calcium levels, which determines the strength of the crosstalk between CRP and L-ficolin. Our study also revealed differential regulatory effects of C4BP. While C4BP delays but does not decrease the classical complement activation, it attenuates but does not significantly delay the lectin pathway activation. We also found that the major inhibitory role of C4BP is to facilitate the decay of C3 convertase. In summary, the present work elucidates the regulatory mechanisms of the complement system and demonstrates how the bio-pathway machinery maintains the balance between activation and inhibition. The insights we have gained could contribute to the development of therapies targeting the complement system.

The complement system, which is the frontline immune defense, constitutes proteins that flow freely in the blood. It quickly detects invading microbes and alerts the host by sending signals into immune responsive cells to eliminate the hostile substances. Inadequate or excessive complement activities harm the host and may lead to immune-related diseases. Thus, it is crucial to understand how the host boosts the complement activity to protect itself and simultaneously establishes tight surveillance to attain homeostasis. Towards this goal, we developed a detailed computational model of the human complement system. To overcome the challenges resulting from the large model size, we applied probabilistic approximation and inference techniques to train the model on experimental data and explored the key network features of the model. Our model-based study highlights the importance of infection-mediated microenvironmental perturbations, which alter the pH and calcium levels. It also reveals that the inhibitor, C4BP induces differential inhibition on the classical and lectin complement pathways and acts mainly by facilitating the decay of the C3 convertase. These predictions were validated empirically. Thus, our results help to elucidate the regulatory mechanisms of the complement system and potentially contribute to the development of complement-based immunomodulation therapies.

The complement system is pivotal to defending against invading microorganisms. The complement proteins recognize conserved pathogen-associated molecular patterns (PAMPs) on the surface of the invading pathogens [1] to initiate the innate immunity response. The complement activity also enhances adaptive immunity [2], [3] and participates in the clearance of apoptotic cells [4] as well as damaged and altered self tissue. The complement proteins in the blood normally circulate as inactive zymogens. Upon stimulation, proteases in the system cleave the zymogens to release active fragments and initiate an amplifying cascade of further cleavages. There are three major complement activation routes: the classical, the lectin and the alternative pathways [5]. Regardless of how these pathways are initiated, the complement activity leads to proteolytic activation and deposition of the major complement proteins C4 and C3, which induces phagocytosis, and the subsequent assembly of the membrane attack complex which lyses the invading microbes. However, complement is a double-edged sword; adequate complement activation is necessary for killing the bacteria and removing the apoptotic cells, while excessive complement activation can harm the host by generating inflammation and exacerbating tissue injury. Dysregulation of the balance between complement activation and inhibition can lead to rheumatoid arthritis [6], systemic lupus erythematosus [7], Alzheimer's disease [8] and age-related macular degeneration [9]. Since the final outcome of complement related diseases may be attributable to the imbalance between activation and inhibition [10], manipulation of this balance using drugs represents an interesting therapeutic opportunity awaiting further investigation. In light of this potential, complement inhibitors such as factor H and C4b-binding protein (C4BP) are critical since they play important roles in tightly controlling the proteolytic cascade of complement and avoiding excessive activation. Therefore, a systems-level understanding of activation and inhibition, as well as the roles of inhibitors, will contribute towards the development of complement-based immunomodulation therapies. Complement is usually initiated by the interaction of several pattern-recognition receptors with the surface of pathogens. C-reactive protein (CRP) [11] and ficolins are two initiators of the classical and lectin pathways, which boost immune responses by recognizing phosphorylcholine (PC) or N-acetylglucosamine (GlcNAc), respectively, displayed on the surface of invading bacteria [12], [13], [14]. Recently, it was discovered that under local infection-inflammation conditions as reflected by pH and calcium levels, the conformations of CRP and L-ficolin change which leads to a strong interaction between them [15]. This interaction triggers crosstalk between classical and lectin pathways and induces new amplification mechanisms, which in turn reinforces the overall antibacterial activity and bacterial clearance. On the other hand, C4BP, a major complement inhibitor is synthesized and secreted by the liver. The estimated plasma concentration of C4BP is 260 nM under normal physiological condition [16] but its plasma level can be elevated up to four-fold during inflammation [17], [18]. Through its α-chain [19], [20], C4BP modulates complement pathways by controlling C4b-mediated reactions in multiple ways [21], [22], [23]. Further, C4BP has been proposed as a therapeutic agent for complement-related autoimmune diseases on the premise that mice models supplemented with human C4BP showed attenuation in the progression of arthritis [24]. Therefore, it is important to understand the systemic effect and the underlying inhibitory mechanism of C4BP. With this background, we constructed a detailed computational model of the complement network consisting of a system of ordinary Differential equations (ODEs). The large model size and the many unknown kinetic rate parameters lead to significant computational challenges. Using the technique developed in [25], we approximated the ODE dynamics as a dynamic Bayesian network [26] and used it to estimate the model parameters. After constructing the model, we investigated the enhancement mechanism induced by local inflammation and its interplay with the inhibition mechanism induced by C4BP. Our studies confirmed and further elucidated the previous experimental findings [15]. Specifically, using our model we established a detailed relationship between the antimicrobial response and the strength of the crosstalk between CRP and L-ficolin as determined by various combinations of the pH and calcium levels. We also found that C4BP prevents complement over-activation and restores homeostasis, but it achieves this in two distinct ways depending on whether the complement activity was initiated by PC or GlcNAc. Finally, the computational model suggested that the major inhibitory effect of C4BP is to potentiate the natural decay of C3 convertase (C4bC2a). These findings regarding the role of C4BP were experimentally validated. An earlier mathematical study [27] of the complement system focused on the classical pathway. This study assumed the dynamics to be linear, which is a severe restriction. A later study by Korotaevskiy et al [28] more realistically assumed the dynamics to be non-linear. It also included the alternative pathway. The main focus was to derive quantitative conclusions regarding the lag time of the immune response as the initial concentrations of the constituent proteins were varied. Relative to [28], our model additionally includes the lectin pathway and the recently identified amplification pathways induced by the crosstalk between CRP and L-ficolin [15]. On the other hand, given our focus on the up- and down- regulation mechanisms of the complement, we do not model the alternative pathway in detail since its role is to maintain a basal level of complement activation. Instead, this basal activity and the effects of other mechanisms such as C2 bypass [29] are implicitly captured by the kinetic parameters in our model. We start with an overview of our results before getting into details. The ODE model we constructed is based on the current knowledge of the complement system. It includes the classical and lectin pathways and the recently identified amplification pathways [15]. Thus it is a more accurate reflection of the complement system in comparison to the earlier models [27], [28]. The model consist of 42 species, 45 reactions and 85 kinetic parameters with 71 of the parameters being unknown. To deal with the large number of unknown kinetic parameters, we applied the technique developed by Liu et al [25] and derived an approximation of the ODE model as a Dynamic Bayesian Network (DBN) [26]. We then performed parameter estimation using the DBN approximation and experimentally generated test data. The resulting model was validated using published observations of bacterial killing rates [15]. We next performed sensitivity analysis of the model, which showed that the interaction between CRP and L-ficolin as well as the decay of the C3 convertase played crucial roles. Since the decay of C3 convertase is one of the regulatory targets of C4BP, which had been identified as an important regulator [22], [23], [24], we decided to investigate the role of C4BP under enhanced complement activity as a part of our study. Next we used the model to investigate the effects of pH and calcium levels on the antibacterial response guided by previous experimental findings [15]. We computed detailed response curves of the antibacterial activity as a function of pH and calcium levels which confirmed the findings in [15] and further elucidated the enhancement of the cross talk between the classical and lectin pathway under varying local inflammation conditions. We then turned to the inhibitory role of C4BP. We found that for PC-initiated complement activation, increasing the C4BP initial concentrations delayed the time taken to achieve the peak amplitude of the response but it did not significantly alter the magnitude of the peak response itself. On the other hand, for GlcNAc-initiated complement activity, C4BP levels affected the magnitude but not the time taken to achieve this peak response. Our subsequent experimental results agreed with these observations. Finally, we investigated how C4BP mediates its inhibitory function. According to current understanding, there are four major mechanisms through which C4BP can inhibit complement activation. Using the model, we determined that facilitating the natural decay of C3 convertase is the most important inhibitory function of C4BP. This finding was also validated experimentally. Given our focus on the amplification and down-regulation mechanisms of complement, we included in our model only the key proteins in the classical and lectin pathways. The basal activity maintained by the alternative pathway and other mechanisms are implicitly captured by the kinetic parameters in our model. A schematic representation of the model structure is shown in Figure 1A. The cascade of events captured by the model can be described as follows. The classical pathway is initiated by the binding of antibodies or CRP to antigens or PAMPs. In our model, in order to decouple the involvement of adaptive immune response, the classical pathway is triggered by the binding of CRP to PC, which is a ligand often displayed on the surface of the invading bacteria [30], [31]. Deposited CRP then binds to C1-complex (formed by C1q, two molecules of C1r, and two molecules of C1s) that is further activated. The activated C1-complex recruits C4 leading to the cleavage of C4 to its fragments, C4b and C4a. After binding of C2 to C4b, the same protease complexes are responsible for generating fragments, C2a and C2b, by cleaving C2. The C2a and C4b then form the C4bC2a complex, which is an active C3 convertase, cleaving C3 to C3a and C3b. The formation of C3b exposes a previously hidden thioester group that covalently binds to patches of hydroxyl and amino groups on the bacterial surface [32]. The surface-deposited C3b plays a central role in all subsequent steps of the complement cascade: (1) it acts as an opsonin that enhances the binding and leads to the elimination of bacteria by the phagocytes, (2) it induces the formation of membrane attack complex leading to the lysis of bacteria. Since the concentration of the deposited C3 reflects the antibacterial activity of complement, we terminated our model at this step to simplify the network. On the other hand, the lectin pathway is initiated by the binding of mannose-binding lectin (MBL) or ficolins to PAMPs on the pathogen surface. In our model, we focused on the lectin pathway initiated by L-ficolin as it can interact with CRP and induce crosstalk between classical and lectin pathways. L-ficolin recognizes various PAMPs on the bacterial surface via the acetyl group on the GlcNAc moiety [33], [34]. Therefore, in our model the lectin pathway was triggered by binding of L-ficolin and GlcNAc onto the bacterial surface. Subsequently, a protease zymogen called MASP-2 is recruited and activated. Activated MASP-2 cleaves C4 and C2 to form C4bC2a which is C3 convertase. At this point, the classical pathway and lectin pathway merge at the cleavage step of the central complement protein, C3, and hence constitutes the endpoint of our model. As discovered in [15], infection-induced local inflammation conditions (slight acidosis and hypocalcaemia) provoke a strong crosstalk between CRP and L-ficolin [15]. This elicits two new complement-amplification pathways, which reinforce the classical and lectin pathways. Since we aimed to study the complement activation and modulation under pathophysiological conditions, we included these two amplification pathways (Figure 1A, purple) in our model. Infection by bacteria containing PC will induce the CRP:L-ficolin mediated amplification pathway: PC→CRP:L-ficolin→MASP2→C4→C2→C3. On the other hand, infection by bacteria containing GlcNAc will induce the CRP:L-ficolin mediated amplification pathway: GlcNAc→CRP:L-ficolin→C1→C4→C2→C3. The complement allows a rapid attack to intruding bacteria while at the same time protecting the host cells from over-activation. C4BP, a major inhibitor of complement activation, was reported to either accelerate the decay of the convertases or aid proteolytic inactivation of key players in the pathway into inactive forms such as factor H [32] but the systemic effect of C4BP has remained unclear. Hence, in our model, we included this major multifunctional inhibitor. Upstream of the complement cascade, C4BP competes with C1 for the immobilized CRP [23]. Downstream to this, C4BP binds to C4b and serves as a cofactor to the plasma serine protease factor I in the cleavage of C4b both in the fluid phase and when C4b is deposited on bacterial surfaces [21]. In addition, C4BP is able to prevent the assembly of the C3 convertase and accelerate the natural decay of the complex [35]. All of the above effects of C4BP are considered in our model and the relevant components are depicted as red bars in Figure 1A. The reaction network diagram of the model is shown in Figure 1B. Processes such as protein association, degradation and translocation are modeled with mass action kinetics and processes such as cleavage, activation and inhibition with Michaelis-Menten kinetics. The resulting ODE model consists of 42 species, 45 reactions and 85 kinetic parameters with 71 unknown. The details can be found in the supporting information (Text S1). Due to the large model size and many unknown kinetic parameters, tasks such as parameter estimation and sensitivity analyses became very challenging. Hence, we applied the probabilistic approximation technique developed by Liu et al [25] to derive a simpler model based on the standard probabilistic graphical formalism called Dynamic Bayesian Networks (DBNs) [26]. Briefly, this approximation scheme consists of the following steps: (i) Discretize the value space of each variable and parameter into a finite set of intervals. (ii) Discretize the time domain into a finite number of discrete time points. (iii) Sample the initial states of the system according to an assumed uniform distribution over certain intervals of values of the variables and parameters. (iv) Generate a trajectory for each sampled initial state and view the resulting set of trajectories as an approximation of the dynamics defined by the ODEs system. (v) Store the generated set of trajectories compactly as a dynamic Bayesian network and use Bayesian inference techniques to perform analysis. A more detailed description of this construction can be found in the Methods section while we explain in the Discussion section how we fixed the number of trajectories to be generated and the maximum time point upto which the trajectories are to be constructed. In the ODE model the PC-initiated and GlcNAc-initiated complement cascades are merged for convenience. By suppressing these two cascades to one at a time (by setting the corresponding expressions in the reaction equations to zero), we constructed two dynamic Bayesian networks; one for the PC-initiated complement cascade and the other for GlcNAc-initiated complement cascade. The range of each variable and parameter was discretized into 6 non-equal size intervals and 5 equal size intervals, respectively. The time points of interest were set to {0, 100, 200,…,12600} (seconds). Each of the resulting DBN approximations encoded trajectories generated by sampling the initial values of the variables and the parameters from the prior, which was assumed to be uniform distributions over certain intervals. The quality of the approximations relative to the original ODEs dynamics was sufficiently high and the details can be found in the supporting information (Figure S1). The values of initial concentrations and 14 kinetic parameters were obtained from literature data (Table S1 and Table S2). To estimate the remaining 71 kinetic parameters, we generated test data by incubating human blood under normal and infection-inflammation conditions with beads coated with PC or GlcNAc followed by immunodetection of the deposited CRP, C4, C3 and C4BP in time series. For PC-beads, the concentration levels of deposited CRP, C4, C3 and C4BP were measured at 8 time points from 0 to 3.5 h (Figure 2A,B, red dots). For GlcNAc-beads, the concentration levels of deposited MASP-2, C4, C3 and C4BP were also measured at 8 time points from 0 to 3.5 h (Figure 2C,D, red dots). To estimate unknown kinetic parameters, a two-stage DBN based method [25] was deployed. In the first stage, probabilistic inference applied to the discretized DBN approximation was used to find the combination of intervals of the unknown parameters that have the maximal likelihood, given the evidence consisting of the test data. As mentioned above, each unknown parameter's value space was divided into 5 equal intervals and the inference method called factored frontier algorithm [35] was used to infer the marginal distributions of the species at different time points in the DBN. We then computed the mean of each marginal distribution and compared it with the time course experimental data. To train the model by iteratively improving fitness to data, we modified the tool libSRES [36] and used its stochastic ranking evolutionary strategy (SRES), to search in the discretized parameter space consisting of 571 combinations of interval values of the unknown parameters. The result of this first stage was a maximum likelihood estimate of a combination of intervals of parameter values. In the second stage we then searched within this combination of intervals having maximal likelihood. Consequently, the size of the search space for the second stage was just 1/571 of the original search space. We used the SRES search method and the parameter values thus estimated are shown in Table S2. In principle, given the noisy and limited experimental data and the high dimensionality of the system, one could stop with the first stage [37] and try to work an interval of values for each parameter rather than a point value. However, in our setting we wanted to use the ODE model too for conducting in silico experiments such as varying initial concentrations including the down and over expression of C4BP. This would have been difficult to achieve by working solely with our current DBN approximation. We address this point again in the Discussion section. Figure 2A–2D shows the comparison of the experimental time course training data (red dots) with the model simulation profiles generated using the estimated parameters (blue lines). The model predictions fit the training data well for most of the cases. In some cases, the simulations were only able to reproduce the trends of the data. This may be due to the simplifications assumed by our model and further refinement is probably necessary. We next validated the model using previously published experimental observations [15]. In particular, normalized concentration level of deposited C3 was used to predict the antibacterial activity since C3 deposition initiated the opsonization process and the lysis of bacteria. We first simulated the concentration level of deposited C3 at 1 h under different conditions. We next normalized the results so that the maximum value among them equals to 95% which is the maximum bacterial killing rate reported in the experimental observations [15]. The normalized values were then treated as predicted bacterial killing rates. The simulation results are shown in Figure 2E and 2F as black bars. Consistent with the experimental data (Figure 2E, grey bars), our simulation showed that under the infection-inflammation conditions, the P. aeruginosa, a clinically challenging pathogen, can be efficiently killed (95% bacterial killing rate) by complement whereas under the normal condition, only 28% of the bacteria succumbed (Figure 2E, black bars). Consistent with experimental data, our simulation results show that in the patient serum, depletion of CRP or ficolin induced a significant drop in the killing rate from 95% to 33% or 25% respectively, indicating that the synergistic action of CRP and L-ficolin accounted for around 40% of the enhanced killing effect. However, in the normal serum, depletion of CRP or ficolin only resulted in a slight drop in the killing rate from 28% to 18% or 10% respectively. Furthermore, simulating a high CRP level (such as in the case of cardiovascular disease) under the normal healthy condition did not further increase the bacterial killing rate. As shown in Figure 2F, the simulation results matched the experimental data. Thus, our model was able to reproduce the published experimental observations shown in both Figure 2E and 2F with less than 10% error. This not only validated our model thus promoting its use for generating predictions, but also yielded positive evidence in support of the hypothesized amplification pathways induced by infection-inflammation condition. It also suggested that the antibacterial activity can be simulated efficiently by the level of deposited C3 and this was used to generate model predictions described in later sections. We performed local and global sensitivity analysis of the model to identify species and reactions that control complement activation during infection, and to evaluate the relative importance of initial concentrations and kinetic parameters for the model output. To identify critical species, we first calculated the scaled absolute local sensitivity coefficients [38] for initial concentrations of major species using the COPASI tool [39]. The model outputs were defined as the peak amplitude (maximum activation) and integrated response (area under the activation curve that reflects the overall antibacterial activity) of C3 deposition. The results are shown in Figure 3A. Both the peak amplitude and integrative response were strongly influenced by initial concentrations of C2 and C3, and were mildly influenced by initial concentrations of C4BP, C1 and C4. In contrast, the low sensitivities of CRP, MASP-2 and L-ficolin indicate that over-expression of these proteins is unlikely to increase the antibacterial activity. Interestingly, it was observed that the integrative response was more sensitive than the peak amplitude to the changes in the initial concentration of PC. Since the concentration of PC is correlated to the amount of invading bacteria, this result implies that the maximum complement response level may not increase as the amount of bacteria increases but the overall response (i.e. the area under the curve obtained by integrating the response level over time) will be enhanced to combat the increased number of bacteria. In order to identify critical reactions, we next computed global sensitivities for kinetic parameters. To reduce complexity, we used the DBN approximations. Multi-parametric sensitivity analysis (MPSA) [40] was performed on the DBN for PC-initiated complement cascade (the details are presented in the Materials and Methods section). The results are shown in Figure 3B. Strong controls over the whole system are distributed among the parameters associated with the immobilisation of C3b with the surface, interaction between CRP and L-ficolin, cleavage of C2 and C4, and the decay of C3 convertase (see Figure 1B, reactions labeled in red). The sensitivity of reactions associated with C3, C2 and C4 is consistent with the local sensitivity analysis, which highlighted the significant role of major complement components. The high sensitivity of interaction of CRP and L-ficolin confirms that the overall antibacterial response depends on the strength of the crosstalk between the classical and lectin pathways. In addition, since the decay of C3 convertase is one of the regulatory targets of C4BP, the sensitivity of the system to a change in the rate of decay of C3 convertase suggested that the regulatory mechanism by C4BP plays an important role in complement. Since the critical reactions identified are common in PC- and GlcNAc-initiated complement cascades, MPSA results using the other DBN will produce similar results and hence this analysis was not performed. We next focused our investigation on the enhancement mechanism by the crosstalk and the regulatory mechanism by C4BP. Under infection-inflammation conditions where PC-CRP:L-ficolin or GlcNAc-L-ficolin:CRP complex is formed, the amplification pathways are triggered. Model simulation showed that if C1 and L-ficolin or CRP and MASP-2 competed against each other, the antibacterial activity of the classical pathway or lectin pathway might be deprived of the amplification pathways (see Figure S2). Therefore, in order to achieve a stable enhancement, C1 and L-ficolin (or CRP and MASP-2) must simultaneously bind to CRP (or L-ficolin). Further, the abilities of CRP and L-ficolin to trigger subsequent complement cascade were not affected by the formation of this complex. This is consistent with the previous experimental observation that two amplification pathways co-exist with the classical and lectin pathways [15]. According to [15], slight acidosis and mild hypocalcaemia (pH 6.5, 2 mM calcium) prevailing at the vicinity of the infection-inflammation triggers a 100-fold stronger interaction between CRP and L-ficolin compared to the normal condition (pH 7.4, 2.5 mM calcium). This can be explained by the fact that the pH value and calcium level influence the conformations of CRP and L-ficolin which in turn govern their binding affinities. Therefore, the overall antibacterial response which is influenced by the binding affinity of CRP and L-ficolin will be sensitive to the pH value and calcium level. To confirm this and further investigate the effects of pH and calcium on the antibacterial response, we simulated the complement system under different pH and calcium conditions. Based on the previous biochemical analysis [15], we first estimated functions using polynomial regression to predict the binding affinity of CRP and L-ficolin for different pH values and calcium levels (Figure 4A,B, right panel). In the right panels of Figure 4A and 4B, the reported binding affinities [15] were normalized and are shown as dots. By curve fitting the dots, we estimated polynomial functions that can be used to predict the binding affinity. The curves of these functions are shown in red. We then simulated the C3 deposition dynamics using the predicted binding affinities at pH ranging from 5.5 to 7.4 in the presence of 2 mM and 2.5 mM calcium. The simulation time was chosen to be 3.5 h which is the time frame of the response peaks. The results are shown in Figure 4A and 4B. Under both 2 mM and 2.5 mM calcium conditions, decreasing pH induces not only the increase of the peak amplitude (maximum activation) but also hastens the peak time (time of maximum activation). To further compare the effects of the two calcium levels, the dose-response curves were generated as shown in Figure 4C. The antibacterial response was predicted by simulating the system for 1.5 h. At 2 mM calcium (blue curve), the antibacterial response was clearly greater than at 2.5 mM calcium (pink curve) indicating that slight hypocalcaemia enhanced the antibacterial activity in a stable manner. In addition, the pH-responses were reaching saturation levels when pH was near 5.5 (Figure 4C), implying that the undesirable complement-enhancement by extreme low pH condition can be avoided. This also suggests that the saturation of the pH-response was influenced by the calcium level in the milieu. We next investigated the complement regulation by the major inhibitor, C4BP, under infection-inflammation conditions. We varied the initial concentration of C4BP and simulated the PC- and GlcNAc- initiated complement under infection-inflammation conditions. The simulation time was chosen to be 5 h which is slightly beyond the largest time point of our training experimental data. The predicted effects of the initial concentration of C4BP on the antibacterial response in terms of C3 deposition are shown in Figure 5A and 5B. For PC-initiated complement activation, when the starting amount of C4BP was perturbed around the normal level of 260 nM [16], increasing C4BP level only delayed the peak time but did not decrease the peak amplitude significantly. In contrast, reducing the initial C4BP level clearly hastened the complement activation and maximized the activity. Interestingly, the GlcNAc-initiated complement activation (Figure 5B) behaved differently from the PC-mediated complement activation (Figure 5A). Around the normal level of 260 nM, perturbing the initial C4BP changed the maximum activity but did not affect the peak time, suggesting that C4BP plays distinct roles in regulating the classical and lectin pathways. To experimentally verify the model predictions, we perturbed the initial amount of C4BP in the patient sera by (i) spiking with purified C4BP (high C4BP) and (ii) reducing it by immunoprecipitation (low C4BP). The resulting C4BP levels in the normal and patient sera are shown in Figure S3. The sera were then incubated with PC- or GlcNAc-beads to initiate complement. Unaltered serum served as the normal control. The time profiles of the deposited C4BP level was measured over 4 h using Western blot (Figure 5C). Comparing the kinetic profiles in the C4BP deposition initiated by both PC and GlcNAc, we observed the following order of peak time: high C4BP>normal C4BP>low C4BP, indicating that the pre-existing initial level of C4BP was indeed the driving force controlling the deposition of complement components onto the simulated bacterial surface. We then measured the time profiles of deposited C3. Figure 5D shows that with PC-beads, high C4BP sera induced an early peak and low C4BP delayed the peak of C3 deposition. The peak amplitude for all three conditions was at a similar level. These observations are consistent with the simulation results shown in Figure 5A. With GlcNAc-beads, reducing C4BP led to a slight increase in the peak height although the peak coincided with the normal condition. In contrast, spiking the sera (high C4BP) delayed and lowered the peak amplitude of C3 deposition. Thus the experimental results broadly agree with our model predictions presented in Figure 5B. We next investigated how C4BP mediates its inhibitory function. As shown in Figure 1A, the inhibitory effects of C4BP target different sites in complement: (a) binding to CRP and blocking C1, (b) preventing the formation of C4bC2a by binding to C4b, (c) acting as a cofactor for factor I in the proteolytic inactivation of C4b, and (d) accelerating the natural decay of the C4bC2a complex, which prevents the formation of C4bC2a and disrupts already formed convertase. To identify the dominant mechanism, we employed in silico knockout of the reactions involved for each mechanism and performed simulations. Figure 6A–6D shows the model predictions. Among the four inhibitory mechanisms, only the knockout of reaction (d) significantly enhanced the complement activation suggesting that facilitating the natural decay of C4bC2a (C3 convertase) is the most important inhibitory function of C4BP. This is consistent with our previous observations derived from sensitivity analysis, which identified the decay of C3 convertase as a critical reaction. In addition, as the inhibitory effect of reaction (d) is stronger than others, knocking out reaction (a) and (b) can even reduce the complement activity, which is counter-intuitive and emphasizes the significance of the systems-level understanding. To confirm our hypothesis that the major inhibitory role of C4BP relies on accelerating the decay of C3 convertase, we measured the C4 cleavage at different time points. Figure 6E (black triangles) indicates the inactive C4b fragments presented from the time points of 20, 30 and 90 min under high, normal, and low C4BP conditions, suggesting that C4BP aided cleavage and inactivation of C4b, and thereby caused the natural decay of the C4bC2a. Here, we developed an ODE-based dynamic model for the complement system accompanied by DBN-based approximations of the ODEs dynamics to understand how the complement activity is boosted under local inflammation conditions while a tight surveillance is established to attain homeostasis. Previously published models of complement system have focused on the classical and alternative pathways [27], [28]. Our model includes the lectin pathway and more interestingly, the recently identified amplification pathways induced by local inflammation conditions [15]. It also encompasses the regulatory effects of C4BP in the presence of enhanced complement activity. The ODE model incorporated both the PC-initiated and GlcNAc-initiated complement together for convenience. By setting the corresponding expressions to zero one at a time, two DBN approximations were then derived; one for the PC-initiated complement cascade and the other for GlcNAc-initiated complement cascade. For constructing the DBN approximation from an ODE model, one needs to fix , the maximal time point upto which each trajectory is to be explored and , the number of trajectories to be generated. is set to be suitably beyond the largest time point for which experimental data is available. In the present study 3.5 h, is the largest time point of our training experimental data. Based on this we set to be 5 h. After constructing the model, we simulated the system upto 10 h and found no relevant dynamics after 3.5 h. As for the choice of , the number of trajectories, ideally one would like to specify the acceptable amount of error between the actual and the approximated dynamics and use to determine . This is however difficult to achieve due to the following: The dynamic Bayesian network we construct is a factored Markov chain. It approximates the idealized Markov chain induced by the ODEs dynamics. This idealized Markov chain is determined by the discretization of the value spaces of the variables and the parameters, the discretization of the time domain and the prior distribution of the initial states. As observed in, Liu et al [25], given an error bound, a confidence level and the transition probabilities of the idealized Markov chain, we can estimate (upper bound) the required to fall within the given error bound with the required confidence level. However, our high dimensional ODE system does not admit closed form solutions and hence the transition probabilities of the idealized Markov chain will not be computable. Hence one must make a pragmatic choice of . Our approach has been to use a sampling method by which we can provide a minimum coverage of at least samples for each possible combination of interval values of the unknown parameters in the equation for each variable. This will ensure that the dynamics, governed by the set of equations (one for each variable) is being sufficiently sampled at least on a per equation basis. To achieve the required coverage, one will need samples, where is the maximal number of unknown parameters appearing in an equation and is the number of variables in the system. In our experience, seems to be an adequate choice. Based on this, we sampled initial points and generated the corresponding trajectories. The quality of the approximations relative to the original ODEs dynamics was sufficiently high and the details can be found in the supplementary information (Figure S1). How to determine with guaranteed error bounds is however a basic problem and we are continuing to study this issue. The study here has involved a tight integration of computational and experimental aspects. First, we used available biological information to form the biochemical network and the corresponding ODE system. We then experimentally generated test data to train the model during the process of parameter estimation. After constructing the model, one part of the computational exploration of the model was guided by the previous experimental study reported in [15]. Specifically, we computed through simulations the antibacterial response curves for varying combinations of pH values and calcium concentrations starting with the data provided in [15]. On the other hand, the second part of the study started from the computational side, namely, sensitivity analysis. Once C4BP was confirmed to be an important inhibitor through sensitivity analysis, we explored its regulatory mechanisms through simulations and generated the hypotheses concerning the differentiated influence of C4BP on PC-initiated and GlcNAc-initiated complement activity as well as the decay of the C3 convertase being the main inhibitory activity of C4BP. These hypotheses were then experimentally validated. At present, we have used the DBN approximation to mainly aid the tasks of parameter estimation and sensitivity analysis. The key idea is to use the DBN approximation and probabilistic inference to first reduce the search space and then apply conventional search techniques to this reduced search space in the second stage. For a -dimensional search space with discretized intervals for each dimension, the first stage can reduce the search space by a factor of . For analyses involving multiple initial conditions (such as the in silico experiments involving C4BP), we found it more convenient to use the ODE model. This is due to the fact that the size of the DBN approximation increases significantly if it must encompass multiple initial conditions. Alternatively, one must construct a separate DBN for each choice of initial conditions. Related probabilistic formalisms such as Multi Terminal Binary Decision Diagrams (MTBDDs) and Probabilistic Decision Graphs (PDGs) are also available for analysis. It is not clear at present how they can be derived directly from the ODE model. One could however try to convert our DBNs to MTBDDs for purposes of model checking [41] or develop statistical model checking methods [42]. As compact representations of the probability distributions, PDGs are, in spirit, similar to Bayesian networks [43] and can be computationally as efficient as Bayesian networks [44]. Further, probabilistic inference can be carried out with a time complexity linear in the size of the PDGs [44]. Thus, it will be an interesting future direction for us to explore the performance of PDGs in our setting. Finally, we are aware that model construction is rarely complete. In the present setting, we included as much of the relevant and available biological information as possible in our model. Once the model was calibrated using the test data and was validated using reported bacterial killing rates, we were reasonably confident that it could be used as a platform for studying the up- and down- regulation mechanisms of the complement under local inflammation conditions. In exploring the model, we were guided by both the previous experimental study [15] and standard techniques such as sensitivity analysis. It is clear that the model will have to be refined and modified as new experimental findings become available. Indeed, we consider the systematic incremental updating of a computational model as new data become available, to be an important task [45]. Turning next to the biological insights gained from this study, we have shown that increase in PC concentration, representing the inoculum size of the invading bacteria, affected the overall classical pathway response time more than the peak amplitude. Our model analysis confirmed that the enhancement of complement activity under infection-inflammation condition was attributable to the synergistic action of CRP and L-ficolin and supported the existence of the amplification pathways. We showed that to achieve a steady enhancement, C1 and L-ficolin (or CRP and MASP-2) should not compete with each other and the activities of CRP and L-ficolin should remain after forming the complex CRP:L-ficolin. Our computationally derived antibacterial response curves corresponding to varying pH values and calcium levels showed that the overall complement response was sensitive to pH and calcium levels. Through model analysis we found that with under PC-activation, perturbation of the initial C4BP level only affected the peak time but not the amplitude of the response. In contrast, in the case of GlcNAc-activation, perturbation of the initial level of C4BP only affected the peak amplitude and not the peak time. These results imply that C4BP regulates the lectin pathway more stringently than the classical pathway, which is consistent with previous experimental findings [46]. Further, for PC-initiated complement cascade, the over-expression of C4BP only delays but does not “turn off” the antibacterial response. In contrast, increased C4BP can efficiently inhibit GlcNAc-initiated complement activation. This may explain previous observations that bacteria such as Yersinia enterocolitica, Streptococcus pyogenes, Neisseria gonorrhoeae, Escherichia coli K1, Moraxella catarrhalis, Candida albicans, Bordetella pertussis [47], [48], [49], [50], [51], [52], [53] can exploit C4BP to evade complement. Through in silico knockouts, we found that, of the four documented inhibitory roles, C4BP mainly aided the natural decay of C3 convertase. As the enhancement mechanism by the crosstalk between CRP and L-ficolin occurs upstream of the cascade, we envisage C4BP acts downstream to ‘quality control’ and modulate C3 convertase activity. Thus our results suggest that efficient regulation of complement can be achieved by targeting the C3 convertase, where the complement pathways merge. In summary, by integrating our computational model and experimental observations we have obtained novel insights into how the complement activation is enhanced during infection and how excessive complement activity may be avoided. This introduces a new level of understanding of the host defense against bacterial infection. It also provides a platform for the potential development of complement-based immunomodulation therapies by exploiting the sensitivities of the perturbations of the pH, calcium and C4BP levels.

10.1371/journal.pcbi.1003258

Predictive Coding of Dynamical Variables in Balanced Spiking Networks

Two observations about the cortex have puzzled neuroscientists for a long time. First, neural responses are highly variable. Second, the level of excitation and inhibition received by each neuron is tightly balanced at all times. Here, we demonstrate that both properties are necessary consequences of neural networks that represent information efficiently in their spikes. We illustrate this insight with spiking networks that represent dynamical variables. Our approach is based on two assumptions: We assume that information about dynamical variables can be read out linearly from neural spike trains, and we assume that neurons only fire a spike if that improves the representation of the dynamical variables. Based on these assumptions, we derive a network of leaky integrate-and-fire neurons that is able to implement arbitrary linear dynamical systems. We show that the membrane voltage of the neurons is equivalent to a prediction error about a common population-level signal. Among other things, our approach allows us to construct an integrator network of spiking neurons that is robust against many perturbations. Most importantly, neural variability in our networks cannot be equated to noise. Despite exhibiting the same single unit properties as widely used population code models (e.g. tuning curves, Poisson distributed spike trains), balanced networks are orders of magnitudes more reliable. Our approach suggests that spikes do matter when considering how the brain computes, and that the reliability of cortical representations could have been strongly underestimated.

Two observations about the cortex have puzzled and fascinated neuroscientists for a long time. First, neural responses are highly variable. Second, the level of excitation and inhibition received by each neuron is tightly balanced at all times. Here, we demonstrate that both properties are necessary consequences of neural networks representing information reliably and with a small number of spikes. To achieve such efficiency, spikes of individual neurons must communicate prediction errors about a common population-level signal, automatically resulting in balanced excitation and inhibition and highly variable neural responses. We illustrate our approach by focusing on the implementation of linear dynamical systems. Among other things, this allows us to construct a network of spiking neurons that can integrate input signals, yet is robust against many perturbations. Most importantly, our approach shows that neural variability cannot be equated to noise. Despite exhibiting the same single unit properties as other widely used network models, our balanced networks are orders of magnitudes more reliable. Our results suggest that the precision of cortical representations has been strongly underestimated.

Neural systems need to integrate, store, and manipulate sensory information before acting upon it. Various neurophysiological and psychophysical experiments have provided examples of how these feats are accomplished in the brain, from the integration of sensory stimuli to decision-making [1], from the short-term storage of information [2] to the generation of movement sequences [3]. At the same time, it has been far more difficult to pin down the precise mechanisms underlying these functions. A lot of research on neural mechanisms has focused on studying neural networks in the framework of attractor dynamics [4]–[6]. These models generally assume that the system's state variables are represented by the instantaneous firing rates of neurons. While quite successful in reproducing some features of electrophysiological data, these models have had a hard time reproducing the irregular, Poisson-like statistics of cortical spike trains. A common assumption is that the random nature of spike times is averaged out over larger populations of neurons or longer periods of time [7]–[10]. However, the biophysical sources of noise in individual neurons are insufficient to explain such variability [11]–[13]. Several researchers have therefore suggested that irregular spike timing arises as a consequence of network dynamics [8], [14]. Indeed, large networks of leaky integrate-and-fire (LIF) neurons with balanced excitation and inhibition can be “chaotic” and generate asynchronous and Poisson-like firing statistics [15]–[18]. While these studies explain how relatively deterministic single units can generate similar statistical properties as random spike generators in rate models, they generally do not clarify how particular computations can be carried out, nor do they fundamentally answer why the brain would be operating in such a regime. Here we show that the properties of balanced networks can be derived from a single efficiency principle, which in turn allows us to design balanced networks that perform a wide variety of computations. We start from the assumption that dynamical variables are encoded such that they can be extracted from output spike trains by simple synaptic integration. We then specify a loss function that measures the system's performance with respect to an idealized dynamical system. We prescribe that neurons should only fire a spike if that decreases the loss function. From these assumptions, we derive a recurrent network of LIF neurons that is able to implement any linear dynamical system. We show that neurons in our network track a prediction error in their membrane potential and only fire a spike if that prediction error exceeds a certain value, a form of predictive coding. Our work shows how the ideas of predictive coding with spikes, first laid out within a Bayesian framework [19], 20, can be generalized to design spiking neural networks that implement arbitrary linear dynamical systems. Such multivariate dynamical systems are quite powerful and have remained a mainstay of control-engineering for real-world systems [21]. Importantly, the networks maintain a tight balance between the excitatory and inhibitory currents received by each unit, as has been reported at several levels of cortical processing [22]–[26]. The spike trains are asynchronous and irregular. However, this variability is not noise: The neural population essentially acts as a deterministic “super-unit”, tracking the variable with quasi-perfect accuracy while each individual neuron appears to behave stochastically. We illustrate our approach and its usefulness with several biologically relevant examples. Our basic model strategy is represented in Fig. 1 A. Let us consider a linear dynamical system describing the temporal evolution of a vector of dynamical variables, :(1)Here is the state transition matrix, and are time-varying, external inputs or command variables. We want to build a neural network composed of neurons, taking initial state and commands as inputs, and reproducing the temporal trajectory of . Specifically, we want to be able to read an estimate of the dynamical variable from the network's spike trains . These output spike trains are given by , where is the time of the spike in neuron . Our first assumption is that the estimate is obtained by a weighted, leaky integration of the spike trains,(2)where the matrix contains the decoding or output weights of all neurons, and is the read-out's decay rate. Whenever neuron fires, a -function is added to its spike train, . The integration of the respective delta-function contributes a decaying exponential kernel, , weighted by , to each dynamical variable, . This contribution can be interpreted as a simplified postsynaptical potential (PSP). The effect of a neuron's spike can be summarized by its weights, , which we call the output kernel of neuron . Note that these weights correspond to the columns of the matrix . The estimate can also be written as a weighted linear summation of the neuron's firing rates, , if we define the time-varying firing rates of the neurons, , as(3) Our second assumption is that the network minimizes the distance between and by optimizing over the spike times , and not by changing the fixed output weight matrix . This approach differs from the “liquid computing” approach in which recurrent networks have fixed, random connectivities while the decoding weights are learnt [27]. In our case, the decoding weights are chosen a-priori. In order to track the temporal evolution of as closely as possible, the network minimizes the cumulative mean-squared error between the variable and its estimate, while limiting the cost in spiking. Thus, it minimizes the following cost function,(4)where denotes the Eucledian distance (or L2 norm), and the Manhattan distance (or L1 norm), which here is simply the sum over all firing rates, i.e., . Parameters and control the cost-accuracy tradeoff. The linear cost term, controlled by , forces the network to perform the task with as few spikes as possible, whereas the quadratic cost term, controlled by , forces the network to distribute spikes more equally among neurons, as explained in Material and Methods. To derive the network dynamics, we assume that the firing mechanism of the neurons performs a greedy minimization of the cost function . More specifically, neuron fires a spike whenever this results in a decrease of , i.e., whenever . As explained in Material and Methods, this prescription gives rise to the firing rule(5)with(6)(7)Since is a time-varying variable, whereas is a constant, we identify the former with the -th neuron's membrane potential , and the latter with its firing threshold . In the limit , the membrane potential of the -th neuron can be understood as the projection of the prediction error onto the output kernel . Whenever this projected prediction error exceeds a threshold, a new spike is fired, ensuring that precisely tracks . For finite , the membrane voltage measures a penalized prediction error. If the neuron is already firing at a high rate , only a correspondingly larger error will be able to exceed the threshold and lead to a spike. To connect this firing rule with the desired network dynamics, Eqn. (1), we take the derivative of each neuron's membrane potential, Eqn. (6), and consider the limit of large networks (see Material and Methods) to obtain the differential equation(8)where is a leak term, is a weight matrix of connectivity filters, explained below, and corresponds to a white “background noise” with unit-variance. The leak-term does not strictly follow from the derivation, but has been included for biological realism. A similar rationale holds for the noise term which we add to capture unavoidable sources of stochasticity in biological neurons due to channel noise, background synaptic input, etc. The differential equation then corresponds to a standard LIF neuron with leak term , external, feedforward synaptic inputs , recurrent synaptic inputs mediated through the weight matrix , and a firing threshold , as specified in Eqn. (7). The weight matrix of connectivity filters is defined as(9)and contains both “fast” and “slow” lateral connections, given by the matrices(10)(11)where corresponds to the identity matrix. Accordingly, the connectivity of the network is entirely derived from the output weight matrix , the desired dynamics , and the penalty parameter . Note that the diagonal elements of implement a reset in membrane potential after each spike by . With this self-reset, individual neurons become formally equivalent to LIF neurons. Whereas the linear penalty, , influences only the thresholds of the LIF neurons, the quadratic penalty, , influences both the thresholds, resets, and dynamics of the individual neurons, through its impact on the diagonal elements of the connectivity matrix. Slow and fast lateral connections have typically opposite effects on postsynaptic neurons, and thereby different roles to play. The fast connections, or off-diagonal elements of the matrix , implement a competition among neurons with similar selectivity. If neuron fires, the corresponding decreases in prediction errors () are conveyed to all other neurons . Neurons with similar kernels ( inhibit each other, while neurons with opposite kernels () excite each other. This is schematized by the blue and red connections in Fig. 1 A. In contrast, the slow connections, , implement a cooperation among neurons with similar selectivity. These connections predict the future trajectory of (term “”) but also compensate for the loss of information due to the decoder leak (term “”). For example, when the variable is static (, ), these connections maintain persistent activity in the network, preventing the variable from decaying back to zero (see below). Note that when the internal dynamics of change on a slower time scale than the decoder (i.e. ), and if we neglect the cost term , slow and fast connections have the same profile, (i.e. ), but opposite signs. The combined effect of fast and slow connections yields the effective PSPs in our network, , with , which can be obtained by integrating Eqn. (8) for a single spike. Two example PSPs are shown in Fig. 1 B. We note that our network model may contain neurons that both inhibit and excite different targets, depending on the kernel sign, a violation of Dale's law. This problem can be addressed by creating separate cost functions for excitatory and inhibitory neurons, as laid out in full detail in Text S1. Here, we simply interpret the resulting connectivity as the effective or functional connectivity of a network, akin to the types of connectivities arising in generalized linear models (GLMs) of neural networks [28]. We now briefly consider how the above equations can be mapped onto realistic physical units. This consideration has the additional benefit that it clarifies how the network parameters scale with the number of neurons (see also Material and Methods). In order to express the network dynamics in biophysically relevant units, the membrane potential , Eqn. (6), and threshold , Eqn. (7), have to be rescaled accordingly. We can obtain proper membrane potential units in mV if we apply the simple transformations and . In turn, we obtain the modified equations(12)(13)and the modified dynamics(14)with a resting potential of . Note that both the feedforward and recurrent connectivities change in this case. Specifically, we obtain and , and a similar expression for the noise, . In turn, we can freely choose and to find realistic units. For instance, we can fix the threshold at , and the reset potential at , which uniquely determines both and for each neuron. In the absence of linear costs (), the reset potential becomes simply . When we increase the network size while keeping the average firing rates and the read-out constant, we need to change the decoding kernels. Specifically, the decoding kernels need to scale with . If we assume that the relative importance of the cost terms is held fixed for each neuron, then the original threshold scales with , and the original connectivities similarly scale with , compare Eqns. (9–11). As a consequence, the rescaled synaptic weights, do not scale in size when the network becomes larger or smaller. When considering the summation over the different input spike trains, we therefore see that all synaptic inputs into the network scale with : the feedforward inputs, the slow recurrent input, and the fast recurrent inputs (the latter two are both contained in the matrix ). The equal scaling of all inputs maintains the detailed balance of excitation and inhibition in the network. An instructive case is given if we neglect the cost terms for a moment () in which case we obtain the following (rescaled) feedforward weights and connectivities:(15)(16)Accordingly, the strength of the lateral connections is independent of the kernel norm. In contrast, the strength of the feed-forward connections scales with the inverse of the kernel norm. Since smaller kernels provide a more precise representation, the precision of the rescaled network, and its firing rates, are controlled entirely by its input gain. Once the dynamics and the decoders are chosen, Eqn. (1) and Eqn. (2), the only free parameters of the model are , , , and . The model presented previously can in principle implement any linear dynamical system. We will first illustrate the approach with the simplest linear dynamical system possible, a leaky integration of noisy sensory inputs where can be interpreted as the sensory stimulus while represents shared sensory noise. The corresponding dynamical system, Eqn. (1), is then given by(17)The integrated sensory signal is a scalar () and represents the leak of the sensory integrator. For a completely homogeneous network, in which the output kernels of all neurons are the same, we can solve the equations analytically which is shown in Text S1. A slightly more interesting case is shown in Fig. 1 C,D, which illustrate network dynamics for two different choices of . Here we used neurons, half of them with positive kernels (), and the other half with negative kernels (). Neurons with positive kernels fire when variable is positive or increases, while neurons with negative kernels fire when the variable is negative or decreases. Moreover, we set the cost terms and at small values, ensuring that our objective function is dominated by the estimation error, compare Eqn. (4). As a consequence, the estimate closely tracks the true variable . Albeit small, the cost terms are crucial for generating biologically realistic spike trains. Without them, a single neuron may for example fire at extremely high rates while all others stay completely silent. The regularizing influence of the cost terms is described in more detail in Text S1. For , the network is a perfect integrator of a noisy sensory signal. The neural activities resemble the firing rates of LIP neurons that integrate sensory information during a slow motion-discrimination task [1]. In the absence of sensory stimulation, the network sustains a constant firing rate (Fig. 1 C after sec), similar to line attractor networks [29]–[31]. In fact, as long as the dynamics of the system are slower than the decoder (), the instantaneous firing rates approximate a (leaky) integration of the sensory signals. On the other hand, if the system varies faster than the decoder (i.e. ), then neural firing rates track the sensory signal, and model neurons have transient responses at the start or end of sensory stimulation, followed by a decay to a lower sustained rate (Fig. 1 D). These responses are similar to the adaptive and transient responses observed in most sensory areas. The overall effect of the lateral connections depends on the relative time scales of the variable and the decoder (Fig. 1 B). For neurons with similar selectivity (or equal read-out kernels, ), the postsynaptic potentials are given by (assuming ),(18)For neurons with opposite read-out kernels, we obtain just a sign reversal. When (), the interplay of fast inhibition with slower excitation results in a bi-phasic interaction between neurons of similar selectivity (Fig. 1 B, left). Moreover, the network activity persists after the disappearance of the stimulus. In the extreme case of the perfect integrator (), the temporal integral of this PSP is exactly zero, which explains how the mean network activity can remain perfectly stable (neither increase nor decrease) in the absence of any sensory stimulation. In contrast, lateral interactions are entirely inhibitory when the network tracks the stimulus on a faster time scale than the decoder (i.e. , Fig. 1 B, right). The dominance of lateral inhibition explains the transient nature of the network responses and the absence of persistent activity. Other response properties of the model units are illustrated in Fig. 2. We define the tuning curves of the neurons as the mean spike count in response to a 1 s presentation of a constant stimulus . Firing rates monotonically increase (for positive kernels) or decrease (for negative kernels) as a function of and are rectified at zero, resulting in rectified linear tuning curves (Fig. 2 A). Since all neurons have identical kernels (i.e. all or ), neurons with the same kernel signs have identical tuning curves. However, such a homogeneous network is rather implausible since it assumes all-to-all lateral connectivity with identical weights, so that all units in the network receive exactly the same synaptic input and have the same membrane potential. To move to more realistic and heterogeneous networks, we can choose randomized decoding kernels . Even then, however, the connectivity matrix is strongly constrained. For negligible costs, , the weight matrix has rank one (since ). A lot more flexibility can be introduced in the network connections by simultaneously tracking variables with identical dynamics and identical control , rather than a single scalar variable. Thus the variable and the kernels have dimensions and . We then define the actual network output, , as the mean of those variables (simply obtained by summing all network outputs). The network estimation error, , is an upper bound on , ensuring similar performance as before (see Fig. 3). Importantly, we can choose the output kernels to fit connectivity constraints imposed by biology. For instance, the output kernels can be made random and sparse (i.e. with many zero elements), resulting in random and sparse (but symmetrical) connection matrices. In such a network, the tuning curves are still rectified-linear, but have different gains for different neurons (Fig. 2 B). Output spike trains of both homogeneous and inhomogeneous networks are asynchronous and highly variable from trial to trial (see raster plots in Fig. 1 C,D and Fig. 2). Fano factors (measured during periods of constant firing rates), CV1, and CV2, were all found to be narrowly distributed around one. The interspike interval (ISI) distribution was close to exponential (Fig. 2 C). Moreover, noise correlations between neurons are extremely small and do not exceed 0.001 (noise correlations are defined as the cross-correlation coefficient of spike count in a time window of 1 s in response to a constant variable ). Finally, analysis of auto and cross-correlograms reveals the presence of high-frequency oscillations at the level of the population (Fig. 2 D). These high frequency oscillations are not visible on Fig. 2 C since the size of the bin (5 ms) is larger than the period of the oscillations (1 ms). Note that if we add a realistic amount of jitter noise (ms) to spike times, these high frequency oscillations disappear without affecting the response properties of the network. In contrast to the output spike trains, the membrane potentials of different neurons are highly correlated, since neurons with similar kernels () receive highly correlated feed-forward and lateral inputs (Fig. 4 A,B). In the homogeneous networks without quadratic cost (), these inputs are even identical, resulting in membrane potentials that only differ by the background noise (Fig. 4 A). Despite these strong correlations of the membrane potentials, the neurons fire rarely and asynchronously. Fig. 4 C illustrates why this is the case: let us consider a population of neurons with identical output kernels , maintaining an estimate of a constant positive (top panel, blue line). Due to the decoder leak , the network needs to fire periodically in order to maintain its estimate at the level of (top panel, red line). However, the exact order at which the different neurons fire does not matter, since they all contribute equally. The period between two spikes can be called an “integration cycle”. Within one integration cycle, the prediction errors and thus the membrane potentials, , rise for all neurons (bottom panel, red line). Since all kernels are identical, however, all neurons compute the same prediction error and will reach their firing thresholds at approximately the same time. Only chance (in this case, the background noise ) will decide which neuron reaches threshold first. This first neuron is the only one firing in this integration cycle (middle panel, colored bars) since it immediately inhibits itself and all other neurons. This starts a new integration cycle. As a result of this mechanism, while the population of neurons fire at regular intervals (hence the high frequency oscillations in Fig. 2 D) only one neuron fires in each cycle, and its identity is essentially random. The resulting variability has no impact on the network estimate, since all spike orders give the same output . In the presence of a quadratic cost (), neurons that did not fire recently have a higher probability of reaching threshold first (their membrane potential is not penalized by ). When the cost term is large compared to the background noise (i.e. when , which is not the case in the example provided here), this tends to regularize the output spike trains and leads to s smaller than 1. However, this regularization is not observed in inhomogeneous networks. The inhomogeneous network behaves similarly, except that all neurons do not receive the same inputs and do not reach threshold at the same time (Fig. 4 B). In this case, we can even dispense of the background noise (i.e. ) since fluctuations due to past network activity will result in a different neuron reaching threshold first in each cycle. The individual ups and downs caused by the synaptic inputs from other neurons will nonetheless appear like random noise when observing a single neuron (Fig. 4 B,D). Furthermore, even in this deterministic regime, the spike trains exhibit Poisson statistics. In fact, changing the timing of a single spike results in a total reordering of later spikes, suggesting that the network is chaotic (as illustrated in Fig. 3). We now apply this approach to the tracking of a 2D point-mass arm based on an efferent motor command. The dynamical variable has dimensions corresponding to the arm positions and the arm velocities . The arm dynamics are determined by elementary physics, so that(19)(20)where is a 2D (control) force exerted onto the arm, and captures possible friction forces. To simulate this system, we studied an arm moving from a central position towards different equidistant targets. This reaching out arm movement was obtained by “push-pull” control forces providing strong acceleration at the beginning of the movement, and fast deceleration at the end of the movement (Fig. 5 A, top panel). As previously, the network predicts the trajectory of the arm perfectly based on the forces exerted on it (Fig. 5 A, bottom panel; we again use relatively small cost terms and ). The resulting spike trains are asynchronous, decorrelated, and Poisson-like, with unpredictable spike times (rasters in Fig. 5 A; Fano factor and CVs close to 1). The membrane potential of neurons with similar kernels are correlated while output spike trains are asynchronous and decorrelated. The effective postsynaptic potentials have biphasic shapes reflecting the integrative nature of the network for small friction forces (). To measure tuning curves in this “center out” reaching task, we varied the speed and direction of the movement, as well as the starting position of the arm. Neural activity was defined as the mean spike count measured during movement. As illustrated in Fig. 5 B,C,D, instantaneous firing rates are modulated by arm position, velocity and force. We found that tuning curves to arm position are rectified linear, with varying thresholds and slopes (as in Fig. 2 B). Such linear-rectified gain curves with posture have been reported in premotor and motor cortical areas [32], [33]. In contrast, tuning curves to circular symmetric variables such as movement direction or arm angle are bell-shaped (Fig. 5 B,C,D). In addition, direction tuning curves are gain modulated by arm speed, such that responses are stronger for larger speed when the arm moves in the preferred direction, and weaker when the arm moves in the anti-preferred direction (Fig. 5 B). Finally, arm positions have both an additive and a gain modulating effect on the tuning curve, and these modulation can be monotonically increasing (Fig. 5 C) or decreasing (Fig. 5 D) with arm position. These observations have a simple geometric explanation, schematized in Fig. 5 E for the velocity space, . A neuron is maximally active (; assuming ) when its kernel (, thick vector in Fig. 5 E) points in the direction of the derivative of the prediction error, . Since the decoder leak is faster than the arm dynamics, this error mostly points in the direction opposite to the leak, (thin vectors). Within the velocity space, the kernel thus defines the neuron's preferred movement direction (dashed line and filled circles). The neurons is less often recruited when the arm moves away from the kernel's direction (empty circles), resulting in a bell-shaped tuning curve. Finally, since the vector gets larger at larger speeds, more spikes are required to track the arm state resulting in a linear tuning to movement speed. The same reasoning applies for the position space . These predictions are independent of the choice of kernels and are in direct agreement with experimental data from the pre-motor and motor cortices [32], [33]. We chose to present a sensory integrator and an arm controller for their biological relevance and simplicity. However, any linear dynamical system can be implemented within our framework, and our networks are not limited to performing integration. To illustrate the generality of the approach, we applied the framework to two additional examples. In Fig. 6 A, we simulated a “leaky differentiator” with a transition matrix . This system of differential equations is designed so that the variable approximates a temporal derivative of a command signal . The command signal, , is shown in the top panel of Fig. 6 A; the input signal is zero. We used neurons with kernels drawn from a normal distribution, and then normalized to a constant norm of . As in the other examples, the firing statistics are close to Poisson, with a . In Fig. 6 B, we simulated a network that implements a damped harmonic oscillator. Here we chose a transition matrix . The oscillator is initially kicked out of its resting state through a force given by the command signal , as plotted on the top panel. The input signal is zero. We used neurons with kernels drawn from a normal distribution, and normalized to a constant norm of . The network tracks the position and speed of the damped oscillator until position and speed are too close to zero to allow a reliable approximation. The firing statistics of single units are again Poisson-like, with . Note that in these two examples, the dynamics implemented by the network are faster than the decoder's time scale . Accordingly, our networks can track changes faster than the time scale of the decoder. This speed-independence relies on a simple scheme: Spikes from neurons with positive kernel weight, , represent instantaneous increases in , whereas spikes from neurons with negative kernel weight represent instantaneous decreases in . Even if the inter-spike interval is much shorter that , the decoder can therefore still provide an efficient staircase approximation for . In conclusion, the temporal accuracy of these networks is not limited by , but by . We have proposed a method for embedding any linear dynamical system in a recurrent network of LIF neurons. The network connectivity and spike generation are entirely derived from a single loss function which seeks to optimally place spikes so that the relevant information can be extracted by postsynaptic integration. Accordingly, the network structure follows exclusively from functional principles, and no extensive parameter searches are required. This approach implies in particular that neurons share information in a smart way, yet fire almost randomly at the level of the single cell. We also included a cost term in the error function, Eqn. (4). Due to this cost term, the network finds a solution minimizing the metabolic cost associated with high spike counts. Both linear and quadratic cost terms regularize the firing rate and make the network robust against artefacts such as high firing rates that may be caused by the greedy spiking mechanism (see Text S1). Further generalizations or modifications of these predictive coding principles may eventually help to explain other biophysical or electrophysiological phenomena of the brain. Our current work both generalizes and modifies our earlier work in which we applied the principle of predictive coding with spikes to a Bayesian inference problem [20]. This model tracked a log-probability distribution and implemented a non-linear drift-diffusion model, rather than a generic linear differential equation. In addition, we here introduced cost terms which provided us with greater flexibility in regulating and controlling the dynamics of the spiking networks. A quite general framework for designing networks of neurons that implement arbitrary dynamical systems has previously been described in the “neuro-engineering” approach [30]. This approach relies on linearly combining the non-linear rate transfer function of LIF neurons. In its essence, the method is therefore based on firing rates, and makes few predictions about the spiking statistics of cortical neurons. A recently developed model, the “ReFiRe network” [34] provides a recipe for designing networks maintaining stable memories, and shares some of the features of our networks. Just as the neuro-engineering framework, however, its design is essentially based on firing rates. Here we have designed a network based on the principle of predictive coding with spikes. Even though indistinguishable from older models on the single cell level, our work is different in several important respects. A first major difference of our approach is that it predicts a detailed balance between excitation and inhibition, rather than imposing it upfront. This balance follows from the attempt of the network to minimize the loss function, Eqn. (4), which in turn implies that the membrane potential of neurons represents a prediction error and that neurons spike only when this prediction error exceeds a certain value—a form of predictive coding. Any increase in excitation causes an increase in prediction error, immediately compensated by an increase in inhibition to bring down the prediction error (and vice versa). This interplay causes a tight temporal correlation between excitation and inhibition at the time scale of the stimulus but also at a much finer time scale, within a single ISI (Fig. 7 A). Note that this balance only holds when considering all inputs. In the leaky integrator, for instance, all lateral connections are inhibitory (Fig. 1 B, right panel). However, the network is still globally balanced when taking into account the contribution from the feedforward connections. Such a tight balance between excitation and inhibition has been observed at several levels of cortical processing [22]–[26]. Accordingly, spike trains in our network usually resemble independent Poisson processes, with rates tuned to the variable . We note that spike trains can also be more regular if the networks are smaller and the noise is not too large. A simple example is a network composed of a single neuron (), for which we provide an analytical solution in Text S1. Such a LIF neuron responds to a constant positive input with a perfectly regular spike train. In practice, regular firing is observed when only a few neurons are simultaneously co-active (e.g. for networks composed of less than neurons). Firing becomes irregular when many neurons are co-active (e.g. for networks of several hundreds of neurons or more). Increasing synaptic background noise tends to make firing less regular, while increasing the quadratic metabolic costs makes firing more regular. However, for large networks, these effects are small and remain within the range of Fano-factors or CVs observed in cortex. The amount of regularity has no impact on the network performance. Despite the variability observed in large networks, one cannot replace or approximate one of our spiking networks with an equivalent rate model composed of Poisson spike generators, a second major difference to other network models. This point is illustrated in Fig. 7 B,C for the homogeneous integrator model, where we removed the fast connections in the network and replaced the integrate-and-fire dynamics by independent Poisson processes (see Material and Methods). The performance of the resulting Poisson generator network is strongly degraded, even though it has the same instantaneous firing rates and slow connections as the LIF network. We can quantify the benefit of using a deterministic firing rule compared to stochastic rate units by considering how the estimation error scales with the network size. The integrator model tracks the dynamical variable with a precision defined by the size of a kernel . Estimation errors larger than are immediately corrected by a spike. As the network size increases, maintaining the same firing rates in single units requires that the kernels, and thus, the estimation error, scale with (see Material and Methods). In contrast, the error made when averaging over a population of independent Poisson neurons diminishes with . Intuitively, the predictive coding network achieves higher reliability because its neurons communicate shared information with each other via the fast synapses, whereas the independent Poisson neurons do not. The communicated information actively anti-correlates all spike trains, which, for networks composed of more than a dozen neurons, will be indistinguishable from the active decorrelation of pairwise spike trains that has recently been observed in vivo [35]. Therefore, the precision of the neural code cannot be interpolated from single-cell recordings in our network, and combining spike trains recorded in different trials results in a strong degradation of the estimate (Fig. 7 D). A third major difference between our network model and those proposed previously concerns the scaling of the network connectivity. Most previous approaches assumed sparse networks and weak connectivity in which the probability of connections (and/or connection strengths) scales as or . This weak connectivity leads to uncorrelated excitation and inhibition and thus neurons driven by random fluctuations in their input [15], [36]. For comparison, the connectivity in our network is finite (once the membrane have been rescaled by the kernel norm to occupy a fixed range of voltage). Our approach is therefore reminiscent of a recent model with finite connection probability [17]. As in our model, stronger connectivity leads to correlation between excitation and inhibition but uncorrelated spike trains. The strong network connectivity in turn swamps the membrane potential of each neuron with currents. The excitatory and inhibitory currents driving the neural response grow linearly with the number of neurons, , and are thus much larger than the membrane potential (prediction error) , which is bounded by the (fixed) threshold. In turn, the leak currents become negligible in large networks. For example, the integrator network in Fig. 1 C has neurons and can maintain information for 100 s (it takes 100 seconds for the network activity to decay by half) despite the fact that the membrane time constant () is only 0.1 s. Thus, according to our model, spiking neurons can fire persistently and thereby maintain information because their leaks are dwarfed by the currents they receive from recurrent connections. There are several non-trivial circumstances under which our network models may fail. First, we notice that the spiking rule that we derive amounts to a greedy optimization of the loss function. Future costs are not taken into account. This may cause problems in real neurons which can only communicate with time delays, but it may also cause problems when neurons have opposing kernels. For instance, two neurons with opposing kernels may become engaged in rapidly firing volleys of spikes, each trying in fast succession to decrease the error introduced by the previous spike from the other neuron (see Text S1), a problem that we call the “ping-pong” effect. This effect can become a serious problem if the network dynamics is corrupted by strongly correlated noise, which may occur in the presence of synaptic failures. However, it is usually possible to diminish or eliminate this effect by increasing the spike count cost (see Text S1). Second, the leak term we introduced in the voltage equation provides only an approximation to the actual voltage equation (see Material and Methods). Specifically, the term we approximate is times smaller than the other terms in the membrane potential dynamics. In practice, we can therefore always increase the network size to reach an acceptable level of performance. For a given network size, however, the approximation may break down when becomes too large or when both and are too small (of order ). Third, the speed at which the linear dynamical system can evolve will be limited from a practical point of view, even in the limit of large networks. While the time scale of the decoder, does not put any strict limitations on the speed (since spikes corresponding to positive and negative kernels can always provide an efficient stair-case approximation to any time-varying function), faster dynamics can only be obtained if the linear dynamical system compensates for the decoder filtering. This compensation or inversion process is a case of deconvolution, and bound to be severely limited in practice due to the noise inherent in all physical systems. Finally, the network requires finely tuned lateral connections in order to balance excitation and inhibition (from feed-forward and lateral connections). In particular, the strength of the fast connections between two neurons corresponds to minus the correlation coefficient of their feed-forward connections (and thus, to their level of shared inputs). Whether such finely tuned motifs exist in biological networks is still an open question. We showed recently that fast lateral connections can be learnt using unsupervised Hebbian learning [37], suggesting that networks with the appropriate form of plasticity would be able to develop and maintain this tight balance. We note that the performance of the networks is quite sensitive to global perturbations of the balance between excitation and inhibition, an issue that we discuss in more detail in Text S1. The most crucial work left to the future will be to test the predictions derived from this theory, three of which are described here. First, one could test how the decoding error scales with the numbers of simultaneously recorded neurons. A single unit in the model network (considered in isolation) is in fact exactly as reliable as a Poisson spike generator with the same rate. As the number of simultaneously recorded neurons increases, the decoding error initially decreases as , similar to a Poisson rate model. However, as the number of neurons reaches a certain threshold (10% for the network models simulated here), the error from the spiking network decreases faster than predicted for a Poisson rate model (Fig. 7 E). So far, single-unit recordings or multi-electrode recordings have only sampled from a very small subpart of the population, making it impossible to see this difference (and in turn, potentially leading to an under-estimation of the precision of the neural code). However, with newer techniques, such as dense multi-electrode arrays or optical imaging, as well as with focusing on smaller networks (such as the oculomotor integrator or insect systems), these model predictions are nowadays within experimental reach. We note that one has to carefully account for the effect of shared sensory noise () to see the predicted scaling effect. Shared noise (absent in Fig. 7 E) introduces correlations between neurons and results in a saturation of the error with . In our network, such a saturation would only be seen if there were limits to the sensory information available in the first place; saturation would not be seen as a consequence of neural noise or correlations (as proposed for example in [38], [39]). Second, one could look at the global interaction between neurons of similar selectivity, for example by applying a GLM model to the data [28]. The model predicts that neurons involved in slow integration tasks or working memory tasks should inhibit and excite each other at different delays. In particular, neurons with similar selectivities should be (paradoxically) negatively correlated at short delays. Thus, applying GLM analysis even on a small sub-population can uncover the effective PSPs caused by the lateral connections and, indirectly, the dynamical equation implemented by the network. Fig. 7 F shows the GLM filters learnt from the inhomogeneous integrator network during working memory (i.e. sustained activity in the absence of sensory input). The analysis recovered the shape of the filters between neurons of similar kernels and opposite kernels, despite the fact that only 10 simultaneously recorded neurons (2.5% of the population) were used in this analysis. Third, the spiking network is by essence self-correcting and will thus be resilient to lesions or many sudden perturbations (an exception being perturbations of the global balance of excitation and inhibition, see above). Equipping neural networks with such resilience or robustness has been a well-studied theoretical problem. For the specific example of the neural integrator, solutions range from constructing discrete attractor states [40], [41], [42], adding adaptation or learning mechanisms to a network [43], [44], or changing the nature of network feedback [45], [46]. In the case of the neural integrator, the robustness of our network could likely be interpreted as a case of derivative feedback [46]. While we know that biological neural networks are quite robust against partial lesions, their response to sudden, yet partial perturbations is less well known. For example, suddenly inactivating half of the active neurons in our sensory integrator increases the firing rates of the remaining neurons but has essentially no effect on the network performance (Fig. 7 G). This instantaneous increase in firing rates without performance loss generates a strong prediction for our network model, a prediction that distinguishes our network from previously proposed solutions to the robustness problem. Indeed, as long as the pool of available kernels remains sufficient to track , and as long as increased firing rates are not affected by saturation, inactivation will not affect the network's computation. This prediction could be tested using for example optogenetic methods. We here derive the network equations using compact matrix-vector notation. In Text S1, we also consider the special case of a homogeneous network and a single neuron, for which the derivations are simpler. We consider the error function, Eqn. (4), which is given by(21)The -th neuron should spike at time if(22)A spike by the -th neuron adds a single delta-function to its spike train. This additional spike enters the right-hand-side of the read-out equation, Eqn. (2). Integration of this extra delta-function amounts to adding a decaying exponential kernel, to the read-out. Hence, if neuron spikes at time , we have(23)(24)where the latter equation describes the instantaneous change in firing rate due to the additional spike. Note that the standard Eucledian basis vector is a vector in which the -th element is one, and all others are zero. Each spike influences the read-out several time intervals into the future. To see whether a spike leads to a decrease of the error function, we therefore need to look into the future (from time onwards). For a future time with , the spiking rule in Eqn. (22) translates into(25)We can expand the terms on the left-hand-side, and then eliminate identical terms on both sides. For that, we remind the reader that the relation holds for the norm, whereas holds for the norm in our case, since all elements (firing rates) are positive by definition. Hence we obtain(26)We rearrange the inequality by moving all terms that depend on the dynamical variables , the estimates , or the firing rates to the left, and all other terms to the right, and we then multiply both sides by minus one, to obtain(27)Moving the kernels to the front of the integrals and noticing that for , we obtain(28)The integral on the left-hand-side weights the influence of the spike, as given by the decaying exponential kernel, , against the future development of the error signal, , and firing rate . These future signals are unknown: while we may be able to extrapolate , given its dynamical equation, we cannot safely extrapolate or , since this would require knowledge of all future spikes. We therefore choose a “greedy” approximation in which we only look a time into the future. For the relevant times , we can then approximate the integrands as constants so that (using for )(29)which is our decision to spike, and corresponds exactly to Eqns. (5–7). We notice that the right-hand-side is a constant whereas the left-hand-side is a dynamical quantity which corresponds to the projection of the prediction error, , onto the output kernel of the -th neuron, , subtracted by a term depending on the firing rate of the -th neuron. Given this threshold rule, it seems only natural to identify the left-hand-side with the membrane voltage of the -th neuron and the right-hand-side with its spiking threshold, , which is what we did in the main text. If we write the voltage of all neurons as one long vector, , then we can write(30)We generally assume that there are more neurons than variables to represent so that . We also assume that the output kernel matrix, , has rank , and that the dynamical variables are not degenerate or linearly dependent on each other. In this case, the left pseudo-inverse of exists and is given by(31)so that . Note that is an -matrix, while has size . In turn, we can solve the voltage equation for by multiplying with the pseudo-inverse from the left so that(32)Taking the derivative of the voltages, we obtain(33)Replacing , , and with their respective equations, Eqns. (1–3), we obtain(34)In turn, we can replace with Eqn. (32) to obtain(35)Sorting some of the terms, and remembering that , we obtain(36) To evaluate the relative importance of the different terms, we consider the limit of large networks, i.e., the limit . First, we impose that the average firing rates of individual neurons should remain constant in this limit. Second, we require that the read-out does not change. Given the scaling of the firing rates, and since , the output kernels must scale with . Accordingly, the pseudo-inverse scales with . Finally, we need to choose how the cost terms and , scale with respect to the read-out error. The linear and quadratic error terms and scale with . To avoid a contribution of the cost term increasing with network size, and should scale (at the least) with . However, even if the cost terms scale with , they will still dominate the network dynamics. For instance, the threshold, Eqn. (7) becomes independent of the output kernel, while the contribution of fast lateral connections becomes negligible. In practice, this causes the performance to degrade quickly with increasing network size. A better choice is to require and to scale with , keeping the relative contribution of the kernel and cost to each neuron's dynamics fixed. With such scaling, large networks can still track the variable while the performance increase with network size. Given the scaling of the output kernels and , the threshold scales with , compare Eqn. (7). In turn, since the voltage is bounded by the threshold from above (and bounded from below due to the existence of neurons with opposing kernels; see also below), the voltage also scales with . Accordingly, in a large network, the first, voltage-dependent term in Eqn. (36) scales with , as do the terms and . In contrast, the terms and represent a sum over all neurons in the population, and thus scale with , similar to the inputs . For large networks, we can therefore neglect the terms that scale with . We note that none of the terms involving delta functions (i.e. ) can be neglected. We keep a generic leak term, , although the term is essentially irrelevant in large networks, and may be detrimental in very small ones (e.g., less than 10 neurons). Hence, we approximate Eqn. (36) by(37)with(38)(39)Since and since , we can define the effective connectivities(40)to obtain the voltage equation(41)which is the vectorized version of Eqn. (8) without the noise term. In the homogeneous integrator network (with low noise and small costs), the membrane potentials of neurons with identical kernels are approximately equal, which allows us to write down an analytical solution (see Text S1). Briefly, the population inter-spike interval, i.e. the interval between two successive spikes from any neuron, corresponds to the time it takes for this “common” membrane potential to rise from the reset potential to the threshold . We call this time period an “integration cycle”. Note that this interval is typically much shorter than the ISI of an individual neuron or the time constant of the decoder. During this short time interval, the leak term can be neglected, and the derivative of the membrane potential, , is approximately constant. The population ISI is thus simply given by the time it takes to integrate from the reset, , to the threshold, , so that . All neurons with the same kernel have identical firing rates, and, since only half of the population is spiking at any value of (in the limit of small noise), the firing rates of individual neurons are equal to the population firing divided by . Thus, the firing rate of each neuron can be approximated as . To construct the Poisson generator network, we removed the fast connections (but not the slow connections) and replaced the LIF neurons by Poisson spike generators with the same instantaneous firing rate, i.e., . The resulting recurrent network roughly matches the instantaneous firing rates (but not the performance) in the LIF network. The match could be enhanced, for example by adding a small baseline firing rate or a refractory period; However, these changes can only decrease the performance of the Poisson rate model. To obtain the filters in the integrator network (Fig. 7 F), we performed the following procedure: The inhomogeneous integrator was driven by an input sampled from Gaussian white noise (with mean , standard deviation ) and convolved by an exponential filter of width ms. The spike trains of the ten “recorded” neurons were modeled as independent Poisson processes with instantaneous firing rates(54)The feed-forward weights and lateral filters were estimated by maximizing the log-likelihood of the spike trains, following the method of [28]. Briefly, the filters were discretized in 500 time bins of , and conjugate gradient ascent of the log likelihood was performed on the value of the filters in each time bin for the equivalent of 5 hours of recording. The of a spike train is defined as(55)where is the total number of spike in the spike train, and is the duration of the inter-spike interval. The reported in the paper are the value of measured in each neuron and averaged over the population.

10.1371/journal.ppat.1002734

Structure and Functional Analysis of the RNA- and Viral Phosphoprotein-Binding Domain of Respiratory Syncytial Virus M2-1 Protein

Respiratory syncytial virus (RSV) protein M2-1 functions as an essential transcriptional cofactor of the viral RNA-dependent RNA polymerase (RdRp) complex by increasing polymerase processivity. M2-1 is a modular RNA binding protein that also interacts with the viral phosphoprotein P, another component of the RdRp complex. These binding properties are related to the core region of M2-1 encompassing residues S58 to K177. Here we report the NMR structure of the RSV M2-158–177 core domain, which is structurally homologous to the C-terminal domain of Ebola virus VP30, a transcription co-factor sharing functional similarity with M2-1. The partial overlap of RNA and P interaction surfaces on M2-158–177, as determined by NMR, rationalizes the previously observed competitive behavior of RNA versus P. Using site-directed mutagenesis, we identified eight residues located on these surfaces that are critical for an efficient transcription activity of the RdRp complex. Single mutations of these residues disrupted specifically either P or RNA binding to M2-1 in vitro. M2-1 recruitment to cytoplasmic inclusion bodies, which are regarded as sites of viral RNA synthesis, was impaired by mutations affecting only binding to P, but not to RNA, suggesting that M2-1 is associated to the holonucleocapsid by interacting with P. These results reveal that RNA and P binding to M2-1 can be uncoupled and that both are critical for the transcriptional antitermination function of M2-1.

Premature termination of transcription by the RNA-dependent RNA polymerase (RdRp) complex of respiratory syncytial virus (RSV) is prevented by the M2-1 protein. This transcription factor interacts with both RNA and viral phosphoprotein P, the main RdRp cofactor, through a specific “core” domain. Using NMR, we solved the 3D structure of this domain and characterized the surface residues involved in P- and RNA-binding. Based on these data, we designed point mutations impairing binding to either RNA or P. We studied the functional implications of these mutations for transcription and co-localization of full-length M2-1 in cytoplasmic inclusion bodies, where viral transcription likely occurs. We found that the RNA- and P-binding surfaces are in close proximity, accounting for competitive binding in vitro. Our results suggest that binding to both RNA and to P is necessary for transcriptional antitermination by M2-1, but that the role of the interaction with P is primarily to recruit M2-1 to the RdRp complex. Finally we show that the M2-1 core domain is homologous to the C-terminal domain of Ebola virus VP30 despite low sequence identity, solidifying the relationship between these two proteins and transcriptional regulation strategies shared by viruses belonging to the Filoviridae family and the Pneumovirinae subfamily.

Human respiratory syncytial virus (RSV), a pneumovirus of the Paramyxoviridae family in the Mononegavirales order, is an important respiratory pathogen and the major cause of bronchiolitis and pneumonia in children [1]. Bovine RSV on the other hand represents an important economic issue due to the high morbidity and mortality of infected calves [2]. Whereas current efforts are mainly focused on the development of safe RSV vaccines for infants, the development of antiviral drugs specifically targeting viral-specific functions such as the RSV RNA-dependent RNA polymerase complex (RdRp) represents a promising alternative for treatment. Four of the 11 proteins (the nucleoprotein N, the phosphoprotein P, M2-1 and the large polymerase subunit L), encoded by the RSV single-stranded negative-sense genomic RNA, are associated with the viral genome to form the holonucleocapsid [3]. The genomic RNA of RSV is maintained as a nuclease-resistant N-RNA ribonucleoprotein complex, which acts as a template for the RdRp that is responsible for both replication and transcription of the genome. Whereas the highly processive replicase generates a complete positive-sense RNA, which acts in turn as a template for genomic RNA synthesis, the transcriptase produces ten different subgenomic capped and polyadenylated mRNAs. Transcription proceeds by a sequential stop-and re-start mechanism in which the polymerase responds to cis-acting signals present in intergenic regions [4]. Transcription is (re)initiated at a highly conserved 9–10 nucleotide transcription promoter (gene start, GS) signal. Semi-conserved 12–13 nucleotide gene ends (GE) signal for polyadenylation and release of the nascent mRNA [5]. The polymerase has a propensity to dissociate from the N-RNA template, but cannot reinitiate at a downstream gene in case of premature termination [6], which leads to a decreasing transcription gradient from the 3′ to the 5′ end of the genome. For all known pneumoviruses, RdRp driven transcription depends on M2-1. RSV M2-1 is a transcription antitermination factor that is important for the efficient synthesis of full-length mRNAs [7] as well as for the synthesis of polycistronic readthrough mRNAs [4], [6], [8], [9]. The latter activity is thought to facilitate polymerase access to promoter-distal regions of the genome, and hence transcription of all genes [3]. It was shown that the M2-1 protein reduced termination at all gene junctions, but that the efficiency in the presence of M2-1 varied at the different gene junctions [6], [9]. However, mechanisms by which M2-1 prevents the polymerase from terminating transcription remain to be clarified. There are at least three different scenarios. (i) M2-1 could bind to the nascent mRNA transcript to facilitate transcription elongation, perhaps by preventing the mRNA from re-hybridizing to the template, or forming secondary structures that might destabilize the transcription complex. This hypothesis is sustained by the finding that RSV mRNA are co-precipitated with M2-1 from RSV infected cells [10]. (ii) The polymerase processivity enhancing effect of M2-1 could be due to an increase of the affinity of the polymerase for the genomic RNA template in a sequence non-specific manner. (iii) M2-1 could recognize GE sequences either on the nascent mRNA or on the RNA template and prevent the release of the polymerase complex from its template, favoring transcription re-initiation at the downstream GS sequences. In RSV-infected cells, M2-1 co-localizes with the other RdRp components in inclusion bodies (IBs) [11], which are regarded as centers of RNA synthesis [12]. The basic 194-residue RSV M2-1 has been shown to be an RNA binding protein, but specificity of M2-1 RNA binding has been debated. It was reported that M2-1 was able to bind to long RNAs without sequence specificity and with an apparent Kd of 30 nM, and that it bound specifically to short (80 nucleotides) but not long (700 nucleotides) RNAs containing the positive-sense antigenomic leader sequence with an apparent Kd of 90 nM [13]. Elsewhere it was demonstrated that M2-1 interacts more specifically with viral mRNAs not containing the leader sequence during infection [10]. In addition M2-1 interacts with P in vitro [14], competitively to RNA [15]. M2-1 is a modular protein that contains four domains. M2-1 forms tetramers in solution [15], [16], and the oligomerization domain was mapped to the region 33–62 [15]. The N-terminal region (residues 1–30) contains a putative zinc binding domain with a Cys3His motif, which is essential for the function of M2-1, but whose exact role is still unknown [17], [18], [19]. The RNA and P binding properties are related to the central part (or core domain) of the molecule (residues 53–177) [13], [15]. The C-terminal tail is predicted to be unstructured. Here we have addressed the molecular basis of the interaction between RSV M2-1 and its partners. We have investigated the structural aspects of P and RNA binding to the core domain of M2-1 by NMR spectroscopy and report the solution structure of RSV M2-158–177, which shows structural homology with the C-terminal domain (CTD) of the Ebola virus (EBOV) VP30 protein. In this α-helical domain we have identified residues that contribute to two adjacent, partially overlapping contact surfaces, with P and RNA respectively. We show that mutations of several of these residues specifically disrupt the M2-1:P or M2-1:RNA interaction in vitro and have a drastic effect on intracellular co-localization of full-length M2-1 with P as well as on the function of M2-1 as a transcription co-factor. The boundaries of the protein fragment M2-158–177 were chosen to focus on the binding regions of RNA and RSV phosphoprotein determined previously, but also to exclude the oligomerization domain and the disordered C-terminus, which are not necessary for the interactions with RNA and P [15]. Line widths of the solution NMR spectra were compatible with a monomeric state, and M2-158–177 was amenable to structure determination by NMR, in contrast to tetrameric full-length M2-1. The resonance assignments were reported elsewhere [20]. M2-158–177 contains a single globular domain spanning residues G75-I171 and comprising six helices: α1 (G75-G85), α2 (K92-E105), α3 (S108-D117), α4 (K124-K140), α5 (K143-R151) and α6 (D155-I171). The N-terminus (S58-L74), which corresponds to the linker to the upstream oligomerization domain of M2-1, is disordered. The α-helix bundle consists of a scaffold, formed by α1, α2, α5 and α6, and an α3–α4 hairpin stacked on α6 (Figure 1A). M2-158–177 displays two oppositely charged faces (Figure 1B). The positively charged face contains a large basic cluster along a grove delimited by helices α2 (K92), α5 (K150, R151) and α6 (K158, K159, K162, K169). Three smaller basic clusters are found on α4 ( K124 and R126), on α3 (K112, K113 and R115), and between α4 and α5 (R139, K140 and K143) as shown in Figure 1B. The putative overall tetrameric domain organization of full-length M2-1 is schemed in Figure 1C. Incubation of M2-158–177 with yeast RNA in a ∼1∶1 molar ratio resulted in simultaneous shifting and broadening of several 1H-15N cross peaks in the 1H-15N HSQC spectrum of M2-158–177 (see Figure S1). Treatment with RNAse A reversed these effects. This experiment confirmed the RNA binding ability of M2-158–177 in vitro. The observed fast to intermediate exchange regime was an indication for a weak interaction. To get rid of the broadening contribution by transversal relaxation of large RNAs twice the molecular weight of M2-158–177, we investigated the RNA:M2-158–177 interaction by NMR by using short synthetic 10–15 nucleotide negative-sense (genomic) RNAs containing selected transcriptional signals [3], [21] as well as their complementary positive-sense sequences. We tested the 3′ polymerase entry site (leader), the U-rich region upstream of the first GS signal, the GS, and the F and SH gene ends (GE_F and GE_SH). Sequences are detailed in Table 1. The oligonucleotides were designed to minimize self-association or formation of secondary structure. Only the leader-neg (5′-CGCAUUUUUUCGCGU-3′) and long U-rich (5′-CCCAUUUUUUUGGUU-3′) sequences were predicted to form hairpins with a poly-U loop and a stem of two or three G-C and A-U pairs with negative free energies. Calculations were carried out on the mfold web server [22]. The absence of self-association was assessed by 1H NMR using 200 µM oligonucleotide solutions in water, except for leader-pos and GE_F-pos, for which two broad imino proton signals were observed in the guanosine and uridine regions respectively, corresponding to partial formation of two G-C and U-A base pairs. Finally formation of an RNA duplex with the two complementary strands GE_F-pos and GE_F-neg was observed by NMR with sharp signals corresponding to base-paired uridine imino protons. Oligonucleotide binding was followed by chemical shift perturbation experiments in 1H-15N HSQC and 1H-13C HSQC experiments, as exemplified for the short polyA sequence in Figure 2 (panels A to C). The chemical shift variation profile in the 1H-15N HSQC spectrum of short polyA is shown in Figure 2D. Chemical shift perturbation profiles were similar for all RNAs and consistent with those observed for yeast RNA. A complete set of profiles for all oligonucleotides is given in Figure S2. For all RNAs the largest chemical shift changes were observed for residues belonging to the main basic cluster, i.e. K92-V97 (α2), L149-L152 (α5) and D155-K159 (α6), as illustrated in Figure 2E. These residues contribute to forming a continuous positively charged surface, which is consistent with an RNA binding surface (Figure 2F). Chemical shift variations in 1H-13C HSQC experiments were observed for residues in the same region, solvent exposed methyl side chains in the hinge between helices α5 and α6 as well as R151-Hδ protons being affected by RNA binding (see Figure 2B and 2C). The fast to moderately fast exchange regime in both 1H-15N and 1H-13C HSQCs allowed to extract apparent dissociation constants for a binding model with a 1∶1 stoichiometry. The fitted chemical shift variation curves for individual residues in a fast exchange regime are given in Figure S3. The values of the apparent dissociation constants (Kd) are recapitulated in Table 1. They range from 2.5 µM to >600 µM and fall into two groups. Except for a short U-rich RNA sequence, Kds for U-rich negative-sense RNAs are in the 75–125 µM range. UGA2, an RSV-unrelated stem-loop RNA comprising a U-stretch and five base pairs, used as a control, binds with similar affinity (Kd = 60 µM). A second control was carried out with a short single stranded DNA equivalent to the GE_F-neg RNA, showing that the same DNA sequence (250 µM) binds with slightly less affinity than the RNA sequence (85 µM). Apparent Kds for A-rich positive-sense RNA sequences (leader-pos, GE_SH-pos, GE_F-pos and short A-rich) are in the 2.5–22 µM range. A last experiment with the double-stranded F gene end (GE_F-ds, 4.5 µM) shows that the affinity is further increased as compared to the affinities of single-stranded RNAs. Taken together, these results indicate that M2-158–177 binds to RNA with lower affinity than reported for full-length M2-1 [10], [13]. Importantly they show that the presence of A-rich stretches increases binding affinity as compared to U-rich stretches and that RNA base-pairing might also play a role. As shown above, helices α2, α5 and α6 form a uniform RNA binding surface. However, residues L74 and G75 in helix α1, which are located on the opposite negatively charged side of the protein, are also affected by RNA binding. For all RNAs, Kds determined for residues in helix α1 were always the same as those measured for residues of the main binding site. This strongly suggests that perturbations in α1 on the one hand and in α2, α5 and α6 on the other hand are related to the same binding event. Moreover as the hypothetical second contact surface with RNA would be very distant from the main binding site, we exclude the possibility of a second binding site at α1. However RNA binding to α2 could be transmitted to α1 by slight alterations of helix-helix packing. To compare RNA and P affinities for M2-158–177, we performed isothermal titration calorimetry (ITC) experiments using this truncated form of M2-1. Results showed that the phosphoprotein P binds to M2-158–177 with a stoichiometry of 1∶1 and a Kd of ∼3 µM (Figure S4). We investigated this interaction further by NMR for atomic details. Perturbations induced by tetrameric full-length P in 1H-15N and 1H-13C HSQC spectra of M2-158–177 were monitored. Due to the size of the complex and the unfavorable exchange regime, addition of 0,5 molar equivalents of P was accompanied by extensive overall line broadening in the 1H-15N HSQC spectra, except for the unstructured N- and C-termini (see Figure S5A). At lower P concentrations, the majority of cross-peaks remained detectable. Residues T130-L165 exhibited larger line broadening, suggesting that P binds to the α4/α5/α6 region. Transferred cross-saturation (TCS) experiments [23] were carried out with 2H15N-M2-158–177 to more specifically identify residues involved in P binding (Figure 3, panels A to D). Saturation of the methyl protons of P resulted in reduction of M2-1 cross peak intensities in the regions V127-S137 and L152-T164 (helices α4 and α6, and α5/α6 hinge). They form a nearly uniform surface on M2-1 (Figure 3D). Since P residues 100–120 were reported to be critical for M2-1 binding [14], [24], we also monitored intensity variations in 1H-15N HSQC spectra in the presence of a truncated form of P, P100–166, comprising this region as well as the oligomerization domain of P [25], [26]. The results are shown in Figure 3 (panels E to H). Line broadening was enhanced for helices α4 and α6 (Figures 3F and 3G), which is consistent with the results obtained with full-length P. In summary, P binds to a region proximal to the RNA anchoring surface. Helices α6 and α5 were sensitive to both P and RNA binding in our NMR experiments, and could thus contribute to both interaction sites, while α4 and α2 appear to be involved specifically in P binding and RNA binding, respectively. Based on these results, we designed single residue mutations of the full-length M2-1 protein targeting the binding sites of P and RNA. Solvent-exposed residues were first substituted by Ala. The effect of these mutations on transcription antitermination by M2-1, which was previously shown to be the same for RSV-specific and heterologous sequences [6], was assessed using an RSV dicistronic subgenomic replicon, pM/Luc. It contains the authentic M/SH gene junction, and the Luc reporter gene downstream of the gene start sequence present in this gene junction. The expression of the Luc gene in this system is absolutely dependent on the presence of a functional M2-1 [15]. The pM/Luc plasmid was co-transfected in BHK-21 BSRT7/5 cells expressing T7 RNA polymerase together with p-β-gal, pL, pP, PN, and pM2-1. Luciferase activities were determined and normalized based on β-galactosidase expression [5], [8], [15], [27], [28]. Except for L148A and N163A, most of the Ala substitutions had only little effect on transcription, as assessed by Luc expression (Figure 4A). As the P- and RNA-binding surface identified by NMR is highly positively charged, residues were substituted by Asp to emphasize the electrostatic effect of mutations. When substituted by Asp, K92 (in helix α2), a residue involved in RNA but not in P binding according to the NMR results, and R126 and T130 (in helix α4), two residues involved in P but not in RNA binding, appeared to be critical for transcription. A similar effect on transcription was observed for four other residues (K150, R151, T160 and N163, in helices α5 and α6), that belong to the dual RNA/P binding surface. In contrast, Ala and Asp mutants of K158, which is also located in the RNA/P binding region, still displayed more than 80% activity as compared to the WT (Figure 4A). As a control, we further verified by NMR that the deactivating mutations did not disrupt the protein fold (Figure S6). Co-expression of P and N proteins (in the absence of other viral proteins) in cells induces the formation of cytoplasmic inclusion bodies containing P-N complexes, as observed in RSV infected cells [11]. When co-expressed with P and N, WT M2-1 also localizes preferentially in these IBs as seen in Figure 4D [11]. We analyzed the intracellular localization of the M2-1 mutants in cells co-transfected with expression vectors for P, N and M2-1 by fluorescence microscopy (Figure 4C and 4D). Contrary to WT M2-1, the mutants R126D, T130D, L148A, T160D and N163D were excluded from the IBs and spread all throughout the cytoplasm. In contrast K92D, K150D and R151D retained their localization to cytoplasmic IBs. These mutants were expressed at comparable levels, as determined by Western blotting (Figure 4B), and were correctly folded when purified from E. coli and analyzed by NMR (Figure S6). To verify that the residues identified using NMR and the minigenome assay are critical for RNA- and/or P-binding, the in vitro RNA and P binding capacities of eight M2-1 mutants selected by the Luc assay were investigated. As M2-1 did not migrate in native agarose gel, it was not possible to obtain electrophoretic mobility shift assays (EMSA) with the GST-free forms. We thus used M2-1 fused to GST for the in vitro binding assays with RNA and P. For RNA binding assays, we used either full-length (tetrameric) or truncated 58–177 (monomeric) forms of M2-1 fused to GST. GST-M2-158–177, incubated with tRNA, was analyzed by EMSA. Formation of GST-M2-158–177:RNA complexes was only impaired by the K92D, K150D and R151D mutations, which did not prevent M2-1 association with IBs (Figure 5A and 5B). The effect of single mutations of the full-length form on in vitro RNA-binding was smaller than with truncated M2-158–177, probably owing to the higher avidity of tetramers for RNA compared to monomers, and the Lys/Arg repetitions that limit the effect of single substitutions (data not shown). In vitro M2-1:P interactions were assessed by GST pull-down assays. As shown in Figure 5C and 5D, a≥50% decrease in P binding was observed for the mutants R126D, T130D, L148A, T160D and N163D, which were excluded from IBs, but could still bind RNA. The two separate binding surfaces for RNA and P, determined from these 8 residues, are illustrated in Figure 5E. Together with the cellular localization experiments, these results show that there is a strong correlation between the reduced capacity of M2-1 to pull-down P in vitro and its exclusion from IBs. Conversely, since residues specifically involved in RNA binding appear not to be crucial for recruitment of M2-1 to IBs, the M2-1:RNA interaction is not required for this process. In contrast to other members of the Mononegavirales order for which transcription is driven by only three proteins (N/NP, P and L), pneumoviruses and filoviruses encode a fourth transcription co-factor (M2-1 and VP30 respectively). Parallels have been drawn between VP30 and M2-1 on the basis of their functions and domain organization [17], [29]. Both were shown to be necessary for efficient transcription in reconstituted minigenome systems [7], [30]. In addition, although VP30 and M2-1 were shown to be dispensable for RNA replication [30], [31], [32], both are required to rescue recombinant EBOV or Marburg virus and RSV [33], [34], [35]. M2-1 and VP30 contain an N-terminal Cys3His1 motif that does not bind RNA directly, but that is essential for VP30 RNA binding [15], [17], [36], [37]. In the case of M2-1 this motif is indispensable for function, but its exact role is still unknown [17], [18], [19]. Like M2-1, VP30 also contains an oligomerization domain downstream of the Cys3His1 motif, which is necessary for its function during transcription [15], [36]. But, despite functional similarities, the sequence identity between the two proteins is very low and it was not possible to know that they were evolutionary related. Here we show that the globular domains M2-158–177 and EBOV VP30CTD [38] are structurally homologous and that they display the same α-helix bundle fold. The structures were aligned on the DALI server [39] which yielded a Z-score of 5.7 and an rmsd of 3.9 Å for 92 aligned residues and 9% sequence identity (see Figure 6). The core helices α1, α2, α5 and α6 align well, with a slight shift of α1M2-1 with respect to α1VP30. The α3–α4 hairpins are skewed relative to each other. Moreover M2-158–177 and VP30CTD contain several identical hydrophobic residues in helices α5 and α6 (indicated in Figure 6), which are involved in stabilizing inter-helical contacts and appear to be semi-conserved among pneumovirus/metapneumovirus M2-1 and filovirus VP30 protein sequences (Figure S7). Although they mainly contribute to inter-helix packing, it is noteworthy that one of them (L148) is also critical for M2-1 transcription antitermination. VP30CTD contains an additional 7th C-terminal helix, which has no counterpart in M2-1. It stabilizes the crystallographic VP30CTD dimer by interaction with α1. Since full-length M2-1 forms tetramers with the oligomerization domain just upstream of the core domain, its monomeric state in isolation suggests that the core domains may at best be loosely associated with each other in the tetramer, in the absence of partner molecules, as schemed in Figure 1C. The overall structural match rationalizes the relationship between M2-1 and VP30. Importantly both proteins associate with the nucleocapsid by means of their globular core domains [10], [14], [38], [40]. Whereas the VP30CTD:nucleocapsid interaction is mediated by the EBOV nucleoprotein but not by RNA [41], the M2-1:nucleocapsid interaction has been proposed to be mediated by RNA [10] and by P [14]. RNA binding to VP30CTD has not been observed [38], in contrast to what we report for M2-158–177. As no direct M2-1:N complex was evidenced in vitro [10], [14], the correlation found between M2-1 localization to cytoplasmic IBs, containing N and P, and the capacity of M2-1 to bind P in vitro indicates that binding of M2-1 to P drives recruitment of M2-1 to the holonucleocapsid. M2-1 was first described as an antitermination factor, preventing cessation of chain elongation and release of the nascent mRNA [7]. Later M2-1 was reported to inhibit transcription termination at the GE signals to produce polycistronic readthrough mRNAs [6], [8], [9], depending on GE sequences. The semi-conserved 12–13 nucleotide GEs fall into three groups with respect to this property: the first (NS1/NS2, NS2/N, M2/L, and L/trailer) contains sequences inefficient for transcription termination and is insensitive to M2-1; the second (N/P, P/M, M/SH, SH/G, G/F) is very efficient in transcription termination but not sensitive to M2-1; finally (F/M2) is highly sensitive to M2-1 [9]. M2-1 did not direct readthrough at the leader-NS1 junction [6]. Thus, it was suggested that M2-1 would not only prevent inappropriate intragenic termination, but also allow the polymerase to access to promoter-distal regions of the genome and to transcribe downstream genes [9]. Cuesta et al. reported that renatured M2-1 bound preferentially to a short positive-sense leader RNA in vitro [13], while Cartee and Wertz co-immunoprecipitated RSV mRNA with M2-1 from infected cells treated with Actinomycin D [10]. These observations suggested that M2-1 could modulate transcription termination by recognizing specific viral RNA sequences, either on transcribed mRNAs or on the genomic RNA template. Contrary to transcribed viral mRNA, genomic RNA is encapsidated. However if the model of RNA synthesis by the vesicular stomatitis virus, a prototype of nonsegmented negative-strand RNA viruses, could be transposed to RSV, genomic RNA would become accessible as the nucleocapsid is being locally disassembled by the transcription complex [42], [43]. By comparing the affinities of M2-158–177 for short 10–15 nucleotide RSV RNA sequences, we found that M2-158–177 bound preferentially to positive-sense RNA sequences such as the 12 first bases of the positive-sense leader (Kd = 2.5 µM), positive-sense GEs (11 and 13 µM) and an A-rich sequence located on the positive-sense leader (22 µM). Affinities of the negative-sense signal sequences were systematically lower by one order of magnitude or more than those of their complementary sequences. But there was no difference between the SH and F gene ends, which are respectively insensitive and highly sensitive to M2-1. The 3′ and 5′ extremities of the leader and the GS displayed a similar behavior to that of GEs of same polarity, indicating that M2-158–177 does not discriminate between transcription signals of same polarity. Sequence specificity of RSV transcription antitermination thus appears not to be linked to the M2-1 core domain. These results are also consistent with the observation that M2-1 is not required for initiation of RNA transcription [7]. Our results show that M2-158–177 displays a preference for purine-rich and especially A-rich RNAs found in positive-sense RSV RNAs over pyrimidine-rich sequences containing U-stretches found in negative-sense RNAs. Thus, they reinforce the hypothesis that M2-1 binds preferentially to positive sense RNA transcripts rather than to the template, in agreement with previous reports [10], [13]. An additional finding is that M2-1 binds to the double-stranded F gene end with similar or better affinity (4.5 µM) than to the positive-sense sequence alone (11 µM). This implies that M2-1 could bind to the nascent mRNA transcript, either still bound to the template or released from it. In both cases, in order to facilitate transcription elongation, M2-1 could prevent formation of mRNA secondary structures that might destabilize the transcription complex, in analogy to the function of N protein that binds to the nascent RNA during replication. Using in vitro experiments, we have previously shown that RNA and P bind to M2-158–177 in a competitive way [15]. Here we have characterized two distinct binding surfaces for P and RNA by NMR (Figure 5E). Their edges are partially overlapping. The P binding epitope, which is buried between helix α4 and the hinge between α5 and α6 (Figure 3D and 3H), might be occluded by RNA binding in the vicinity, and conversely RNA binding might be hampered by bound P due to steric hindrance. Both binding sites are located on the positively charged face of M2-158–177, underlining the importance of electrostatic interactions for association with RNA and with P. This hypothesis is also in agreement with the predicted negatively charged surface in computer models of the M2-1 binding region of the P tetramer, spanning residues 100–120 of P [44]. Redundancy of positive charges in the P and RNA binding region certainly accounts for the weak effect on Luc expression observed for Ala mutants of residues in this region in the minigenome assay, as compared to Asp substitution, which introduced opposite charges. Electrostatic interactions could be further emphasized in the M2-1 tetramer if the core domains were arranged to provide an extended binding surface consisting of two adjacent positively charged clusters. In addition to charge effects, specific hydrophobic interactions may also contribute, as suggested for the interaction between P and M2-1 in [14]. This could apply to RNA as well. Although the range of RNA affinities for M2-158–177 (micromolar to millimolar) is an indicator for non-specific binding mediated by the RNA phosphate backbone, corroborated by binding of a DNA sequence equivalent to GE_F-neg, the difference between A-rich and U-rich RNA affinities indicates that the nature of the bases also comes into play. The large M2-158–177:RNA binding surface determined by NMR coincides with the main positively charged cluster of M2-158–177 (Figure 1B and 2F), which is well conserved among Pneumovirinae (Figure S7). We confirmed by mutagenesis that the three basic residues K92, K150 and R151, included in this epitope, are crucial for in vitro RNA binding to M2-158–177 and for transcription enhancement in vivo by full-length M2-1. However they did not prevent association with cytoplasmic IBs or P binding. In contrast, mutants R126D, T130D, L148A, T160D and N163D, for which M2-1-controlled transcription was impaired and which still bound RNA in vitro, had lost their ability to bind P in vitro and did not co-localize with N-P complexes in cytoplasmic IBs. These residues are not well conserved in Pneumovirinae M2-1 proteins (Figure S7A), but this would be consistent with the sequence variability of the M2-1 binding region of P(100–120), which has co-evolved with M2-1. Altogether the NMR and mutagenesis results provide a rationale for the competitive binding previously observed between P and RNA [15]. They highlight the role of P for the recruitment of M2-1 to cytoplasmic IBs that contain N and P and where viral RNA synthesis takes place, by analogy to Rhabdoviridae [12]. Possible cooperative effects in full-length tetrameric M2-1, involving the N-terminal Cys3His motif and interactions between core domains, may induce an increased affinity for both P and RNA as well as sequence specific recognition of RNA, with respect to the core domain of M2-1. However the overall lower affinities of RNA for M2-158–177, as compared to P, suggest that M2-1 binds P preferentially in the absence of viral RNA so that M2-1 would be recruited as an RNA-free M2-1:P complex to the IBs. This hypothesis is in agreement with our fluorescence microscopy observations, since WT M2-1 and mutants that do not bind RNA are found likewise in IBs containing N and P (Figure 4C and 4D). Indeed, higher affinity to RNA would result in sequestration of M2-1 in the cytoplasm, where cellular RNA is highly accessible, and M2-1 would not be recruited to the IBs. The relatively low Kds are in the same range as those measured for the Measles virus P-N interaction [45], this intrinsically weak association being probably required for movement of the viral polymerase during RNA transcription. In summary, our results indicate that not only P:M2-1, but also RNA:M2-1 interactions are required for efficient transcription activation by M2-1. Association with P is strictly required for recruitment to the viral RNA synthesis site. As suggested by the proximity of the binding sites, it is likely that the P:M2-1 interaction is displaced in favor of other interactions, RNA:M2-1 interactions in particular, in the context of the holonucleocapsid. The higher affinity of M2-158–177 for the 5′ end of the positive-sense Leader RNA (2.5 µM), which is in the same range as the affinity for P (3 µM), suggests that M2-1 could be loaded onto the polymerase during transcription initiation. It can further be speculated that an L:M2-1 interaction also takes place when the P:M2-1 interaction breaks down, thus altering the sensitivity of the polymerase to transcription termination signals. A similar L:VP30 interaction was recently reported for EBOV [46]. This hypothesis is currently under investigation for RSV. Recombinant P, P100–166, M2-1 and M2-158–177 (WT and mutants) as well as 15N-,13C- and/or 2H-labeled M2-158–177 were expressed in E. coli BL21(DE3) strain. Full-length M2-1 and M2-158–177 were produced as GST fusion proteins as described in [15] and [25] with a thrombin cleavage site. M2-1 amino acid substitution mutants were obtained with the Quickchange site-directed mutagenesis kit (Stratagene) by using pGEX-M2-1, pGEX-M2-158–177 and pM2-1 as templates. Isotopically labeled proteins for NMR were produced in minimal M9 medium supplemented with 1 g/L 15NH4Cl and 4 g/L 13C- or unlabeled glucose (Cortecnet). Protocols are detailed as Supplemental Materials and Methods (Text S1). BSRT7/5 cells stably expressing T7 RNA polymerase [47] were maintained in EMEM supplemented with 10% FCS/l-glutamine/penicillin-streptomycin solution. The cells were grown in an incubator at 37°C under 5% CO2. pN, pP, pM2-1 and pL plasmids coding for HRSV (strain Long) N, P, M2-1 and L proteins respectively, under the control of the T7 promoter, have been described previously [15], [48]. An encephalomyocarditis virus internal ribosome entry site (IRES) sequence was placed between the T7 promoter and the inserted ORF to enhance protein expression in BSR/T7-5 cells, as previously described [48], [49]. The pM/Luc subgenomic replicon was derived from the pM/SH subgenomic replicon [8] and has been described previously [15]. It contains two transcription units, the second encoding the firefly luciferase (Luc) gene under the control of the M/SH intragenic sequence. The expression of the Luc gene in this system is absolutely dependent on the presence of M2-1 [15]. BSRT7/5 cells were transfected with pN, pP, pL, pM2-1, p-β-Gal coding for beta-galactosidase under the control of the Rous sarcoma virus promoter (Promega) and pM/Luc, where luciferase expression is controlled by the RSV M/SH intergenic region [15]. Luciferase activity was determined in triplicate 24 h post-transfection as previously described [15]. Cells were lysed in luciferase lysis buffer (30 mM Tris, pH 7.9, 10 mM MgCl2, 1 mM dithiothreitol [DTT], 1% [vol/vol] Triton X-100, and 15% [vol/vol] glycerol), and luciferase activities were evaluated twice for each cell lysate with an Anthos Lucy 3 luminometer (Bio Advance). BSRT7/5 cells were transfected with pP (0.4 µg), pN (0.4 µg) and pM2-1 (0.2 µg) containing either WT or mutant M2-1 by using Lipofectamine2000 (Invitrogen). Samples were fixed after 24 h in 4% paraformaldehyde, and permeabilized in PBS containing 0.1% Triton X-100 and 3% BSA. Each coverslip was incubated with primary antibodies: anti-N (1∶100 dilution) and anti-P (1∶500 dilution) rabbit polyclonal sera, and 37M2 and 22K4 anti-M2-1 monoclonal antibodies (1∶40 dilution) [50]. These samples were incubated for 1 h at room temperature, washed, and then incubated for an additional hour with Alexa Fluor 488 goat anti-mouse and Alexa Fluor 594 goat anti-rabbit (1∶1000) IgG (Invitrogen). Cells were observed with a Nikon TE200 inverted microscope equipped with a Photometrics CoolSNAP ES2 camera. Images were processed using MetaVue software (Molecular Devices). GST-M2-158–177 fusion proteins (WT or mutants, final concentration 100 µM) were eluted by GSH and incubated with yeast tRNA (Sigma, final concentration 50 µM) in a final volume of 10 µl. Complexes were resolved by 1.5% agarose gel electrophoresis in 1× Tris-Glycine buffer at 4°C, stained with ethidium bromide and amido black. GST pull-down of purified recombinant P by full-length GST-M2-1 fusion proteins (WT and mutants) was performed by incubating 10 µl aliquots of a 50% slurry of Glutathione-Sepharose 4B beads (GE Healthcare) containing ∼25 µM GST-M2-1 in PBS with a 3-fold molar excess of P for 1 h at 20°C under agitation. Beads were washed extensively with PBS, boiled in 25 µl Laemmli buffer and analyzed by SDS-PAGE and Coomassie blue staining. Bands were quantified using ImageJ software. Affinity between P and M2-1 was determined by isothermal titration calorimetry. Raw ITC data were processed and quantitative analysis of the P-M2-158–177 interaction was performed on a MicroCal ITC200 microcalorimeter (Microcal, Northampton, MA). Samples were dialyzed against 1× PBS for 15 h. The experiments were carried out at 20°C. The P concentration in the microcalorimeter cell (1.4 mL) was of 33 µM. In total, 20 injections of 2 µL of M2-158–177 solution (concentration 335 µM) were carried out at 180 s intervals, with stirring at 1000 rpm. The experimental data were fitted to a theoretical titration curve with software supplied by MicroCal (ORIGIN). This software generates titration curves based on the relationship between the heat generated by each injection and ΔH (enthalpy change in kcal/mol), Kd (dissociation constant), n (number of binding sites), the total protein concentration and free and total ligand concentrations. 10–15 nucleotide RNAs corresponding to viral RNA sequences of negative and positive polarity were synthesized on a Pharmacia Gene Assembler Plus using phenoxyacetyl β-RNA phosphoramidites and Universal Support resin (Glen Research) with a protocol adapted for RNA from DNA solid-phase synthesis [51]. The products were HPLC-purified (Beckman) by anionic chromatography on a DEAE-Sepharose column (Pharmacia) in 10 mM phosphate buffer pH 6.8 and eluated with a linear gradient from 0.55 to 1 M NaCl in 80 min. The oligonucleotides were extensively dialyzed in water. If necessary the pH was adjusted to 7 with 0.5 M NaOH and the concentration determined using the A260 value. Aliquots were then lyophilized and stored at −20°C. The oligonucleotide sequences are given in Table 1. Folding properties of RNAs were estimated by using the mfold web server [22]. NMR measurements were carried out at 293 K (or 298 K) on Bruker Avance 600, 800 and 950 MHz spectrometers equipped with triple resonance cryoprobes. All samples were in 50 mM sodium phosphate buffer pH 6.8, 150 mM NaCl, 1 mM DTT and 7% D2O. Resonance assignment was reported elsewhere [52]. 15N-NOESY-HSQC (800 MHz) and 13C-NOESY-HSQC (950 MHz) experiments, simultaneously edited for aliphatic and aromatic 13C, with 80 and 100 ms mixing times, were recorded using 120 µM 15N- M2-158–177 and 100 µM 13C-M2-158–177 samples respectively. Data were processed with Topspin 2.1 or NMRPipe [53]. Spectra were analyzed with Sparky [54]. Residual dipolar couplings (RDCs) were collected on 15N-labeled M2-158–177 in two alignment media. The first sample contained 120 µM 15N-M2-158–177 in a stretched 6% acrylamide/bisacrylamide (37.5∶1) gel (deuterium splitting Δ2H = 3 Hz). The second sample contained 200 µM 15N-M2-158–177 in a 5% hexanol/C12E5 phase with r = 0.96 molar ratio (Δ2H = 26,7 Hz). 1DNH RDCs were extracted from spin-state selective InPhase-AntiPhase 1H-15N HSQC experiments acquired in interleaved fashion. M2-158–177 structures were calculated with CYANA 2.1 [55], using distance restraints obtained from 15N- and 13C-NOESY-HSQC spectra and backbone torsion angles generated with TALOS. They were refined in Xplor-NIH [56] using 1DHN residual dipolar couplings measured in a stretched polyacrylamide gel and a hexanol/C12E5 phase. Structure statistics are summarized in Table S1. More details are provided in Text S1. Graphic representations were performed with PyMOL [57]. Chemical shift and cross-peak intensity perturbations experiments were carried out with yeast RNA (Roche) and short synthesized oligonucleotides (see Table 1). 0.25–20 equivalents of lyophilized RNA were added stepwise to 50–60 µM 15N-,13C- or 15N13C-labeled M2-158–177. 1H-15N and 1H-13C HSQC spectra were recorded at 293 K at each step, at magnetic fields of 14.1 T and/or 22.3 T. Interaction with phosphoprotein was probed by cross-peak intensity perturbations in 1H-15N and 1H-13C HSQC experiments, with full-length P (P1–241) and truncated P100–166. Samples with P1–241 were prepared by mixing concentrated solutions of P1–241 (30 µM) and 15N13C-labeled M2-158–177 (275 µM) in the same phosphate buffer to a final concentration of 40 µM M2-158–177 and P∶M2-1 molar ratios of 0.25∶1:and 0.5∶1. Samples with P100–166 were prepared by mixing solutions of P100–166 (600 µM) and 15N13C-labeled M2-158–177 (140 µM) to a final concentration of 45 µM M2-158–177 and a P∶M2-1 molar ratio of 2∶1. Spectra were typically recorded at 14.1 T and 293 K. For each RNA sequence, addition of RNA was followed by 1H-15N HSQC experiments of 15N-M2-158–177. Weighted averaged chemical shift differences were calculated from 1H and 15N chemical shifts according to following equation, where the 1/10 scaling factor for 15N chemical shifts corresponds to the ratio of gyromagnetic ratios of 15N and 1H: . The apparent dissociation constant Kd was obtained by fitting 1H or 15N chemical shift differences at each titration point with two parameters in a binding model with 1∶1 stoichiometry and with a user-defined function in Origin 7 software as follows: where , , . Samples were prepared with a 30 µM solution of P1–241 centrifuged at 15 krpm in a TLA55 rotor (Beckman) for 15 min and exchanged into D2O buffer with biospin columns according to the manufacturer's instructions (Biorad). Transferred cross-saturation experiments [23] were conducted with 150 µM 2H15N labeled M2-158–177 in 91% D2O in the presence of 15 µM (10%) unlabeled P1–241 protein. Broadband proton saturation (2 s) was achieved with a 1.8 kHz Wurst pulse centered at 0.8 ppm. Spectra were recorded in interleaved fashion with and without saturation, with 1.5 s recycling delay, at 14.1 T and 293 K. Intensity ratios were determined based on experiments with and without saturation. Control experiments were carried out without P to account for spin diffusion and effects of residual aliphatic protons in M2-158–177. Atomic coordinates and structural constraints have been deposited in the Protein Data Bank (PDB accession code 2L9J). Swiss-Prot P04545.1 UniProt P28887.1

10.1371/journal.pntd.0002447

A Review of Factors That Influence Individual Compliance with Mass Drug Administration for Elimination of Lymphatic Filariasis

The success of programs to eliminate lymphatic filariasis (LF) depends in large part on their ability to achieve and sustain high levels of compliance with mass drug administration (MDA). This paper reports results from a comprehensive review of factors that affect compliance with MDA. Papers published between 2000 and 2012 were considered, and 79 publications were included in the final dataset for analysis after two rounds of selection. While results varied in different settings, some common features were associated with successful programs and with compliance by individuals. Training and motivation of drug distributors is critically important, because these people directly interact with target populations, and their actions can affect MDA compliance decisions by families and individuals. Other important programmatic issues include thorough preparation of personnel, supplies, and logistics for implementation and preparation of the population for MDA. Demographic factors (age, sex, income level, and area of residence) are often associated with compliance by individuals, but compliance decisions are also affected by perceptions of the potential benefits of participation versus the risk of adverse events. Trust and information can sometimes offset fear of the unknown. While no single formula can ensure success MDA in all settings, five key ingredients were identified: engender trust, tailor programs to local conditions, take actions to minimize the impact of adverse events, promote the broader benefits of the MDA program, and directly address the issue of systematic non-compliance, which harms communities by prolonging their exposure to LF. This review has identified factors that promote coverage and compliance with MDA for LF elimination across countries. This information may be helpful for explaining results that do not meet expectations and for developing remedies for ailing MDA programs. Our review has also identified gaps in understanding and suggested priority areas for further research.

Lymphatic filariasis (LF, also known as “elephantiasis”) is a deforming and disabling disease that is caused by roundworm parasites that are transmitted by mosquitoes. The Global Programme to Eliminate Lymphatic Filariasis is the largest public health intervention program attempted to date based on mass drug administration (MDA). MDA does not cure filarial infections, but it can reduce or interrupt transmission of new infections by clearing larval parasites from human blood so that they are not available for mosquitoes. High levels of participation are required for this strategy to work; guidelines from the World Health Organization call for at least 65% of the eligible population to take the medications annually for four to six years. MDA presents logistical challenges that require cooperation between donors, health ministries, and communities. The success of MDA depends on coverage (drug delivery) and compliance (people ingesting antifilarial drugs), which depends on individual interactions between drug distributors and the people who live in LF-endemic areas. This paper focuses on this last step of implementation with a comprehensive review of published and unpublished information on factors that affect compliance with MDA at the level of the individual. We have also provided an outline of current gaps in understanding and recommendations for further research.

The Global Programme to Eliminate Lymphatic Filariasis (GPELF) is one of the most ambitious, exciting, and challenging public health programs of our time. Part of the challenge is in the large numbers of people that either have lymphatic filariasis (LF) (an estimated 120 million) or who are at risk for the disease (1.39 billion) [1], [2]. Like many other neglected tropical diseases (NTDs), LF disproportionately affects vulnerable populations and perpetuates existing relationships between disease and poverty [3]. The economic effects of LF can be devastating, as sufferers with disfigurement and disability due to lymphedema, hydrocele, and elephantiasis have reduced work capacity and household income [4]. The ripple effects of this loss of income limit the ability to pay for healthcare, education, and basic household expenses. The social effects of the disease can be equally devastating, potentially ostracizing people from their communities and families [5]. In 1997, the World Health Assembly targeted LF for elimination as a public health problem by the year 2020 [6]. The main tool for GPELF is repeated, annual mass drug administration (MDA) of antifilarial drugs to people living in endemic areas [7], [8]. According to guidelines put forward by the World Health Organization (WHO), at least 65% of the at-risk population should comply with annual MDA so that elimination targets can be reached within four to six years (WHO-GPELF). Implementation of MDA requires cooperation and coordination of activities by donors, national and local health officials, non-governmental organizations (NGOs), and communities. GPELF is currently active in 53 of 73 countries that are endemic for LF [9]. More than 536 million people received MDA in 2011, and this success depended on the coordinated efforts of the donor community, health ministries, community volunteers, NGOs, and research institutions [9]. GPELF guidelines require high compliance with MDA, but in practice this can be difficult to achieve and sustain. It can be challenging to convince people who feel well to take repeated doses of medicines that may cause adverse events [10]. Emerging challenges to GPELF include program fatigue, knowing what to do when elimination targets have not been reached after six years of MDA, identifying and targeting systematic non-compliers, and maintaining momentum and focus for LF elimination during the new era of integrated NTD control programs. Securing participation with MDA is essential if GPELF is to achieve the 2020 target for LF elimination. As the program enters its 14th year, the relevance of understanding how to reach individuals and convince them to comply with MDA has become increasingly important. While several countries have stopped MDA and started verification procedures and post-MDA surveillance [9], many LF-endemic countries are still providing MDA or preparing to start. A better understanding of factors that affect compliance with MDA at the level of the individual could have far-reaching effects as programs strive to adapt their campaigns to more effectively reach their target populations. Many articles have been published that address the issue of compliance with MDA for LF. This paper reports results of a thorough review of publications and unpublished information on this important subject. The goal of this review was to attempt to identify factors and patterns that are associated with compliance with MDA that apply across countries and cultures. We have also attempted to identify high-priority topics for additional research on compliance. Improved understanding of factors that affect an individual's compliance should be helpful not only for LF elimination activities but also for integrated NTD programs that employ MDA and/or preventive chemotherapy. A systematic search of 13 databases including PubMed, Medline, EMBASE, and CAB Abstracts was performed using the key words “compliance,” “non-compliance,” “predictor,” “factor,” “acceptance,” or “refusal” and “lymphatic filariasis,” “elephantiasis,” or “filarial” and “MDA,” “treatment,” “chemotherapy,” or “treatment coverage.” Papers published in English or French from 2000 through March 2012 were considered. The search generated 404 citations, and a preliminary review of title and abstract was performed using a flowchart for the first screening. In the second review, 86 papers were read in full. Those papers that met one of the following criteria were included in the final dataset: i) reviewed the literature on compliance with MDA for LF; ii) described or assessed factors associated with compliance with MDA for LF; iii) analyzed, observed, or documented compliance rates with MDA and/or provided an explanation or discussion of the rates; and iv) were identified from reference lists of primary papers. In addition to the published literature, informal interviews were conducted with five senior LF scientists who were selected based on their broad international experience and knowledge of issues surrounding compliance. These phone discussions provided the opportunity to solicit unpublished data and reports for inclusion in this review. A data extraction form was created using Microsoft Excel to enable a systematic analysis of the relevant themes arising from the literature. For the purposes of this paper, coverage is defined as the percentage of targeted persons who receive MDA medications, and compliance refers to the percentage of a targeted population who swallow the medications. Unless otherwise stated, MDA refers to mass drug administration at the community level; and low or inadequate compliance is described as <65%. See Table 1 for more details of definitions used in this paper. The search of published literature produced 79 papers that met criteria for inclusion in this review. Unpublished literature and data were provided from studies performed in four different countries (Haiti, Dominican Republic, Cameroon, and Egypt) for review. Additional data on compliance were contributed from small surveys conducted in Malaysia, Togo, Mali, and Niger. Types of study design included self-reported questionnaires, coverage surveys, household surveys, and intervention studies. Countries where published research was performed included (in order of most to least frequent): India (n = 40), Haiti (n = 7), Sri Lanka (n = 5), Papua New Guinea (n = 5), Kenya (n = 4), Indonesia (n = 3), Tanzania (n = 2), Sierra Leone (n = 1), Vanuatu (n = 1), Ghana (n = 1), Togo (n = 1), American Samoa (n = 1), Egypt (n = 1), and the Philippines (n = 1). Six papers included in the review were from multi-country or non-specific settings. The reviewed publications used varied definitions for coverage and compliance. Fifteen studies provided quantitative results for factors demonstrated to be associated with compliance (Table S1). However, the majority of the papers reviewed provided qualitative results and anecdotal data for factors associated with compliance (Table S2). Low MDA compliance may be related to the MDA delivery system and/or to characteristics of targeted recipients. For the purposes of this paper, the delivery system refers to those who distribute the MDA drugs at various levels (national, provincial, district, community, and individual drug distributors), and recipients are those individuals who are targeted to receive MDA in an endemic area. Results are presented according to their relation to the delivery or recipient side of the MDA equation. Problems with availability of drugs and/or promotional materials were cited as reasons for inadequate coverage in several studies [11]–[14]. For example, inadequate drug supply contributed to low MDA coverage rates in Vanuatu [15] and in East Godavari district in India [16]. Inappropriate distribution time was sometimes cited as negatively affecting coverage. In one Indian study, MDA was postponed and rescheduled to take place during a major Hindu fasting festival, and this made it difficult for distributors to reach individuals for directly observed treatment (DOT) [12]. In another Indian study from Kerala state, repeated postponement of the MDA resulted in poor coverage and high rates of non-compliance [17]. In other cases, insufficient time and personnel adversely affected coverage and compliance [13], [14], [18], [19]. The absence of eligible recipients of the drugs during MDA due to short- or long-term migration provided challenges to coverage in some areas [12], [19], [20]. However, absence at the time of drug distribution was also a commonly reported cause for non-compliance by persons who were not migrants [11], [12], [15], [21]–[31]. It seems that the time allocated for MDA and for mopping-up activities was sometimes not sufficient to reach even those people who were not away at the time of drug distribution [16], [18], [32]. Good coverage in one year sometimes had a positive effect on coverage in subsequent years. One study reported that health workers in the Philippines were especially motivated to reach their coverage target as a result of seeing a reduction of LF cases in the community [33]. In the same study, those individuals who had awareness about LF (p = 0.01) and awareness about MDA (p = 0.02) as well as awareness that MDA was associated with LF (p = 0.01) were more likely to have received the drug. Other factors that were positively associated with good coverage included good coordination among health staff [11] and the introduction of Filaria Prevention Assistants (FPA) or community drug distributors (CDD) to augment the distribution potential of the health services [13], [18], [34], [35]. A better understanding of factors that influence compliance with MDA for LF may improve results for current and future MDA campaigns. This paper has attempted to provide a thorough review of published information on this topic in the global context, and that information was supplemented by unpublished data and interviews with key informants. It is important to emphasize the importance of definitions and the distinction between coverage (delivery of the medicines) and compliance (ingestion of the pills). For example, the WHO collects reported coverage data from national LF elimination programs. However, GPELF publications sometimes use reported coverage data as if they were the same as “coverage with ingestion of drugs” [9]. Another illustration of the importance of these definitions comes from the computer simulation literature. The LYMFASIM simulation model uses “the fraction of people treated per round” (equivalent to “epidemiological drug coverage” as defined in Table 1) as a key input for estimating the impact of MDA on filarial infection rates and transmission [76]. Important differences between “epidemiological drug coverage” and “reported coverage” may explain the persistence of LF after several years of MDA in countries despite high reported coverage rates. One of the limitations of this review is that its primary focus is on the factors influencing compliance decisions at the individual level. Little published information is available on how factors such as program design, management, or implementation might affect coverage and compliance. Hence, further research is needed in this area. The reviewed studies were quite varied in terms of size, location, and design. Results also varied, and our challenge was to tease out the common themes. Our review did not identify a specific compliance recipe that will work in any country or context. However, the study has identified five key “ingredients” that appear to be essential for encouraging compliance at the individual level for successful MDA programs. The issue of trust permeates many of the factors and specific experiences that are associated with compliance or non-compliance. Specifically, individuals need to trust the person(s) delivering the drugs to them and communities need to trust the health workers and governments promoting and directing the MDA. As a result, elimination programs need to ensure that elements of trust are built into campaigns so as to effectively engage with communities. Some of these elements are outlined in Figure 1 and discussed below. It is impossible to overemphasize the importance of the face-to-face interaction between the drug distributor and the individual who receives the medicine. Recipients may be reassured, motivated, discouraged, angered, or confused during this critical interaction. We found that the identity of the drug distributor was consistently linked to compliance. The following factors regarding distributors may build trust in MDA: selection of distributors known to the community and with reputations or other credentials to indicate that they can be trusted, adequate training for communicating knowledge related to MDA, motivation and belief in the importance of the program, and adequate time for MDA and a willingness to answer questions. While it is recognized that the role of the drug distributor is key to the efforts to eliminate LF, more work is needed to develop protocols to motivate and sustain the commitment of drug distributors over the long course of MDA programs. Government health workers serve as drug distributors in some communities, and as stated above, many of the same issues apply to their success at the point of distribution. Since community members trust their advice, they tend to be successful promoters of MDA if they provide motivation, have the time to reassure participants about adverse events, and if they are enthusiastic about the campaign. Compliance suffers when they are overburdened, unmotivated, and undertrained. Private health providers could provide a similar service, however they are usually not systematically involved in the promotion and administration of the MDA. Engagement of private providers could reduce misinformation or confusing messages from these workers about MDA programs and improve community compliance. Published literature and WHO guidelines do not describe proven strategies to enhance motivation and participation by public or private health care workers in MDA programs. At the onset of MDA programs, consideration of several programmatic components is warranted in order to cultivate trust between the health services and the recipient communities—namely timely promotion of MDA, adequate management of adverse events, and recognition of other pressing health needs in the community. Each of these elements has been shown to enhance individual compliance with MDA. Across international contexts, advance knowledge of MDA was positively associated with ingestion of the pills. In order to promote MDA effectively and to encourage the establishment of the social norm of compliance, information about the pills and their distribution needs to be made frequently available before MDA begins to foster an environment of trust between the delivery system and the recipient. For communities to be effectively engaged in the elimination of LF, evidence shows that they need to hear about the MDA not only from the health services, but also from local leaders (cultural, religious, village), household authority figures, and from others in their personal networks. The involvement of these persons in MDA enhances confidence in the MDA program and further promotes its individual and collective benefits. In 2000 when the GPELF was launched, it stood on two pillars—MDA and morbidity management. In the majority of the papers reviewed here, there was little discussion or evidence that morbidity management had been implemented as a complement to MDA. From evidence in India, Haiti, Indonesia, and Zanzibar, it is clear that lymphedema management programs and hydrocele surgeries lend credibility to MDA programs. Aside from the benefit to MDA, there is clear evidence that these programs also alleviate the suffering caused by the long-term manifestations of LF and tangibly improve the lives of patients and their families [77]. Many articles mentioned that fear motivates many people not to take the tablets. If the components discussed above can be put into place before MDA begins, then this can help to alleviate fears and anxieties by establishing an environment of trust to increase chances for a successful program. MDA occurs in many different contexts around the world and also within countries and regions. These contexts reflect varied histories, local culture, and the impact of local leaders. Programs that strive to work together with leaders and that try to understand and respect their beliefs, trust networks, and fears will have better chances for success. In addition, programs should be planned so that MDA begins only after drugs, promotional materials, and personnel are available and in place. Delays in distribution can confuse communities and provide opportunities for rumors to arise and persist that may hamper compliance with MDA when it finally arrives. Social scientists have recommended comprehensive anthropological studies as prerequisites to MDA [78]. While the results of these studies may well enhance the planning of MDA, time, money, and expertise are not always available to carry them out. When such studies are not feasible, programs can look to other health agencies working in the same area to learn from their experiences (immunization drives, polio campaigns, maternal and child health programs) and adapt the MDA accordingly. While several studies have emphasized the importance of DO-MDA, it is more the exception than the rule within GPELF. More effort in this area is needed to ensure that individuals actually swallow the tablets that are distributed to reduce the gap between coverage and compliance. Planners should consider the habits of community members, their work, and social schedules including meal times. Drug distributors must be trained on procedures to follow when they encounter non-compliers. Options might include referral to a supervisor or scheduling a time for a second visit when there is more time for questions and discussion. This type of activity increases the time and effort for drug distribution, and this should be taken into account when planners consider time and budgets for drug distribution. Adverse events (AEs) following MDA are generally mild, and the frequency of AEs decreases after the first round of MDA [79]. However, our literature review showed that fear of AEs is one of the major reasons people do not comply with MDA. Furthermore, residual apprehension about the possibility of AEs may persist in an area even after the frequency of such events has declined [23]. As a result, programs should be vigilant when planning MDA and in the early years of MDA to ensure that health services are prepared to handle minor AEs as they arise. Management of AEs in these first years of treatment will benefit the program in subsequent years, as this will reassure communities that the drugs are safe and that the health services can be trusted. In the same vein, community drug distributors need be aware of how AEs will be managed so they can clearly communicate plans for this with communities and individuals as they distribute the pills. The positive side effects (so-called “ancillary benefits”) of MDA need to be communicated and reiterated during the pre-MDA socialization as well as at the time of pill distribution. The clearance of soil-transmitted helminths is a major associated benefit of MDA, particularly in children. Drug distributors should convey this information to community leaders and to families so that they can use this information to increase acceptance of MDA. Traditionally, awareness campaigns stress the functional benefits of a cure or treatment that heals sickness or prevents disease. Generally speaking, health officials and workers consider health to be an adequate motivation to achieve participation or compliance. However, the promise of health may not be a strong motivator in areas where clinically evident LF is uncommon. In many of the papers reviewed, the prospect of being cured or preventing LF was infrequently named as the primary reason people complied with treatment. While the health benefits of MDA may indeed influence some to take the treatment, the non-health benefits may be equally influential on behavior. Therefore campaigns should make every effort to promote intangible benefits associated with the treatment. These might include: feeling modern by accepting non-traditional medicines, being perceived as someone who cares about his family and health, being seen as someone who fits in with others in the community, being a good citizen and following the government program, being smart and preventing future economic loss, protecting future generations from LF, or feeling safer. Promotion of additional benefits of MDA may stimulate demand for the tablets. To illustrate this further, the association between LF and a cycle of poverty has been well documented [80]. Those living with chronic filariasis suffer from acute attacks that require them to forgo work for a period of time and seek medical assistance. Those with elephantiasis, lymphedema, or hydrocele may also suffer from associated stigma [5], [81]; the emotional burden of these conditions is only just starting to be understood [82]. With this knowledge in mind, promotion of MDA should address the economic and social benefits of treatment. Understanding the everyday economic benefits of LF elimination at the household level may persuade some individuals to comply. Furthermore, when discussing LF elimination with government health officials, the importance of these economic savings over the long term for endemic districts should be part of the promotion and advocacy activities. Children under five years of age were identified as systematic non-compliers in several studies from areas where MDA is provided to children starting as young as two years of age. Publicity regarding the deworming effects of MDA (with the associated improvements in general health, school attendance, and growth) may be useful for countering this problem. If parents understand that MDA has these benefits in addition to anti-LF properties, this may tip the balance in favor of participation. It may also be helpful to explain that the time-limited MDA program aims to provide a healthier and safer environment for all future children and generations. For example, a paper published in 2011 stated that MDA had already protected 66 million newborns from acquiring LF [83]. Systematic non-compliers have been identified as persons who persistently refuse or do not ingest antifilarial medications over the multi-year course of an MDA program. Non-compliers can serve as a continued source of infection that may place their community at continued risk for transmission of LF. This has been demonstrated in Haiti and in Egypt, where those who reported never taking MDA had higher infection rates than compliant persons [23], [47]. Therefore further work is needed to identify systematic non-compliers and their motivations. Explanations will vary in different locations, but special attention should be paid to the issues of seasonal migration and low compliance in young children. When groups with high rates of systematic non-compliance are identified, specific approaches are needed to counter factors that result in non-compliance. There is an urgent need to develop guidelines for managing this problem, as none currently exist. There is a sense of urgency to some of the issues raised in this paper. The clock is ticking for the elimination of LF by 2020. Through unprecedented pharmaceutical donations, the drugs for MDA are ready and available for use by national programs. The focus should now move from the issue of supply to the question of how to best deliver precious donated drugs into the mouths of those living in endemic areas. The success of the global program hinges on sufficient and sustained compliance, which is achieved one person at a time. The risks of insufficient compliance are too great to ignore. These include the possible emergence of drug resistance [85], the potential need for additional rounds of treatment [76] with their associated costs, and the risk of program fatigue at the community and health service levels. With seven years remaining, focused attention is needed to optimize MDA in countries that are just starting their programs and to improve ongoing campaigns in countries where compliance has been less than adequate. Finally, in light of the London Declaration of 2012 [86] and the increased global commitment to eliminate or control NTDs, it will be important to share and integrate lessons learned from various NTD programs regarding compliance and behavioral change to maximize the benefit of these interventions for at-risk populations.

10.1371/journal.pntd.0006891

The ectodomains of the lymphocyte scavenger receptors CD5 and CD6 interact with tegumental antigens from Echinococcus granulosus sensu lato and protect mice against secondary cystic echinococcosis

Scavenger Receptors (SRs) from the host’s innate immune system are known to bind multiple ligands to promote the removal of non-self or altered-self targets. CD5 and CD6 are two highly homologous class I SRs mainly expressed on all T cells and the B1a cell subset, and involved in the fine tuning of activation and differentiation signals delivered by the antigen-specific receptors (TCR and BCR, respectively), to which they physically associate. Additionally, CD5 and CD6 have been shown to interact with and sense the presence of conserved pathogen-associated structures from bacteria, fungi and/or viruses. We report herein the interaction of CD5 and CD6 lymphocyte surface receptors with Echinococcus granulosus sensu lato (s.l.). Binding studies show that both soluble and membrane-bound forms of CD5 and CD6 bind to intact viable protoscoleces from E. granulosus s.l. through recognition of metaperiodate-resistant tegumental components. Proteomic analyses allowed identification of thioredoxin peroxidase for CD5, and peptidyl-prolyl cis-trans isomerase (cyclophilin) and endophilin B1 (antigen P-29) for CD6, as their potential interactors. Further in vitro assays demonstrate that membrane-bound or soluble CD5 and CD6 forms differentially modulate the pro- and anti-inflammatory cytokine release induced following peritoneal cells exposure to E. granulosus s.l. tegumental components. Importantly, prophylactic infusion of soluble CD5 or CD6 significantly ameliorated the infection outcome in the mouse model of secondary cystic echinococcosis. Taken together, the results expand the pathogen binding properties of CD5 and CD6 and provide novel evidence for their therapeutic potential in human cystic echinococcosis.

Scavenger Receptors (SRs) are constituents of host’s innate immune system able to sense and remove altered-self and/or pathogen components. Data on their interaction with helminth parasites is scarce. In this work, we describe that CD5 and CD6 -two lymphoid SRs previously reported to interact with conserved structures from bacteria, fungi and viruses- recognize tegumental components in the cestode parasite Echinococcus granulosus sensu lato (s.l.). Moreover, both receptors differentially modulate the cytokine release by host cells exposed to E. granulosus s.l. tegumental components. Importantly, the infusion of soluble forms of CD5 or CD6 improve infection outcomes in a murine model of secondary cystic echinococcosis. In summary, our results expand the pathogen binding properties of CD5 and CD6 and suggest their therapeutic potential against helminth infections.

The mammalian innate immune system relies on a limited number of germline-encoded and non-clonally distributed receptors for pathogen recognition, which have evolved to identify the so called pathogen associated molecular patterns (PAMPs): conserved microbial structures, essential for their survival and not shared by the host, such as lipopolysaccharide (LPS) from Gram-negative bacteria, lipotheichoic acid (LTA) from Gram-positive bacteria, lipoarabinomannan from mycobacteria, mannan from fungi, chitin from parasites, and viral RNA [1]. Such kind of receptors are collectively named pattern recognition receptors (PRRs), and can be grouped into structurally diverse classes according to the protein domains involved in pathogen recognition (e.g., C-type lectin domains or leucine-rich repeats) [1,2]. This is well exemplified by the Scavenger Receptors (SRs), a large group of cell surface and soluble protein receptors that are structurally diverse and participate in a wide range of biological functions (endocytosis, phagocytosis, adhesion, and signaling) following binding to multiple non-self or altered-self ligands [3,4]. Some SR (namely SR-A and SR-I) are characterized by the presence of one or multiple repeats of an ancient and highly conserved cysteine-rich protein domain named SRCR (for scavenger receptor cysteine-rich) and constitute a superfamily (SRCR-SF) comprising more than 30 different cell-surface and/or secreted proteins present from lower invertebrates to mammals, as well as in algae and plants [5,6]. Despite the high degree of structural conservation among SRCR-SF members, a common single unifying function has not been reported. However, a steadily growing bunch of SRCR-SF members is known to interact with diverse microbial (bacterial, fungal, parasitic and/or viral) structures [6,7]. This is the case of the functionally and structurally highly homologous lymphocyte SR-I receptors CD5 and CD6. These two receptors are encoded by contiguous genes thought to derive from duplication of a common ancestral gene and are mainly expressed on all T cells, and a minor subset of B cells (B1a cells) [6]. The extracellular regions of both receptors are exclusively composed of 3 consecutive SRCR domains showing extensive sequence identity [8]. Their diverging cytoplasmic tails are devoid of intrinsic catalytic activity but both display several structural motifs compatible with a signaling transduction function [6]. Importantly, CD5 and CD6 are physically associated with the clonotypic antigen-specific receptor complex present on T and B1a cells (TCR and BCR, respectively) [9,10] and are involved in the fine tuning of the activation and differentiation signals generated by such relevant receptors through still incompletely understood and complex signaling pathways [11]. In addition to their immunomodulatory properties, CD5 and CD6 also exhibit PRRs activities. Available data indicate that soluble and membrane-bound forms of CD6, but not of CD5, bind to Gram-negative and Gram-positive bacteria through recognition of LPS and LTA, respectively [12,13]. In contrast, soluble and membrane-bound forms of CD5, but not of CD6, recognize and bind to saprophytic and pathogenic fungal species through β-glucans [14]. More recently, CD5 has been reported as a key receptor for human hepatitis C virus (HCV) entry into T lymphocytes [15], and preliminary observations indicate that CD6 may interact with human immunodeficiency virus 1 (HIV-1) [16]. It remains to be explored, however, whether the broad PRR activity exhibited by CD5 and CD6 also includes other groups of pathogen besides bacteria, fungi and viruses. Helminths -a diverse group of metazoan parasites able to produce long-lasting infections in immunocompetent hosts- currently affect one third of the world population [17]. Helminthiases are usually chronic infections due to the pathogens’ ability to adapt to the defense mechanisms triggered by infected hosts. Therefore, in most cases host immune responses are ineffective in parasite elimination, and are often associated with polarized and stereotyped Th2-type responses, with rare to no levels of Th1-type components [18]. In most helminthiases, such an early response bias does not associate with protective immunity [18–20], and therefore identification of innate receptors able to recognize and respond to parasite-derived components during early infection stages is highly relevant. Among helminthiases, cystic echinococcosis (CE) -formerly known as hydatidosis- is a zoonotic disease caused by the larval stage of the cestode Echinococcus granulosus sensu lato (s.l.), which shows a cosmopolitan distribution with high prevalence worldwide [21–23]. E. granulosus s.l. is composed of numerous variants initially called genotypes/strains (G1-G10), which nowadays are recognized as new species: E. granulosus sensu stricto (s.s.) (G1/G2/G3), E. equinus (G4), E. ortleppi (G5), E. canadensis (G6/G7/G8/G10) and E. felidis (‘lion strain’). Among them, the G1 genotype of E. granulosus s.s. is the most frequently found worldwide in livestock and humans [24]. Primary CE occurs in intermediate hosts (domestic and wild ungulates; accidentally humans) via ingestion of eggs containing oncospheres, which later develop into metacestodes -or hydatid cysts- mainly in the liver and lungs of the infected host. Secondary CE occurs after protoscolex (PSC) spillage from a fertile hydatid cyst within an infected intermediate host. This kind of CE derives from PSC developmental plasticity, which allows them to develop either into new cysts within intermediate hosts or into adult worms if ingested by a definitive host (usually dogs) [25]. Human secondary CE is an important medical problem associated with the surgical removal of primary cysts. In fact, although actual percentages of secondary CE cases post-surgery are debatable, recent studies have reported rates of 10–35% depending mainly on the type of surgery, the geographical location, and the follow-up time [26–28]. The murine model of secondary CE (inoculation of viable PSC into mice) has been widely used to study both the basic aspects of E. granulosus s.l. immunobiology [29–35], and the new chemotherapeutics or therapeutical protocols [36–38], novel vaccine candidates [39–41], and diagnostic or follow-up tools [42–44]. In this model, secondary CE can be divided into two stages: an early pre-encystment stage (until day 20–30 post-inoculation) with PSC developing into hydatid cysts [45], and a late or post-encystment stage in which differentiated cysts grow and eventually become fertile cysts [46]. Such a sequential developmental process is associated with a strong local control of inflammation during the initial phase of PSC differentiation into hydatid cysts [32,47]. The present report extends PRRs activities of both CD5 and CD6 receptors to helminth parasites, using E. granulosus s.l. as a case study. The data we provide indicate that ectodomains from both receptors recognize specific parasite components present in the tegument of PSC. Additionally, the prophylactic potential of CD5 and CD6 ectodomains infusion is shown using the murine model of secondary CE. Experimental animal procedures were performed in compliance with the Spanish Animal Experimentation Ethics Committee of Universitat de Barcelona School of Medicine, and the Uruguayan Comisión Honoraria de Experimentación Animal (Universidad de la República) according to the Canadian Guidelines on Animals Care and the National Uruguayan Legislation No.18.611. Protocols were approved by Comité de Ética en el Uso de Animales (Facultad de Química—Universidad de la República) and were given the Protocol Approval Number 101900-000361-16 (www.expe.edu.uy). Wild-type Balb/c and C57BL/6N mice (8–12 weeks old female) were obtained from DILAVE-MGAP (Uruguay) or Charles River (France), and housed under specific pathogen-free (for in vitro studies) or conventional (for experimental infections) conditions at the animal facilities of Instituto de Higiene (Universidad de la República, Uruguay) and of Universitat de Barcelona School of Medicine (Spain). CD5-deficient (CD5-/-) (provided by C. Raman, University of Alabama, Birmingham, AL) [48] and CD6-deficient (CD6-/-) mice [49] on C57BL/6 background were maintained at the animal facilities of the Universitat de Barcelona School of Medicine (Spain) under specific pathogen-free conditions. For tegumental antigens extraction, E. granulosus s.l. PSC were obtained by aseptic puncture of fertile bovine hydatid cysts from Uruguayan abattoirs, washed several times with phosphate buffered saline (PBS) pH 7.2 containing gentamicin (40 μg/mL), and their viability assessed [29]. Tegumental proteins were extracted from PSC (viability ≥80%) using an extracting solution consisting of PBS plus 1% (w/v) MEGA-10, 5 mM EDTA, and 2 mM PMSF [39]. Briefly, 125,000 viable PSC/mL of extracting solution were incubated for 2 h at RT with gentle shaking. Then, PSC were allowed to settle down and the supernatant was removed and extensively dialyzed against PBS through a cellulose membrane (MW cut-off: 12,000 Da). Protein content of the obtained antigens (termed PSEx) was assessed using BCA Protein Assay Reagent (Pierce). PSEx were stored at -20°C until used. Treated PSC were washed thrice with PBS, and their physical integrity was confirmed by observation under a light microscope. For experimental infections, E. granulosus s.s. PSC were obtained by aseptic puncture of fertile bovine or ovine hydatid cysts provided by Uruguayan abattoirs and Dr. Raúl Manzano-Román (IRNASA-CSIC, Salamanca, Spain), respectively. In both cases -Spanish as well as Uruguayan PSC- only parasite batches with ≥95% viability were used for experimental infections, and E. granulosus s.s. genotype was confirmed to belong to the G1 strain by sequencing a fragment of the gene coding for mitochondrial cytochrome c oxidase subunit 1 (CO1), as previously described [50]. Production of purified recombinant soluble proteins encompassing the whole ectodomains of human CD5 (rshCD5; from R25 to D345) and CD6 (rshCD6; from D25 to R397) receptors (in PBS with 10% glycerol, pH 7.4) was performed based on previously reported methods [51] but using SURE CHO-M Cell line clones from the Selexis SURE-technology Platform (Geneva, Switzerland) and subjecting serum-free supernatants to size exclusion chromatography protocols developed at PX’Therapeutics (Grenoble, France). Bovine serum albumin (BSA) was from Sigma-Aldrich. Proteins were biotinylated with EZ-Link PEO-maleimide-activated biotin (Pierce) following the manufacturer’s instructions. Binding of biotin-labelled recombinant proteins to PSC was assessed according to [51] with slight modifications. Briefly, 5,000 PSC (viability ≥90%) were incubated in 600 μL of biotinylated rshCD5, rshCD6 or BSA protein solutions (20 μg/mL) in binding buffer (veronal buffer saline plus 5 mM CaCl2). After 1 h of incubation at RT with gentle orbital rotation, PSC were pelleted and washed thrice with binding buffer, and 125 μL of each solution was stored for further analyses. A new aliquot of 5,000 PSC was added to the remaining solutions and the same procedure was performed. Sequential incubations were performed with four PSC aliquots. Then, PSC pellets and 25 μL of stored supernatants (from the original solution and the last incubation), were mixed with SDS-PAGE reducing sample buffer and heat-denatured during 10 min at 100°C. Biotin-labelled proteins were developed by Western blotting (see below) following sample resolution in 12% SDS-PAGE and electro-transfer to PVDF membranes (Bio-Rad). The binding ability of rshCD5 and rshCD6 proteins to PSC tegumental antigens was assessed by using 96-well microtiter plates (Nunc, Roskilde, Denmark) coated ON at 4°C with 100 μL/well of PSEx in PBS (10 μg/mL), and further blocked for 1 h at RT with 200 μL/well of PBS containing 1% (w/v) BSA. Increasing concentrations (0–40 μg/mL) of biotinylated rshCD5, rshCD6 or BSA (100 μL/well) were then added to the wells and incubated ON at 4°C. Bound protein was detected by the addition of 100 μL/well HRP-labelled streptavidin (1:5,000—Sigma) for 1 h at 37°C. Between every incubation step, unbound proteins were washed out thrice with PBS containing 0.05% (v/v) Tween-20. Enzymatic activity was developed at RT by adding 100 μL/well of 3,3’,5,5’-tetramethylbenzidine (TMB) substrate (Sigma). After stopping the reaction with H2SO4 0.5 M (50μL/well), absorbance values were read at 450 nm. To assess whether antigens recognized by rshCD5 and rshCD6 within PSEx were carbohydrates, a similar ELISA was performed including a step of PSEx oxidation with NaIO4 [33]. Briefly, PSEx-coated and BSA-blocked plates were incubated during 1 h with 20 mM NaIO4 in 50 mM acetate buffer pH 4.5 (200 μL/well) at RT. After three washings with acetate buffer, treated-wells were incubated for 30 min with 50 mM NaBH4 in PBS (250 μL/well), and 100 μL/well of biotin-labelled rshCD5 (20 μg/mL) or rshCD6 (10 μg/mL) was added to treated and untreated wells. The remaining ELISA protocol was performed as described above. Binding to NaIO4-resistant antigens was assessed as the percentage of absorbance values in treated wells respect to untreated wells. ELISA competition was performed to explore potential overlapping between rshCD5 and rshCD6 for binding to PSEx. Briefly, PSEx-coated and BSA-blocked plates were incubated ON at 4°C with 100 μL/well of either a mixture made of a fixed amount of biotin-labelled rshCD5 (20 μg/mL) or rshCD6 (10 μg/mL) and increasing concentrations of unlabeled rshCD6 (0–40 μg/mL) or rshCD5 (0–20 μg/mL), respectively. After washing out unbound proteins, the remaining ELISA protocol was performed as described above. Ligand overlapping was assessed as the percentage of absorbance values in competed wells respect to non-competed wells (0 μg/mL of unlabeled protein). Analysis of PSEx by 2D SDS-PAGE was performed following standard protocols. Briefly, 300 μg of PSEx antigens were first precipitated by ON incubation at -20°C in 300 μL ice-cold acetone containing 20% of trichloracetic acid (TCA) and 0.07% dithiothreitol (DTT) to remove insoluble proteins and lipids. After centrifugation for 15 min at 10,000g and 4°C, the supernatant was discarded and 300 μL of ice-cold acetone containing 20% dimethyl sulfoxide (DMSO) and 0.07% DTT were added and incubated for 1 h at -20°C. Then, samples were centrifuged for 15 min at 10,000g and 4°C, the supernatants discarded, and 300 μL ice-cold acetone containing 0.07% DTT were added. This step was repeated twice. Finally, the pellet was lyophilized, rehydrated in immobilized pH gradient (IPG) buffer (GE Healthcare) and frozen at -80°C for 24 h to improve solubilization. For the first dimension, 7 cm linear pH gradient (pH 3–10) Immobilie DryStrips (GE Healthcare) were re-hydrated with the sample and run on an IPGphore isoelectric focusing system (9.5 h run and a total of 35.5 KV/h), and stored at -80°C until use. Strips were then soaked for 15 min in equilibration buffer (50 mM Tris-Cl pH 8.8, 6 M urea, 30% glycerol, 2% SDS, and traces of bromphenol blue) containing 10 mg/mL DTT, further soaked for 15 min in equilibration buffer containing 25 mg/mL iodoacetamide, and sealed to 10% acrylamide gels using 0.5% agarose in standard Tris-glycine electrophoresis buffer. Second dimension SDS-PAGE was run at 50 V for the first 15 min and then raised to 150 V until ending. Finally, replicates of 2D SDS-PAGE gels were subjected either to mass-spec compatible silver nitrate staining or to electro-transference to PVDF membranes (Bio-Rad) for Western blotting. Electro-transferred PVDF membranes either from PSC binding assays or from PSEx 2D SDS-PAGE were blocked with 1% (w/v, in PBS) BSA for 2 h at RT. Membranes from PSEx 2D SDS-PAGE were additionally incubated ON at 4°C with solutions of biotin-labeled rshCD5 or rshCD6 (15 μg/mL). All membranes were then incubated for 1 h at 37°C in a PBS solution of 0.1% (w/v) BSA and 0.05% (v/v) Tween-20 containing HRP-streptavidin (1:5,000—Sigma). Finally, membranes were extensively washed with PBS plus 0.05% (v/v) Tween-20, and blots were developed by chemo-luminescence (SuperSignal West Pico Substrate, ThermoScientific) in a G-Box equipment (Syngene). Clean spots observed in PSEx 2D SDS-PAGE transferred PVDF membranes Western blotted with biotinylated rshCD5 and rshCD6 were manually back-mapped on gels for mass spectrometry identification at the Proteomic Facility of Pasteur Institut (Montevideo). Briefly, spots were excised, faded and tryptic digestions were performed using sequencing-grade modified trypsin (Promega). After gel extraction into 60% acetonitrile containing 0.1% TFA, the excess of acetonitrile was removed by speed vacuum. Peptide samples were then combined with an equal volume of matrix, spotted onto a MALDI sample plate, and allowed to dry at RT. Mass spectra were acquired on a 4800 MALDI TOF/TOF Mass Analyzer (Abi Sciex) operating in the positive ion reflector mode. Protein identifications were performed using an in-house Mascot v.2.3 search engine by searching a custom database that includes the full proteome of E. granulosus s.l. and E. multilocularis, composed of 20,787 sequences (10,310,548 residues) obtained from the Sanger Helminth Database. Additionally, every mass spectrum was also analyzed using NCBI database to discard possible host related proteins. The search criteria used were cystein carbamidomethylation and methionine oxidation as variable modification, and mass deviation <200 ppm with peptide fragment tolerance of 0.45 Da. Scores >56 were considered significant (P<0.05). Assessment of fluorescein isothiocyanate (FITC)-labelled PSEx binding to membrane-bound CD5 or CD6 was performed by flow cytometry analyses of parental 2G5 cells (a Jurkat cell derivative selected for deficient CD5 and CD6 expression [52]) and stable 2G5-transfectants expressing wild-type human CD5 (2G5-CD5.wt) or CD6 (2G5-CD6.wt) [10]. FITC labeling of PSEx was done as previously reported [53] using fluorescein isothyiocyanate (Sigma). Briefly, 1 mg of PSEx was dialyzed against 100 mM NaHCO3 buffer pH 9, and then 500 μg of FITC (Sigma) in DMSO were added. After 8 h of vigorous shaking in the dark, the mixture was extensively dialyzed against PBS and stored at 4°C until use. Binding assays were performed by incubating 2x105 2G5, 2G5-CD5.wt and 2G5-CD6.wt cells with increasing amounts of FITC-labelled PSEx in blocking buffer (PBS plus 10% human AB serum, 2% FCS and 0.02% NaN3) for 30 min at 4°C. Then, cells were washed thrice with washing buffer (PBS plus 2% FCS and 0.02% NaN3) and analyzed on a FACSCalibur flow cytometer using CellQuest software (Becton Dickinson). Additionally, competition assays were performed in a similar way, but incubating cells with a fixed amount (10 μg) of FITC-labelled PSEx in the presence of increasing amounts (5–20 μg) of unlabeled rshCD5, rshCD6 or BSA. The influence of the interaction between PSC tegumental antigens and CD5/CD6 on the PSEx-induced cytokine profile was first assessed by cell cultures of spleen and peritoneal cells from naïve CD5-/- [48], CD6-/- [49] and their corresponding C57BL/6 wild-type littermates in the presence of increasing concentrations of PSEx (0–40 μg/mL). Secondly, peritoneal cells from naïve C57BL/6 wild-type mice were stimulated with a fixed concentration of PSEx (20 μg/mL) in the presence of increasing amounts of rshCD5, rshCD6 or BSA (0–40 μg/mL). Spleen cells were obtained by mechanically disrupting spleens with a syringe plunger through a cell strainer. Harvesting of peritoneal cells was done by repeated washings (4 times with 2 mL/washing) of peritoneal cavities with cold PBS plus 2% FCS. Both procedures were performed under sterile conditions. Cell pellets -either from spleen or peritoneal cells- were treated with red blood cell lysing buffer (Sigma) following manufacturer’s instructions, and then suspended in complete culture medium (RPMI 1640 plus 10% FCS, 50 μM 2-mercaptoethanol, 100 μg/mL streptomycin and 100 U/mL penicillin, all from Sigma) and counted. Cells were seeded in 96-wells U-bottom plates at 2x105 cells/well in 200μL of complete culture medium and then incubated for 72 h at 37°C in a 5% CO2 atmosphere. Stimulation of cells with 10 μg/mL LPS (Sigma) was used as a positive control. Mouse cytokine levels in culture supernatants were determined by commercially available ELISA kits following the manufacturer’s instructions. The IL-17A ELISA kit was from R&D Systems. The IL-1β, IL-2, IL-4, IL-5, IL-6, IL-10, IL-12p40, TNF-α and IFN-γ BD OptEIA-Mouse ELISA Sets were from BD Biosciences Pharmigen. To assess whether rshCD5 or rshCD6 could modulate CE outcome, secondary infections were performed in Balb/c mice [33,35]. Mice were administered with rshCD5, rshCD6 or BSA in 200 μL sterile PBS (25 μg, i.p.) one hour before (-1h) and after (+1h) i.p. inoculation of 2,000 PSC (viability ≥95%) in 200 μL of sterile PBS. Mice were euthanized 14 months post-challenge and peritoneal cysts were recovered. Groups were compared in terms of (i) frequency of infection (proportion of mice harboring at least one cyst), (ii) number of developed cyst within each mouse, and (iii) total mass of cyst within each mouse (cyst wet-weight). Depending on the characteristics of the values, statistical analyses were assessed by either Student’s t-test (parametric values), Mann-Whitney U-test (non-parametric values) or Fisher’s exact test (non-parametric contingencies). Differences were regarded as significant when P <0.05. In order to determine whether the human CD5 and CD6 ectodomains are able to directly bind to the surface of viable E. granulosus s.l. PSC, pathogen-binding assays previously used for exploring their putative interaction with fungal and bacterial cell wall components, respectively, were performed [12,14]. Thus, biotinylated rshCD5 and rshCD6 proteins were sequentially incubated with viable PSC suspensions, and SDS-PAGE and Western blotting of pellets against streptavidin-HRP further assayed their adsorption to PSC. The results showed that both rshCD5 and rshCD6 (but not BSA, used as a negative control) bound to viable PSC (Fig 1A), indicating that they possess helminth-parasite binding activity. Next, it was investigated whether the observed binding of rshCD5 and rshCD6 to viable PSC involved tegumental components. To that end, increasing concentrations of biotin-labeled proteins were assayed on ELISA plates coated with the antigenic fraction termed PSEx, composed of PSC tegumental antigens. Results depicted in Fig 1B show that both, biotinylated rshCD5 and rshCD6 (but not BSA), interact with structures present in the PSEx fraction in a dose-dependent manner. Additionally, when assayed the supernatants resulting from sequential incubations of PSC with biotin-labeled rshCD5 and rshCD6 depicted in Fig 1A, reactivity against PSEx decreased as the number of incubations increased, in accordance with a sequential co-precipitation phenomenon (Fig 1C). Once evidenced the interaction of rshCD5 and rshCD6 with the PSEx fraction, the biochemical characterization of the PSEx components involved was addressed. To that end, an ELISA-based assay was first performed to determine whether rshCD5 and rshCD6 interactors were metaperiodate-sensitive (i.e. carbohydrates) or -resistant (i.e. proteins/lipids) compounds. Results depicted in Fig 2A indicate that all rshCD6- and most rshCD5-mediated interactions were metaperiodate-resistant, suggesting they are of protein and/or lipid nature. Then, in order to assess if the ligand patterns are similar or different for each molecule, we performed competition experiments in PSEx-coated ELISA plates with a fixed concentration of biotin-labeled rshCD5 incubated with increasing amounts of unlabeled rshCD6, and vice versa (i.e. fixed biotin-labeled rshCD6 and increasing amounts of unlabeled rshCD5). The results obtained indicate that rshCD5 and rshCD6 exhibit little overlapping regarding their PSEx interactions (Fig 2B). This was further supported by Western blotting the 2D SDS-PAGE resolved PSEx fraction with biotinylated rshCD5 or rshCD6 and HRP-labeled streptavidin. As illustrated by Fig 2C, rshCD5 and rshCD6 differed regarding their “spot” pattern reactive with the PSEx fraction. Accordingly, MALDI-TOF/TOF analyses identified parasite thioredoxin peroxidase as a potential interactor for rshCD5, and parasite peptidyl-prolyl cis-trans isomerase (cyclophilin) and endophilin B1 (antigen P-29) in the case of rshCD6 (Table 1). Summing up, these results indicate that both rshCD5 and rshCD6 molecules exhibit binding capacity to different structures -mainly proteins and/or lipids- present in the tegument of E. granulosus s.l. PSC, expanding their known spectrum of pathogen recognition. Next, it was questioned whether membrane-bound forms of human CD5 and/or CD6 lymphocyte receptors also retain their PSEx-binding activity. To that end, binding of FITC-labeled PSEx to parental 2G5 cells (a Jurkat T cell derivative selected for deficient CD5 and CD6 expression [52]) and to stable 2G5 transfectants expressing wild-type CD5 (2G5-CD5.wt) and CD6 (2G5-CD6.wt) surface receptors [10] was analyzed by flow cytometry. As shown in Fig 3A, mean fluorescence intensity (MFI) was significantly higher for 2G5-CD5.wt and 2G5-CD6.wt transfectants compared with parental untransfected 2G5 cells. The specificity of these interactions was confirmed by competition binding experiments, in which binding of a fixed amount of FITC-labeled PSEx to 2G5-CD5.wt and 2G5-CD6.wt cells was competed in a dose-dependent manner by unlabeled rshCD5 and rshCD6, respectively (Fig 3B). By contrast, unlabeled BSA (included as a negative control protein) did not inhibit binding of FITC-PSEx to 2G5-CD5.wt nor 2G5-CD6.wt (S1 Fig). Taken together, this evidence indicates that cell surface-expressed CD5 and CD6 retain PSEx-binding activity as well. The influence of cell surface CD5 or CD6 expression on PSEx-induced cytokine production by spleen and peritoneal cells (PECs) from either naïve CD5-/- or CD6-/- mice, as well as their respective wild-type controls was first analyzed. To that end, cells were cultured for 72 h in the presence of increasing amounts of PSEx, and then cytokine production in supernatants was analyzed by capture ELISA. Spleen cells showed no significant PSEx-induced cytokine production over background (S2 Fig), in agreement to previous reports for other PSC-derived antigens [29,30,54]. By contrast, PSEx simulation of PECs resulted only in significant IL-10, TNF-α and IL-6 cytokine responses. Therefore, our further analyses focused on those cytokines within supernatants of cultured PECs. Since levels of spontaneous cytokine secretion usually differed between PECs from knockout and wild type mice (S3 Fig), results were further displayed in terms of fold changes for an easier interpretation. As illustrated by Fig 4A, while no differences were observed regarding IL-10 induction, PSEx-stimulated PECs from CD5-/- mice underwent significant higher fold-increases for TNF-α and lower for IL-6 than their wild-type mice counterparts. Regarding CD6 surface expression, the results depicted in Fig 4B showed that PSEx-stimulation of PECs from CD6-/- mice underwent significant higher fold-increases for IL-6 but not TNF-α or IL-10 than their wild-type mice counterparts. On the other hand, LPS-stimulated PECs (included as a positive stimulation control) from CD5-/- and CD6-/- mice exhibited higher fold-increases for IL-10 and TNF-α and lower fold-increases only for IL-10, respectively, than their wild-type counterparts (Fig 4A and 4B), suggesting an overall antigen-independent difference in stimulation threshold. To exclude possible CD5/CD6-independent alterations in knockout mice, TNF-α, IL-6 and IL-10 cytokine levels in supernatants of PSEx-stimulated PECs from wild-type C57BL/6 mice in the presence of increasing amounts of rshCD5 or rshCD6 were also assessed. This strategy might reduce the interaction of PSEx with membrane-bound CD5 and CD6 through direct competition with the soluble forms of the receptor recombinant ectodomains. Regarding PSEx-induced IL-10 secretion, no variations due to rshCD5 addition was observed, while rshCD6 significantly reduced IL-10 levels (Fig 5A and 5B, top bar charts). On the other hand, rshCD5 as well as rshCD6 both modified TNF-α and IL-6 (Fig 5A and 5B, middle and bottom bar charts) production in response to PSEx, but in opposite ways. Thus, while rshCD5 addition increased PSEx-induced TNF-α and IL-6 production by wild-type PECs, rshCD6 decreased the secretion levels of both cytokines. BSA addition (included as a negative control) did not affect PSEx-stimulated cytokine responses (Fig 5A and 5B). Taken together this set of results indicates that either the absence of cell surface CD5 and CD6 receptors or the presence of both receptors in soluble form modulate the cytokine responses induced by tegumental antigens from E. granulosus s.l. PSC in different ways. Interestingly, the blockade of PSEx components by rshCD5 seems to up-regulate pro-inflammatory cytokine responses (i.e. increasing TNF-α and IL-6 secretion, without affecting IL-10 production levels), while blockade through rshCD6 seems to overall down-regulate the PSEx-induced cytokine response (i.e. decreasing the production of the three induced cytokines). In light of the observed modulation by rshCD5 and rshCD6 of PSEx-induced cytokine responses in PECs from wild-type mice, it was further assessed whether rshCD5 or rshCD6 administration would modify the infection outcome in a mouse model of secondary CE. To that end, rshCD5 and rshCD6 (25 μg/mouse) were i.p. infused 1 h before and after i.p. inoculation of viable PSC (2,000/mouse) into Balb/c mice, a highly susceptible mouse strain to secondary CE [33]. The i.p. route for rshCD5/rshCD6 administration was chosen because the peritoneal cavity is the natural anatomical site for infection establishment and hydatid cyst development in this infection model. At 14 months post-infection, mice were euthanized for infection inspection, and the peritoneal hydatid cysts within each mouse were counted and weighted. As illustrated by Fig 6A, rshCD5 infusion exhibited a remarkable prophylactic potential against secondary CE, since it significantly reduced the proportion of infected individuals, as well as the number of hydatid cysts per mouse (Fig 6B), and the total wet weight of hydatid cysts per mouse (Fig 6C). On the other hand, rshCD6 infusion also exhibited some degree of prophylactic potential in secondary CE, since a trend towards reduction in the proportion of infected mice (Fig 6A) and the number of hydatid cysts per mouse (Fig 6B) was observed. Infusion of equivalent doses of BSA did not affect any parasitological parameter of infection outcome (Fig 6A–6C). Finally, our results showed that rshCD5 -and to a lesser extent rshCD6 as well- exhibit prophylactic potential in the murine model of secondary CE. Effective mammalian immune responses rely on the early recognition of pathogen-derived components by innate immunity related receptors, otherwise named PRRs. Data on key early steps of helminth parasitic infections is scarce. Functional approaches suggest the involvement of different TLRs (namely TLR4, TLR3, and TLR2) [55–59], and SRs in the recognition of helminth components. Current mammalian SRs include 10 different classes (SR-A to SR-L; excluding SR-C only present in Drosophila melanogaster), being class E (SR-E) the most important group in helminth-derived antigens recognition, including Dectin-2 [60], Mannose Receptor/CD206 [61–63], CLEC4F/CLECSF13 [64], and DC-SIGN/CD209a [65]. Our work expands the group of SRs interacting with helminth pathogens to CD5 and CD6, two lymphoid members of the class I (SR-I). SR-I ectodomains harbor several tandem repeats of the SRCR protein module. In addition to CD5 and CD6, SR-I members include CD163A/M130, CD163B/M160, SCART1, SCART2 and WC1 [3,4]. Macrophages and lymphocyte subsets expressing CD163 or WC1, respectively, play a role against certain parasite infections (e.g. Theileria parva [66], Leishmania braziliensis [67], Trypanosoma vivax [68], or Neospora caninum [69]), but no direct interaction with parasite helminth structures has been previously reported. A single C-terminal SRCR domain characterizes class A SRs (SR-A), including the SR-AI, MARCO, and SCARA5 receptors [3]. While some evidence supports the involvement of SR-AI and MARCO in parasite infection (Schistosoma japonicum and Leishmania major, respectively) [70,71], the ability of SR-AI to directly recognize helminth components has only been shown for Heligmosomoides polygyrus calreticulin [72]. The present study shows that CD5 and CD6 should be added to the list of SRs able to sense helminth components. Soluble CD5 and CD6 physically bind to different components within the tegument of E. granulosus s.l. PSC, modulate their induced cytokine profiles in naïve peritoneal cells, and protect mice from secondary CE. Tegumental components are crucial for helminth physiology (i.e. nutrients up-take and waste disposal) and for their ability to modulate immune responses leading to chronic parasite establishments. They are the first parasite structures recognized by soluble and/or membrane-bound host receptors. Therefore, it is highly relevant to identify which receptors hold the ability to recognize them and, if possible, the involved parasite structures. We determine that human CD5 and CD6 ectodomains bind to the surface of viable PSC (Fig 1A). More specifically, CD5 and CD6 interact with components of an antigenic fraction termed PSEx, which is mainly composed of tegumental antigens from PSC (Fig 1B). Such PSEx components are metaperiodate-resistant compounds, indicating their protein and/or lipid nature (Fig 2A). Additionally, bound components differed depending on the receptor analyzed (Fig 2B). 2D SDS-PAGE and MALDI-TOF/TOF analyses of the PSEx fraction provides a differentiated spot pattern (Fig 2B and 2C), and a distinctive set of potential parasite ligands for each receptor: thioredoxin peroxidase for CD5, and peptidyl-prolyl cis-trans isomerase (cyclophilin) as well as endophilin B1 (P-29 antigen) for CD6 (Table 1). These results show that CD5 and CD6 ligands within PSEx do not fully overlap. The CD5 and CD6 ligands besides being present in the PSEx fraction, have been found in other sources of E. granulosus s.l. antigens. Thioredoxin peroxidase has been detected in different PSC antigen sources [73–75], in hydatid fluid [76], in adult worms [74], and in extracellular vesicles obtained from fertile cysts [77]. Endophilin B1 (P-29 antigen) has been detected in somatic antigens of PSC and adult worms [74], in nuclear and cytosolic extracts of PSC [75], and in hydatid fluid [76]. Finally, peptidyl-prolyl cis-trans isomerase (cyclophilin), has been detected in PSC excretion/secretion products [73], in nuclear and cytosolic extracts of PSC [75], and in hydatid fluid [78,79]. Implying that CD5/CD6 parasite sensing would span the different stages of E. granulosus s.l. life cycle and not be limited to tegumental structures from PSC. PSEx interaction with CD5 and CD6 is also extendable to their membrane-bound forms, as demonstrated by binding (and competition binding) experiments of FITC-labelled PSEx to parental and stably CD5- and CD6-transfected 2G5 cells -a Jurkat cell derivative deficient for surface CD5 and CD6 expression (Fig 3A and 3B). The basal interaction observed between PSEx and parental 2G5 cells suggests other cell surface receptors in parasite interaction (Fig 3A). Jurkat cells (a human CD4+ T cell lymphoma line) have been shown to express most TLRs [80], and 2G5 cell line in particular has been previously checked for surface expression of TLR2 and TLR4 [12,14]. Therefore, membrane-expressed CD5 and CD6 receptors specifically interact with PSEx components, even if other surface molecules may concomitantly act as PSEx-binding receptors. Specific recognition of parasite structures by membrane-bound CD5 and CD6 may have relevant functional consequences, since both receptors induce intracellular signalling (namely MAPK cascade activation) when ligated by PAMPs [12,14]. PECs harbor CD5 and/or CD6 expressing immune cells involved in helminth infection protection. They include all T cells (comprising Tγδ and iNKT) and B1a cells as well as a macrophage subset [81]. Interestingly, CD5+ B1a cells are an important source of polyreactive natural IgM antibodies [82] and of IL-10 [83]. Thus, CD5- and CD6-mediated signalling by immune PECs may contribute to relevant biological effects, like modulation of cytokine production and release. Accordingly, PSEx-induced stimulation of PECs from CD5- and CD6-deficient mice resulted in different cytokine responses compared with their wild type controls (Fig 4). Interestingly, such differences were also observed following LPS-stimulation, suggesting an overall difference in stimulation thresholds (Fig 4). Indeed, data from knockout mice have shown the involvement of membrane-bound CD5 and CD6 in the fine-tuning of T (and likely B1a) cell subset responses [11,84]. Thus, the possibility that differences observed in PSEx-induced cytokine production by CD5- and CD6-deficient mice could be due to effects beyond direct recognition of parasite ligands by membrane-bound CD5 or CD6, was further excluded by similar PSEx stimulation studies in wild type mice in the presence of increasing amounts of soluble CD5 or CD6. The results showed that rshCD5 up-regulated PSEx-induced pro-inflammatory cytokine responses (i.e. increased TNF-α and IL-6 secretion, without altering IL-10 production rates) (Fig 5A), while rshCD6 down regulated the overall (TNF-α, IL-6 and IL-10) cytokine response (Fig 5B). Such findings are highly relevant either in a basic immunological sense, as well as from a potential prophylactic point of view. Cytokine profiles are relevant for host susceptibility/resistance to E. granulosus s.l. infection. Thus, pro-inflammatory responses have been associated with host protection either in experimental infection models [30,34,85,86] as well as in human patients [87–91], being nitric oxide-mediated mechanisms involved in such protection [89–92]. Our results indicate that CD5 or CD6 binding to tegumental antigens from PSC contribute to cytokine induction associated with parasite establishments, and is not optimal for parasite killing and clearance. From a prophylactic point of view, in vivo infusion of rshCD5 or rshCD6 during the early stages of parasite establishment may be a useful strategy for immunomodulating the host into an anti-parasite state. Accordingly, the assessment of in vivo administration of rshCD5 or rshCD6 in a mouse model of secondary CE resulted in a remarkable prophylactic potential of the former, since it reduced not only the proportion of infected mice, but also the number of developed hydatid cysts per mouse and their parasite loads (Fig 6). A trend towards reduction in the proportion of infected individuals and the number of developed hydatid cysts was observed in rshCD6 infusion (Fig 6). The prophylactic potential of rshCD5 may correspond to pro-inflammatory cytokine (TNF-α and IL-6) upregulation (Fig 5A), while CD6 lesser efficiency parallels down-modulating cytokine production, especially IL-10 (Fig 5B), a cytokine usually associated with increased susceptibility to E. granulosus s.l. infection [31,34,90]. In this sense, our preclinical results might be useful for designing/proposing a novel strategy to reduce secondary infection rates in CE patients. For example, once a hydatid cyst is surgically removed, a concomitant intraperitoneal infusion of rshCD5 -or rshCD6- would help in preventing remaining PSC to develop into new hydatid cysts. Interestingly, besides modulating cytokine production, the potential ligands identified for both CD5 and CD6 receptors are highly relevant for E. granulosus s.l. physiology. Thioredoxin peroxidase is a key enzyme for reactive oxygen species detoxification in E. granulosus s.l. [93,94]; while peptidyl-prolyl cis-trans isomerase (cyclophilin) has been associated with parasite sensitivity to lethal effects of cyclosporine A [95], and endophilin B1 (P-29 antigen) revealed significant protective activity against secondary CE in mice [96], as well as against primary infection in sheep [97], when used as a vaccine antigen. Therefore, the binding of such parasite components by rshCD5 or rshCD6 may also contribute to a better host parasite control. In conclusion, we provide the first evidence for direct recognition of tegumental PSC structures from E. granulosus s.l. by SR-I CD5 and CD6, in addition to bacterial, fungal, and/or viral components binding. Moreover, we prove the prophylactic potential of soluble CD5 (and CD6) infusion in the mouse model of secondary CE. In this sense, such a prophylactic potential could be ascribed either to: (i) a cytokine-modulating activity through the competition with the interaction between tegumental antigens and host membrane-bound forms of CD5/CD6; (ii) the blockade of key tegumental components highly relevant for PSC physiology; or (iii) a mixture of both situations finally contributing to a better parasite control. Additionally, although CE has a cosmopolitan distribution and represents a major public health problem in regions of South America, Mediterranean, Central and Western Asia, and East Africa [98], our PSC experimental infections were performed with E. granulosus s.s. G1 genotype, which shows the highest cosmopolitan distribution and is responsible for most human CE cases worldwide [24,99]. Further work is required to ascertain CD5 and CD6 prophylactic potential in other helminth-driven pathologies and to explore if additional SRCR-SF members share such interactions.

10.1371/journal.pntd.0000154

Household Transmission of Leptospira Infection in Urban Slum Communities

Leptospirosis, a spirochaetal zoonotic disease, is the cause of epidemics associated with high mortality in urban slum communities. Infection with pathogenic Leptospira occurs during environmental exposures and is traditionally associated with occupational risk activities. However, slum inhabitants reside in close proximity to environmental sources of contamination, suggesting that transmission during urban epidemics occurs in the household environment. A survey was performed to determine whether Leptospira infection clustered within households located in slum communities in the city of Salvador, Brazil. Hospital-based surveillance identified 89 confirmed cases of leptospirosis during an outbreak. Serum samples were obtained from members of 22 households with index cases of leptospirosis and 52 control households located in the same slum communities. The presence of anti-Leptospira agglutinating antibodies was used as a marker for previous infection. In households with index cases, 22 (30%) of 74 members had anti-Leptospira antibodies, whereas 16 (8%) of 195 members from control households had anti-Leptospira antibodies. Highest titres were directed against L. interrogans serovars of the Icterohaemorrhagiae serogroup in 95% and 100% of the subjects with agglutinating antibodies from case and control households, respectively. Residence in a household with an index case of leptospirosis was associated with increased risk (OR 5.29, 95% CI 2.13–13.12) of having had a Leptospira infection. Increased infection risk was found for all age groups who resided in a household with an index case, including children <15 years of age (P = 0.008). This study identified significant household clustering of Leptospira infection in slum communities where recurrent epidemics of leptospirosis occur. The findings support the hypothesis that the household environment is an important transmission determinant in the urban slum setting. Prevention therefore needs to target sources of contamination and risk activities which occur in the places where slum inhabitants reside.

Leptospirosis has emerged to become an urban slum health problem. Epidemics of severe leptospirosis, characterized by jaundice, acute renal failure and haemorrhage, are now reported in cities throughout the developing world due to rapid expansion of slum settlements, which in turn has produced the ecological conditions for rodent-borne transmission of the spirochete pathogen. A survey was performed in the city of Salvador, Brazil, to determine whether the risk of Leptospira infection clustered in households within slum communities in which a member had developed severe leptospirosis. We found that members of households with an index case of leptospirosis had more than five times the risk of having serologic evidence for a prior infection than members of neighbourhood households in the same communities. Increased risk of infection was found among all age groups who resided in these households. The finding that Leptospira infection clusters in specific slum households indicates that the factors associated with this environment are important determinants for transmission. Further research is needed to identify the sources of contamination and risk exposures which occur in the places where slum inhabitants reside such that effective community-based prevention of urban leptospirosis can be implemented.

Leptospirosis is an important zoonotic health problem because of its life-threatening clinical manifestations, Weil's disease and severe pulmonary haemorrhage syndrome, for which fatality is 10 to 50% [1]. Moreover there has been growing awareness of the large under-recognized disease burden that leptospirosis imparts in developing countries [2]. Leptospirosis is an environmentally-transmitted disease. Pathogenic spirochetes of the genus Leptospira establish chronic carriage in the kidney tubules of wild and domestic mammalian reservoirs and persist for weeks in the environment after excretion from the host [3]. The major mode of transmission to humans is indirect contact with water or moist soil contaminated with the urine of animal reservoirs [3]. Leptospirosis is associated with a spectrum of environmental settings and risk exposures. Recreation, travel and water sports have become significant risk factors in industrialized countries [3],[4],[5], as exemplified by outbreaks during triathlon and adventure tourisms events [6],[7]. In developing countries situated in tropical climates, leptospirosis is an endemic disease of rural-based populations engaged in subsistence farming, sharecropping and animal husbandry [3],[8]. Furthermore leptospirosis has emerged to become an urban slum health problem in developing countries [9],[10]. The rapid expansion of slum settlements, in which 1 billion of the world's population reside [11], has produced the ecological conditions for rodent-borne transmission [9],[12]. Epidemics of severe leptospirosis are now reported in cities throughout the developing world [1]. In Brazil alone, more than 10,000 cases of severe leptospirosis are reported each year [1] during outbreaks that occur in major urban cities [9],[13],[14],[15],[16]. During these outbreaks, leptospirosis cases cluster in slum settlements which lack adequate sewage systems and refuse collection services [9],[13],[17]. Public health responses to urban leptospirosis require an improved understanding of the specific exposures in slum communities which lead to epidemic transmission. Urban outbreaks are associated with heavy seasonal rainfall and flooding [9],[13],[16],[18]. Environmental surface waters in slum communities, as found in a study in Peru [12], contain high concentrations of pathogenic Leptospira serovars which are associated with acquiring severe disease forms. Leptospirosis is traditionally considered an occupational disease, since work-related activities are frequently identified as risk exposures [3]. However slum inhabitants reside in close proximity to environmental sources of contamination, such as open sewers, flood areas and trash collections [19]. Determining whether transmission occurs in the household environment will be essential to designing and implementing effective community-based interventions for which at present, none are available. In Salvador, a city of 2.4 million inhabitants in Northeast Brazil, outbreaks of leptospirosis occur annually during the seasonal period of heavy rainfall [9],[19]. A case-control investigation found that residence in proximity of open sewers and peridomicilary sighting of rats to be risk factors for acquiring severe leptospirosis [19], suggesting a role for household-related environmental exposures in transmission. In this study, we surveyed members of households in which an index case of severe leptospirosis resided and control households that were situated in the same slum communities. A serologic evaluation was performed of samples obtained from this survey to determine whether the risk for Leptospira infection clustered within households. Active surveillance consecutively identified patients that fulfilled a clinical case definition for severe leptospirosis [9],[19] and were admitted to the infectious disease hospital during an outbreak that occurred in Salvador, Brazil in 2001. According to health secretary protocols, suspected cases of leptospirosis from the metropolitan region are referred to this site. Acute and convalescent-phase serum samples were collected and evaluated in the microscopic agglutination test (MAT) according to protocols previously described [9]. A laboratory-confirmed case of leptospirosis was defined as the presence of a four-fold rise in the MAT titre between paired acute and convalescent-phase serum samples or a titre of >1∶800 in a single sample [9]. A survey was performed between May and October 2001 of members of case households, in which index cases of confirmed leptospirosis resided at the time of illness, and neighbourhood control households selected from the same slum communities. The study team visited the residences of the first 22 cases that were identified to have a confirmed diagnosis of leptospirosis. Cases resided in 19 densely-populated slum neighbourhoods, most of which were situated in the periphery of Salvador. Although the majority (>90%) of the households have access to potable water in these communities, more than 30% are served by open sewage systems. Study communities were built on poor land quality which are at risk for flooding during the seasonal period of heavy rainfall. Due to the lack of refuse collection services, refuse accumulates in open deposits and open sewers. Dogs, cats and chickens are domestic animals which are encountered in households and the peridomicilary environment. Due to the poor overall sanitation infrastructure, the environment surrounding households in these communities is infested with domestic rats, in particular Rattus norvegicus. Neighbourhood control households were selected according to the sampling scheme used in a previous case-control investigation [19]. Households were surveyed which were located a distance of five domiciles from the case household, and at every household thereafter, until a neighbourhood control household was identified which did not have a member who was diagnosed at a health care facility as having leptospirosis in 2001 and agreed to participate in the study. This strategy was chosen to avoid sampling control subjects who resided in close proximity to households and may therefore have had shared risk exposures. Subjects were interviewed to determine whether they had a history of a recent diagnosis of leptospirosis. In addition, we screened the database of cases identified during active hospital-based surveillance to exclude the possibility that subjects were recent leptospirosis patients. A total of four control households were selected for each of the first four identified case households by sampling domiciles in perpendicular directions. Two control households were selected for each of the remaining 18 case households by sampling domiciles in opposite directions. Eligible subjects were defined as those who were greater than five years of age and resided in case and control households for more than one month prior to identification of the index case of leptospirosis. The study team identified and recruited subjects from case households and their respective control households during the same community site visit. All subjects were enrolled into the study according to written informed consent protocols approved by the Ethical Committee in Research of the Oswaldo Cruz Foundation and IRB Committee of Weill Medical College of Cornell University. The study team performed interviews to obtain demographic information and obtained blood samples from which sera was separated and stored for serologic analyses. The MAT was performed to determine the presence of anti-Leptospira antibodies in subject samples [20] and used a panel that included seven pathogenic reference strains, L. interrogans serovars Autumnalis, Canicola, Copenhageni, Icterohaemorrhagiae and Saxkoebing, L. kirschneri serovar Grippotyphosa, and L. borgpetersenii serovar Ballum, one non-pathogenic reference strain, L. biflexi serovar Patoc (WHO/FAO Collaborating Centre for Reference and Research on Leptospirosis, Royal Tropical Institute, Holland), and two clinical isolates of L. interrogans serovars Canicola and Copenhageni [9]. The use of the reduced panel of ten strains in MAT evaluations had the same predictive value in identifying positive agglutination reactions among confirmed leptospirosis cases in Salvador [9] as did the WHO recommended battery of 18 reference serovars [20]. Operators were blinded to whether samples were from case and control household members. Each serum sample was tested at three dilutions, 1∶25, 1∶50 and 1∶100, and when agglutination was observed at the 1∶100 dilution, was titrated to determine the highest agglutination titre. Epidemiological and laboratory information for subjects were entered into an EpiInfo version 6.04 software system (Centres for Diseases Control and Prevention) database. Chi-square and Wilcoxon rank sum tests were used to compare categorical and continuous data, respectively, for eligible subjects who were and were not enrolled in the study and members of case and control households and to evaluate differences between highest agglutination titres for members of case and control households. A P value of less than 0.05 in two sided testing was used as criteria for a statistically significant difference. Analyses evaluated a range of reciprocal MAT titres against pathogenic serovars as threshold criteria for a prior Leptospira infection. The presumptive infecting serovar was defined as the pathogenic serovar against which the highest agglutination titre was directed. The chi-square or Fisher's exact tests were used to evaluate for significant associations between residence in a household with an index case of leptospirosis and the risk for acquiring anti-Leptospira antibodies. The CSAMPLE program of the EpiInfo version 6.04 was used to obtain estimates for the OR and 95% CI which were adjusted for the sampling design effect and weighted for the number of eligible household members. Serologic results from index cases of leptospirosis were excluded from the analyses. Active surveillance detected an outbreak of severe leptospirosis in Salvador during the seasonal period of heavy rainfall in 2001. Between March and October, 124 suspected cases were identified which resided within the city whereas 16 cases were identified during the preceding four month period. Leptospirosis cases were residents of 70 slum neighbourhoods (bairros) situated within the city. Cases were mostly adults (mean age±standard deviation, 35.2±13.5 years) and males (86% of 124 cases) and hospitalized with manifestations of Weil's disease such as jaundice (77% of 124 cases) and acute renal failure (73% with serum creatinine >2.0 mg/dL). Overall case fatality was 10% (12 of 124 deaths). Among the 124 suspected cases, 89 (72%) had a laboratory-confirmed diagnosis of leptospirosis. A survey was performed of 22 case households in which an index case of confirmed leptospirosis resided at the time of the outbreak and 52 control households situated in the same slum communities as case households. Households were enrolled from 19 (27%) of the 70 slum neighbourhoods in which leptospirosis cases were identified during the outbreak. Among 79 and 229 eligible subjects who resided in respectively, case and control households, 74 (94% of 79) and 195 (85% of 229) subjects were enrolled into the study. Subjects were healthy at the time of identification and did not report symptoms or signs of leptospirosis in the previous one-year period. One member of an index case household was previously hospitalized for leptospirosis. Neighbourhood control households were located between 15 and 45 meters from their respective index case household. Case and control households did not differ with respect to number of household members (median 3 vs. 3) and median monthly income (US$64 vs. 84) of the head-of-the-household. Subjects from cases and control households were similar with respect to median age (22 vs. 26 years, respectively) and gender (males, 40 vs. 42%). Table 1 shows the distribution of highest agglutination titres against pathogenic Leptospira serovars among subjects. Members of case households had significantly higher agglutination titres than those from control households located in the same slum communities (Wilcoxon rank sum test, P<0.001). Among members of case households, 18%, 23% and 30% had MAT titres ≥1∶100, 1∶50, and 1∶25, respectively, against a pathogenic serovar (Table 2). Among members of control households, 5%, 7% and 8% had titres ≥1∶100, 1∶50, and 1∶25, respectively. Across the range of highest MAT titres, the majority of agglutination reactions recognized L. interrogans serovars of serogroup Icterohaemorrhagiae, serovars Copenhageni and Icterohaemorrhagiae (Table 1). Among members of case and control households with MAT titres ≥1∶25, 95% (21 of 22 subjects) and 100% (16 of 16 subjects), respectively, were directed against serovars Copenhageni and Icterohaemorrhagiae. Members of case households were 4.42 times (95% CI 1.53–12.76) more likely than members of control households to have agglutinating antibodies against a pathogenic Leptospira serovar with an titre of ≥1∶100 (Table 2), a criteria commonly used to define probable infection in clinical cases of leptospirosis [21]. However, significant risk associations were found among members of case households when lower titres were used as threshold values for prior infection (OR 5.29 and 3.71 for ≥1∶25 and 1∶50, respectively, P<0.05). This risk association was specifically found when agglutinating antibodies against a pathogenic serovar were used as the outcome marker for prior Leptospira infection. The prevalence of agglutinating antibodies against a non-pathogenic serovar, L. biflexi serovar Patoc, did not differ between members of case and control households, irrespective of the titre used to define a positive result (Table 1). Within households with index cases of leptospirosis, children and young adults had significantly increased risk of having anti-Leptospira antibodies (Table 3). Significantly higher proportions of children in index case households had anti-Leptospira antibodies, as defined by a MAT titre ≥1∶25, than children in control households (19 vs. 0%, P = 0.008). Older adults with ≥55 years of age and adults with 15–34 years of age had as well a significantly increased risk for acquiring anti-Leptospira antibodies (OR 5.60 and 4.87, respectively, P<0.05) while adults with 35–54 years of age had a increased, albeit non-significant risk (OR 3.18, P = 0.32). This study, performed in an endemic region for urban leptospirosis, identified household clustering of Leptospira infection within slum communities. Among members who resided in the same household as an index case of leptospirosis, 30% had evidence of a previous infection as determined by the presence of anti-Leptospira antibodies in the MAT. These individuals had more than five times the risk for acquiring anti-Leptospira antibodies as compared with members of neighbouring households. Household clustering of Leptospira infection has not been found in surveys which evaluated for this phenomenon in rural endemic settings [22]. A previous serologic survey, which was also performed in Salvador and used IgM ELISA to detect anti-Leptospira antibodies, found that high (41%) proportions of children had antibodies among those residing in households with index cases of leptospirosis [23]. However, a control group was not evaluated in order to assess whether individuals with anti-Leptospira antibodies specifically aggregated in index case households. This study, as far as we are aware, is the first to describe household clustering of Leptospira infection. We used the presence of anti-Leptospira agglutinating antibodies as a marker of previous infection. The MAT is the standard assay to measure seroprevalence [3],[20]. Although a MAT titre ≥1∶100 is often employed [3], the performance of different MAT titre criteria have not been systematically evaluated in community-based investigations and a consensus does not exist with respect to a standard cutoff titre. In this study, we found that subjects from case households had significantly increased risk (OR 4.42, 95% CI 1.53–12.76) of having an agglutination titre ≥1∶100 against a pathogenic Leptospira serovar than subjects from control households. We evaluated lower cut-off criteria of titres ≥1∶50 and ≥1∶25 and found that the use of these criteria predicted the same epidemiological relationship as did the criterion of a titre ≥1∶100 (Table 2). The observed household clustering of individuals with anti-Leptospira antibodies does not therefore appear to be an artefact of serologic criteria used. These findings also indicate that the use of lower cutoff titres (i.e., 1∶25) is a more sensitive yet still specific serologic criterion for a previous Leptospira infection in our setting. Isolation studies showed that urban leptospirosis in Salvador is due to the transmission of a single serovar, L. interrogans serovar Copenhageni [9],[24]. In this study, more than 95% of subjects with positive MAT titres had agglutinating antibodies which specifically reacted against serovars Copenhageni and Icterohaemorrhagiae. This phenomenon was observed across the range of titre values, including titres of 1∶25 (Table 1). Individuals infected with serovar Copenhageni are expected to have serological cross-reactions against serovar Icterohaemorrhagiae since the both serovars belong to same serogroup [8]. Investigations performed in endemic areas in which several serogroups are circulating found that the predictive value of serology is low with respect to identifying the infecting serogroup [25],[26]. Our findings suggest that in a site where a single agent, serovar Copenhageni, is circulating, previously-infected individuals have agglutination titres which, even when low titres are encountered, are specifically directed against this serovar and serovars from the same serogroup. Non-pathogenic serovars, such as L. biflexi serovar Patoc, have been used in MAT strain panels to detect potential cross-reactive antibodies in individuals previously infected with a pathogenic serovar. We found that the titres against non-pathogenic serovars among household members were not significantly associated with residence in a household with an index case of leptospirosis (Table 2), indicating that antibodies that agglutinate non-pathogenic serovars is not be a specific marker for a previous Leptospira infection in our setting and may have been induced by unrelated exposures. Together these findings emphasize the need to validate MAT criteria for the specific epidemiological situation in which seroprevalence surveys are being performed. The study selected households based on identification of an index case of severe leptospirosis. Therefore risk estimates observed in this study may not pertain to households in which a member acquired mild leptospirosis. Furthermore previous infection among subjects occurred over an extended time period since agglutinating antibodies may persist for more than five years [27],[28]. It is therefore likely that severe leptospirosis occurred in households in which a significant proportion of the members had already been exposed to Leptospira. Infection among individuals with anti-Leptospira antibodies was likely to have been mild or asymptomatic since almost all subjects did not report a history of having leptospirosis. These findings are consistent with those observed in other studies which found asymptomatic infection to be common in endemic areas [29],[30]. Migration may have influenced the estimates of the risk associations. In a cohort investigation of residents from one slum community in Salvador, we found that the annual out-migration rate is approximately 12% (unpublished data). Reliable estimation of the risk attributable to the household will therefore require prospective evaluation of infection and severe disease outcomes in slum community settings. However, the strength of the association observed for household clustering of Leptospira infection risk indicates that household factors play an important role in transmission of leptospirosis in the urban slum setting. Members who reside in the same household may have had similar risk exposures outside the household environment. This explanation may be less likely since children who resided in households with index cases had significantly increased risk for acquiring anti-Leptospira antibodies in comparison to neighbourhood control subjects of the same age group (Table 2). Host susceptibility determinants shared among household members are a possible explanation for the observed clustering of individuals with anti-Leptospira antibodies. HLA gene polymorphisms have been reported to be associated with the risk of acquiring leptospirosis during an outbreak associated with a triathlon event [31]. Alternatively, the observed clustering of infection risk may reflect shared exposures among household members to transmission sources located in the environment where they reside. Ecological studies [9],[13],[17] and case-control investigations [19] of urban leptospirosis found environmental attributes of slum households to be risk factors for acquiring severe leptospirosis. In these studies, two of which were performed in the same slum communities in Salvador which were investigated in this study [9],[13],[17],[19], risk factors were related to open sewers, poor rainwater drainage systems and associated flooding, lack of refuse collection services, and sighting of rats in the household environment. These environmental sources of contamination are not uniformly distributed in slum settlements and may have contributed to the observed household clustering of infection risk. In addition, our findings suggest that infection risk varies over short distances within these communities. Members of households with index cases of leptospirosis had significantly increased risk of having a previous Leptospira infection than their neighbours who resided a distance of 15 to 45 meters from case households. These risk differences may relate to differences between households with respect to rodent population densities and proximity to environmental sources of contamination, such as open sewers, open refuse deposits and flood risk areas. Molecular tools have been recently developed to detect pathogenic Leptospira in environmental samples [12],[32],[33] and used to identify that surface waters in slum communities contain high concentrations of pathogenic Leptospira which are agents for severe leptospirosis [12]. The use of these tools may enable more refined analyses aimed at identifying specific sources of contamination in the household environment which promote transmission of urban leptospirosis. Although our findings indicate that household factors are determinants of Leptospira transmission in slum communities, work-related exposures may contribute as well to infection in this population. A previous case-control study found workplace-related exposures to contaminated environment to be a risk factor for severe leptospirosis in addition to attributes associated with the household environment [19]. We did not evaluate infection risk at the workplace in this study. Occupations may have been shared among household members and have been a possible confounding factor. Slum residents often engage in informal work, such as small-scale construction and food preparation for vending, in the same environment in which they reside. Targeting exposures in the household environment may therefore have a potential beneficial effect in reducing potential work-related exposures. The study findings may not be generalizable to other epidemiological situations such as rural leptospirosis. Furthermore, the findings do not apply to urban leptospirosis in industrialised countries, which is a sporadic disease associated with inner city homeless populations [34]. However, the finding that leptospirosis is transmitted in the household environment will likely be relevant to the large slum population residing in developing countries which are subjected to similar conditions of poverty and social marginalization. In conclusion, this study identified significant household clustering of Leptospira infection in slum communities, indicating that the household environment and related factors are important determinants for transmission of urban leptospirosis. These findings need to be confirmed in prospective studies of Leptospira infection in slum communities. Research is needed to determine whether specific attributes of the slum household environment, such as open sewers, flooding, open refuse deposits, serve as transmission sources and evaluate the role of rats, dogs and other urban reservoirs which are commonly encountered in the places where slum residents reside. Elucidation of the risk factors in slum settlements may lead to targeted community-based interventions for leptospirosis. Moreover, efforts need to be made to raise awareness among slum residents of the risks that occur in their household environment. Implementation of effective community-based interventions will require further studies aimed at identifying the specific activities in the household setting which place slum residents at risk for leptospirosis and developing health education strategies to prevent these risks.

10.1371/journal.ppat.1003441

Multifunctional Double-negative T Cells in Sooty Mangabeys Mediate T-helper Functions Irrespective of SIV Infection

Studying SIV infection of natural host monkey species, such as sooty mangabeys, has provided insights into the immune changes associated with these nonprogressive infections. Mangabeys maintain immune health despite high viremia or the dramatic CD4 T cell depletion that can occur following multitropic SIV infection. Here we evaluate double-negative (DN)(CD3+CD4−CD8−) T cells that are resistant to SIV infection due to a lack of CD4 surface expression, for their potential to fulfill a role as helper T cells. We first determined that DN T cells are polyclonal and predominantly exhibit an effector memory phenotype (CD95+CD62L−). Microarray analysis of TCR (anti-CD3/CD28) stimulated DN T cells indicated that these cells are multifunctional and upregulate genes with marked similarity to CD4 T cells, such as immune genes associated with Th1 (IFNγ), Th2 (IL4, IL5, IL13, CD40L), Th17 (IL17, IL22) and TFH (IL21, ICOS, IL6) function, chemokines such as CXCL9 and CXCL10 and transcription factors known to be actively regulated in CD4 T cells. Multifunctional T-helper cell responses were maintained in DN T cells from uninfected and SIV infected mangabeys and persisted in mangabeys exhibiting SIV mediated CD4 loss. Interestingly, TCR stimulation of DN T cells from SIV infected mangabeys results in a decreased upregulation of IFNγ and increased IL5 and IL13 expression compared to uninfected mangabeys. Evaluation of proliferative capacity of DN T cells in vivo (BrDU labeling) indicated that these cells maintain their ability to proliferate despite SIV infection, and express the homeostatic cytokine receptors CD25 (IL2 receptor) and CD127 (IL7 receptor). This study identifies the potential for a CD4-negative T cell subset that is refractory to SIV infection to perform T-helper functions in mangabeys and suggests that immune therapeutics designed to increase DN T cell function during HIV infection may have beneficial effects for the host immune system.

SIV infection of sooty mangabeys is generally characterized by maintained CD4 T cell levels and a lack of disease progression despite active virus replication. We have previously shown however, that dramatic loss of CD4 T cells can occur during SIV infection of mangabeys. This study investigates the potential for double negative (DN) T cells (which lack CD4 and CD8, and are refractory to SIV/HIV infection) to perform helper T cell functions. In our study, sooty mangabey DN T cells exhibited a memory phenotype and a diverse repertoire in their T cell receptors. Once stimulated, the DN T cells expressed multiple cytokines, indicating that they have the potential to function as helper T cells (a function normally undertaken by CD4+ T cells). In SIV infected mangabeys, DN T cells continue to function, proliferate in vivo, and maintain expression of homeostatic cytokine receptors on their surface. It is therefore likely that DN T cells have the potential to compensate for the loss of CD4 T cells during SIV infection. These studies indicate that increasing DN T cell levels and/or function during pathogenic HIV infection may provide one tangible component of a functional cure, and inhibit progression to clinical disease and AIDS

While simian immunodeficiency virus (SIV) infection of Asian macaques generally results in progression to simian AIDS, SIV infection of African monkey species is typically associated with a nonpathogenic outcome. These African monkeys, including sooty mangabeys, are found naturally infected with SIV and are thought to have evolved with their species-specific viruses [1]. Studies of SIV infections of mangabeys and African green monkeys have provided key insights into the evolutionary mechanisms enabling these monkey species to remain free of disease. A number of studies have established that plasma viral levels are similar between natural SIV hosts and pathogenic infections observed in Rhesus macaques and HIV infected patients [2]–[6]. The SIV-specific antibody and cytotoxic T-lymphocyte (CTL) levels are also generally similar between the hosts of natural and pathogenic infections, indicating that stronger or more effective adaptive immune responses are not responsible for the non-pathogenic disease course [7], [8]. A major difference between natural and pathogenic infection is the lack of systemic immune activation (measured by proliferation/activation of immune cells, plasma cytokine levels) in the natural hosts during the chronic phase of the infection [4], [9]–[11]. How the natural hosts are able to suppress activation following the acute infection phase and why the natural hosts do not progress to simian AIDS despite high levels of viral replication are current avenues of research to understand SIV/HIV infection and host response. While it is clear that helper T cells play a key role in the immune responses elicited in sooty mangabeys and other natural host infections [12], [13], there is evidence that preservation of specific helper T cell subsets may be associated with improved disease outcomes [14]. Recent studies have indicated that low levels of direct virus infection in central memory CD4+ T cells, correlate with the lack of disease progression in SIV infected mangabeys [13]. Although CD4 T cells are critical for immune health, healthy SIV+ mangabeys with extremely low CD4 T cell levels are found in the Yerkes mangabey colony [11]. In addition, studies from our laboratory have identified a dramatic CD4 T cell loss (<50 cells/µl of blood) within a cohort of experimentally SIV infected mangabeys, which have remained free of clinical AIDS for greater than 11 years despite this paucity of CD4+ T cells [15], [16]. Here, these SIV+ mangabeys are termed “CD4-low” to indicate their unique CD4 status. Throughout the text, the terms “CD4-healthy (healthy range for CD4 T cells, 400–1500 cells/µl of blood)” and “CD4-low” are used to indicate the CD4 status, although both groups remain asymptomatic for simian AIDS. It is likely that, similar to HIV infection, the expansion in co-receptor usage in the replicating virus of the SIV+ CD4-low mangabeys enables the infection and depletion of a greater number of CD4+ T cells [15], [16]. Despite this depletion, the CD4-low SIV infected mangabeys maintain low levels of systemic immune activation, preserve lymphoid architecture, preserve function of non CD4 T cells and most importantly, show no clinical signs of simian AIDS [15], [16]. We have also previously shown that these SIV infected CD4-low animals have robust SIV specific T cell responses, SIV specific antibodies and antigen specific responses to neo-antigens present in vaccines [15], [16]. We postulate that helper T cell function in SIV infected mangabeys is compensated for by a SIV-resistant T cell subset that lacks CD4 expression, the double negative T cell. Double negative (DN) T cells are defined by expression of the CD3 protein, and a lack of both CD4 and CD8. Although the precise path of peripheral DN T cell development is not known, there are three models that have been proposed to explain how these cells arise and are maintained in the periphery [17]. One model proposes that immature DN thymocytes acquire expression of the T-cell receptor (TCR), bypass the subsequent double-positive (DP) and single-positive (SP) stages of classical T cell maturation, and migrate directly to the periphery. A second model suggests that a pre-T cell experiences all the normal development in the thymus but due to strong TCR∶MHC binding (sufficient to evade apoptosis) during negative selection, does not experience the CD4 or CD8 single positive stage. A third model suggests T cells proceeding all the way to the single-positive CD4 T cell, then experiencing a subsequent down-modulation of expression of the CD4 mRNA transcription and surface protein expression (reviewed in [17]). Peripheral DN T cells have important functions in a wide range of different disease states in both mice and humans [15], [18]–[28]. In infectious disease mouse models, DN T cells produce IL17 early during pulmonary Francisella tularensis live vaccine strain infection and also secrete IFNγ necessary for controlling intracellular bacterial growth [20]. In humans, DN T cells play T helper roles during parasitic infection, expressing IFNγ, TNFα and IL17 as a component of the immune response to Trypanosoma cruzi [25]. In addition to T helper functional roles, the regulatory potential of DN T cells has been identified through their ability to inhibit antigen specific CD4 and CD8 T cell proliferation in vitro in healthy humans [28], [29]. Furthermore, higher DN T cells numbers early in HIV infection is associated with decreased chronic immune activation later in infection, suggesting a regulatory role for DN T cells in HIV+ patients [30]. The cytotoxic potential of DN T cells has also been demonstrated by their ability kill allogeneic as well as antigen-loaded syngeneic DCs [31], autoreactive CD8+ T cells [32], and activated allogeneic and syngeneic B cells [21]. Together, these data suggest that DN T cells in mice and humans exhibit functions similar to other T cell subsets. In natural host monkey species, two different T cell subsets that lack a CD4 molecule have been described: the first is CD3+CD4−CD8αdim cells [33], [34] and the second is CD3+CD4−CD8− DN T cells [15], [34] (these are distinct from invariant chain NKT cells [35]). In addition to peripheral blood, DN T cells are also present in different immunological tissue sites including lymph nodes, lungs, and rectal mucosa [15], [16], [36], [37]. These tissue sites also maintain DN T cell numbers as CD4+ T cells are depleted during both pathogenic and non-pathogenic SIV infections [15], [16], [36], [37]. Vinton et al., performed a cross sectional analysis of DN T cells in different natural hosts to elucidate their function and revealed that DN T cells are found in larger proportions (10–40% of lymphocytes) in natural hosts (sooty mangabeys, African green monkeys and patas monkeys) than in pathogenic host species (Rhesus macaques) [34]. In addition, there is limited apoptosis in DN T cells during SIV infection of natural hosts (sooty mangabeys) compared to SIV infected Rhesus macaques [38]. In these studies, DN T cells in the peripheral blood were predominantly memory cells with the majority of these cells having a central memory phenotype (expressing CD28, CD95 and CCR7). Assessment of purified populations also found very low, or undetectable, levels of SIV nucleic acid within the DN T cell subset, indicating that these cells are refractory to SIV infection [34]. These key pieces of data led us to hypothesize that DN T cells in natural hosts may be important for providing helper T cell responses during SIV infection, as they are more abundant in natural hosts, and do not become depleted during lentiviral infections. The findings presented herein provide an assessment of the functional attributes of DN T cells in uninfected and SIV infected sooty mangabeys. Our previous studies have indicated that DN T cells may have specific T helper roles in the CD4-low mangabeys [15]. Using comprehensive transcriptome analysis of DN, CD4 and CD8 T cells, we now demonstrate that DN T cells in sooty mangabeys share a functional profile closer to CD4 T cells than to CD8s. In addition, by comparing SIV infected mangabeys with different CD4 T cells levels (CD4-low and CD4-healthy), we demonstrate that neither SIV infection nor SIV induced CD4-T cell loss markedly altered overall DN T cell function in this natural host model. We further demonstrate that, like CD4 T cells [13], DN T cells have a diverse T cell receptor repertoire and are predominantly effector memory cells that can be restricted by either MHC-I or MHC-II molecules (with a multi-functional response being associated with MHC-II utilization). The similarities between DN and CD4 T cells during SIV infection of sooty mangabeys indicate a key importance of this T cell subset and underscore the potential for this cell population as an immunotherapeutic target to prevent HIV-induced disease progression. Double negative (DN) T cells express CD3 protein but do not express either CD4 or CD8 proteins and can be identified through flow cytometric analysis of peripheral blood cells (Figure 1A) as well as other tissues [15], [16] (Figure S2). This DN phenotype (absence of CD4 and CD8) can be observed at both the protein (Figure S1A) and mRNA levels (Figure S1C, depicts an absence of CD4 mRNA) and is maintained even after the cells are stimulated through their T-cell receptor (TCR). These cells are predominately not NKT cells, as staining with an canonical NKT cell-specific TCR antibody (anti-Vα24) indicated that less than 1% of DN T cells express this TCR (Figure S1A). In addition, mangabey DN T cells predominately express an αβTCR, with 17% expressing the γδTCR (similar between SIV infected and uninfected mangabeys (Figure S1B). In addition to peripheral blood, we also assessed the levels of DN T cells in mucosal tissues. The relative proportion of DN T cells (as a percentage of CD3+ cells) increased within the rectal biopsy samples during the times when the CD4 T cell levels decline within the SIV-infected CD4-low mangabeys (Figure S2A). In contrast, the proportion of DN T cells within the bronchoalveolar lavage (BAL) samples remain constant during this time, possibly due to an increased percentage of CD8 T cells in this site (Figure S2B). The naïve and memory phenotypes of the DN T cells can be categorized based on the expression of CD95. Although previous studies have further characterized central and effector memory subsets using the differential expression of the co-stimulatory marker CD28 [15], we have now adapted a more recent definition of identifying central memory (CD95+/CD62L+) and effector memory (CD95+/CD62L−) cell subsets [13]. CD62L expression on a T cell induces lymph node homing for the central memory T cells, whereas effector memory cells lose the CD62L expression upon recognition by cognate antigen in order to travel to the site of injury [13], [39], [40]. In the peripheral blood, DN T cells were predominantly memory cells (CD95+) with 75%±9% (mean±SD) of memory cells demonstrating an effector memory (EM: CD95+CD62L−) phenotype, and 25%±9% (mean±SD) with a central memory (CM: CD95+ CD62L+) phenotype. During chronic SIV infection, we observed an increase in the relative percentage of EM DN T cell population, possibly due to a conversion of CM to EM phenotype. Together, these findings identify a predominately memory phenotype for mangabey peripheral blood DN T cells, similar to human DN T cells (defined by the expression of CD45RO) [41]. The T cell repertoire represents the potential for a T cell population to recognize a broad range of antigens, and provide protection against more potentially pathogenic organisms. To evaluate the diversity of purified mangabey DN T cells we utilized PCR primer sets based on Rhesus macaque TCR designed to potentially amplify 23 Vβ sequences. Multiple Vβ sequences were amplified from DN T cells from uninfected, SIV infected CD4-healthy and SIV infected CD4-low mangabeys (Table 1). A further evaluation of the length heterogeneities of the CDR3 region was undertaken by spectratyping [42], [43] to determine the intra-Vβ diversity (Figure S3). Five Vβ genotypes were not detected in DN and CD4 T cells irrespective of SIV infection status. This lack of detection occurred even though these five Vβ sequences share 100% homology in the Rhesus and human genomes (therefore indicating a high likelihood of nucleotide similarity in mangabeys as well) and may thus indicate a the absence of these Vβs in the mangabey repertoire (although the potential for primer mismatch to the mangabey sequences can not be ruled out). In the detected Vβ genotypes, polyclonal repertoire was observed in the DN T cells with 11 Vβ amplified in DN T cells from uninfected, 17 Vβ in SIV infected CD4-healthy and 14 Vβ for the SIV infected CD4low mangabeys (Table 1). This repertoire was similar to the repertoire of CD4 T cells in the same animals (data not shown). Junctional diversity within the majority Vβ CDR3s was identified through a visualization of DNA length differences (3 bps intervals), with the most prevalent CDR3 length generally at the central position (Figure S3, Table S1). Junctional diversity was observed for each amplified Vβ with the exception of Vβ7 and Vβ20, which manifest as single peaks (indicating clonality) in the uninfected mangabeys (Table 1). Interestingly, 4 Vβ genotypes undetectable in the uninfected mangabeys (Vβ 8, 15, 21 and 24) were present in SIV infected mangabeys (both CD4-healthy and CD4 low. Table 1) providing evidence for expansion of T cells with TCR containing these Vβ regions after SIV infection. These findings indicate that DN T cells express a diverse polyclonal TCR repertoire in both uninfected and SIV infected mangabeys, during both CD4-healthy and CD4 T cell depleting conditions. Gene expression analysis of purified CD4, CD8 and DN T cells from uninfected mangabeys were analyzed by both array and real-time PCR analysis. Microarray analysis undertaken to evaluate TCR stimulation (anti-CD3/anti-CD28) determined that CD4 T cells differentially regulated (over 2 fold with p<0.05) 2005 genes, whereas CD8 and DN T cells differentially regulated fewer genes, 1254 genes and 1123 genes respectively. Some of the immunomodulatory genes are depicted in the heat-map (Figure 2A) highlighting the impact of the TCR stimulation on DN T cells (lane 1), CD4 T cells (lane 2) and CD8 T cells (lane 3). Among the 122 genes differentially regulated in each cell type, there were numerous immune modulatory transcripts including cytokines and chemokines (IFNγ, CCL4L1 (CCR5 agonist) and XCL1) as well as transcription factors (RGS8 and BATF). The similarity between DN, CD4 and CD8 T cells can be visualized via a Venn diagram (Figure 2B), which depicts 509 genes regulated to the same extent in DN and CD4 T cells and 248 genes regulated similarly between DN and CD8 T cells. Additionally, a hierarchical cluster analysis (Pearson's algorithm) of highly upregulated genes (over 4 fold with p<0.05) determined that genes upregulated by DN T cells more closely cluster with CD4 T cells rather than CD8 T cells (Figure 2A, phylogenetic tree above the heat map). Genes differentially regulated similarly in the DN and CD8 T cells subsets were often involved in ion channel and signaling function (KCNK5, RIPK2 and ERG2) as well as the costimulatory molecules such as ICOS. Interestingly, many of the differentially regulated genes shared by CD4 and DN T cells were associated with T helper function including IL4, IL17F, IL22 and IL2RA; chemokines such as CXCL9, CXCL10 and CXCL11 and other immune genes such as MIP3α (CCL20) and IRF8 (Figure 2) In order to quantify and compare these shared T helper cytokines between CD4 and DN T cells, we performed qPCR analysis on purified CD4 (open symbols) and DN T cells (filled symbols) from 10 uninfected mangabeys focusing on seven key cytokines. These cytokines included those specific for Th1 (IFNγ), Th2 (IL4), Th17 (IL17), pro-inflammatory (TNFα), anti-inflammatory (IL10), Treg (TGFβ) and antiviral (IFNα) cytokines. Following stimulation through the T cell receptor (anti-CD3/CD28) DN T cells exhibited the strongest upregulation in three canonical T helper cytokines IFNγ (700 fold, Th1), IL4 (1000 fold, Th2) and IL17 (900 fold, Th17), with slightly less upregulation for the pro-inflammatory cytokines TNFα (30 fold) and anti-inflammatory IL10 (10 fold) (Figure 3, filled symbols). No upregulation was observed in TGFβ (Treg) or IFNα (antiviral) following TCR stimulation, although DN T cells did express these cytokines at a basal level. Similarities were clear in the pattern of cytokine expression between DN and CD4 T cells (Figure 3), although DN T cells demonstrated a greater upregulation of IFNγ (2.4 fold, p = 0.003), IL4 (3 fold, p = 0.03) and IL17 (5 fold, p = 0.002) than CD4 T cells (Figure 3). Stronger mitogenic stimulus with PMA/Ionomycin also revealed a similar functional profile in DN and CD4 T cells (Figure S4); however, the significant differences between CD4 and DN T cells were no longer evident, indicating that DN T cells respond in a more similar manner to CD4 T cells when provided a strong stimulus. We next analyzed whether SIV infection or SIV-induced CD4 T cell loss impacts the functional potential of DN T cells. Microarray profiles of DN T cells were evaluated from 2 SIV+ CD4-low, 2 SIV+ CD4-healthy and 2 uninfected mangabeys. Following TCR-stimulation (anti-CD3/CD28), DN T cells differentially regulated 263 genes over 4 fold (p<0.05) in the SIV+ CD4-low and SIV+ CD4-healthy animals compared to unstimulated DN T cells (Figure 4A, Blue circles). Several genes were also differentially regulated only in the CD4-low (Figure 4A, Purple circles n = 215) or CD4-healthy mangabeys (Figure 4A, green circles, n = 90). Interestingly, SIV+ CD4-low mangabeys exhibited a greater number of genes that are either up- or down-regulated when compared to SIV+ CD4-healthy mangabeys (478 versus 353 respectively) indicating that the function of DN T cells may be altered following a prolonged depletion of CD4 T cells (Figure 4A). Differentially regulated genes were assessed for enrichment of Gene Ontologies using the DAVID Bioinformatics Database [44], [45], and genes with related ontologies clustered into functional groups. Based on this unbiased assessment, upregulation of immune response genes was the most prominent among the ontologies (enrichment for immune genes) in addition to a number of more specific functional ontologies, including chemotaxis, apoptosis, T-cell proliferation and transcription (Figure 4B). Guided by these ontologies, we further classified the data, using heat maps to represent differential expression upon TCR stimulation, separating the immune response ontology into innate and adaptive responses, and generating a broad category of immune modulatory signaling and transcription. From several hundred differentially regulated genes, 111 genes with known immune modulatory function are presented (Figure 4C). Transcription factors and signaling molecules are the largest group of genes upregulated by DN T cells after TCR stimulation with over 4 fold upregulation in ZBED2, TNIP3, RGS4, IRF8, GADD45B and SMAD1 (Figure 4C). Interestingly, a number of genes, including signaling molecules such as RGS4 and TNIP3, were upregulated after SIV infection and differentially upregulated between CD4-low and CD4-healthy mangabeys (5 and 2 fold higher respectively; Table 2). However, some transcription factors such as SMAD7 were downregulated in the DN T cells particularly in the CD4-low mangabeys. Strong upregulation was observed for genes associated with classical Th1 function in both CD4-low and CD4-healthy mangabeys, including the signature cytokine IFNγ as well as interferon stimulated chemokines CXCL9, CXCL10 and CXCL11 (with the highest CXCL10 expression in DN T cells from CD4-low mangabeys, Table 2). This chemokine expression signature was maintained in DN T cells from both uninfected and SIV infected mangabeys (Figure 4C). DN T cells also have a potential for Th2 function as indicated by the upregulation of CD40LG as well as cytokines IL4, IL5, IL13, and IL10. CD40LG, IL4 and IL10 expression was maintained irrespective of SIV infection although the expression of IL5 and IL13 was increased most dramatically in DN T cells from SIV infected mangabeys (Figure 4C). Cytokine mRNAs associated with canonical Th17 function IL17, IL6, TNFα, IL22 were maintained after SIV infection and exhibited a greater upregulated in CD4-low compared to CD4-healthy mangabeys (Table 2). TFH (T follicular helper) function (IL21, ICOS, CD40L, IL6) was assessed in DN from infected and uninfected mangabeys. Although ICOS, CD40L and IL6 function remained similar irrespective of SIV infection, IL21 was strongly induced in the DN T cells from SIV infected mangabeys (Figure 4C). Interestingly, an unsupervised in silico pathway analysis based on all significantly upregulated genes also indicated with a high degree of confidence (p = 5.94E−12) that the expression patterns of DN T cells is consistent with helper T-cell attributes and include Th1, Th2, Th17 and TFH functionality (Figure 5A). Gene expression data were integrated with extant literature to generate a network of direct interactions between key cytokines, chemokines and related molecules, revealing interesting regulatory features (Figure 5B). For example, the expression of IL2 as well as its cognate receptor IL2RA and the enrichment of the lymphocyte proliferation ontology points to an ability of DN T cells to proliferate in response to this T cell survival cytokine. Conversely, although the pro-inflammatory cytokine TNFα and anti-inflammatory cytokine IL10 were upregulated, their specific receptors, IL10R, IL10RA and TNF receptor were downregulated in the DN T cells, indicating that DN T cells buffer their own pro- and anti-inflammatory capacity, yet aid in providing a cytokine milieu to potentiate immune responses upon TCR stimulation. (Figure 5B). Taken together, these expression data and systems-level analyses demonstrate that DN T cells from both CD4-low and CD4-healthy SIV infected mangabeys signal through their T cell receptors in the absence of CD4 and CD8 and express cytokines/chemokines associated with Th1, Th2, TFH and Th17 helper T cells. The ability of DN T cells to compensate for CD4+ T cell loss during SIV infection can be further inferred by evaluating DN from SIV+ CD4-healthy (n = 7) and SIV+ CD4-low mangabeys (n = 6) to the functional profile to CD4 T cells from the CD4-healthy mangabeys (SIV+ CD4-low mangabeys had too few CD4 T cells to isolate and analyze in this manner). Similar to our evaluation of cells from uninfected mangabeys (Figure 3), cytokine transcripts associated with Th1 (IFNγ), Th2 (IL4), Th17 (IL17), pro-inflammatory (TNFα), anti-inflammatory (IL10), Treg (TGFβ) and antiviral (IFNα) were evaluated (utilizing quantitative real-time PCR) following stimulation through the T cell receptor (anti-CD3/CD28 – Figure 6) or through a mitogenic stimulus (PMA/Ionomycin – Figure S5). A paired evaluation of the mRNA expression of these cytokines following TCR stimulation in DN and CD4 T cell was undertaken in SIV+ CD4-healthy mangabeys. The mean upregulation of the cytokines IFNγ (500 fold), IL17 (800 fold), IL10 (20 fold) and TNFα (50 fold) was similar in the DN and CD4+ T cells (Figure 4, black symbols compared to open symbols). The one difference observed was that IL4 upregulation was higher in the DN compared to the CD4 T cells, indicating that DN T cells were particularly adept at making this cytokine (Figure 4). Following a more robust stimulation with the mitogen PMA/Ionomycin, a significant elevation in cytokine mRNA levels (compared to TCR stimulation) was observed in both DN and CD4 T cells however no significant difference in was observed between CD4 and DN T cells (Figure S5). IFNα and TGFβ mRNA were detectable, but their levels were not upregulated in DN or CD4 T cells following either TCR or mitogen stimulation. Importantly, we observed that DN T cell function was similar in both SIV+ CD4-low mangabeys (Figure 6, red symbols) and SIV+ CD4-healthy mangabeys (black filled symbols) following either TCR or PMA/Ionomycin stimulation (Figures 6 and S5). There was no significant difference between cytokine mRNA expression of DN T cells between CD4-low and CD4-healthy mangabeys in these signature T helper cytokines indicating a maintained function despite CD4 loss. Differential expression of cytokines in the SIV infected and uninfected mangabeys have the potential to uncover key insights into DN T cell function. Comparison of IFNγ expression by DN T cells in SIV infected and uninfected mangabeys identified a two-fold higher upregulation of IFNγ in the uninfected mangabeys following CD3/CD28 TCR stimulation (Figure 7). This difference was statistically significant at both the protein (1.7% in uninfected and 0.7% in infected - Figure 7C) and mRNA (700 fold in uninfected and 350 fold in infected - Figure 7D) levels. This difference was not observed following a stronger mitogenic PMA/Ionomycin stimulation (Figure 7D, S4, S5). Further, intracellular cytokine production analysis revealed that DN T cells expressed either a Th1 (IFNγ) or Th2 (IL4) cytokine profile (Figure 7A) [46], [47]. These findings provide evidence that, like CD4 T cells, DN T cells exist as different functional subpopulations and that SIV infection impairs the ability of these cells to produce IFNγ, a cytokine that has the ability to direct the immune system primarily toward a Th1 response. No differences were evident between infected and uninfected mangabeys in any of the other cytokine mRNA measured by qPCR or cytokine protein levels measured by intracellular cytokine staining. We also assessed the ability of DN T cells to respond to SIV Env specific peptides in the context of MHC binding. Blocking experiments with anti-MHCI or anti-MHC-II antibodies demonstrated that IFNγ secretion by DN T cells is mediated through either MHC-I (54% inhibition) or MHC-II (46% inhibition) dependent pathways. In contrast, Env peptide stimulation of IL2 secretion was mediated primarily through an MHC-II dependent manner (40% inhibition) and was not inhibited by MHC-I antibodies (Figure 7B). Taken together, these data demonstrate that with regard to IFNγ secretion, DN T cells can recognize antigens presented by either MHC-I or MHC-II molecules. However, DN T cells the express the proliferation cytokine IL2 are driven primarily through a MHC-II mechanism. These data may indicate that one DN T cell can respond to both MHC types, or rather that one population responds to antigen presented by MHC-I and a different DN subpopulation responds to antigens presented by MHC-II. To evaluate DN T cell proliferation and turnover, 3 uninfected and 5 chronically SIV infected mangabeys were treated with the thymidine analogue 5-bromo-2′-deoxyuridine (BrdU) for 14 days (Figure 8). BrdU incorporation was monitored throughout the labeling period and for a period of 90 days following treatment. The BrdU uptake by the DN T cells during labeling period and the kinetics of BrdU during the wash-out was similar in both SIV infected and uninfected mangabeys (Figure 8). This finding was further verified through expression of the Ki67 protein by flow cytometry, which was similar in the mangabey DN T cells irrespective of their infection status (data not shown). In addition, we assessed the expression of the IL2 receptor (CD25) and IL7 receptor (CD127) on DN T cells from SIV infected and uninfected mangabeys to evaluate their potential to respond to the homeostatic cytokines (Figure 9). CD25 was expressed on 8% of DN T cells from uninfected mangabeys and at a significantly higher level (14%) in SIV infected mangabeys. Additionally, a higher percentage of DN T cells expressed CD25 compared to CD4 T cells in both infected and uninfected mangabeys (Figure 9). CD127 was consistently expressed at a high level in DN T cells from both SIV infected and uninfected mangabeys with a mean expression of 62% (Figure 9), albeit at lower levels than CD4 T cells (mean expression 81%). Taken together these data show that mangabey DN T cells proliferate during SIV infection and express levels of homeostatic cytokine receptors that demonstrate a potential for responding to both IL2 and IL7. This study demonstrates that (CD3+CD4−CD8−) DN T cells possess Th1, Th2, Th17 and TFH function in sooty mangabeys irrespective of SIV infection. We expand upon previous findings by performing an in-depth analysis of DN T cell functions using transcriptomic as well as flow cytometric analyses, and now define how these functions are altered during SIV infection and SIV mediated CD4 T cell loss. DN T cells are refractory to SIV-infection [34], which likely aids their ability to maintain function and proliferative capacity during SIV infection and the subsequent CD4 depletion observed in the CD4-low mangabeys [11], [15], [16]. We identify DN T cells as a polyclonal T cell subset with a predominantly EM cell phenotype, maintained proliferative capacity, and with a specific increase in the proportion of EM DN T cells. Further, we show that they have the ability to respond to SIV specific antigens presented in the context of both MHC-I or MHC-II. Previous studies have demonstrated that during SIV infection of mangabeys, CD4 T cells retain a CM phenotype and this preservation of CM T cells has been implicated as one of the mechanisms by which mangabeys are able to inhibit progression to simian AIDS [13]. Since EM CD4 T cells are infected at greater frequency than CM CD4 T cells in the sooty mangabeys, the DN T cells, which are refractory to SIV infection, may therefore perform T helper effector functions and serve as workhorses for the mangabey immune system, with a potential to function during SIV infection. Whole transcriptome microarray analysis was previously utilized in non-sorted peripheral blood cells from natural hosts to demonstrate that natural hosts can produce robust innate immune responses and upregulate interferon stimulated genes (ISGs) following SIV infection [48]. Our transcriptome analysis of DN T cells demonstrated that TCR stimulation results in an upregulation of numerous transcription factors, signaling proteins and immune modulatory proteins despite the lack of CD4 or CD8 proteins. When CD4 and CD8 T cells respond to antigens in the context of either MHCII or MHC I respectively, the protein tyrosine kinase Lck (p56) associated with the intracellular domains of CD4 or CD8 is responsible for initiating the signal transduction through the TCR [49], [50]. One possibility for DN signaling in the absence of these critical molecules is that DN T cells are less dependent on Lck signaling [27], [51], [52] and activation of Fyn along with LAT and Erk kinases may be sufficient to activate ITAMs in DN T cells [52]–[54]. Our analysis has shown that DN T cells do respond to SIV specific peptides in both a MHC I and MHC II dependent manner (through production of IFNγ). We hypothesize that DN T cells are able to respond to antigens presented by either MHC molecule due to the absence of either a CD4 or CD8 protein in these cells. IFNγ is a Th1 cytokine that was highly expressed in TCR-stimulated DN T cells (500 fold upregulation by real-time PCR). IFNγ plays a role in clearance of intracellular pathogens by macrophages and is critical for antigen specific T cell responses. IFNγ stimulated genes CXCL9, CXCL10 and CXCL11 were also strong upregulated and maintained in the DN T cells irrespective of SIV infection or SIV-induced CD4 T cell loss. The strong upregulation of IFNγ by DN T cells provides evidence that they can influence both innate and adaptive immune response by enhancing macrophage antimicrobial activity and CD8 T cell function. Here we demonstrate that SIV infection is associated with a decrease in IFNγmRNA and protein expression in the DN T cells (Figure 7C,D). We hypothesize that this SIV related IFNγ down-modulation may play a role in the suppressed immune activation observed during the chronic phase of the infection in mangabeys compared to SIV infected macaques and HIV-infected humans. A functional humoral immune response requires upregulation of TFH (follicular helper) and Th2 cytokine expression in order to provide B cell help. Unsupervised in silico analysis (Figure 5) of DN cytokine expression indicated that DN T cells possess Th2 function, as evidenced by upregulation of signature cytokines IL4, IL5, IL10 and IL13 (Figure 2A, 4C) necessary for activation of antigen-specific B cells and the production of antibodies. It is interesting that SIV infection augments the ability of DN T cells to upregulate IL5 and IL13 while maintaining IL4 and IL10 levels (Figure 4C). In addition, our analysis also indicates potential TFH (T follicular helper) function in DN T cells. Peripheral blood TFH cells generally have a memory phenotype [55], [56], express the cytokine IL21, and express high levels of CD40 ligand (CD40L) as well as inducible costimulator (ICOS) required for B cell activation. They also secrete IL6, augmenting IL21 production to assist in generating antibody responses [57]. Transcriptome analysis (Figure 2, 4) determined that DN T cells in SIV infected mangabeys upregulate IL21, ICOS, CD40L and IL6 indicating the potential for TFH functionality. This pattern of IL4, IL5, IL10 and IL13 expression in addition to IL21, ICOS and CD40L suggests that DN T cells can provide B cell help to maintain adaptive humoral responses when CD4 levels are low due to SIV infection. Indeed, SIV+ CD4-low mangabeys are able to generate recall antibody responses to vaccination against Influenza [16] and DN T cells have the potential to assist in facilitating this helper T cell function. Preservation of Th17 function in natural hosts is critical to a non-pathogenic disease course following SIV infection, as Th17 T cells are involved in maintaining gut mucosa integrity [58]–[61]. Th17 cells are selectively depleted during pathogenic SIV infection of macaques and in HIV+ humans [14], [58], [59]. Th17 function in DN T cells was demonstrated by the upregulation of Th17 cytokines (IL17 and IL22 [60]–[62]) as determined by both arrays and real-time PCR analysis indicate that this Th17 function was maintained in DN T cells irrespective of SIV infection. The maintenance of Th17 function by DN T cells during SIV-infection may allow the natural host immune system access to the Th17 cytokines critical for maintaining mucosal integrity to prevent microbial translocation that would otherwise contribute to chronic immune activation. A key feature of SIV infected natural hosts in limiting disease progression is their ability to limit chronic phase immune activation, the mechanism for which is still under investigation [3], [5], [10], [63]. The cytokine profile of maintained Th17 function, reduced IFNγ expression upon SIV infection (discussed above), and upregulation of the classic anti-inflammatory cytokine IL10 by DN T cells may contribute towards controlling this systemic immune activation. Additionally DN T cells also strongly express the Th17 cytokine IL22 (Figure 2 and 4), which can have anti-inflammatory properties when acting in the presence of with IL10 [64]. Interestingly, IL10 upregulation in DN T cells occurs concurrently with a down regulation of IL10 receptors – IL10R, IL10RA, raising the possibility that the DN T cells suppress inflammation in other cells, while inhibiting their own response to IL10. This may allow DN T cells to remain fully activated and functional while contributing to lower immune activation in the mangabeys. Recent studies have demonstrated that in primary HIV infection there is a strong correlation between higher DN T cell levels early in infection and lower HLADR+CD38+ CD8 T cells later in infection [30]. This reduced immune activation was associated with the expression of anti-inflammatory cytokines IL10 and TGFβ by DN T cell subset that does not express FoxP3 [30] (a marker for Treg function in CD4 T cells). Previous studies of DN T cells in sooty mangabeys demonstrate that DN and CD4 T cells express similar levels of FoxP3 [34]. Our findings also demonstrate the expression of FoxP3 by DN T cells, although no upregulation was detected after TCR stimulation. Our transcriptomic analysis of TCR stimulated DN T cells also identified a strong upregulation in transcription factor TNIP3 (TNFAIP interacting protein 3) an anti-inflammatory gene that limits NF-κB signaling and signaling molecule RGS4 (also a negative regulator [65]), thus contributing to control of excessive/sustained immune activation during chronic SIV infection. This functional profile of DN T cells in mangabeys, characterized by higher DN T cell number and IL10 expression supports our hypothesis that DN T cells aid in the control of immune activation during the chronic phase of SIV infection of natural hosts. These multi-functional roles of DN T cells suggest that DN T cells are functional (including their anti-inflammatory role) irrespective of CD4-T cell levels or SIV infection status. It is therefore likely, that similar to CD4 T cells, a subpopulation of DN T cells has a regulatory function. These findings characterize DN T cell functionality in sooty mangabeys, providing evidence for DN T cell subset function that is similar to CD4 T helper cells, and a maintained ability of these cells to proliferate and function following the SIV infection. As a population recalcitrant to SIV infection, DN T cells have the potential to contribute to the maintenance of an effective immune and may be evolution's answer to preserve helper T cell functions during a CD4 T cell depleting infection. We propose that DN T are a candidate immune therapeutic target to decrease disease progression in HIV+ patients. One approach to increase DN T cell functional potential would be to increase their levels, as they exist at a higher frequency in non-pathogenic hosts compared to humans [34], One candidate therapy is to promote the expansion of the DN T cell populations is interleukin 7 (IL7), a cytokine involved in the homeostasis and maintenance of T cell populations [66], [67]. The presence of the IL7 receptor (CD127) on over 60% of DN T cells even during SIV infection (Figure 8) suggests that these cells have the potential to respond to IL7 therapy. Several clinical trials utlizing IL7 therapy in HIV-infected patients and SIV infected Rhesus macaques have demonstrated promising increases in both CD4+ and CD8+ T cell pools in the peripheral circulation [66]–[72], including the naïve and central memory CD4+ T cell compartments [66], [73]. A second approach could be expanding function of DN T cells from pathogenic hosts to encompass Th1, Th2, Th17, and TFH functions (as observed in SIV+ mangabeys). Analysis of DN T cell function in the humans have generally focused on the regulatory potential [28], [29] of these cells and other specific functions are currently being investigated. These studies indicate the potential for utilizing DN T cells as future immune therapeutic targets (potentially utilizing IL7, IL2 or a combination of IL2/IL7 therapy) to increase their levels and/or function during pathogenic HIV infection as one component of a functional cure to inhibit progression to clinical disease and AIDS. All animals involved were cared for in accordance with NIH guidelines as well as by approved protocols with the Seattle Biomedical Research Institute's Institutional Animal Care and Use committee (IACUC) committee (#DS-NHP-Yerkes) and Yerkes National Primate Research Center's IACUC approved protocol (IACUC #2000280). Appropriate measures were taken to assure that discomfort, distress, pain and injury was limited to that which is unavoidable in the conduct of the research plan. Ketamine (10 mg/kg) and/or Telazol (4 mg/kg) were used for sedation and analgesics were used when appropriate as determined by the veterinary medical staff. We obtained mangabey blood from colony bred sooty mangabeys (Cercocebys atys) at the Yerkes National Primate Research Center. The SIV infected mangabeys utilized in this study were either naturally infected in the Yerkes colony or experimentally infected as previously described [16]. Five SIV infected CD4-low mangabeys are part of the previously published Sodora Lab study [16] and we have also included one mangabey present within the naturally SIV infected group within the Yerkes colony that also had a CD4-low phenotype. All antibodies were obtained from BD unless specified. Peripheral blood mononuclear cells (PBMC) were isolated from blood of SIV infected and uninfected mangabeys via Ficoll-Hypaque gradient. PBMCs were stained with CD3 (SP34 clone) APC-Cy7 or Alexa 700; CD4 (L200) Pacific Blue or PerCP-Cy5.5; CD8 (SK1 clone) PerCP-Cy5.5 or (RPA-T8 clone) Pacific Blue; CD28 (28.2 clone) PE-Cy7 or FITC, CD95 (DX2 clone) PE-Cy5 or APC, CD62L PE to perform identify naïve and memory phenotypes. Analysis of proliferating cells were performed using Ki67 (B56 clone) FITC and BrDU (3D4) APC along with CD8 (SK11) APC-Cy7, CD3 (SP34-2 clone) Alexa 700 and CD4 (OKT4 clone) Pacific Blue. Cytokine receptor analysis utilized CD25 PE-Cy7 and CD127 PE antibodies. IFNγ (4S.B3 clone) PE and IL4 (8D4-8 clone) PE-Cy7 were utilized to identify cytokine secretion. For intracellular cytokine staining, CD95 and CD28 antibodies were stained extracellularly followed by a permeabilization step and CD3, CD4, CD8, IFNγ and IL4 included as intracellular stains. Data was acquired on a BD LSR II and we analyzed stained cell populations using FlowJo (TreeStar). PBMCs from SIV infected mangabeys were pretreated with antibodies against MHC-I (G46-2.6, BD Bioscience) or MHC-II (Tu39, BD Bioscience) for 2 hrs at 37°C, followed by an incubation with Env peptides (2 µg/ml) in the presence of 2.5 µg of CD28 monoclonal antibody (CD28.2, BioLegend) and 10 µg/ml brefeldin A (Sigma) for 6 hrs. Following stimulation, the cells were then stained with Live-Dead Aqua and surface markers using anti-CD3-APC-Cy7 (clone SP34-2), anti-CD4-PE (clone L200), anti-CD8-PacBlue (clone RPA-T8), anti-CD95-PE-Cy5 (clone DX2) (all from BD Pharmingen) and anti-CD28-ECD (clone CD28.2, Beckman Coulter). After permeabilization, cells were stained intracellular antibodies against anti-Ki67-Alexa700 (clone B56), anti-IL21-Alexa Fluor647 (clone 3A3-N2.1), anti-IFNγ-PE-Cy7 (clone B27) (all from BD Pharmingen); anti-IL17-Alexa Fluor488 (clone eBio64DEC17, eBioscience); anti-IL-2-Brilliant Violet 605 (clone MQ1-17H12, Biolegend). Data was acquired on an LSRII cytometer using the FACS DiVa software and analysis was performed using FlowJo software (TreeStar). To assess proliferation and turnover of DN T cells in-vivo, three SIV uninfected and five chronically SIV infected mangabeys were assessed. All animals tested negative for STLV. Infected animals exhibited viral loads greater than 10,000 copies/ml and CD4+ T-cell counts greater than 500 cells/µl. All animals exhibited comparable distributions of T-lymphocyte counts. All animals were treated with intravenously with 60 mg/kg of the thymidine analogue 5-bromo-2′-deoxyuridine BrdU (Sigma-Aldrich) diluted in HBSS (Invitrogen) per weekday and 120 mg/kg orally of BrdU per weekend-day over 14 consecutive days. Cessation of BrDU treatment is the beginning of BrDU washout period. Pure populations of DN and CD4 T cells were isolated by a step-wise magnetic bead isolation (Miltenyi) using beads specific for non-human primate antigens. Briefly, we stained PBMC with CD4-Microbeads to isolate CD4 cells and treated the flow-through with CD8-PE and CD16-PE followed by anti-PE Microbeads for CD8 and CD16 depletion. This CD4/CD8/CD16 depleted flow-through was then positively selected for CD3+ cells with CD3-Biotin and anti-biotin Microbead to obtain, CD3+CD4−CD8− (double negative T cells). Purity was checked after each isolation to guarantee >98% DN T cell pure populations. Cells were rested overnight prior to functional analyses. We stimulated 1 million purified T cells or PBMC with either anti-CD3 (3 mg/ml) and anti-CD28 (1 mg/ml); or PMA (25 ng/ml) and Ionomycin (1 mg/ml) in 200 ul of complete RPMI at 37°C for 4 hrs (real time PCR and microarray) or 6 hrs (ICS). For Microarray and real time PCR analysis, the cells were collected via centrifugation post-stimulation and immediately lysed in RLT buffer (Qiagen) for RNA extraction. For ICS, the cells were washed and stained appropriately. We isolated RNA (RNeasy Mini Kit – Qiagen) from unstimulated and stimulated CD4 or DN T cells from the same animal and quantified the obtained RNA. cDNA was synthesized using Superscript II First Strand cDNA synthesis kit (Invitrogen) and normalized to 2 ng/ml of starting RNA. Real time PCR was performed utilizing gene-specific primers and probes [74] using the TaqMan system (Applied Biosystem) on the ABI 7300 (Applied Biosystem). We estimated the changes in gene expression between unstimulated and stimulated cells for IFNγ, IL4, IL17, IL10, TGFβ, TNFα and IFNα gene expression (GAPDH - housekeeping gene control) using the ΔΔCt method. Briefly, the Ct value of the specific gene was normalized to the Ct value of GAPDH thus generating a ΔCt value for the gene-of-interest. The ΔCt was calculated by subtracting ΔCt of gene of interest from the unstimulated cells from the ΔCt of gene of interest from the stimulated cells. Fold change for the gene of interest in each cell type was calculated by 2−ΔΔCt. Fold change of 1 indicated no upregulation of mRNA upon stimulation. Purified DN T cells (magnetic bead isolation) were stimulated for 4 hours with anti-CD3/CD28. Total RNA was isolated from unstimulated and stimulated DN T cells (RNeasy, Qiagen) and genomic DNA was removed on-column. cRNA was synthesized from total RNA and Cy3/Cy5 labeled using the Agilent Low Input Quick Amp Labeling procedure, as per the manufacturer's instructions. Labeled cRNA was hybridized to the Rhesus Macaque (V2) Gene Expression 4×44K Microarray (Agilent) in the appropriate combinations, including dye-flip and self-versus-self controls. After scanning the arrays, data was background subtracted and the probe signal normalized both within the array and between the arrays, using the limma package in Bioconductor [75]. The limma averages probe intensity values and fits the values to a linear model, then computes statistics for the data. An adjusted p-value is also computed to correct for multiple testing. A stringent criterion of an adjusted p-value of ≤0.05, in combination with a minimum 4 fold change (log2 of 2 or greater) in gene expression between stimulated and unstimulated DN T cells, was used to identify significant differentially regulated genes. Hierarchical cluster analysis was performed using the hierarchical clustering function (Pearson algorithm) in the TM4 Microarray Software Suite [76], [77]. Differentially regulated genes were examined for functional categories using the Database for Annotation, Visualization and Integrated Discovery (DAVID) [78], [79] to elucidate the differential responses between stimulated and unstimulated samples. Network analysis was done using Ingenuity Pathway Analysis (Ingenuity Systems, Inc.). cDNA from purified double negative T cells were amplified with 23 primers specific for the Vβ region of the T cell receptor [42]. Primers were blasted against the Rhesus genome to ensure putative TCR match, we optimized primers 2, 3, 7A, 8, 11, 12B, 14, 17 and 20 by nucleotide substitutions (Primer sequence in Table S1) to ensure complementarity. A single FAM labeled C region (TCR-βC) primer was used as a 3′ primer for each Platinum Taq amplification (Invitrogen). The primers were multiplexed with 4–5 primers per set [42]. Analysis of amplified fragments was done on an ABI fragment analyzer at the Seattle Biomed Sequencing Core and data analyzed on the Peak Scanner. Peak ranges for each Vβ region of the sooty mangabey TCR was identified (Table S1) and expressed Vβ regions along with the extent of junctional diversity was tabulated. Graphs were generated using Prism (GraphPad). Data was analyzed identifying Gaussian distribution of differences between sets and using non-gaussian comparison where indicated. A Wilcoxon paired sum test was utilized for paired samples and Mann-Whitney U test for unpaired groups and a p<0.05 considered significant.

10.1371/journal.pcbi.1003792

VASP-E: Specificity Annotation with a Volumetric Analysis of Electrostatic Isopotentials

Algorithms for comparing protein structure are frequently used for function annotation. By searching for subtle similarities among very different proteins, these algorithms can identify remote homologs with similar biological functions. In contrast, few comparison algorithms focus on specificity annotation, where the identification of subtle differences among very similar proteins can assist in finding small structural variations that create differences in binding specificity. Few specificity annotation methods consider electrostatic fields, which play a critical role in molecular recognition. To fill this gap, this paper describes VASP-E (Volumetric Analysis of Surface Properties with Electrostatics), a novel volumetric comparison tool based on the electrostatic comparison of protein-ligand and protein-protein binding sites. VASP-E exploits the central observation that three dimensional solids can be used to fully represent and compare both electrostatic isopotentials and molecular surfaces. With this integrated representation, VASP-E is able to dissect the electrostatic environments of protein-ligand and protein-protein binding interfaces, identifying individual amino acids that have an electrostatic influence on binding specificity. VASP-E was used to examine a nonredundant subset of the serine and cysteine proteases as well as the barnase-barstar and Rap1a-raf complexes. Based on amino acids established by various experimental studies to have an electrostatic influence on binding specificity, VASP-E identified electrostatically influential amino acids with 100% precision and 83.3% recall. We also show that VASP-E can accurately classify closely related ligand binding cavities into groups with different binding preferences. These results suggest that VASP-E should prove a useful tool for the characterization of specific binding and the engineering of binding preferences in proteins.

Proteins, the ubiquitous worker molecules of the cell, are a diverse class of molecules that perform very specific tasks. Understanding how proteins achieve specificity is a critical step towards understanding biological systems and a key prerequisite for rationally engineering new proteins. To examine electrostatic influences on specificity in proteins, this paper presents VASP-E, a software tool that generates solid representations of the electrostatic potential fields that surround proteins. VASP-E compares solids with constructive solid geometry, a class of techniques developed first for modeling complex machine parts. We observed that solid representations could quantify the degree of charge complementarity in protein-protein interactions and identify key residues that strengthen or weaken them. VASP-E correctly identified amino acids with established experimental influences on protein-protein binding specificity. We also observed that solid representations of electrostatic fields could identify electrostatic conservations and variations that relate to similarities and differences in binding specificity between proteins and small molecules.

Software for comparing protein structures is widely used to make inferences about protein function. These methods assist in function annotation by revealing proteins that perform similar biological functions despite vast evolutionary differences. Many methods focus on the discovery of subtle structural similarities among very different molecules [1], [2] using the superposition of catalytic residues [3]–[5] or the comparison of binding cavities [6]–[9]. By aligning polypeptide backbones [10]–[14], distance matrices [15] or geometric graphs [16]–[18], related methods can reveal similarities in tertiary structure that are not evident from sequences alone. Most approaches use atom coordinates or molecular surfaces [19]–[22] as digital representations of protein geometry. Other characteristics, such as evolutionary significance [23]–[25], hydrophobicity [26] and electrostatic potential [5], [23], [27] are attached to this geometric representation as labels. Comparisons of these data often generate a score, such as the root mean squared distance (RMSD), that summarizes structural, biological, and chemical similarities among two or more structures. Proteins with very different sequences sometimes exhibit unusually similar RMSDs, revealing shared origins in antiquity [28]–[30]. An emerging second type of comparison algorithm is designed to find subtle differences among very similar proteins. These methods seek to annotate protein specificity by proposing structural causes for different binding preferences among proteins that perform the same function [31]–[36]. For example, specificity annotation software might identify a cleft inside the ligand binding cavity of one protein that does not exist in a close homolog. That cleft might accommodate substrates that the homolog cannot bind. To find structural features like these, RMSD, and other scores for function annotation, are inappropriate because they obscure meaningful individual variations, like the cleft, behind summary scores. Instead, volumetric representations [37], which describe protein structure based on spatial regions occupied by the atoms of a protein, can be used to identify individual structural differences that may alter the binding preferences of ligand binding cavities [31]–[34]. To date, few comparisons focused on finding subtle electrostatic differences among closely related proteins have been reported, even though electrostatic fields are widely used to infer protein function [38]–[47] and specificity [48]–[53]. To fill this gap, this paper proposes a novel volumetric representation and comparison algorithm for finding electrostatic influences on binding specificity. The problem we are specifically addressing is the case where several closely related proteins have already been structurally aligned and we seek to identify spatially conserved and varying regions in their potential fields that might cause differences in binding specificity. Conserved regions, where the fields have similar potentials, might stabilize a molecular fragment attracted by all proteins (Fig. 1g), while differences in specificity could arise from regions where the fields vary (Fig. 1h,i). Software for identifying conservation and variation in charged regions can thus suggest how such regions may play a role in molecular recognition, and how they might be changed to achieve different binding preferences. Our approach identifies regions like these by representing electrostatic isopotentials with volumetric solids generated by the new program VASP-E (Volumetric Analysis of Surface Properties with Electrostatics). VASP-E computes conserved and varying regions using techniques from Constructive Solid Geometry (CSG) (Fig. 1). Developed originally for computer aided design [54] and computer graphics [55], CSG enables unions, intersections, and differences of volumetric representations to be calculated as if they are three dimensional solids. When used to analyze fields, CSG intersections can approximate regions that are common to isopotentials from several aligned proteins, thereby identifying regions with conserved potentials. CSG differences identify regions inside the isopotential of one protein but not inside that of another, isolating a region where potentials vary. Together, CSG operations provide a novel mechanistic approach to the analysis of electrostatic fields because the approximation of conserved and varying fields is not possible with existing structure comparison methods. The solid representations employed by VASP-E differ in kind from existing electrostatic analyses. While VASP-E deconstructs the electrostatic field to identify conserved and varying electrostatic phenomena, existing methods summarize and quantify the field with comparison scores [56], [57] and biophysical energies [58]–[61]. These numerical values cannot point to specific regions in the field with electrostatic similarities or differences, and so they cannot suggest how a protein might be altered to engineer different binding preferences. A second fundamental difference is that solid representations have the additional capability to represent the region inside molecular surfaces. Using CSG, we can therefore integrate both types of data to focus on electrostatic fields within binding sites. For example, the CSG difference of an isopotential minus the molecular surface at a binding site represents a three dimensional charged region in the solvent that can be occupied by potential binding partners (e.g. Fig. 1c,d). In contrast, representations used in function annotation methods generally represent electrostatic fields at or near the molecular surface only. Sampling a three dimensional field along this curving two dimensional surface cannot describe the electrostatic field as it extends outwards from the protein and influences other molecules. Third, while existing methods characterize fields at all potentials, solid representations describe fields at selected isopotential thresholds only. This feature enables comparisons to focus on ranges of potential that are especially relevant to binding. It can also be used to measure electrostatic complementarity between binding partners, as we will demonstrate later, by identifying interface regions where oppositely charged isopotentials overlap. To our knowledge, VASP-E is the first application of CSG to the volumetric comparison of electrostatic isopotentials, although tree-based methods that summarize topological differences in electrostatic isopotentials [57] have also been developed. This paper explores two applications of VASP-E as it might be applied in support of research in structural biology. One objective in many investigations is to discover electrostatic influences on protein-ligand or protein-protein binding specificity. Given the long range nature of electrostatic interactions, many amino acids could potentially be influential, and it could be impractical to create all possible mutants and determine their binding preferences. Here, a first application of VASP-E is to suggest amino acids that create differences between the electrostatic fields of two ligand binding cavities or to suggest amino acids that enhance or diminish electrostatic complementarity between two interacting proteins. Because amino acids are suggested in tandem with a hypothetical electrostatic influence on binding, VASP-E provides reasons to produce and test certain mutants first, where no reason might have existed before. The second application of VASP-E examined in this paper is the classification of protein-ligand binding cavities based on their electrostatic fields. This application can support efforts to discover patterns of electrostatic similarities or differences among related binding sites. In studies seeking to identify a possible ligand, electrostatic classification can reveal similarities to other proteins that may have known binding partners. Together, these applications of VASP-E represent two of many capabilities that become possible by combining CSG and volumetric representations of electrostatic isopotentials. We validate these capabilities in the results section against established experimental observations. The underlying observation exploited by VASP-E is that geometric comparisons of electrostatic potential fields can focus on biologically relevant regions and specific potential ranges by using CSG. Constraining the comparison of potential fields in this manner ensures that comparisons reflect aspects of electrostatic fields that influence binding, rather than spurious variations that occur by random chance or outside of binding sites. To achieve this kind of focus, comparisons always begin with a multiple structure alignment of whole proteins [10]–[18], [62], where ligand binding cavities or protein-protein interfaces are defined on each structure using cavity detection algorithms [63]–[67] or manual design. Structures aligned in this manner are then used to generate solid representations of electrostatic isopotentials and protein structure. To represent electrostatic isopotentials, we first solve the potential field of a given structure using DelPhi [68]. Using the field, isopotential surfaces are approximated using Marching Cubes [69], an algorithm first applied to visualize electrostatic isopotentials in GRASP [70]. This method is paraphrased below. Solids representing molecular surfaces are generated using the Trollbase library [12], which implements the classical rolling-probe method [71]. The resulting solids, regardless of their origin, are basic inputs for CSG operations, which we described earlier [37]. Below, we use the symbols , and to denote intersection, union, and difference operations, which are the basic CSG operations used in this work. VASP-E uses CSG to integrate solid representations of electrostatic isopotentials and molecular surfaces to create solid representations of the electrostatic field within ligand binding cavities (cavity fields) and protein-protein interfaces (interface fields). These procedures are detailed below. Cavity fields and interface fields are the constrained representations used by VASP-E to focus the comparison of electrostatic fields on biologically significant regions. To quantify similarities, we compute the CSG intersection of two regions and then evaluate the volume of the resulting intersection region. To quantify differences, we measure the volume of the CSG difference. Large volumes of intersection imply similar fields while large differences are characteristic of fields that vary. To estimate the volume of any region , including outputs from CSG operations, we use the Surveyor's Formula [72], which we described earlier [37]. Further CSG operations permit deconstructive comparisons of cavity and interface fields that identify similarities in some regions and differences in other regions within the fields they describe. While many applications this kind are possible with VASP-E, we describe two below: First, we can use VASP-E to trace differences in electrostatic fields to individual amino acids that contribute to these differences, thereby predicting residues that influence binding specificity. Second, we can integrate multiple electrostatic similarity measurements between a family of cavity fields to reveal patterns of ligand binding specificity. As input, marching cubes begins with a molecular structure from the Protein Data Bank (PDB) [73], its electrostatic potential field , a desired isopotential threshold , and the user's choice of representing the region with potential greater than or less than . The overall procedure (Fig. 2) approximates the solid region on one side of the isopotential at , which we refer to as a solid isopotential (Fig. 2a). In this work, when generating electrostatic isopotentials at kT/e, we always represent the region with potential greater than when is positive, and the region with potential less than , when is negative. Regions on the other sides of these potentials are infinite in volume, and thus their comparison is not well defined. Below, we use a negative value for and represent the region on the lower-potential side of , as an example. First, we protonate the PDB structure using the reduce component of MolProbity [74]. The resulting structure is passed to DelPhi [68], which computes numerical solutions to the nonlinear Poisson-Boltzmann equation, yielding an approximation of at every point within a bounding box surrounding the protein. Using , Marching Cubes outputs a polyhedral approximation of the isopotential surface at k kT/e, which we interpret as the exterior boundary of a three dimensional solid. Marching Cubes begins by establishing a regular lattice of cubes around the protein, whose borders fall within the bounding box (Fig. 2b). The lattice as a whole can be interpreted as a collection of lattice points at the corners of each cube, lattice edges connecting adjacent corners, lattice faces between cubes, or as simply a collection of lattice cubes. The resolution of the lattice, defined by the length of a lattice edge, is specified by the user and can be changed to accommodate structures of different sizes in system memory. Once the lattice is initialized, we evaluate the potential of the field at every lattice point . If , we mark as being inside the isopotential. Otherwise, we mark as being outside (Fig. 2c). The evaluation of is made possible using the Trollbase library [12], which evaluates the field to determine the potential at . Next, we select every lattice edge that connects an inside lattice point to one outside. Since isopotentials are topologically closed surfaces, the selected edge must intersect the desired isopotential (Fig. 2d). On each selected edge, we estimate the intersection point between the segment and the isopotential using linear interpolation between the electrostatic potentials at the endpoints (Fig. 2e). Finally, we consider every lattice cube joined to at least one lattice edge with an intersection point. On the cube, the intersection points collectively approximate the places where the isopotential passes through the cube. In two dimensions, this can be drawn as a shape passing through a square (Fig. 2f1), and approximated with a line through the cube. In three dimensions, when the isopotential passes through the cube, it's boundary is a surface that intersects the segments at the intersection points calculated earlier. This surface can be approximated inside each cube with triangles connecting triplets of intersection points. We use a lookup table to specify each triangle layout based on the possible combinations of selected edges in a cube (Fig. 2f2). Assembling the triangles into a single surface generates the exterior boundary of the solid isopotential (Fig. 2g,h). A cavity field is a solid representation of the region inside a ligand binding cavity that is also inside a solid isopotential. To generate a cavity field, we require the solid isopotential and a solid representation of the ligand binding cavity (Fig. 3a,b). We generate the solid isopotential using the method above and represent the ligand binding cavity using VASP and a volumetric approach based on SCREEN [65], described earlier [37]. Computing a CSG intersection of these two regions generates the cavity field (Fig. 3c,d). We compare cavity fields to detect local electrostatic differences that might affect specificity. Our approach follows the assumption that the user has selected solid isopotentials at a threshold that is relevant for ligand binding. For example, if a negative potential is influential for the selection of positively charged substrates, comparing regions of negative potential in several cavities could reveal electrostatic causes for different binding preferences. We discuss the selection of these potentials in Supplemental Text S1. Our comparison begins by structurally aligning two proteins, and , and generating their cavities, and . Because and are regions that are outside the molecular surface, we say that they are solvent accessible regions. Using and , we generate their cavity fields, and at the same side of the electrostatic potential . Next, we generate the intersection . is the region that is solvent accessible in both cavities (Fig. 3e). By comparing electrostatic fields inside , we are guaranteed that our comparison is not influenced by steric differences. For this reason, we next compute the intersection and (Fig. 3f,g). and are regions within the solid isopotentials of and that are solvent accessible in both cavities. Computing and permits several useful comparisons. First, the intersection (Fig. 3h), is solvent accessible in both cavities and also inside both solid isopotentials. This region of structural and electrostatic similarity might stabilize molecular fragments that are common to substrates of both proteins. Second, the difference regions and (Fig. 3i,j) are solvent accessible in both cavities but different in electrostatic character, because they lie inside the solid isopotential of one cavity and not the other. Molecular fragments in this region may thus be accommodated by one protein and electrostatically destabilized in the other. We quantify differences by measuring the volume of and . If and are small, we say that the cavity fields are similar. If one or both volumes are large, we say that and are electrostatically dissimilar, and that (or ) is evidence supporting the hypothesis that could attract or stabilize a ligand that cannot. This computation enables a systematic categorization of all electrostatic differences in the binding cavities of and . An interface field is a solid representation of a region of electrostatic complementarity between two proteins and , in complex. Given a potential threshold , we define a region of electrostatic complementarity to be the spatial region where the field of , independent of , has potential greater than , and the field of , independent of , has potential less than . We refer to this region as an interface field. To generate an interface field, we require three inputs: A solid representation of the interface region, an electrostatically significant isopotential of alone at kT/e (), and an electrostatically significant isopotential of alone at kT/e (). Because we use interface fields to analyze the specificity of interacting proteins, and because VASP-E is not designed to determine how two proteins interact, unbound structures are not used for the generation of interface fields. To generate the interface region, we first identify amino acids at the interface (Fig. 4a). These are the amino acids of that have an atom within 5 Å of , and the amino acids of that have an atom within the same distance of . Next, we generate spheres with radius 5 Å, centered at every atom of these amino acids (Fig. 4b). Finally, we compute the interface region with the CSG union of these spheres (Fig. 4c). In identifying amino acids that are part of the interface region, we do not include or exclude amino acids based the fraction of their surface exposed to the solvent, because some influential “hot spot” amino acids may have low solvent exposure [75]. Electrostatically significant isopotentials and (Fig. 4d,e) are generated with the marching cubes method described above. Using and , we compute , the electrostatically significant region of the field of protein within the interface, and , the oppositely charged electrostatically significant region of the field of protein within the interface (Fig. 4f,g). The intersection of these two regions is the interface field, (Fig. 4h). Since the interface field represents electrostatic complementarity in a given complex, we can use interface fields to compare electrostatic complementarity in two complexes. For two complexes, and , and , the user's threshold for electrostatic significance, we generate four interface fields: , , , and . Comparing the interface fields at and yields a more complete representation of electrostatic complementarity in both complexes. We evaluate the difference between two complexes using the following expression: Where denotes the volume of a given region . The two interface fields for each complex express the degree of complementarity on the positive and negative parts of the electrostatic potential spectrum. The interface fields for the same complex are summed, to represent the total degree of complementarity on that complex. The difference between the two sums expresses the difference in complementarity between the two complexes, on both sides of the potential spectrum. Large absolute values of indicate large differences in electrostatic complementarity between the two complexes, while values close to zero point to similar degrees of complementarity. DelPhi [68] is able to solve the electrostatic field of a given protein structure while omitting the electrostatic contribution of a individual amino acid. This process, which we refer to as nullification, has the unique property of leaving the structure of the amino acid intact while eliminating its electrostatic contribution. Maintaining the structure of the protein is important in an electrostatic analysis because the nullified amino acid still displaces solvent, creating a region of low dielectric. That region can enhance the electrostatic potentials of amino acids that were not nullified because of an effect called electrostatic focusing. Electrostatic focusing is known to play a considerable role in function and specificity [48], [76], [77]. Below, we use nullification in different ways to suggest amino acids that may influence specificity in ligand binding cavities and protein-protein interfaces. Calibration of both nullification techniques is discussed in Supplemental Text S1. Cavity fields based on a given family of proteins were clustered based on the Jaccard distance .where and are cavity fields, and and are the volume of the CSG intersection and CSG union of and , respectively. By normalizing the volume of the intersection by the volume of the union, the Jaccard distance permits cavity fields to be compared without bias relating to total volume. Cavity fields that have a low Jaccard distance have greater volumetric similarity than cavity fields with higher Jaccard distances. Using the neighbor program from Phylip [79], we summarized the pattern of volumetric similarities and variations between all pairs with UPGMA clustering (unweighted pair group method with arithmetic mean). Members of a given family of proteins were also clustered based on amino acid sequence alignments and backbone structure alignments. ClustalW 2.0.7 was used to compute multiple sequence alignments. The resulting alignments were passed to the protpars tool from Phylip [79], to generate a maximum parsimony clustering of the protein sequences. Ska [80] was used to compute backbone structure alignments, which we used to generate a pairwise superposition of every structure onto a selected individual. The root mean squared distance (RMSD) between aligned alpha carbons was clustered via UPGMA, using the neighbor tool from Phylip [79]. Finally, Clustal Omega [81] was used to compute multiple sequence alignments and generate a neighbor joining tree. Because VASP-E is designed to identify electrostatic influences on specificity, we validate it using families of proteins for which the mechanisms that achieve specificity are well understood and fundamentally electrostatic. The serine protease and cysteine protease superfamilies were selected for validating that VASP-E finds amino acids that influence protein-ligand binding specificity because many mutational studies confirm the role of specific residues in achieving specificity. The same studies permit the validation of VASP-E as a method for clustering proteins based on ligand binding preferences. The protein data bank (PDB) [73] contains the structures of 681 serine proteases from the trypsin and chymotrypsin subfamilies, and 859 cysteine proteases from the cathepsin B, cathepsin L, and papain subfamilies. From each subfamily, we first removed mutants and functionally undocumented structures. Then we removed structures with greater than 90% sequence identity, creating a nonredundant subset of 12 serine proteases and 4 cysteine proteases. Filtering in this order maximized the number of diverse representative structures. Serine proteases averaged 51% sequence identity and cysteine proteases averaged 40% sequence identity. We used ska [80] to structurally align the serine proteases to bovine chymotrypsin (pdb: 8gch) and the cysteine proteases to papaya papain (pdb: 1pad). Chymotrypsin and papain were selected because they are in complex with a peptide substrate. Using a method described earlier [37], substrate residues in the S1 subsite of the serine proteases and the S2 subsite of the cysteine proteases were used to generate a solid representation of the binding cavity in all structures. The binding cavity representation and the electrostatic field in each structure was then used to create cavity fields with the method in Section 2.2. We demonstrate the comparison of interface fields on two protein complexes: barnase-barstar (pdb: 1brs) and rap1A-RAF (pdb: 1c1y). We selected these complexes because electrostatic potential is known to affect their binding preferences and because detailed experimental studies have established how binding preferences are affected by mutations on both sides of the interface. These studies create a well-defined gold standard for evaluating how accurately VASP-E can predict amino acids that alter binding preferences. The data set is summarized in Fig. 5. VASP-E was developed in ansi C/C++ using gcc (the Gnu Compiler Collection) version 4.4.7, on 64 bit linux-based computing platforms. Experimentation was performed on Corona, a cluster at Lehigh University with 1056 Opteron cores (model 6128) running at 2.0 Ghz. Each compute node on corona had 16 cores with access to either 2 or 4 GB of random access memory (RAM) per core. VASP-E is a single-threaded process that runs on one core and approximately 1 GB of random access memory. All experimentation was conducted at .5 Å resolution, which permitted accurate results and practical runtimes. Visualization for some figures was performed with SURFview, a tool written using the OpenGL library and running on Intel Core i7 and Nvidia Geforce GTX 660 chipsets, in Microsoft Windows 7. Trees representing clusterings were visualized using Newick Utilities [82]. The performance of VASP-E depends on the volume and resolution of the molecular surfaces or electrostatic isopotentials analyzed. On our dataset, generating solid isopotentials for entire proteins required approximately 9.5 seconds on average, to process an average of 1,337,083 lattice cubes. Comparing cavity fields required 1.06 seconds on average, to process an average of 41,984 cubes via CSG, while interface fields from two complexes required 23.4 seconds on average, to process an average of 729,321 cubes. The website http://www.cse.lehigh.edu/~chen/software.htm hosts the software and primary data associated with this paper for public download. Serine proteases exhibit affinity for amino acids at specificity subsites called S4, S3, …, S1, S1′, …, S3′, S4′ [83]. Each subsite recognizes substrate residues P4, P3, …, P1, P1′, …, P3′, P4′, enabling the protease to selectively cleave the peptide bond between P1 and P1′. Trypsins are digestive serine proteases that narrowly prefer positively charged amino acids [84] at . Their selectivity is assisted by the strongly negative electrostatic character of S1. In contrast, chymotrypsins hydrolyze peptide bonds following large hydrophobic amino acids [85] and exhibit considerably less electrostatic potential at their S1 subsite. Using VASP-E, we identified amino acids that create electrostatic differences between trypsins and chymotrypsins at S1. Fig. 6 reports the average volumetric difference between cavity fields from all trypsins in our dataset and the cavity field of bovine chymotrypsin (pdb: 8gch), where each trypsin residue has been nullified individually. Volume differences were computed for cavity fields generated at −2.5, −5.0, −7.5, and −10.0 kT/e. Volumetric differences between nullified trypsin and chymotrypsin cavity fields varied most at −10 kT/e, so a prediction threshold was computed for differences at this level. The average volumetric difference between trypsin and chymotrypsin cavity fields remained nearly constant for almost all residue nullifications and all four thresholds. Nullifying almost all trypsin residues does not make the very different electrostatic environments of the trypsin and chymotrypsin S1 pockets more similar. One notable exception stands out. Nullifying aspartate 189 in all trypsins results in a large reduction in the average electrostatic difference with chymotrypsin at all potential thresholds, suggesting that the presence of aspartate 189 makes their S1 pockets electrostatically different. Fig. 7 illustrates the effect that nullifying aspartate 189 has on the electrostatic difference between chymotrypsin and trypsin, using bovine chymotrypsin and atlantic salmon trypsin as examples. VASP-E examines only the volumetric intersection of their S1 cavities, where the binding cavities have no steric differences (Fig. 7b). In unmodified trypsin (Fig. 7d), the intersection region exhibits a 152 Å3 region with electrostatic potential less than or equal to −10 kT/e. Once D189 is nullified, the region with potential less than or equal to −10 kT/e drops to 32 Å3 (Fig. 7e). In comparison, regions of negative electrostatic potential in chymotrypsin, where the S1 cavity overlaps with that of trypsin, is small and remains small when S189 is nullified (Fig. 7f,g). Similar effects were observed with other trypsins. These indications predict experimentally established observations that the negatively charged aspartate 189, at the bottom of the S1 pocket, creates the specificity of trypsin for positively charged amino acids [48], [86]. Fig. 8 illustrates a UPGMA clustering of cavity fields from trypsin and chymotrypsin S1 cavities, generated at −10 kT/e. The topology of the tree, which reflects electrostatic similarities and differences measured with the Jaccard distance, correctly separated the chymotrypsins as outliers from the trypsins. This result indicates that the electrostatic characteristics measured by VASP-E correlate with similarities and differences in serine protease binding preferences. Clusterings based on cavity fields generated at −2.5, −5.0, −7.5, or −10.0 kT/e (Fig. S1) illustrate that the classification is correct for a range of isopotential thresholds, though they becomes less accurate as thresholds approach neutral charges. Also, relative to comparisons of protein sequences and structures, separated trypsins and chymotrypsins less well (Fig. S2). Cathepsin B is involved in the onset of pancreatitis [87] and the malignant progress of tumors [88]. Following the same subsite/substrate numbering scheme as serine proteases, Cathepsin B cleaves a peptide bond following two positively charged amino acids that bind in its S1 and S2 subsites [89]. The S2 subsite exhibits a strong negative potential that enables the recognition of positively charged side chains. In contrast, cathepsin L and papain prefer bulky hydrophobic amino acids at [90], [91], and both exhibit an uncharged S2 subsite [91]. We used VASP-E to identify amino acids that create electrostatic differences between cathepsin B and cathepsin L. Fig. 9 illustrates the average volumetric difference between cavity fields representing S2 in cathepsin B and human cathepsin L (pdb: 1icf) generated at −2.5, −5.0, −7.5, and −10.0 kT/e. Volumetric differences between cavity fields with different nullified amino acids were greatest at −2.5 kT/e, so a prediction threshold was computed for differences at this level. The average volumetric difference was nearly constant for almost all residue nullifications. Insignificant fluctuations in the volumetric difference were approximately the same magnitude as in serine proteases. The nullification of two amino acids, glutamic acid 171 or glutamic acid 245, reduced electrostatic differences between cathepsin B and cathepsin L beyond the prediction threshold. This observation suggests that both amino acids create electrostatic differences between the S2 subsites of cathepsin B and L. Indeed, glutamic acid 245 has been shown to cause Cathepsin B to bind arginine residues at the S2 cavity [90], while cathepsin L prefers phenylalanines. In glutamic acid 171, one of the carboxylate oxygens is involved in a hydrogen bond and the other is free to form other interactions in the S2 pocket. Such interactions have been observed with positively charged inhibitors [92], [93], again in contrast with cathepsin L. Fig. 10 plots a UPGMA clustering of cavity fields based on S2 subsites from cysteine proteases in our data set. Cavity fields were generated at −2.5 kT/e. The topology of the tree describes electrostatic similarities and differences measured with the Jaccard distance. It is apparent that the tree structure clusters cathepsin B cavities, setting them apart from those of cathepsin L and papain, which have different binding preferences. Cavity fields produced at −2.5, −5.0, −7.5, and −10.0 kT/e, shown in Fig. S3, cluster in a similar manner. This pattern of separations demonstrates that VASP-E is correctly identifying electrostatic conservations and variations that correlate to binding preferences in the S2 subsite. Global sequence and structure alignments separated the cysteine proteases as well as the Jaccard distance clustering (Fig. S4). Barnase is an guanine-preferring endo-ribonuclease expressed by Bacillus amyloliquefaciens [94] whose activity, without inhibition by barstar, can be lethal to the cell [95]. Barstar inhibits barnase by forming an extremely tight complex with close steric and electrostatic complementarity at many amino acids across the binding site [96]. We used VASP-E to identify mutations that enhance or diminish electrostatic complementarity. Ras is a master regulator that transmits a wide range of signals via protein-protein interactions. Downstream, its effectors are involved in many crucial systems, including cell cycle progression, cell division, apoptosis, lipid metabolism, DNA synthesis, and cytoskeletal organization [102]–[104]. While the structure of ras in complex with these effectors is unknown, rap1a, a homolog of ras (50% sequence identity), can serve as a substitute. Rap1a has an essentially identical binding interface and binds competitively with the same downstream effectors [105], such as raf, an oncogene involved in ERK 1/2 signaling [106]. Here, we use VASP-E to examine the effect of charge nullification on the rap1a-raf interface to make predictions on the effect of mutation on ras-raf binding. By collecting the predictions made on our dataset, we can measure the prediction performance of VASP-E. We begin by counting true positives (TPs), false positives (FPs), true negatives (TNs), and false negatives (FNs). TPs are defined as amino acids that are both predicted by VASP-E to have an influence on specificity and also published in experimental findings to have such an influence. The predictions detailed earlier in this section cite these findings as specific validation for the predictions made with VASP-E. FPs are amino acids that are both predicted by VASP-E to have an influence on specificity and are documented in the literature to not have an effect on specificity. TNs are amino acids that are predicted to not have an influence on specificity that are also documented in the literature to not have an effect on specificity. FNs are amino acids predicted to not have an influence on specificity but are established in the literature as having a role in specificity. Finally, we VASP-E made two predictions that were neither confirmed nor denied in the literature. We leave these two observations as open predictions and do not include them in our evaluation of prediction performance. Of these statistics, TNs cannot be fully counted because no studies categorically classify the role of every amino acid in specificity, including those that are distant from the binding site. For this reason, we describe the number of true negatives as unknown. Nonetheless, we do not require TNs in order to compute precision and recall, two fundamental statistics used to evaluate the accuracy of a predictor. Precision is the fraction of predictions that are verified in experimental studies and recall is the fraction of verified experimental results that are correctly predicted. Using our conservatively defined prediction thresholds, every prediction made with VASP-E was verified, giving perfect precision, and most verified results were correctly predicted, giving strong recall. Precision and recall are reported together in Fig. 16. We have presented VASP-E, a new program for the comparison of electrostatic isopotentials. To our knowledge, VASP-E is the first program capable of comparing isopotentials using CSG, enabling a new unified approach to the characterization of protein-ligand and protein-protein binding specificity. In an application to the serine and cysteine proteases, we demonstrate that VASP-E is capable of reproducing known ligand binding preferences and of detecting differences in electrostatic potential among proteins that, based on global sequence and structure similarity, might have been expected to be similar. Subtle differences like these, which can arise from variations in single amino acids, can still be detected by VASP-E because they are reflected in differently shaped isopotentials. Central to our approach is a novel solid representation of electrostatic isopotentials that can also represent regions within molecular surfaces. This seamless integration of two nearly orthogonal aspects of protein structure enables analytical capabilities that were not possible before. One capability is the identification of amino acids that create differences in electrostatic isopotentials at binding cavities. Using the molecular surface to exclude electrostatic variations outside the binding cavity, we identified three amino acids in trypsin and cathepsin B that create electrostatic differences in binding specificity. These predictions correctly reflected experimentally established observations regarding their electrostatic influence. VASP-E also finds amino acids that change electrostatic complementarity in protein-protein interfaces. In an analysis of the barnase-barstar and rap1a/raf complexes, VASP-E predicted 22 amino acids that either increase or decrease affinity upon mutation, all in agreement with established experimental results. Solid representations enable a deconstructive analysis of electrostatic fields that permits the discovery of individual residues that influence binding preferences in protein-ligand and protein-protein binding sites. As the first approach to the comparison of electrostatic isopotentials with CSG, VASP-E exhibits novel potential for useful experimental applications. In experimental settings, identifying mutants that may alter binding specificity can be a time consuming and expensive effort with many possible mutants to consider. VASP-E identifies amino acids that might play a role in specificity, and, in addition, it suggests a biophysical mechanism for that amino acid: It may increase or decrease electrostatic complementarity. This additional information, beyond simply identifying an important amino acid, provides utility beyond the identification of important amino acids because it suggests how that amino acid might be tested, such as by mutation to an uncharged or oppositely charged residue. When comparing protein-ligand binding cavities, pointing out amino acids that create electrostatic differences can inform experimental design. VASP-E has the potential to serve broad applications. For example, identifying groups of amino acids that work together to achieve specificity can be an especially difficult problem, because of the combinatorial space of variants that must be considered. Nullification, as applied to individual amino acids in this paper, could be exhaustively applied to many combinations of residues to assist in experimental design. Given the rapid performance of VASP-E and the expanding availability of parallel computing, examining combinations of influential amino acids would also be very practical. Furthermore, the analysis of influential amino acids at protein-protein interfaces is not limited to dimers; the approach described here could be logically extended to higher order interactions. For such applications, interfaces between specific chains could be considered individually or in groups, to reflect the order in which the complex associates. Finally, while VASP-E is designed to identify subtle variations among highly similar proteins, VASP-E could in principle be used to analyze electrostatic similarities and differences among binding sites from very different proteins, as long as structural alignments could be correctly generated and binding cavities can be properly defined. These diverse applications suggest that the integrated representation and comparison of structure and electrostatics may offer an important new tool in the study of drug resistance and algorithms for specificity annotation.

10.1371/journal.pbio.1001923

Inhibitor of the Tyrosine Phosphatase STEP Reverses Cognitive Deficits in a Mouse Model of Alzheimer's Disease

STEP (STriatal-Enriched protein tyrosine Phosphatase) is a neuron-specific phosphatase that regulates N-methyl-D-aspartate receptor (NMDAR) and α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor (AMPAR) trafficking, as well as ERK1/2, p38, Fyn, and Pyk2 activity. STEP is overactive in several neuropsychiatric and neurodegenerative disorders, including Alzheimer's disease (AD). The increase in STEP activity likely disrupts synaptic function and contributes to the cognitive deficits in AD. AD mice lacking STEP have restored levels of glutamate receptors on synaptosomal membranes and improved cognitive function, results that suggest STEP as a novel therapeutic target for AD. Here we describe the first large-scale effort to identify and characterize small-molecule STEP inhibitors. We identified the benzopentathiepin 8-(trifluoromethyl)-1,2,3,4,5-benzopentathiepin-6-amine hydrochloride (known as TC-2153) as an inhibitor of STEP with an IC50 of 24.6 nM. TC-2153 represents a novel class of PTP inhibitors based upon a cyclic polysulfide pharmacophore that forms a reversible covalent bond with the catalytic cysteine in STEP. In cell-based secondary assays, TC-2153 increased tyrosine phosphorylation of STEP substrates ERK1/2, Pyk2, and GluN2B, and exhibited no toxicity in cortical cultures. Validation and specificity experiments performed in wild-type (WT) and STEP knockout (KO) cortical cells and in vivo in WT and STEP KO mice suggest specificity of inhibitors towards STEP compared to highly homologous tyrosine phosphatases. Furthermore, TC-2153 improved cognitive function in several cognitive tasks in 6- and 12-mo-old triple transgenic AD (3xTg-AD) mice, with no change in beta amyloid and phospho-tau levels.

A series of recent studies have found that the levels of the enzyme striatal-enriched protein tyrosine phosphatase (STEP) are raised in several different neuropsychiatric and neurodegenerative disorders, including Alzheimer's disease, fragile X syndrome, and schizophrenia. STEP normally opposes the development of synaptic strengthening, and these abnormally high levels of active STEP disrupt synaptic function by removing phosphate groups from a number of proteins, including several glutamate receptors and kinases. Dephosphorylation results in internalization of the glutamate receptors and inactivation of the kinases—events that disrupt the consolidation of memories. Here we identify the benzopentathiepin 8-(trifluoromethyl)-1,2,3,4,5-benzopentathiepin-6-amine hydrochloride (known as TC-2153) as a novel inhibitor of STEP. We show that the mechanism of action involves the formation of a reversible covalent bond between the inhibitor and the catalytic cysteine residue of STEP, and we demonstrate the activity of TC-2153 both in vitro and in vivo. TC-2153 shows specificity towards STEP compared to several other tyrosine phosphatases and shows no toxicity to cultured neurons. Importantly, the compound reversed cognitive deficits in a mouse model of Alzheimer's disease in a way that did not involve changes in the usual pathological signs (p-tau and beta-amyloid).

STriatal-Enriched protein tyrosine Phosphatase (STEP) (PTPN5) is a brain-enriched protein tyrosine phosphatase (PTP) targeted in part to postsynaptic terminals of excitatory glutamatergic synapses [1]–[4]. Recent studies indicate that STEP is overactive in Alzheimer's disease (AD), schizophrenia, and fragile X syndrome (FXS) [5]–[9]. The emergent model based on these findings suggests that the increase in STEP activity interferes with synaptic strengthening and contributes to the characteristic cognitive and behavioral deficits present in these disorders. Elevated levels of STEP activity disrupt synaptic function by dephosphorylation of STEP substrates [10]. These include mitogen-activated protein kinase (MAPK) family members ERK1/2 and p38 [11],[12], the tyrosine kinases Fyn and Pyk2 [13],[14], the glutamate receptor GluN2B subunit of NMDARs (formerly termed NR2B) [6],[15],[16], and the GluA2 subunit of AMPAR (formerly termed GluR2) [16]–[18]. STEP dephosphorylates the kinases at regulatory tyrosine residues within their activation loop and thereby inactivates them. Dephosphorylation of GluN2B promotes internalization of GluN1/GluN2B receptors, whereas dephosphorylation of GluA2 promotes internalization of GluA1/GluA2 receptors. To test the hypothesis that the observed overexpression of STEP disrupts synaptic strengthening in AD, we crossed STEP KO mice with the 3xTg-AD and Tg2576 AD mouse models. Six-month-old progeny null for STEP displayed significant decreases in biochemical and cognitive deficits, despite continued elevated levels of Aβ [16],[19]. These data validated STEP as a target for drug discovery. Herein we describe the search for small-molecule STEP inhibitors. We performed a high throughput screen that culminated in the identification of the benzopentathiepin 8-(trifluoromethyl)-1,2,3,4,5-benzopentathiepin-6-amine hydrochloride (known as TC-2153) as a novel STEP inhibitor. TC-2153 exhibited specificity for STEP in vitro, in cell-based assays, and in vivo, and also reversed cognitive deficits in 6- and 12-mo-old 3xTg-AD mice. We initially screened ∼150,000 compounds from the Laboratory for Drug Discovery in Neurodegeneration library using para-nitrophenyl phosphate (pNPP) as the target substrate (see Text S1 for more information on assay development and secondary screens). Eight compounds were selected for further characterization based on chemical structure and IC50 values, which ranged between 1 µM and 9.7 µM (Table S1), and studies of these molecules indicated potent inhibition of STEP activity in neuronal cultures and cortical tissue after intraperitoneal (i.p.) injections in WT mice. However, following resynthesis of several of the lead compounds, we found that they all exhibited essentially no inhibitory activity towards STEP (Figure S1). We therefore tested the possibility that a “contaminant” in the commercial preparations of the lead compounds was inhibiting STEP activity. To address this issue, we performed preparative HPLC on Compound 3 and tested eluted fractions for activity against STEP in the pNPP assay (Figure 1A). Compound 3 appeared as a major peak (fraction 7) on the reverse-phase HPLC preparation and had no inhibitory activity against STEP compared to a second peak that appeared as a late minor peak (fraction 32) that was a potent inhibitor of STEP. Given the high apparent lipophilicity of the unknown, the supplied material was extracted with hexane and recrystallized from methanol. Small pale yellow needle-shaped crystals (0.5–1 cm in length) were obtained in approximately 1% yield. The isolated crystalline material displayed the same HPLC retention, UV absorbance, and STEP inhibitory properties as the initially collected late-eluting peak. The crystalline compound was characterized by X-ray crystallography and found to be sulfur (S8) (Figure 1B). S8 is poorly soluble in aqueous solution and cannot easily be modified to improve physicochemical properties, redox activity, binding affinity, and selectivity. We therefore sought to identify more conventional inhibitor structures that would improve solubility and enable further refinement through analog preparation and evaluation. We identified the benzopentathiepin core structure present in a number of natural products as the most promising for further investigation (Figure 1B). Natural products incorporating the benzopentathiepin core motif have been reported to have antifungal and antibacterial activity in cell culture as well as cytotoxicity against human cancer cell lines [19],[20]. Moreover, amino-substituted derivatives such as varacin and TC-2153 have reasonable solubility in aqueous solution [21],[22]. TC-2153 reportedly has a low level of acute toxicity (LD50>1,000 mg/kg) and was proposed to cross the blood brain barrier as evidenced by anxiolytic and anticonvulsant effects in mice [23]. We therefore chose to evaluate the STEP inhibitory activity of TC-2153. We first compared the inhibitory activities of S8 and TC-2153 against recombinant STEP using pNPP assays at several concentrations of the inhibitors. Both S8 and TC-2153 inhibited STEP potently, with IC50s of 17.2±0.4 nM and 24.6±0.8 nM, respectively (Figure 1C–D). We treated cortical neurons for 1 h with S8 or TC-2153 and determined the Tyr phosphorylation of residues that STEP dephosphorylates on GluN2B (Y1472), Pyk2 (Y402), and ERK1/2 (Y204/187). For S8, there was a significant increase in the Tyr phosphorylation of all three STEP substrates at doses above 0.05 µM, with 1 µM showing maximum inhibition (Figure 2A and Figure S2A for representative blots) (1 µM dose, pGluN2B, 1.33±0.08, p<0.05; pPyk2, 1.49±0.12, p<0.05; pERK1/2, 1.67±0.14, p<0.01). For TC-2153, there was also a significant increase in the Tyr phosphorylation at these sites (Figure 2B and Figure S2B for representative blots) (1 µM dose, pGluN2B, 2.07±0.15, p<0.001; pPyk2, 1.81±0.21, p<0.001; pERK1/2, 2.39±0.18, p<0.001). The decrease in Tyr phosphorylation in the presence of the highest dose of TC-2153 (10 µM) may be due to off-target effects on positive regulatory PTPs. We found similar inverted-U dose–response curves on Tyr phosphorylation of direct PTP targets in previous work with PTP inhibitors [24],[25]. We next tested whether S8 and TC-2153 inhibited STEP activity in WT mice in vivo. Six-month-old male mice (C57BL/6) were injected with vehicle or S8 (0.5, 1, 3 mg/kg, i.p.) and cortices were removed and processed 3 h postinjection. S8 led to a significant increase in the Tyr phosphorylation of GluN2B, Pyk2, and ERK1/2 (at 1 mg/kg, pGluN2B, 1.31±0.11, p<0.05; pPyk2, 1.46±0.14, p<0.05; pERK1/2, 1.57±0.13, p<0.05) (Figure 2C and Figure S2C for representative blots). Similar results were obtained with TC-2153 (1, 3, 6, 10 mg/kg) (at 10 mg/kg, pGluN2B, 1.66±0.28, p<0.01; pPyk2, 1.80±0.30, p<0.05; pERK1/2, 2.52±0.16, p<0.01) (Figure 2D and Figure S2D for representative blots). Together, these results demonstrate that both S8 and TC-2153 increase the Tyr phosphorylation of three STEP substrates in intact neurons in culture and in vivo in the cortex of WT mice. In an attempt to evaluate possible selectivity of TC-2153, we performed activity assays using the catalytic domain of STEP, and the catalytic domains of two highly related PTPs: He-PTP and PTP-SL. TC-2153 showed no apparent selectivity among these PTPs. Several studies have shown that regions outside of the catalytic domain contribute to the susceptibility of PTPs to the action of selective inhibitors [26],[27]. Thus, we repeated our assays using several full-length PTPs (Table 1). Indeed, we found TC-2153 was more potent against the two major isoforms of STEP, STEP61 (IC50 = 93.3±1.1 nM) and STEP46 (IC50 = 57.3±1.1 nM), compared to HePTP (IC50 = 363.5±1.2 nM) and PTP-SL (IC50 = 220.6±1.3 nM). It displayed even greater selectivity over PTP1B (IC50 = 723.9±1.2 nM) and SHP-2 (IC50 = 6896.0±1.2 nM). These results suggest TC-2153 shows a degree of selectivity toward full-length STEP in in vitro assays. To further address possible off-target inhibition by TC-2153 in cells, cortical cultures from either WT or STEP KO mice were treated with TC-2153. Similar to the rat neuronal cultures, we observed an increase in the Tyr phosphorylation of STEP substrates in WT mouse cortical neurons (Figure 3A, black bars). Consistent with previous findings [12],[16],[18],[28], STEP substrates have higher basal Tyr phosphorylation levels in STEP KO cultures. TC-2153 failed to increase the phosphorylation of STEP substrates in the KO cultures (Figure 3A, grey bar with 0.1 µM and 1 µM; see Figure S3A for representative blots). To exclude a possible ceiling effect, we added a generic tyrosine phosphatase inhibitor, sodium orthovanadate (Na3VO4), which further increased the Tyr phosphorylation of these substrates. These results suggest that TC-2153 is relatively specific towards STEP compared to the generic tyrosine phosphatase inhibitor sodium orthovanadate. There are three highly related PTPs (STEP, HePTP, and PTP-SL) that all dephosphorylate ERK1/2. Only STEP is found in cortex, whereas HePTP is present in spleen, and PTP-SL is present in cerebellum, both tissues that lack STEP. In addition, ERK1/2 and Pyk2 are dephosphorylated by other tyrosine phosphatases outside of the CNS. We examined the specificity of TC-2153 by injecting WT and STEP KO mice with TC-2153 or vehicle, and determined the Tyr phosphorylation of ERK1/2 (Y204/187) and Pyk2 (Y402) in different organs (Figure 3B–G and Figure S3B for representative blots). There was a significant increase in pERK1/2 and pPyk2 in the frontal cortex and hippocampus, but not in the cerebellum or in all tissues tested outside the brain. These results suggest that TC-2153 does not target homologous PTPs known to dephosphorylate ERK1/2 and Pyk2 when tested in vivo. We also performed toxicity studies with TC-2153 in cortical cultures (Figure 4). We measured the release of lactate dehydrogenase (LDH) from the cultures for up to 48 h at various TC-2153 concentrations. Even at the high dose of 100 µM, TC-2153 had no significant effect on LDH release compared to the positive controls. We next examined the mechanism by which TC-2153 inhibited STEP. Because the catalytic cysteine in PTPs is prone to sulfhydration, nitrosylation, and oxidative modifications that cause inhibition of phosphatase activity [27]–[32], we first examined the effect of a reducing agent on STEP inhibition by TC-2153. The addition of reduced glutathione (GSH, 1 mM) decreased the inhibitory activity of TC-2153 by two orders of magnitude in in vitro assays (IC50 = 8.79±0.43 µM compared to 24.6±0.8 nM) (Figure 5A). These results suggested an oxidative mechanism for the inhibition of STEP. We established that TC-2153 was stable and did not degenerate in the assay conditions by sensitive 19F NMR monitoring (Figure S4) and was not acting through generation of reactive oxygen species (ROS), which was tested by the addition of catalase or superoxide dismutase to the in vitro assay (Table S2). To confirm that ROS are also not released in cortical cultures with TC-2153 treatment, we performed H2O2 colorimetric assay and fluorescence assay with 2,7-dichlorofluorescein diacetate (DCF) and did not observe any significant differences in H2O2 or ROS levels between the TC-2153 treated compared to nontreated control groups (Figure S5). To evaluate the mode of inhibition, we incubated STEP with TC-2153, subjected the sample to dialysis to remove excess inhibitor, and monitored enzyme activity (Figure 5B). After 24 h of dialysis, STEP remained inhibited, suggesting that TC-2153 acts as an irreversible inhibitor under the conditions used. Using the progress curve method [29], inhibition was also found to be irreversible and the second order rate of inactivation was determined (Figure 5C). A kobs was determined for pNPP in the presence of varying initial inhibitor concentrations (n≥4). Values were then analyzed with nonlinear regression to obtain the kinetic constants: kinact = 0.0176±0.0007 s−1; Ki = 115±10 nM; kinact/Ki = 153,000±15,000 M−1s−1. However, STEP activity could be recovered following incubation with GSH or DTT (Figure 5D). Aliquots of STEP were incubated with DMSO control or TC-2153 and were then added to assay buffer containing 1 mM GSH, 1 mM DTT, or water control and allowed to incubate for up to 1 h prior to testing for enzymatic activity. STEP activity was rapidly recovered by both reductants, with DTT showing a greater recovery of activity (75% recovery after 1 h, where DMSO control represents 100% activity) compared to GSH (29% recovery after 1 h). We then performed LCMS analysis to determine the intact protein mass of STEP and STEP+TC-2153. Our intact protein analyses suggest a covalent adduct to STEP. Although we were able to obtain the accurate mass for STEP, we were unable to mass spectrally resolve the heterogeneous mixture of intact STEP+TC-2153 and its covalent adducts with sufficient accuracy to fully interpret the results. Therefore, we next used high-resolution tandem mass spectrometry to focus upon whether TC-2153 might modify the active site cysteine of STEP. For these experiments, we used WT STEP as well as a STEP mutant in which the catalytic cysteine was changed to serine. Greater than 90% of the primary amino acid sequences were identified by LC-MS/MS for WT STEP or for the STEP mutant, following in-gel tryptic digestion of STEP from nondenaturing (native) preparations. We initially analyzed the catalytic cysteine at position 472 of STEP in the absence of TC-2153 and found a disulfide bridge between Cys465 and Cys472 that presumably forms following tryptic digestion given the positions of Cys465 and Cys472 in the three-dimensional X-ray crystal structure of STEP [30]. This modification was not observed when the catalytic site cysteine (Cys472) was mutated to serine. Incubation of WT STEP with TC-2153 resulted in the presence of a de novo trisulfide within the Cys465/Cys472 bridge, which was not observed for WT STEP alone or when the catalytic site cysteine (Cys472) was mutated to serine (Figure 5E and Figure S6). The precursor monoisotopic mass of the trisulfide-containing peptide had a mass error of 4 ppm (∼0.011 Da) based on theoretical mass calculation, which is within the 5 ppm external mass calibration expected for MS/MS data collected by the linear ion trap instrument used. These results indicate that the active site cysteine is likely modified by TC-2153 and suggest that following tryptic digestion a sulfur from the benzopentathiepin core is retained, giving rise to the trisulfide identified by mass spectrometry. We next tested the efficacy of TC-2153 to reverse cognitive deficits in an AD mouse model. We first used the Y-maze to evaluate spatial working memory function in 3xTg-AD mice. AD or WT mice were injected with vehicle or TC-2153 (10 mg/kg, i.p.) 3 h prior to the test. Spontaneous alternations and total arm entries were calculated. There was no significant change in arm entries in TC-2153–treated mice, suggesting no drug-induced effect on general motor activity (Figure 6A). The main effect of genotype [F(1, 29) = 5.240, p<0.05], treatment [F(1, 29) = 5.895, p<0.05], and Genotype × Treatment interaction [F(1, 29) = 9.751, p<0.01] were significant for spontaneous alternations, with the AD mice making more incorrect choices (i.e., fewer alternations) (Figure 6B). Compared to their respective vehicle controls, TC-2153 increased percentage alternation in the AD mice (TC-treated, 71.13±4.58 versus Veh-treated, 58.94±2.46, p<0.05), but not in WT mice (TC-treated, 73.45±3.19; Veh-treated, 74.98±2.19, p>0.05). We next investigated whether TC-2153 improved performance in the novel object recognition (NOR) task. WT or 3xTg-AD mice had no significant differences in baseline locomotor activity as measured during the habituation phase. Mice received an acute injection of vehicle or TC-2153 (10 mg/kg) 3 h prior to the training phase. Twenty-four h later, mice were subjected to the test phase. Discrimination indexes (DIs) were compared for group differences in object memory. The main effect of genotype [F(1, 23) = 4.342, p<0.05], treatment [F(1, 23) = 5.895, p<0.01], and Genotype × Treatment interaction was significant [F(1, 23) = 4.362, p<0.05]. Post hoc analysis indicated that the DI in the AD-TC group was significantly higher than those of the AD-Veh group (TC-treated, 0.354±0.094 versus vehicle-treated, −0.259±0.104, p<0.001). In the WT groups, the DI in the TC-2153–treated mice did not differ from the Veh-treated mice (vehicle-treated, 0.166±0.057; TC-treated, 0.304±0.095, p>0.05) (Figure 6C). We then tested the effects of TC-2153 in the reference memory version of the Morris water maze (MWM). A three-way ANOVA analysis revealed a significant Genotype × Treatment × Training Day interaction (p<0.05). Daily injection of TC-2153 3 h prior to training reversed memory deficits in 3xTg-AD mice on days 5 and 6 of the acquisition phase (p<0.01) (Figure 6D). The longer escape latency of 3xTg-AD mice injected with vehicle was not attributed to slower swimming speed, as no significant differences were found between groups (p>0.05; two-way ANOVA) (Figure 6E). To confirm memory status, the number of entries in a circular zone located around the previous platform location (target zone) and in the opposite quadrants was evaluated during the probe trial 24 h after the last acquisition day. A three-way ANOVA analysis revealed a significant Genotype × Treatment × Quadrant interaction (p<0.004). The 3xTg-AD mice treated with TC-2153 spent as much time as WT mice in the target zone, whereas AD mice injected with vehicle showed no preference for the target zone (Figure 6F). All groups had similar escape latencies during the cued trial when the platform was visible, indicating the absence of sensorimotor or motivational deficits to escape from water (WT-Veh, 15.1±1.7 s; WT-TC, 15.6±1.7 s; AD-Veh, 15.3±3.0 s; AD-TC, 16.0±2.3 s; mean ± s.e.m.; p>0.05; two-way ANOVA). There were no differences in thigmotaxic swimming patterns between any of the tested groups (Figure S7). Taken together, these results demonstrate that TC-2153 significantly improved cognitive functioning in 6-mo-old 3xTg-AD mice. We next determined whether inhibition of STEP in 12-mo-old 3xTg-AD mice affected beta amyloid or phospho-tau levels. We first needed to confirm that TC-2153 was effective in attenuating cognitive deficits at 12 mo, as these mice have more robust increases in phospho-tau and Aβ levels. We tested the mice with the NOR task and once again found a significant improvement of memory in AD mice treated with TC-2153 during the choice phase (10 mg/kg, i.p.; TC-2153-AD, familiar versus novel, p<0.05). TC-2153 did not affect cognitive function in WT mice (Figure S8A). There were no significant changes in Aβ or phospho-tau levels after administration of TC-2153 (Figure S8B–C). STEP function is disrupted in several neurological disorders in addition to AD, including FXS [9], Parkinson's disease [31], and schizophrenia [8]. The increase in STEP expression in these illnesses is due to either an increase in its translation (in FXS) or a decrease in its degradation (in AD, Parkinson's disease, and schizophrenia). In contrast, STEP levels or activity are lower in several other disorders, including stress-related conditions [32],[33], excessive EtOH consumption [34], and cerebral ischemia [35]. Thus, the current model is that STEP activity must be within an optimal range and that either high or low levels of STEP disrupt synaptic plasticity. Disruption in STEP function has also been implicated in seizures [36], ethanol abuse [37], amphetamamine-induced stereotypies [38], and Huntington's disease [39],[40], although the basis for these changes remain to be determined. In terms of STEP dysfunction, most is known about its role in AD. Aβ binding to the α7 nicotinic receptor leads to calcium influx and calcineurin activation [6]. Calcineurin activates protein phosphatase 1 (PP1), which dephosphorylates a regulatory serine residue within the STEP substrate-binding domain, enabling STEP to interact with and dephosphorylate its substrates [41]. In addition, STEP is normally ubiquitinated and degraded by the proteasome after NMDAR stimulation [15], and Aβ inhibition of the proteasome [42],[43] results in a build-up of active STEP [7]. Based on these results, STEP was genetically lowered by crossing STEP KO mice with triple transgenic mice to produce progeny that still had the AD mutations, but were null for STEP [16]. These progeny had improved cognitive function and led to the current study to discover STEP inhibitors. We initially developed an HTS assay for STEP using the generic phosphatase substrate pNPP and screened ∼150,000 compounds to identify STEP inhibitors. However, when lead compounds were resynthesized, they had significantly lower activity against STEP, suggesting that an impurity in the commercial substances was likely inhibiting STEP activity. We isolated and identified this “contaminating” compound as elemental sulfur in the form of octasulfur (S8), which led in turn to the identification of the lead compound TC-2153. It should be noted that we did not include the reductant DTT in the initial library screen, in contrast to many screens for PTP inhibitors [44]. This allowed us to discover S8 and identify TC-2153 as a potent inhibitor of STEP and the mechanism of action by which TC-2153 inhibits STEP. The specificity of TC-2153 in in vitro, as well as cell and animal models, was explored. Interestingly, TC-2153 was more selective in in vitro assays against full-length STEP, but showed little specificity when tested against the truncated phosphatase domains of the PTPs. Although the exact mechanisms need to be clarified, these results are consistent with other recent findings with PTP inhibitors [45],[46]. TC-2153 treatment of neuronal cultures and WT mice increased the Tyr phosphorylation of STEP substrates. These results are consistent with previous studies of STEP KO mice that showed that a loss of STEP results in an increase in the Tyr phosphorylation of all STEP substrates identified to date [10],[12],[14],[28]. The Tyr phosphorylation of the three STEP substrates tested was not significantly changed by TC-2153 in STEP KO neuronal cultures, but was increased by the general PTP inhibitor sodium orthovanadate, indicating that a ceiling effect did not explain the results. Similarly, administration of TC-2153 to WT mice led to an increase in the Tyr phosphorylation of the STEP substrates ERK1/2 and Pyk2 in the hippocampus and frontal cortex (where STEP is present). However, treatment of WT mice with TC-2153 did not increase the Tyr phosphorylation of these two substrates in regions with no STEP expression (i.e., the cerebellum and organs outside of the CNS). An important finding was that treatment of STEP KO mice did not increase the Tyr phosphorylation of these substrates over baseline levels in the hippocampus and frontal cortex. The Tyr phosphorylation of ERK1/2 and Pyk2 did not alter in the cerebellum or in organs outside of the CNS (regions were STEP is not expressed). These results suggest a significant degree of in vivo specificity for inhibition of STEP by TC-2153, although additional studies are needed to expand on these initial findings. These results are consistent with an emerging body of research that suggests that oxidative regulation of the catalytic cysteine residue of PTPs is an important regulatory mechanism in vivo that links tyrosine phosphorylation signaling and the redox status of cells [47],[48]. PTPs contain a catalytic cysteine with an SH-group that exists in a thiolate state (S–) and facilitates removal of phosphate groups from substrates. The pKa values of these cysteine residues are in the range of 4–6, making these sites more likely to be oxidized compared to other cysteines that typically have pKa values in the range of 8–9 [49],[50]. Thus, a high reducing environment in cells, either through reduced production of reactive oxygen species or elevated activity of reactive oxygen scavengers, is proposed to decrease PTP oxidation and increase PTP activity. In contrast, a high oxidizing environment would increase PTP oxidation, reduce PTP activity, and increase tyrosine kinase signaling [49],[51],[52]. Although all PTPases are likely to modulate qualitatively in a similar fashion by cell redox status, several studies have shown that PTPs can be differentially inhibited by oxidation. For example, stimulation of T-cell receptors results in a selective oxidation of SHP2, but not SHP1, whereas both PTPs show similar sensitivity to oxidation in vitro. In contrast, SHP1 is more prone to oxidation than SHP2 after treatment of EOL-1 cells with H2O2 [26]. Whether STEP is regulated in cells by a similar balance of reactive oxygen species and oxygen scavengers, and whether this might explain the sensitivity of STEP to TC-2153, is under investigation. In the initial studies presented here, we explored the mechanism for STEP inhibition. Irreversible inhibition in the absence of reducing conditions was suggested for TC-2153, as dialysis did not restore activity. However, enzyme activity could be recovered by incubation with DTT or GSH, consistent with enzyme inactivation by oxidative modification of the active site cysteine. Moreover, the inhibitory activity of TC-2153 was considerably reduced by addition of DTT or GSH to the assay buffer. The oxidative mechanism for inactivation observed for TC-2153 is in agreement with established oxidative mechanisms for regulating PTP activity in vivo [53]–[55]. In support of this, mass spectroscopy analysis indicated that the catalytic Cys residue (Cys472) was modified by a trisulfide bridge that included Cys465, but only in the presence of TC-2153. In summary, we have discovered that the pentathiepin TC-2153 potently inhibits STEP activity. Although the results from the STEP KOs and the biochemistry are suggestive of a direct action of TC-2153 on the STEP active site, we cannot exclude an indirect mechanism in vivo. An important finding is that the new platform represented by TC-2153 reverses cognitive deficits in 6- and 12-mo-old 3xTg-AD mice, a reflection of the suggested role of STEP in the initial synaptopathology of AD [6],[16],[56]. The administration of TC-2351 did not affect the Aβ and tau brain pathology of the mice, although the question of the efficacy of TC-2153 at advanced stages of pathology remains open. Longitudinal studies with TC-2153 as well as administration of TC-2153 to other AD models will help address this question. Longitudinal studies will also address the long-term preventive effects of STEP inhibition on cognitive decline. Finally, it is important to determine whether TC-2153 is effective in other animal models of neuropsychiatric diseases in which STEP activity is elevated and these studies have begun. The Yale University Institutional Animal Care and Use Committee approved all proposed use of animals. All animal work was carried out in strict accordance with National Institutes of Health (NIH) Guidelines for the Care and Use of Laboratory Animals. pNPP, 2-(N-morpholino) ethanesulfonic acid (MES), sodium orthovanadate, ATP, and all buffer components were purchased from Sigma-Aldrich (St. Louis, MO). Malachite Green reagent kit was purchased from Bioassay system (Hayward, CA). 6,8-difluoro-4-methylumbelliferyl phosphate (DiFMUP) and EnzChek phosphatase assay kit were purchased from Invitrogen (Carlsbad, CA). The 96- and 384-well clear polypropylene plates were purchased from VWR (Radnor, PA), and 384-well white plates were purchased from Nalge Nunc International (Rochester, NY). Full-length STEP46 was used in the initial library screen. STEP46 cDNA was cloned into pGEX2T and transformed into BL21 (DE3) E. coli cells. STEP (20 mg) was purified on a glutathione sepharose column to immobilize the GST-tagged protein [14]. The column was loaded, washed, and bound protein eluted using Fast Protein Liquid Chromatography. For some of the biochemical experiments, we purified WT TAT-STEP46 and TAT-STEP46 (C to S) proteins, the latter containing a mutation at its catalytic cysteine within the active site that renders the enzyme inactive [15]. The assay development for the HTS is shown in Figure S9, biochemical characterization of STEP is in Figure S10, and the synthesis of TC-2153 (benzopentathiepin 8-(trifluoromethyl)-1,2,3,4,5-benzopentathiepin-6-amine hydrochloride) is described in detail in Figure S11. Primary cortical neurons were isolated from Sprague Dawley rat embryos (E18) (Charles River Laboratories, Wilmington, MA) as previously described [15]. In some experiments, cortical neurons were made from WT and STEP KO mouse embryos (E18). Neurons were allowed to grow for 18–21 d in CO2 incubator before addition of compounds at indicated doses for 1 h. Immediately following treatments, neurons were lysed in RadioImmuno Precipitation Assay (RIPA) buffer supplied with protease inhibitor cocktail (Roche Applied Science, Indianapolis, IN) and phosphatase inhibitors (NaF and Na3VO4). All experiments were replicated four times with four independent batches of cultures. Wild-type, male C57BL/6 mice (3–6 mo) were used for all studies. An initial dose–response curve was carried out using S8 (0.5, 1, and 3 mg/kg, i.p.) or TC-2153 (1, 3, 6, and 10 mg/kg, i.p.). Pilot studies were conducted to optimize the time after i.p. injection when STEP substrates showed maximum Tyr phosphorylation (1–3 h). Cortical tissues were dissected out 3 h postinjection and processed for subcellular fractionation. We homogenized brain tissue in buffer containing (in mM): 10 Tris-HCl, pH 7.6, 320 sucrose, 150 NaCl, 5 EDTA, 5 EGTA, 20 NaF, 1 Na3VO4, and protease inhibitors (TEVP). Homogenates were centrifuged at 800 × g to remove nuclei and large debris (P1). Synaptosomal fractions (P2) were prepared from S1 by centrifugation at 9,200 × g for 15 min. The P2 pellet was washed twice and was resuspended in TEVP buffer. In some experiments, mice were injected with S8 (1 mg/kg, i.p.) or TC-2153 (3 mg/kg, i.p.), and cortex, cerebellum, and spleen were removed to test for the in vivo inhibition of the highly related PTPs, HePTP, and PTP-SL [57]–[60]. Samples were prepared and resolved by SDS-PAGE, transferred to nitrocellulose membrane, and incubated with phospho-specific antibodies (anti-pY204/187 ERK1/2, anti-pY402 Pyk2, anti-pY1472 GluN2B) or total protein antibodies (anti-ERK2, anti-Pyk2, and anti-NR2B) overnight at 4°C. All antibodies used are listed in Table S3. Immunoreactivity was visualized using a Chemiluminescent substrate kit (Pierce Biotechnology, Rockford, IL) and detected using a G:BOX with the image program GeneSnap (Syngene, Cambridge, UK). All densitometric quantifications were performed using the Genetools program. Compound 3 was extracted with hexane and the residue obtained after rotary evaporation, then recrystallized from methanol. Small pale yellow needle-shaped crystals (0.5–1 cm) were obtained in approximately 1% yield. The isolated crystalline material displayed the same HPLC retention, UV absorbance, and STEP inhibitory properties as the initially collected late-eluting peak. The crystalline compound was characterized by the X-Ray Crystallographic Facility of the Yale University Department of Chemistry and found to be sulfur (S8). Reaction volumes of 100 µL were used in 96-well plates. We added 75 µL of water to each well, followed by 5 µL of 20× buffer (stock, 1 M imidazole HCl, pH 7.0, 1 M NaCl, 0.02% Triton-X 100). We added 5 µL of the appropriate inhibitor dilution in DMSO, followed by 5 µL of phosphatase (stock, 0.2 µM, 10 nM in assay). The assay plate was then incubated at 27°C for 10 min with shaking. The reaction was started by addition of 10 µL of 10× pNPP substrate (stock, 5 mM, 500 µM in assay), and reaction progress was immediately monitored at 405 nm at a temperature of 27°C. The initial rate data collected were used for determination of IC50 values. For IC50 determination, kinetic values were obtained directly from nonlinear regression of substrate–velocity curves in the presence of various concentrations of inhibitor using one site competition in GraphPad Prism v5.01 scientific graphing software. The Km value of pNPP in this system was determined to be 745 µM and was used in the kinetic analysis. For experiments with catalase or superoxide dismutase (SOD), 10 µL of the appropriate enzyme stocks (catalase, 800 U/mL stock, 80 U/mL in assay; SOD, 1,000 U/mL stock, 100 U/mL in assay) were added prior to addition of the inhibitor and STEP. For the experiments with glutathione reducing agent, 10 µL of glutathione (stock, 10 mM, 1 mM in assay) or water control was added before the inhibitor stocks, and only 65 µL of water was added initially to maintain the 100 µL assay volume. Once the inhibitor stocks were added, the assay plate was allowed to incubate 10 min at 27°C with shaking. This was followed by addition of phosphatase (stock, 0.4 µM, 20 nM in assay) and another 10-min incubation at 27°C prior to addition of pNPP substrate. Purification of GST-tagged STEP61 and STEP46 constructs was as previously described [13],[14]. The GST-SHP-2 construct was a generous gift from Dr. A. M. Bennett (Yale University). The GST-PTP1B construct was purchased from Addgene (Cambridge, MA). Constructs were transformed into BL21 (DE3) E. coli cells. Fusion proteins were purified on a glutathione sepharose column. Full-length HePTP and PTP-SL proteins were purchased from Abnova (Taipei, Taiwan). All proteins were dialyzed in 1,000-fold volume of buffer, which was repeated three times. Assays were carried out in 96-well plates with 10 nM of each phosphatase and various doses of TC-2153 in triplicates. After 10 min preincubation of enzyme and inhibitor at 27°C, 500 µM of pNPP was added and incubated for 30 min. Absorbance was taken at 405 nm using a Biotek plate reader. Percent of inhibition by TC-2153 at each dose was calculated and plotted using GraphPad Prism v5.01 to obtain IC50. To monitor the stability of TC-2153 in the imidazole buffer, 20 µL of 20 mM TC-2153 stock in DMSO was added to an Eppendorf tube. The solution was diluted to 400 µL (1 mM TC-2153 final concentration, 5% final DMSO) with either water or the pH 7.0 imidazole buffer. The tube was allowed to incubate at ambient temperature with shaking for 1 h. The mixture was diluted with 150 µL of DMSO-d6 and transferred to an NMR tube containing a capillary of trifluoroacetic acid as an external standard (−76.55 ppm). The stability of the compound in the buffer was confirmed by observing no differences in the sensitive 19F-NMR spectra (Figure S4). As a control for compound modification, the experiment was repeated with the addition of 1 mM GSH in the incubation buffer. STEP was diluted into 1× assay buffer with either inhibitor or DMSO control (final volume, 2.9 mL; final concentration, 1 µM STEP, 5 µM TC-2153; 50 mM imidazole HCl, pH 7.0, 50 mM NaCl, 0.001% Triton-X 100, 5% v/v DMSO). The samples were shaken at room temperature for 1 h to inhibit STEP. Each sample was then transferred to a separate Thermo Scientific Slide-A-Lyzer dialysis cassette with a 10,000 MW cutoff and 0.5–3.0 mL sample volume and was dialyzed into 1 L of 1× assay buffer over 24 h in a 4°C cold room. Aliquots of approximately 100 µL were removed from the dialysis cassette at 0, 4, and 24 h time points. Protein concentration was determined by reading absorbance at 280 nm compared to a standard curve for STEP. The samples were diluted to 100 nM in 100 µL of 1× assay buffer. The reaction was started by addition of 10 µL of 10× pNPP substrate (stock, 20 mM, 1.81 mM in assay; total assay volume, 110 µL), and reaction progress was immediately monitored at 405 nm at a temperature of 27°C. The initial rate data collected were used to determine enzyme activity standardized to the DMSO control zero time point. STEP was diluted to 200 nM in water, and aliquots of this stock were mixed with DMSO (5% by volume) or TC-2153 (5 µM final concentration, 5% DMSO by volume) and incubated at ambient temperature on a shaker for 10 min. Each sample was aliquoted out and 50 µL was transferred to wells of a 96-well microtiter plate containing 40 µL of 2× assay buffer with added reductant (GSH or DTT, 1 mM final concentration) and shaken for 0, 15, 30, or 60 additional minutes at ambient temperature. The reaction was started by addition of 10 µL of 10× pNPP substrate (stock, 20 mM, 2 mM in assay; total assay volume, 100 µL), and reaction progress was immediately monitored at 405 nm at a temperature of 27°C. The initial rate data collected were used to determine enzyme activity standardized to the DMSO controls. The second-order rate constant of inactivation for TC-2153 was determined under pseudo–first-order conditions using the progress curve method [29]. Assay wells contained a mixture of the inhibitor (800, 400, 200, 100, 50, 0 nM) and 745 µM of pNPP (Km = 745 µM) in buffer (50 mM imidazole pH 7.0, 50 mM NaCl, 0.01% Triton-X 100). Aliquots of STEP were added to each well to initiate the assay. The final concentration of STEP was 10 nM. Hydrolysis of pNPP was monitored spectrophotometrically for 30 min at an absorbance wavelength of 405 nm. To determine the inhibition parameters, time points for which the control ([I] = 0) was linear were used. A kobs was calculated for each inhibitor concentration via a nonlinear regression of the data according to the equation P  =  (vi/kobs)(1-exp(-kobst)) (where P, product formation; vi, initial rate; t, time) using Prism 5 (GraphPad). Because kobs varied hyperbolically with [I], nonlinear regression was performed to determine the second-order rate constant, kinact/Ki, using the equation kobs  =  kinact[I]/([I] + Ki (1 + [S]/ Km)). Assays were done in quadruplicate on two separate occasions. The average and standard deviation of the assays is reported. To explore the protein modification(s) of STEP upon TC-2153 inhibition, reduced and nonreduced gel-purified STEP (WT or C472S mutant) proteins were analyzed by high-resolution tandem mass spectrometry. Briefly, purified STEP WT or C-S mutant proteins (10 µg each) were incubated with vehicle (1% DMSO) or TC-2153 (10 µM in 1% DMSO) in assay buffer (50 mM imidazole, pH 7.0) at room temperature (25°C) for 30 min. Samples were resolved on 8% SDS-PAGE or nondenaturing PAGE, and proteins were visualized by Coomassie Blue staining. Gel bands were excised and kept at −80°C until use. Excised gel bands corresponding to the mutant and WT STEP with and without TC-2153 were in-gel trypsin digested under native conditions (w/o reducing agent) overnight. Peptides were extracted from the digested samples with 80% acetonitrile containing 0.1% trifluoroacetic acid, and then dried under SpeedVac. Samples were then reconstituted in minimum solution containing 0.1% TFA, and loaded onto a RP C18 nanoACQUITY UPLC column (1.7 µm BEH130 C18, 75 µm×250 mm, with a 5 µm Symmetry C18 2G-V/M Trap [180 µm×20 mm]). Eluted peptides were directly infused into an Orbitrap Elite LC MS/MS system running data-dependent acquisition. Acquired data were processed utilizing Progenesis LCMS software (Nonlinear Dynamics) and MASCOT Search engine with user-defined possible modification(s) search criteria. A previous study showed that genetic reduction of STEP significantly reversed cognitive deficits in 6-mo-old 3xTg-AD mice [16]. Here we were interested in testing whether pharmacologic inhibition of STEP with TC-2153 had a similar beneficial effect in this AD mouse model. We also wanted to test whether TC-2153 had any effects on cognition in WT mice. Mice completed all tests in the following order: Y-maze alternation, NOR, and MWM. For all behavioral tests, WT or 3xTg-AD mice were randomly allocated to treatment with either vehicle or TC-2153. To assess locomotor activity and exploratory behavior, mice were placed in a square box (60×60×60 cm) and habituated for 5 min. Mice were treated with vehicle or TC-2153 (10 mg/kg, i.p.) 3 h prior to the exploration phase of the experiment. A video camera mounted directly above the box recorded the trials and ANY-maze software analyzed the distance traveled and time spent in the center of the box. A crossover design was used in the Y-maze and NOR tasks, such that mice initially treated with vehicle (or TC-2153) were retested following a 15-d drug-free period and received TC-2153 (or vehicle). The Y-maze apparatus consisted of three dark gray arms (42×4.8×20 cm). Each mouse was treated with vehicle or TC-2153 (10 mg/kg, i.p.) 3 h prior to the experiment, after which they were placed at the end of one arm (the designated “start arm”) and allowed to freely explore the maze for 5 min. The total number of arm entries was recorded, as was the number of entries representing alternation behavior (i.e., sequential entry into all three arms). All four paws of the mouse had to enter an arm for it to count as an arm entry. Percentage spontaneous alternation  =  (number of alternations)/(total arm entries – 2). A crossover design was used after a drug-free period of 15 d, with groups previously treated with vehicle then receiving TC-2153 and vice versa. A total number of 20 WT and 11 AD mice were used in the Y-maze task. Mice were first habituated to the task by allowing them to explore an empty white open field box (60 cm×60 cm) for 5 min. Twenty-four hours later, mice were treated with vehicle or TC-2153 (10 mg/kg i.p.) 3 h prior to the sample phase. After the elapsed time, the mice completed the sample phase in which they were placed into the open field box with two identical objects located in the right and left corners. Mice were allowed to freely explore until they had accumulated a total of 30 s of object exploration (i.e., contact with the object with the nose and/or front paws), at which point the trial ended. The time spent with each object was recorded. Mice that were unable to complete the 30 s exploration within 20 min during the sample phase were excluded from the study (WT = 1 and AD = 3). Twenty-four hours later, mice completed the choice phase that was conducted in an identical manner to the sample phase except that one of the objects was substituted by a novel object and trial duration was set at 5 min. No drug treatments were given during the choice phase. Fifteen WT mice from the initial cohort were used to optimize the novel object conditions (to identify object pairings of inherent equal interest). Location of the novel object (left or right side) was counterbalanced to minimize possible bias. A crossover design was used, with a different set of objects after a 15-d drug-free period. DI was used to evaluate the effects of the TC-2153 compound on object memory in 6-mo-old 3xTg-AD mice. The DI was calculated for each subject by using the following formula: DI  =  (time spent exploring novel object – time spent exploring familiar object)/(total time spent exploring both objects). A DI of 0 is indicative of chance performance (i.e., no preference for one object compared to another), whereas a positive index (ranging from 0 to 1) indicates preference for novel object compared to familiar. In order to achieve greater statistical power, a second cohort of AD mice (n = 7) was run in the novel object test using a crossover design. Any value lower or higher than two times standard deviation away from the mean was considered an outlier and was excluded from the study (AD = 1). A total number of 9 WT and 16 AD mice were used in the NOR task. The reference memory version of the MWM task was performed as described previously [61]. A crossover design was not used in the MWM task, as the mice were randomly assigned to each treatment condition and can be exposed to the task only once. Briefly, animals were trained to swim in a 1.4 m diameter pool to find a submerged platform (14 cm in diameter) located 1 cm below the surface of water (24°C), rendered opaque by the addition of nontoxic white paint. Animals were pseudo-randomly started from a different position at each trial and used distal visual-spatial cues to find the hidden escape platform that remained in the center of the same quadrant throughout all training days. Training measures included escape latency to reach the platform, swim speed, and thigmotaxis. When animals failed to find the platform, they were guided to it and remained there for 10 s before removal. At 24 h after the acquisition phase, the platform was removed and a probe trial of 90 s was given to evaluate the number of entries in a circular zone (three times the platform diameter) positioned around the previous platform location (target zone) and in the opposite quadrants. To assess visual deficits and motivation to escape from water, the probe test was followed by a cued task (60 s; three trials per animal) during which the platform was visible. The visible platform was moved to different locations between each trial. After each trial, animals were immediately placed under a warming lamp to dry to prevent hypothermia. The experimenter was blind to mouse genotype when administering TC-2153 or vehicle to AD mice (AD-TC, n = 7; AD-Veh, n = 6) or WT mice (WT-TC, n = 13; WT-Veh, n = 12). Behavioral data from training, probe, and cued trials were acquired and analyzed using the ANY-maze automated tracking system (Stoelting, IL, USA). A two-way analysis of variance (ANOVA) with genotype as the between-subject factor and treatment as the within subject factor was used for the Y-maze and object recognition tasks. Percent alternation (Y-maze) and DI (object recognition) were the dependent measures. Post hoc analyses were carried out using Bonferroni's multiple comparison tests as appropriate (GraphPad Prism, La Jolla, CA). In an older (12 mo) cohort of WT and 3xTg-AD mice, the exploration time (NOR task) did not meet the assumption of normality and equal variance, and raw data (seconds) were converted using square-root transformation followed by t test. For the MWM training and probe sessions, a three-way repeated measures ANOVA with two between-subject (Genotype, Treatment) and one within-subject (training day or quadrant) factor was used. Escape latency (training) and number of entries (probe) were the dependent measures (StatView, Cary, NC). Swim speed and escape latency during the probe and cued trials, respectively, were analyzed using a two-way ANOVA with genotype and treatment as the between-subject factors. Post hoc analyses were conducted on significant results. For cell-based assays, one-way ANOVA with post hoc Bonferroni test was used to determine statistical significance. All data were expressed as mean ± s.e.m.

10.1371/journal.ppat.1002579

HMOX1 Gene Promoter Alleles and High HO-1 Levels Are Associated with Severe Malaria in Gambian Children

Heme oxygenase 1 (HO-1) is an essential enzyme induced by heme and multiple stimuli associated with critical illness. In humans, polymorphisms in the HMOX1 gene promoter may influence the magnitude of HO-1 expression. In many diseases including murine malaria, HO-1 induction produces protective anti-inflammatory effects, but observations from patients suggest these may be limited to a narrow range of HO-1 induction, prompting us to investigate the role of HO-1 in malaria infection. In 307 Gambian children with either severe or uncomplicated P. falciparum malaria, we characterized the associations of HMOX1 promoter polymorphisms, HMOX1 mRNA inducibility, HO-1 protein levels in leucocytes (flow cytometry), and plasma (ELISA) with disease severity. The (GT)n repeat polymorphism in the HMOX1 promoter was associated with HMOX1 mRNA expression in white blood cells in vitro, and with severe disease and death, while high HO-1 levels were associated with severe disease. Neutrophils were the main HO-1-expressing cells in peripheral blood, and HMOX1 mRNA expression was upregulated by heme-moieties of lysed erythrocytes. We provide mechanistic evidence that induction of HMOX1 expression in neutrophils potentiates the respiratory burst, and propose this may be part of the causal pathway explaining the association between short (GT)n repeats and increased disease severity in malaria and other critical illnesses. Our findings suggest a genetic predisposition to higher levels of HO-1 is associated with severe illness, and enhances the neutrophil burst leading to oxidative damage of endothelial cells. These add important information to the discussion about possible therapeutic manipulation of HO-1 in critically ill patients.

HO-1 is an important anti-inflammatory enzyme induced by several stimuli associated with critical illness. In humans, the amount of HO-1 produced is influenced by a genetic polymorphism in the gene promoter region. Using Plasmodium falciparum malaria that can cause a sepsis-like syndrome as an example, we characterize the associations between the (GT)n polymorphism, HO-1 protein levels and HMOX1-mRNA expression with severity of malaria in 307 Gambian children. Our results support the functionality of this polymorphism, demonstrate that P. falciparum infections increase HO-1 levels, and indicate that a genetic predisposition to strongly upregulate HO-1 is associated with severe forms of malaria and increased risk of dying. We identify neutrophils as the main HO-1-producing blood cells, and provide evidence that hemin-mediated induction of HMOX1 in neutrophils in vitro enhances the oxidative burst. In this way sequestered neutrophils may contribute to oxidative damage of endothelial cells, which may be part of a causal pathway explaining the association between short (GT)n repeats and increased disease severity. Our findings imply that the beneficial effects of HO-1 may be limited to a narrow window of concentrations, which should be born in mind when considering the therapeutic potential of manipulating HO-1 induction in critically ill patients.

Heme oxygenase (HO) is the rate limiting enzyme that catabolizes free heme into carbon monoxide (CO), ferrous iron, and biliverdin/bilirubin [1]. To date, two functional isoforms (HO-1, HO-2) have been described. While HO-2 is constitutively produced by most cells, HO-1 protein is induced by its substrate heme and a broad array of acute stress stimuli, many of which are associated with critical illnesses [2]. HO-1 induction produces cytoprotective and anti-inflammatory effects by reducing intracellular heme availability, through generation of CO and bilirubin, through stimulation of ferritin synthesis [3], and possibly, by heme-independent mechanisms of transcriptional regulation [4]. HO-1 is an essential enzyme in humans and mice; deficiency in humans is deleterious, predominantly affecting endothelial cells and the reticuloendothelial system, and results in a greatly reduced life expectancy [5]. However, much of what is known about HO-1 function is derived from experiments in animal models or in in vitro experiments. The impact of HMOX1 over- or under-expression, silencing or knockout and the concomitant changes in protein levels in a physiological or homeostatic context [6] or in humans [7] is less clear. The blood stage of malaria infection is characterized by hemolysis and consequent release of hemoglobin and its heme moiety [8]. Elegant mechanistic studies in mice have shown that free heme has a profound pro-inflammatory and cytotoxic effect in malaria, increasing susceptibility to experimental cerebral malaria (ECM), and hepatic failure. These adverse events can be prevented by HO-1 induction or administration of CO that can reduce the levels of free heme [9], [10]. Marked differences amongst mouse strains in the kinetics of HO-1 in response to P. berghei ANKA infection appeared to determine susceptibility to ECM, suggesting that regulation of HMOX1 expression is a crucial factor in this model. However, higher HO-1 levels in the murine liver appeared to allow the development of liver stage parasites by reducing the host inflammatory response, indicating that optimal regulation of HO-1 must balance control of pathogen replication with protection from inflammatory damage during infection [11]. As expected, evidence of increased expression and activity of HO-1 has been observed in human malaria [12], [13], [14], [15], but its functional relevance has been far more difficult to establish. Other than inbred mice, the amount of HO-1 produced in response to a defined stimulus in humans may be influenced by a (GT)n repeat length polymorphism in the promoter region of the HMOX1 gene [16]. In various chronic inflammatory conditions and other diseases, long HMOX1 (GT)n repeats, associated with lower HO-1 protein have been identified as disease risk factors [17]. This has led to the hypothesis that the ability to mount a strong HO-1 response is beneficial for people living in malaria endemic areas, and that the disease may have applied selective pressure for shorter (GT)n repeats [17]. The potential role of HO-1 and CO has also been recognized in other critical illnesses [18], and in a murine model free heme clearly contributed to the pathogenesis of severe sepsis [19]. If the protective effect of CO or other products of the enzymatic reaction catalyzed by HO-1 can be established in man, this could provide novel avenues for treatment using therapeutic CO inhalation, or systemic administration of CO releasing molecules (CORM), capable of releasing CO in a controlled fashion [20]. HO-1 has thus moved to “center stage” for a variety of infectious diseases, not just for malaria [20], [21]. A recent report on critically ill patients measuring CO bound hemoglobin (COHb) levels of which 85% can be ascribed to HO-1 mediated heme metabolism [22] indicates that both excessively low or high levels of COHb appear to be associated with death [23]. This indicates that the protective effects of HO-1 are limited to a narrow range of HO-1 concentrations [18]. While HO-1 uses the highly cytotoxic heme as a substrate, one of the products of the enzymatic reaction, ferrous iron, is released into the endoplasmic reticulum (ER). In this form, iron is redox active and can catalyze the formation of organic and inorganic reactive oxygen species (ROS) [24]. HO-1 thus has both anti-oxidant and oxidant properties. In malaria its induction may be particularly enhanced by a pronounced intravascular hemolysis liberating considerable amounts of heme [25] that require degradation by HO-1 in endothelial cells, resulting in an increase of ferrous iron. Indeed, in vitro studies indicate that the equimolar production of anti-oxidant bilirubin and ferrous iron by HO-1 results in an overall pro-oxidant effect [26], and that high HO-1 levels can lead to tissue damage [27]. In light of this, it has been hypothesized with regard to the effect of the (GT)n repeat length polymorphism in the HMOX1 gene promoter that – in contrast to what has been observed for chronic inflammatory conditions - short repeat array alleles may cause susceptibility to severe malaria in humans [28]. Consistent with this, short alleles were found to be associated with risk of cerebral malaria in Myanmar [29], and Angola [30]. The functional duality is problematic with respect to developing adjuvant therapies for severe malaria based on induction of HO-1 or administration of CO, and highlights the need to understand better the regulation and function of HO-1 in humans in relation to both promoter polymorphisms and malaria. In the present study we characterized in detail the genetic and functional associations between HMOX1 promoter polymorphisms, HO- 1 inducibility, HMOX1 expression and severity of malaria in Gambian children exposed to seasonal malaria. We show that short (GT)n repeat alleles in the HMOX1 gene are associated with higher HMOX1 expression in white blood cells of this population, and that short repeat alleles are strongly associated with severe disease and death, whilst high HMOX1 mRNA and HO-1 protein levels are associated with severe disease. We establish that neutrophils are the main HO-1 expressing cell type in peripheral blood ex vivo, and demonstrate in vitro that HMOX1 mRNA expression in purified neutrophils can be upregulated further by lysed erythrocytes, or hemin. We provide mechanistic evidence that hemin-mediated HMOX1 expression potentiates the neutrophil respiratory burst, and propose that this may be part of a causal pathway driving the association between short (GT)n repeats and increased disease severity. The study was reviewed and approved by the Gambian Government/MRC Joint Ethics Committee and the Ethics Committee of the London School of Hygiene & Tropical Medicine (London, UK). Between September 2007 and January 2010, after written informed consent was obtained from the parents or guardians, a total of 153 severe and 154 uncomplicated malaria cases were enrolled. (see Table S1 in Text S1 for detailed information). Subjects enrolled in this study were recruited from an ongoing health centre based study comparing children with uncomplicated and severe malaria disease resident in a restricted peri-urban area of the Gambia described in more detail previously [31]. Uncomplicated malaria (UM) was defined as an episode of fever (temperature >37.5°C) within the last 48 hours with more than 5000 parasites/µl detected by slide microscopy. Severe malaria (SM) was defined using modified WHO criteria [32]: severe anaemia (SA), defined as Hb<6 g/dl; severe respiratory distress (SRD) defined as serum lactate >7 mmol/L; cerebral malaria (CM) defined as a Blantyre coma score ≤2 in the absence of hypoglycaemia or hypovolaemia, with the coma lasting at least for 2 hours; severe prostration (SP) defined as inability to sit unsupported (children>6 months) or inability to suck (children≤6 month). The term “disease severity” refers to comparisons between UM and SM, and, where indicated, to a comparison across disease entities grouped according to increasing severity. To avoid the confounding effects of other pathogens in children with concomitant systemic bacterial infections, children with clinical evidence of infections other than malaria were not enrolled into the study. For some experiments, healthy children (HC, n = 6) of the same age were enrolled as controls. On admission (D0, also referred to as “acute disease”) and after 4 weeks (D28±3 days, also referred to as “convalescence”) one ml of blood was collected in RNA stabilizing agent (PAXgene Blood RNA system, Pre-AnalytiX) and a maximum of 4 mls of blood (mean: 3.2 mls CI 95%: 3.1–3.3 mls) were collected into heparinized vacutainers (BD). Four buccal swabs were performed using sterile mouth brushes (Cytobrush plus, Henley's Medical, UK) and stored in a DNA-stabilizing buffer containing 10 mM Tris, 10 mM EDTA, 0.5% Sarkosyl for subsequent DNA extraction. All patients received standard care according to the Gambian Government Treatment Guidelines, provided by the health centre staff. The children's health was reviewed 7 days after admission. Healthy adult volunteers were bled for the in vitro experiments after informed consent was obtained. P. falciparum parasites were identified by slide microscopy of 50 high power fields of a thick film. Full differential blood counts were obtained on days 0 and 28 using a Medonic instrument (Clinical Diagnostics Solutions, Inc). Blood samples were processed within 2 hours of collection. Flow cytometry was performed on 300 µl of whole blood collected into heparinized tubes. HO-1 induction assays were performed on 200 µl of whole blood collected on D28. From the remaining sample, plasma was removed, stored at −80°C and replaced by an equal volume of RPMI 1640 (Sigma-Aldrich). PBMC were isolated after density centrifugation over a 1.077 Nycoprep (Nycomed, Sweden) gradient (800 g, 30 min) and washed twice in RPM 1640. PBMCs were used for other studies [31]. The remaining PBMC deficient blood suspension underwent a further density centrifugation over Histopaque 1119 (Sigma Aldrich) to isolate granulocytes that were used for Western blood analysis of HO-1 expression after a microscopic purity check with Giemsa stain. Whole blood was incubated for 35 min at 4° in the dark with the following cocktail of surface antibodies: 5 µl each of PE anti-CD16b, Per CP anti-CD14, PE-Cy7 anti-CD4, APC anti-CD19 (all Becton Dickinson), Pacific blue anti-CD3 and 4 ul of APC-AF 750 anti-HLA-DR (both Ebioscience), or a cocktail of manufacturer matched isotype controls. Thereafter, erythrocytes were lysed using FACS lysing buffer (Becton Dickinson), and the remaining cells were fixed and permeabilized (Cytofix/Perm reagent; Becton Dickinson). After a blocking step with 5% of mouse serum (4°C, 15 min) intracellular staining (4°C, 30 min) for HO-1 was performed with 3.5 µl of FITC anti-HO-1 (Abcam). Samples were acquired on a 3 laser/9 channel CyAn ADP flowcytometer and analysed using FlowJo 7.25 (Tree Star Inc.). For quantitative reverse transcription-polymerase chain reaction (qRT-PCR), total RNA was extracted from PAX tubes, collected from study patients following the manufacturer's instructions and reverse transcribed into cDNA using TaqMan reagents for reverse transcription (Applied Biosystems), according to the manufacturer's protocol. In addition, whole blood used for HO-1 induction assays from samples obtained on day 28 were collected into Trizol LS (Invitrogen) and the RNA precipitated by a chloroform/ethanol step. Isolation of RNA from neutrophils or whole blood used in the in vitro assays was performed with the RNeasy Mini kit (Qiagen) after collection and storage of the cells into RLT buffer. Gene expression profile for IL-10 was measured previously on a subset of samples from the clinical study and were used for correlation analysis [31]. HMOX1 gene expression was determined by qRT-PCR on a DNA Engine Opticon (MJ Research) using the TaqMan Probe kit with primers (all Metabion) as previously published [33]. 18S rRNA, amplified using a commercially available kit (rRNA primers and VIC labeled probe, Applied Biosystems), was assayed as a housekeeping gene with a stable expression profile in this setting regardless of disease severity or time point [31]. Data were analysed using Opticon Monitor 3 analysis software (BioRad) and are expressed as the ratio of the transcript number of the gene of interest over the endogenous control, 18S rRNA. Levels of soluble HO-1 were determined in plasma (1∶50 diluted) or cell culture supernatants (neat) by ELISA in duplicate wells of Immunolon HX4 plates, using the HO-1 human ImmunoSet Kit (Stressgen). A commercial HRP-2 ELISA kit (CELISA, Cellabs, Australia) was used to quantify HRP-2 in plasma samples diluted 1∶20, in duplicate wells of Immunolon HX4 plates. Some samples were out of range and were repeated at a 1∶2 dilution if below the bottom of the standard curve, or at a 1∶100 dilution if above the top of the standard curve. Free heme in plasma was quantified using a published method [9]. Briefly, plasma was centrifuged at 1000 g for 5 min and the supernatant passed through a Microcon YM-3 column (Millipore, 14,000 g for 100 min at RT) to remove proteins. Free heme from protein depleted plasma and heme content in lysates of infected and uninfected RBC as well as a solution of uninfected intact RBC was quantified by a chromogenic assay (QuantiChrom Heme Assay Kit, BioAssay Systems). Immunodetection of HO-1 protein in lysates of isolated neutrophils was performed using a protocol adapted from [4]. Polyclonal rabbit anti-HO-1 antibodies obtained from StressGen Biotechnologies Corp. (Victoria, BC, Canada) were used. Polyclonal goat anti-actin (Santa Cruz) antiserum was used for staining as loading controls. Briefly, 10 µg of cell lysate proteins was separated by reducing sodium dodecyl sulphate polyacrylamide gel electrophoresis using precast Nupage gels and MOPs buffer in the X-cell mini electrophoresis chamber (Life Technologies). Separated proteins were then transferred onto methanol treated PVDF membranes using the X-cell mini blotting system. Blotted membranes were rinsed in 1×PBS and blocked overnight at 4°C in blocking buffer containing 5% non-fat milk in 1×PBS and 0.1% Tween 20 (Sigma Aldrich). After the blotted membranes were washed three times in PBS-Tween, they were incubated with constant shaking for 2 h at RT with anti-HO-1, or anti-actin diluted 1∶1,000 in blocking buffer. The membranes were then washed with three changes of PBS-Tween and further probed with horseradish peroxidase-conjugated donkey anti-goat or goat anti-rabbit IgG (Santa Cruz) at a dilution of 1∶10,000 for 2 hours at room temperature with constant shaking. Following three washes in PBS-Tween, membranes were rinsed in 1×PBS and bound antibodies were revealed by chemiluminescent detection performed with the Amersham ECL detection kit according to the manufacturer's instructions. Material from mouth brushes was eluted into transport buffer and incubated with Proteinase K, guanidine hydrochloride, and ammonium acetate (Sigma Aldrich, UK, at final concentrations of 262 µl/ml, 1.57M and 0.59M, respectively), for 1 hr at 60°C. Ice cold chloroform was added to each sample at a ratio of 1 to 1.9 followed by a 5 min centrifugation at 1000 g. The upper layer was transferred onto 10 mls pure ethanol and kept at −20°C for 1 hour to precipitate the DNA. After 15 min centrifugation at 1200 g the pellet was resuspended in 70% ethanol, washed again (1200 g, 5 min), resuspended in 100 µl 1X TE buffer (Sigma Aldrich, UK), and stored at −20°C until processing. The 5′-flanking region of the HMOX1 gene containing a (GT)n repeat was amplified by PCR using a fluorescein-conjugated sense primer (HMOX1_microsat_ (9/10) 5′-AGAGCCTGCAGCTTCTCAGA- 3′) and an antisense primer HeOP-1/R HMOX-1mi (1/3) 5′-ACAAAGTCTGGCCATAGGAC-3′) previously described [29], [34]. Samples that did not amplify with these primers were amplified using a second set of primers; HMOX-2micro-Fwd (2/3) 5′ CTTTCTGGAACCTT CTGGGAC 3′ designed according to the published sequence [35] and the above antisense primer. Thirty-five cycles were performed under the following conditions; 96° for 1 minute, 95° for 30 seconds, 60° for 30 seconds, 72° for 3 minutes. The sizes of the microsatellites were determined by the use of a laser based automated DNA sequencer; ABI genetic analyser 3130xl (Applied Biosystems, Forster City, Calif, USA), with a cloned fragment of 28 bp that was used as a size marker. Allele scoring was performed using GeneMapper version 4.0 (Applied Biosystems, Forster City, Calif, USA), by 2 investigators blinded to the disease status of donors. For analysis purposes, alleles were subsequently divided into “S” (for short <27 repeats), “M” (for medium 27 to 32 repeats), and class “L” (for long >32 repeats), using an established classification [33], [34], [36]. To determine the frequency of Glucose 6 Phosphate Dehydrogenase (G6PD) deficiency, genomic DNA was genotyped for SNPs A376G (rs1050829) (G6PD A), G202A (rs1050828) (G6PD A-), and T968C (G6PD A-) [37] that are mutations causing reduced enzyme activity [38]. In order to infer the frequency of individuals with blood group O, rs8176719 was typed to identify the frame shift deletion at this position that encodes the O allele [39]. Genotyping was performed on a Sequenom MassArray platform [40]. For each reaction 20 ng of gDNA was used and each genotype was replicated three times. Sickle cell status was determined by metabisulfite test and the genotype was confirmed by cellulose acetate electrophoresis [31]. From a subset of participants (12 SM, 20 UM), 200 µl of whole blood collected at D28 were kept for 3 hours at 37°C, at 40°C (water bath), or stimulated with hemin (10 µM, Sigma Aldrich) at room temperature, to determine inducibility of HO-1 mRNA. An additional 6 samples from healthy controls were processed similarly. After the incubation, samples were diluted 1∶1 in RNAse free water, and transferred into Trizol LS reagent (Life Technologies). RNA processing and HMOX-1 gene expression was carried out as described under qRT-PCR. P. falciparum parasites (3D7 clone) were cultured in vitro as described [31], and were routinely shown to be mycoplasma free by PCR (Bio Whittaker). Schizont-infected erythrocytes were harvested from synchronized cultures by centrifugation through a Percoll gradient (Sigma-Aldrich). P. falciparum schizont extracts (PfSE) was prepared by three rapid freeze-thaw cycles between liquid nitrogen and a 37°C water bath. Lysates of uninfected erythrocytes (uRBC lysate) were prepared in the same way. Neutrophils were isolated from whole blood using CD15 beads (Miltenyi, Germany). The purity of neutrophils was assessed by flow cytometry and found to be 95.4% (95%CI: 93.6% to 97.1%). Neutrophils were cultured for various times either with intact, uninfected red blood cells (intact uRBC, containing 6.6 µM heme), uRBC lysate or PfSE (containing 95.3 µM and 99 µM of heme, respectively), growth medium (GM), or 100 µM hemin (Sigma Aldrich). Cell supernatants were harvested and assayed for HO-1 by ELISA, and cells were collected into RLT buffer and processed for HMOX1 mRNA as described under qRT PCR. To investigate the impact of hemin-induced HO-1 on the neutrophil respiratory burst, 500 uL whole blood of 4 healthy adult donors was diluted 1∶1 in RPMI (Gibco) and incubated for 18 hours at 37°C, in 5% humidified CO2 atmosphere, with different concentrations of hemin (0 to 200 µM). Half of the cells were used to measure hemin-induced induction of HO-1 mRNA. A small aliquot of cells was stained with a neutrophil marker (anti CD15ab labeled with APC, Miltenyi) and the ‘live dead cell stain’ (Invitrogen) to assess the viability of neutrophils by flow cytometry after pre-incubation with hemin. The oxidative burst was measured using a validated flow cytometric assay on the remainder of the cells [41]. Briefly, samples were stimulated by adding PMA (Sigma) to a final concentration ranging from 0 to 1000 nM for another 15 minutes. Thereafter, dihydrorhodamine 123 (DHR 123) (final concentration 5 ug/ml) was added for 5 min. Red cells were lysed (with ammonium chloride lysis buffer) and the remaining cells were stained with anti CD15 APC (Miltenyi) and the median fluorescence intensity of rhodamine, the fluorescent oxidation product of DHR 123, was measured in CD15+ cells by flow cytometry. In a separate series of experiments 500 µl of whole blood from another 4 healthy adult donors was incubated with 0–200 µM hemin, either with or without addition of tin protoporhyrin IX dichloride (SnPP; final concentration 10 µM), a non-substrate inhibitor of HO-1 activity [42]. A viability check was performed as described above, and the oxidative burst was induced by stimulation with PMA at a final concentration of 100 nM for 15 min, and the burst was measured as described above. Flow cytometric results, HMOX1 mRNA and HO-1 plasma levels obtained on D0 and D28 were compared using linear regression based on ranks, with a random effect to allow for repeated measurements over time. Significance (measured at the 5% level) tests for the effects of malaria group (SM, UM), time (D0 and D28) and their interaction were adjusted for the possible confounding effects of age, gender, duration of prior symptoms and Hb levels, as indicated. Further adjustment for neutrophils was performed for the analysis of WBC, and HO-1 mRNA. Where there was no significant malaria group and time interaction, p-values for the overall comparison of D0 vs D28 are given. Comparisons of SM vs UM are given for each time point separately if the malaria group and time interactions were significant. To allow for the multiple tests resulting from multiple responses and multiple comparisons within a response performed in the model, a false discovery rate (FDR) of 5% was assumed. Using the Benjamini and Hochberg approach [43] only tests with a p-value below 0.012 have an FDR of ≤5%. Comparison of HMOX1 mRNA and HO-1 plasma levels between different disease entities during acute disease was performed using linear regression based on ranks, adjusting for the confounding effects of age, gender, duration of symptoms, and for neutrophil counts and Hb levels where indicated. A multinomial logistic regression model was employed to explore the association between the exposure ‘L allele containing genotype’ and the outcome of different disease entities. Pearson Chi-squared tests were used to compare proportions amongst more than 2 groups. The magnitude of the differences of the long L allele frequencies reported for African, Asian and European populations was explored with fixation (FST) indices, using FSTAT [44]. For the analysis of in vitro induction of HMOX1 mRNA in neutrophils, pairwise comparisons using Wilcoxon matched pairs test were performed, with p values adjusted for multiple comparisons using Holm's step down procedure. Where more than two groups were compared, non-parametric one way ANOVA (Friedman test for paired samples, Kruskal Wallis test for unpaired samples) was used with Dunn's post test adjustment for multiple comparisons. Linear regression was performed to assess whether the magnitude of the oxidative burst induced by a given concentration of PMA increases with increasing concentrations of hemin used for pre-incubation. Analyses were performed using Stata version 10, and Graph Pad PRISM version 5.01. In total, 154 uncomplicated (UM), and 153 severe malaria (SM) cases were recruited into this study. The proportions followed up at day 28 were 85.7% for the UM group and 83% for the SM group. Ten children (6.5%) in the SM group died, and an additional 15 (9.8%) were either lost to follow up after hospital discharge or withdrew consent. Nine of the ten deaths occurred in children who were classified as having severe respiratory distress (SRD), defined as lactic acidosis, that has been recognized as the single most important determinant of mortality in severe malaria [45]. Children suffering simultaneously from SRD plus cerebral malaria (CM) had the highest mortality (29.4%), followed by cases with SRD (16.7%). This is consistent with a large study from Kenya, demonstrating that mortality decreases in the following order for different disease entities: [SRD+CM+severe anaemia (SA)]>[SRD+CM]>[SRD]>[SRD+SA]>[CM]>[CM+SA]>[SA]>[severe prostration (SP)] [46]. Where appropriate, cases were grouped according to disease entities and analyzed in the order of increasing severity (Table S1 in Text S1). To identify which leucocyte subsets express HO-1 protein, whole blood collected on days 0 and 28 from 16 SM and 21 UM cases was stained for lineage markers and intracellular HO-1 (Figure 1A–D). In SM, both the proportion and the total number of white blood cells (WBC) expressing HO-1 was 2.2 and 1.3 fold higher during the acute phase compared to convalescence (both with p<0.0001, both adjusted and unadjusted). For UM cases, a smaller difference was observed between time-points that became non-significant after adjustment for percentage (number) of neutrophils. (%WBC UM: p = 0.280 [unadjusted: p<0.0001]; WBC number UM: p = 0.31 [unadjusted: p = 0.01]). Both the percentage and total number of WBC expressing HO-1 were significantly higher in SM than UM cases on day 0 (% WBC: p = 0.004 [unadjusted: p = 0.001]; total WBC count: p = 0.009 [unadjusted p = 0.008]), while no differences were observed on day 28. Almost all neutrophils stained positive for HO-1, both on D0 (median: 93%, CI 95%: 90.5–96%) and D28 (median: 97%, CI95%: 97–99%), with no difference between SM and UM. Similarly, a median of 98% (CI 95%: 86–99.8%) of neutrophils from healthy controls (HC) stained positive for HO-1 (Figure 1E). Western blot of isolated neutrophils from 4 cases confirmed the presence of HO-1 in purified cells (Figure 1F). Monocytes, B cells, T cells and DCs also expressed HO-1, albeit at lower levels, rarely exceeding 4% of the lymphocyte subset (Figure 1E). In both SM and UM cases the proportions of monocytes, T cells and DCs expressing HO-1 were slightly but significantly higher on D0 compared to D28 (p = 0.002 [unadjusted: p<0.0001], <0.0001 [unadjusted: p<0.0001], <0.0001 [unadjusted: p<0.0001], respectively; Figure 1E). The total number of HO-1 positive neutrophils, monocytes and DCs was significantly higher on Day 0 compared to D28 for both SM and UM cases (p<0.0001 [unadjusted: p<0.0001], p = 0.005, [unadjusted: p = 0.001], p<0.0001[unadjusted: p<0.0001], respectively; Figure S1). Irrespective of disease status (SM, UM, HC) or time of sampling, a median of 98% (CI95%: 97.3–98.6%) of HO-1 expressing cells were neutrophils, whereas the other cell subsets accounted for less than 1% of HO-1 producing cells in peripheral blood (Figure 1G). RNA was extracted from 128 SM and 134 UM cases at both D0 and D28 from blood collected into PAX tubes to assess HMOX1 mRNA levels by qRT-PCR. Considering that neutrophils were the major source of HO-1 in peripheral blood and that their numbers are slightly higher in acute malaria compared to convalescence [31], the random effects model (from which the p values are derived) additionally adjusted for neutrophil counts to rule out that the observed difference merely reflects different numbers of HO-1 producing neutrophils. For both SM and UM cases a geometric mean 4.3 and 3.7 fold higher HMOX1 mRNA/18s rRNA ratio was found during acute disease compared to convalescence (p<0.0001 [unadjusted: p<0.0001], for both SM and UM, Figure 2A), while no significant difference was observed between SM and UM on D0 or D28. When HMOX1 mRNA expression of peripheral blood cells was assessed for different disease entities during acute disease (Figure 2B), linear regression with UM as a baseline group adjusting for age, gender duration of symptoms and neutrophil counts revealed a trend towards higher levels with increasing disease severity. This reached borderline significance for children classified as having SRD plus CM (p = 0.04) [unadjusted: p = 0.19]. In the UM, SP, SA, CM and SRD groups HMOX1 mRNA levels at D0 were elevated similarly (3.1 to 4.8 fold higher [geometric means] than on D28, p = 0.36, for comparison among groups, Kruskal Wallis test), but a significantly higher, 10.1 fold increase was measured for the SRD+CM group (p<0.05 compared to UM, Dunn's post test, adjusting for multiple comparisons). Considering that previous studies report increased HO-1 protein in plasma of critically ill patients, [47], we measured HO-1 levels in plasma on days 0 and 28 for 138 SM and 137 UM cases. Similar low levels of HO-1 were observed during convalescence for both SM and UM, but HO-1 concentrations measured during acute disease (D0) were 5.7 fold (SM) and 3.3 fold (UM) higher than on D28 (p<0.0001 for both groups [unadjusted: p<0.0001 for both groups], Figure 2C), with HO-1 levels in SM being significantly higher than in UM at D0 (p<0.0001) [unadjusted: p<0.0001], but not day 28. Using uncomplicated cases as the baseline group, plasma concentrations of HO-1 measured on day 0 were significantly associated with disease severity in a linear regression model based on ranks adjusting for age, gender, and duration of symptoms (p<0.0001 [unadjusted: p<0.0001]). In particular, patients with SA, SRD and SRD plus CM had significantly higher HO-1 plasma concentrations (SA: p = 0.004 [unadjusted: p = 0.001], SRD: p<0.0001 [unadjusted: p<0.0001], and SRD+CM: p = 0.001 unadjusted: p<0.0001]). After additional adjustment for Hb concentration the overall association between severity and HO-1 levels remained significant (p = 0.004). However, while HO-1 remained significantly elevated in cases with SRD (p = 0.003) and SRD plus CM (p = 0.007), the difference previously seen in the SA group was lost (p = 0.12; Figure 2D). While the origin of plasma HO-1 remains unclear, the latter observation supports the hypothesis that it is derived at least in part from damaged tissues [47]. By definition, SA cases have lower Hb levels, and a higher degree of hemolysis, which is associated with considerable damage of endothelial cells due to release of iron and heme-containing moieties from hemolysed RBC [48]. Adjusting for Hb may even out the effect of hemolysis-driven damage of endothelial cells that may lead to HO-1 release into plasma. However, additional factors may be responsible for the high HO-1 levels found in patients with SRD, where a significantly higher increase of HO-1 plasma levels was observed between D0 and D28 in patients with SRD or SRD plus CM compared to other entities (p<0.0001, Kruskal Wallis test). While patients with SP, SA or CM had 4.8, 7.7 and 5.5 fold higher values on D0 compared to D28, a 10.2 and 8.9 fold difference between D0 and D28 was measured for patients with SRD plus CM or SRD, respectively. This was significantly higher than the 3.3 fold increase observed for UM (p<0.05, Dunn's post test). In summary, the data indicate that HO-1 production is induced during acute malaria in peripheral blood cells and probably in various other tissues, and the effect is greatest in cases with SRD. Interestingly, for both SM and UM, HO-1 levels in plasma correlated well with indirect bilirubin, one of the end products of the reaction catalysed by HO-1, that is usually seen as an indirect measure of HO-1 activity, and has been established as a marker for disease severity [49] (r: 0.69, p<0.0001 [SM] and r: 0.77, p = 0.0012 [UM]; Figure S2). Considering that parasitaemia [%] correlated with HO-1 in plasma (r: 0.5, p<0.0001) and the observation that the majority of neutrophils contained HO-1, we investigated whether encounter with P. falciparum antigens can induce HMOX1 expression in neutrophils. To this end, HMOX1 mRNA expression was determined by qRT-PCR in neutrophils purified from whole blood of 7 donors using magnetic beads and cultured for 3 hours with either growth medium (GM) (negative control), 100 µM hemin (positive control), intact uninfected red blood cells (uRBC) at 1×108/ml, or freeze – thaw lysates of either uRBC or P. falciparum Schizont extract (PfSE) at a concentration equivalent to 1×108 cells/ml. Heme concentrations were measured in all RBC preparations and were found to be 6.6 µM (intact uRBC), 95.3 µM (uRBC lysate) or 99 µM (PfSE). Culture with lysates of both uRBCs and PfSE resulted in a significant increase in median HMOX1 mRNA expression compared to culture in GM (2.3 and 2.7 fold with uRBC (p = 0.04) and PfSE (p = 0.01) lysate, respectively), while HMOX1 mRNA remained at baseline levels in neutrophils cultured with intact uRBCs (Figure 3). Culture in the presence of hemin led to a significant, 4.4 fold median increase in HMOX1 mRNA in neutrophils (p = 0.023; all results adjusted for multiple comparisons). To evaluate whether neutrophils could contribute to plasma HO-1 levels by releasing HO-1, we cultured bead-purified neutrophils from an additional 4 donors using the above described conditions for 3, 6, 12, 24 and 36 hours, respectively, and tested the supernatants for HO-1 protein by ELISA. RNA was isolated from neutrophils for determination of HMOX1 mRNA by qRT-PCR. Although HMOX1 mRNA expression increased up to 6 hrs in response to uRBC, PfSE and hemin no significant amount of HO-1 could be measured in the supernatants for any of the conditions tested (data not shown). Taken together, these data demonstrate that HMOX1 mRNA can be induced in neutrophils in response to hemin as well as RBC lysates containing significant amounts of heme, and suggest that heme released during RBC lysis rather than parasite-derived molecules contribute to this increase. Further, we demonstrate that neutrophils do not release HO-1 in response to these stimuli within 36 hours, and thus are unlikely to contribute to plasma HO-1. In order to explore a correlation between free heme and HO-1 in plasma from our clinical samples, we attempted to quantify non-protein bound (free) heme. After filtration of the samples heme concentrations were barely measurable (median 1.42 µM, CI95%: 1.37–1.48 µM), being 4.6 fold lower than that in washed preparations of intact uninfected RBCs. We therefore excluded these data from further analysis. However, both RBC counts and Hb levels that may be regarded as surrogates for the degree of hemolysis, and therefore free heme in acute malaria, showed a negative correlation with plasma HO-1 (r = −0.34, p<0.0001 in both cases). The observation that HMOX1 expression can be increased in neutrophils in response to heme prompted us to investigate the role of HO-1 for the neutrophil respiratory burst. The oxidative burst in neutrophils is essential for the host's ability to kill ingested microorganisms and parasites [50], but intense oxidative stress has also been associated with severe forms of malaria, triggering unspecific tissue damage [10]. To assess the impact of hemin-mediated induction of HO-1 on the neutrophil function, the respiratory burst in response to PMA stimulation (0, 0.05, 0.1 and 1 µM) was measured using a validated flow cytometric whole blood assay [41] after overnight incubation of blood from 4 healthy donors with several concentrations of hemin. As expected, pre-incubation with hemin induced HMOX1 mRNA expression in a dose dependent manner (r2lin regression = 0.9, p = 0.0052, Figure 4A), and did not significantly affect the viability of neutrophils (Figure 4B). While hemin pre-incubation alone did not induce the oxidative burst, stimulation with 0.05, (0.1) or [1.0] µM PMA reliably induced an oxidative burst in 94.4%, 96.5% and 96.9% of neutrophils, respectively (p = 0.052, paired measures ANOVA). For each PMA concentration, pre-incubation with different hemin concentrations did not affect the proportion of neutrophils responding with an oxidative burst (Figure 4C). However, the magnitude of the oxidative burst induced with 0.05 µM and 0.1 µM PMA in neutrophils increased significantly with increasing concentrations of hemin used during the pre-incubation (r2lin regression 0.92 [0.05 µM PMA] and 0.84 [0.1 µM PMA], with p = 0.037 [0.05 µM PMA] and 0.04 [0.1 µM PMA], Figure 4D). When 1 µM PMA was used, the burst could be maximally stimulated without preincubation with heme. This suggests that hemin-mediated induction of HO-1 may prime neutrophils to mount a stronger oxidative burst. To further investigate whether the observed effect is mediated by hemin-induced HO-1, we repeated the experiment in the presence and absence of 10 µM tin protoporphyrin (SnPP), an inhibitor of HO-1 activity, using whole blood from another 4 healthy volunteers in separate experiments. At this concentration, SnPP did not significantly affect neutrophil viability compared to pre-incubation with hemin alone (p = 0.125 [0 µM hemin], 1.0 [0.05 µM hemin], 0.125 [0.1 µM hemin] and 0.625 [1 µM hemin], Wilcoxon signed rank test). Further, the addition of 10 µM SnPP did not affect the proportion of neutrophils responding with an oxidative burst to stimulation with 0.1 µM PMA, compared to pre-incubation with hemin alone (p = 0.375 [0 µM hemin], 0.25 [0.05 µM hemin], 0.875 [0.1 µM hemin] and 0.625 [1 µM hemin], Wilcoxon signed rank test). However, hemin pre-incubation in the presence of 10 µM SnPP abrogated the hemin dose-dependent increase of the oxidative burst (r2lin regression 0.88, p = 0.035 [no inhibitor] and r2 = 0.22, p = 0.23 [with inhibitor], Figure 4E). Taken together, the data indicate that heme-mediated induction of HO-1 primes the oxidative burst in neutrophils. Apart from the availability of its substrate heme, other factors may induce HO-1. In animal models [51], and human hepatoma cell lines [52] HO-1 could be induced by heat exposure. However, in human alveolar macrophages or erythroblastic cell lines [53], as well as in PBMC [54] thermal stress failed to induce HO-1. Since we observed a weak but significant positive correlation between temperature on admission to the clinic and HO-1 plasma levels (r: 0.266, p<0.0001), we explored whether a temperature of 40°C maintained over 3 hours would induce HMOX1 mRNA in human whole blood, using samples collected on D28 from 12 SM, 20 UM cases and 6 HC. Samples kept at 37°C or cultured with hemin served as negative and positive controls, respectively. In all three groups (SM, UM, HC) both incubation at 40° as well as with hemin resulted in a significant upregulation of HMOX1 mRNA compared to cells kept at 37°C, with no significant differences observed between groups (Figure 5A). In a separate experiment using blood from 5 healthy donors we verified that incubation at 40°C for 3 hours did not lead to a significant change in hemolysis markers such as heme, haptoglobin or LDH, compared to incubation at 37°C (data not shown). For murine macrophages, IL-10 has been shown to induce HO-1 [55]. We therefore correlated HMOX1 mRNA to IL-10 mRNA from D0 samples for 58 SM and 59 UM cases for which IL-10 mRNA measurements were available from a previously reported study [31]. For both SM and UM a positive correlation (SM: r = 0.59, p<0.0001; UM: r = 0.37, p = 0.0037) was found in whole blood, compatible with a role for IL-10 as an inducer of HO-1 in human blood cells (Figure 5B, C). The extent to which HMOX1 is upregulated in an individual in response to a defined stimulus may be influenced by genetic polymorphisms in the promoter region of the HMOX1 gene of which several have been described, (reviewed by [16]). Of particular interest, a (GT)n repeat length polymorphism regulates the promoter activity and gene expression, with short repeats (<27 repeats) resulting in an increased transcription of HMOX-1 compared to alleles with long repeats (>32 repeats) [33], [56]. Associations of the (GT)n polymorphism with disease outcomes have been explored in numerous association studies for various diseases, recently reviewed in [17]. To investigate whether the (GT)n polymorphism is associated with disease severity in malaria, we genotyped this microsatellite for 142 SM and 151 UM cases for whom DNA samples were available. In the study population (GT)n variation ranged from 13 to 45 repeats, and the allele frequency distribution was trimodal with peaks at 26, 30 and 39 repeats (Figure 6A). Using a previously established classification [33], [34], [36], alleles were divided into 3 groups: “S” (<27 repeats), “M” (27 to 32 repeats), and “L” (>32 repeats). The frequency of “S” alleles was significantly higher in SM cases (0.50 vs. 0.37, p = 0.0021), whereas the frequency of “L” alleles was significantly higher in UM cases (0.36 vs 0.26, p = 0.009, Figure 6B). The frequency of different alleles according to disease entities is shown in Figure 6C. Based on the “S-M-L” classification, six genotypes (SS, SM, MM, ML, LL and SL) were defined. As shown in Figure 6D, the “SS” genotype is significantly more frequent in SM cases compared to UM cases (27.5% vs. 8%, p<0.00001). Conversely, the “SL” genotype is more prevalent in UM cases (35% vs. 21.8%, p = 0.012). Figure 6E depicts the frequency of the 6 genotypes for each disease entity. When dichotomized as described previously [34], [56] into genotypes containing at least one “L” allele (LL, ML, SL = L-carriers; labeled green in Figure 6E) versus non-L carriers (SS, SM, MM; labeled red in Figure 6E), SM patients were 53% less likely to be L carriers than UM cases (OR: 0.47, CI 95%: 0.29 to 0.75, p = 0.002). When L carrier status was analysed in relation to different disease entities using multinomial logistic regression, cases with SRD and SRD plus CM were significantly less likely to be L carriers (85% and 74% less likely, respectively) compared to uncomplicated cases (Table 1). Of note, 9 out of the 10 individuals who succumbed to malaria were non-L carriers, compared to 53.4% non-L carriers within the remaining severe cases (p = 0.043, Fisher's exact test), or 37% non-L carriers within UM cases (p = 0.003, Fisher's exact test). We further examined whether ethnicity was associated with either disease outcome, frequency of L alleles or L allele containing genotypes. We found that ethnicity was not associated with being a SM or UM case (p = 0.294), or with any of the particular disease entities (p = 0.112). There was also no difference in the frequency of the L allele (p = 0.063) or of the L allele-containing genotypes (p = 0.088) among ethnic groups. To determine a possible confounding effect of some of the major factors known to determine disease severity, we measured the frequency of sickle cell trait, blood group O and G6PD deficiency, and the level of HRP-2 in our study population. Hemoglobin S (HbS) confers protection from severe malaria in humans [57], and was recently shown to induce HO-1 in murine hematopoietic cells [58]. Blood group O has been associated repeatedly with reduced risk of severe malaria [59], and so has been G6PD deficiency [60], [61], as hypothesized by Allison [62]. In agreement with two previous studies using samples from this geographic area [37], [63], we confirmed the 968C/376G allele as the most common G6PD A- deficiency allele in our study population (6.26%). The latter study from the Gambia [37] suggested that heterozygous females and hemizygous males are relatively protected from severe disease. The histidine-rich-protein 2 (HRP-2) has been proposed as a surrogate for parasite biomass and is considered to be associated with disease severity [64]. Figure S3 shows the frequency of these factors according to disease group and HMOX1 genotype. In our study population, carriage of the sickle cell trait, G6PDA−, or blood group O were neither associated with disease severity, nor with HMOX1 genotype. As expected, HRP-2 was associated with disease severity, but showed no association with HMOX1 genotypes. To examine whether the genotype of the (GT)n polymorphism was associated with the magnitude of HMOX1 mRNA induction in peripheral blood leucocytes in response to a defined stimulus, the data for hemin and heat-mediated induction of HMOX1 mRNA in whole blood collected on D28 shown in Figure 3B were plotted according to L carrier status (Figure 7). Both L and non L carriers had similar HMOX1 expression at baseline (p = 0.65), and showed a significant 2.5 fold (L carriers, p = 0.0002), or 4.3 fold (non L carriers, p = 0.002) increase in HMOX1 mRNA levels in response to heat. When hemin induced HMOX1 mRNA levels were compared to baseline a 4.7 fold (L carriers, p = 0.0002), or 17.1 fold (non L carriers, p = 0.0039) increase was observed. Importantly, the median HMOX1 mRNA measured in non-L carriers after hemin stimulation was 3.9 fold higher (p = 0.0028) than was observed in L carriers (Figure 7). In response to heat stimulation, non L carriers had 1.9 fold higher mRNA levels compared to L carriers (p = 0.183). After Bonferroni's adjustment for multiple comparisons the significance threshold for these tests becomes 0.007. The differential HMOX1 mRNA expression of L and non L carriers in response to hemin is in line with what has been reported from human lymphoblastoid cell lines treated with H2O2 [33], and demonstrates that in our study population the magnitude of HMOX1 mRNA expression in peripheral blood leucocytes in response to a defined stimulus is associated with the length of the (GT)n repeat. To investigate the possibility that, in West Africans, the (GT)n microsatellite is tightly linked to (or “tagging”) unexplored functional variants in alternative genes, we analyzed the extent of linkage disequilibrium (LD) at and around the HMOX1 locus, using HapMap data for Nigerian Yoruba. In this West African population, the (GT)n microsatellite in the promoter lies in a 6KB LD block that extends from rs2071746 (5′) to rs11912889 (3′), encompassing approximately half of the HMOX1 gene. According to these data, the (GT)n microsatellite would effectively tag only variants within HMOX1, but no other genes. Moreover, the two genes immediately flanking the HMOX1 gene (TOM1 33 kb 5′ and MCM5 6 kb 3′ of HMOX1) are not obvious candidates for malaria susceptibility (Figure S4). Future comparative studies of LD would be ideally conducted in the Gambian ethnic groups, and will benefit from whole genome sequence data likely to emerge from initiatives including the 1000 Genomes Project [65]. The frequency of the long (>32 copy) repeat alleles was similar to that recently reported from Angola [30], but significantly higher than was reported from populations living in areas where malaria is not endemic, such as Europe (French [66], German [67]), and North America (Caucasians, [68]) or Asia (Japanese [34], [36], [69], Karen people from Myanmar [29]; see Table S2 in Text S1). The magnitude of the differences was therefore explored with fixation (FST) indices. Differences between the Gambian and North American (FST = 0.29), or European populations (Gambia vs France, FST = 0.21; vs Germany, FST = 0.14) were slightly higher than the average for a genome-wide sample of polymorphisms (FST = 0.126, [70]), whereas differences between Gambian and Asian populations (FST for Gambia vs Myanmar, or vs Japan) were slightly less than the genome-wide average. In the Angolan population allele frequency of the long repeat alleles was higher than in The Gambia, and therefore slightly more divergent from the non-African population samples (Table S3 in Text S1). When available data from each continent were pooled, the largest FST values were observed for comparisons between Africa and America (FST = 0.33 CI 95%: 0.29 to 0.38), Africa and Europe (FST = 0.25 CI 95%: 0.2 to .29), and Africa and Asia (FST = 0.21 CI 95%: 0.18 to 0.25; Table S4 in Text S1). Inspired by elegant studies in murine malaria models clearly demonstrating that the induction of HO-1 helps prevent severe forms of malaria [9], [10], and the intriguing possibility either to manipulate HO-1 activity pharmaceutically [71], [72] or to mimic its effect by administering CO [20], we explored the role of HO-1 in children with severe and uncomplicated P. falciparum infection. During acute disease, the number of WBC staining positive for HO-1, the HMOX1 mRNA levels, and the HO-1 protein concentrations in plasma were significantly higher than during convalescence, being highest in the most seriously ill patients presenting with SRD. While the association between elevated HO-1 and severe illness we and others [47], [68], [73] observed might merely reflect an appropriate response insufficient in magnitude and/or occurring too late, the association between short (GT)n repeat alleles and increased inducibility of HO-1 in vitro, and more severe disease suggests that HO-1 levels above a certain threshold may be part of the causal pathway leading to severe disease and death. The association between short (GT)n repeat alleles in the HMOX1 gene promoter region (resulting in enhanced HMOX1 mRNA expression) with CM observed in a small study in Myanmar [29], and more recently, in Angola [30] supports this notion. Intriguingly, both our study and the study carried out in Angola [30] observed a distinct peak around 39 (GT)n-repeats, which is in contrast to previous data from populations from non-malaria endemic areas [34], [36], [66], [67], [68], [69]. We have noted that the FST indices for comparisons between these two African populations and those from non-malaria endemic areas were slightly above the average for a genome-wide sample of polymorphisms [70], although more detailed study of polymorphism in this gene would be needed to test a neutral hypothesis. The possibility that the relatively high frequency of long (GT)n repeats in Africa may have resulted from a survival advantage from P. falciparum should encourage investigation to prospect more powerfully for evidence of selection on this locus, given that malaria has been one of the most powerful selective forces acting on the human genome [74]. The fact that increased levels of indirect bilirubin and COHb – both end products of the reaction catalyzed by HO-1 – are widely recognized as independent markers for mortality in critically ill patients [23], [49], and that long (GT)n alleles were associated with less frequent multi organ dysfunction in European ICU patients irrespective of the specific diagnosis [47], make it tempting to speculate that a high HO-1 response is disadvantageous for acute inflammatory conditions in general. In fact, HO-1-induced CO may reduce oxygen carrying capacity in the blood and tissue oxygenation, ultimately leading to metabolic acidosis. Furthermore, increased HO-1 may result in low nitric oxide (NO) levels [75]. This may constitute another pathway by which over-expression of HO-1 contributes to severe disease based on the beneficial effects ascribed to inhaled NO on endothelial function in patients with adult respiratory distress syndrome (ARDS) [76], and the accumulating evidence that depletion of NO contributes to severe malaria [77]. Considering that HO-1 overexpression in the liver leads to an increase in parasite liver load [11], and that the major benefit of RTS,S (a malaria vaccine that partially reduces the parasite burden in the liver) is the reduction of severe disease, it is tempting to speculate that particularly high HO-1 levels in the liver might contribute to severe malaria in man. These findings differ from the role HO-1 plays in preventing severe disease in mice [8], [10]. An attempt to reconcile these observations needs to take into account that mice in contrast to humans lack the (GT)n repeat polymorphism [18]. The functional relevance of the human (GT)n promoter length polymorphism indicated here and elsewhere [33], [56] suggests that humans might have a greater genetically determined variability of HMOX1 expression than exists among inbred mouse strains. The infection of BALB/c mice with P. berghei ANKA, for example, results in a fairly homogeneous 3–4 fold upregulation of HMOX1 mRNA 6 days post infection [9], comparable to what we observed in uncomplicated cases. However, a more than 10 fold difference was measured in the most seriously ill patients of the SRD plus CM group. In fact, HO-1 has both pro-and anti-oxidant properties [24], and dependent on its amount, diametrically opposed effects have been described: Several in vitro studies have shown that moderate (less than 5 fold) induction of HO-1 is associated with protection against heme-mediated damage [78], while higher levels (greater than 10 fold) resulted in loss of cytoprotection [79]. Using HO-1 transfected hamster fibroblasts with either low, moderate, or high HO-1 activity, Suttner et al. demonstrated how HO-1 related cytoprotection turns into HO-1 mediated oxidative injury with increasing HO-1 expression [27]. Importantly, ferrous iron accumulated in high HO-1 expressing cells, and the addition of iron chelators or specific HO-1 inhibitors significantly reduced all measures of oxidative tissue injury [27]. The notion that high levels of HO-1 activity may potentiate, rather than attenuate ROS toxicity, and that this is related to the increased availability of ferrous iron is further supported by several in vitro studies [80], [81], as well as studies in animals [82], [83]. Thus, we hypothesize that, up to a certain level, induction of HO-1 is protective, while excessive upregulation of HO-1 in response to an inflammatory stimulus is deleterious. The clinical relevance of free iron in severe malaria infections has been investigated previously, and high transferrin saturation (which indicates mobilization of ferrous iron) was associated with delayed recovery from coma in CM patients [84]. While HO-1 was not measured in this trial, a more recent study in patients with ARDS established that transferrin saturation increased in parallel to HO-1 [85], strengthening the idea that in vivo high HO-1 levels may result in a clinically relevant increase of ferrous iron. However, results of studies on the usefulness of iron chelation therapy with desferroxiamine in malaria patients have been inconclusive [86], [87]. We also provide mechanistic evidence that hemin-mediated HMOX1 mRNA expression in neutrophils potentiates the magnitude of the neutrophil oxidative burst, and propose that a genetic predisposition to high levels of HO-1 may cause an otherwise protective response to become deleterious. By demonstrating how the neutrophil oxidative burst may influence disease severity, our data help to determine the role of this cell subset in the pathogenesis of severe malaria, which is currently ill-defined. Earlier in vitro studies established that iRBCs can be phagocytosed by neutrophils [88], and can activate them to produce ROS [89], which can kill parasites [50], [90]. In line with this, the amount of ROS produced by neutrophils from children with P. falciparum infection was associated with faster parasite clearance [91], and clinical protection from P falciparum correlated with neutrophil respiratory burst induced by merozoite antigens opsonized by antibodies [92], proposing neutrophils as an efficient defense mechanism. However, in murine models of severe malaria, early depletion of neutrophils prevents experimental CM [93], as well as sequestration of neutrophils to the lungs and reduces mortality [94], demonstrating that neutrophil effector mechanisms are capable of contributing to severe disease. Taken together, this suggests that an early neutrophil oxidative burst may benefit the host by contributing to initial parasite control, while a genetic predisposition towards an enhanced oxidative burst as suggested by our data may result in enhanced damage of endothelial cells, especially in conditions where neutrophils become sequestered in capillaries. Consistent with our data, a genome-wide analysis of the host response to malaria recognized neutrophil-related gene expression responses as the principal pattern distinguishing convalescent from acute malaria patients, and the HMOX1 gene was amongst the genes showing a stepwise increase with increasing severity [14]. The nature of the association between HO-1 in plasma and severe disease, and whether or not soluble HO-1 has a causal role in the pathogenesis of severe disease isn't entirely clear, but we consider it unlikely that HO-1 is functional in plasma. HO-1 is an intracellular enzyme [95], [96], and a molecule transporting HO-1 into the extracellular compartment has not been described. Furthermore, enzymatic functionality of HO-1 requires its C terminal end to be located in the membrane of the endoplasmic reticulum (ER) [97], and several electrons to be provided by an ER-bound NADPH cytochrome p-450 [98], [99]. Like Saukkonen et al. [47], we speculate that plasma HO-1 leaks from damaged tissue, and the association between plasma HO-1 and severe disease is primarily driven by the degree of tissue damage and not the degree of HO-1 induction itself. It is therefore not surprising that the HMOX1 genotypes show no clear association with plasma HO-1 (Figure S5A). Based on the negative association we observed between plasma HO-1 and RBCs or Hb (both can be seen as markers of hemolysis in malaria), and considering that the release of free iron and heme-containing moieties that occurs during hemolysis leads to considerable damage of endothelial cells [48], [100], [101], we propose damaged endothelial cells as an important source of plasma HO-1. This study had several limitations. In order to study HO-1 protein levels in WBCs according to disease entities or (GT)n repeat polymorphisms, flow cytometric examination of blood from more participants would have been required. In view of the results we obtained for the comparison of in vitro stimulated whole blood showing a significant difference in HMOX1 mRNA expression between L and non L carriers in response to a defined amount of hemin (Figure 7), the lack of an association between HMOX1 genotype and HMOX1 mRNA expression ex vivo (Figure S5B) or the lack of a clear difference in HMOX1 mRNA levels between SM and UM cases may be surprising. However, it is important to note that in contrast to the in vitro experiment where the nature and the strength of the stimulus is known, in vivo HMOX1 mRNA expression may be driven by a variety of stimuli. To explore this further, it would have been necessary to measure various factors known to induce HMOX1 mRNA expression in vivo. In this regard, our inability to measure free heme in plasma, known to be a major stimulus for HO-1, was an unforeseen limitation. We can therefore only speculate that inter-individual differences in the nature and amount of HMOX1 mRNA inducing factors may have obscured the effect of the HMOX1 promoter polymorphism on HO-1 levels in vivo. With this caveat in mind, our data do indicate that a genetic factor affecting high HO-1 levels in response to heme is associated with more severe disease and death from malaria. We identified neutrophils as the predominant source of HO-1 in peripheral blood, provide evidence that increasing HMOX1 mRNA expression in these cells enhances the oxidative burst, and suggest that this may constitute a mechanism by which sequestered neutrophils cause tissue damage, thereby contributing to severe pathology. Considering that similar associations between high HO-1 and illness severity have been observed in other conditions [47], [73], limiting HO-1 activity pharmacologically with tin protoporphyrin IX [71] or other inhibitors may be an interesting therapeutic option worth considering. An alternative therapeutic strategy might alter the distribution of HO-1 induction to particular cell types by therapeutic administration of haptoglobin or hemopexin, which both might limit the toxicity of free heme and restrict the uptake of cell free-hemoglobin and heme, and consequently upregulation of HO-1, to those cells bearing receptors for these molecules.

10.1371/journal.pgen.1000069

An Inducible and Reversible Mouse Genetic Rescue System

Inducible and reversible regulation of gene expression is a powerful approach for uncovering gene function. We have established a general method to efficiently produce reversible and inducible gene knockout and rescue in mice. In this system, which we named iKO, the target gene can be turned on and off at will by treating the mice with doxycycline. This method combines two genetically modified mouse lines: a) a KO line with a tetracycline-dependent transactivator replacing the endogenous target gene, and b) a line with a tetracycline-inducible cDNA of the target gene inserted into a tightly regulated (TIGRE) genomic locus, which provides for low basal expression and high inducibility. Such a locus occurs infrequently in the genome and we have developed a method to easily introduce genes into the TIGRE site of mouse embryonic stem (ES) cells by recombinase-mediated insertion. Both KO and TIGRE lines have been engineered for high-throughput, large-scale and cost-effective production of iKO mice. As a proof of concept, we have created iKO mice in the apolipoprotein E (ApoE) gene, which allows for sensitive and quantitative phenotypic analyses. The results demonstrated reversible switching of ApoE transcription, plasma cholesterol levels, and atherosclerosis progression and regression. The iKO system shows stringent regulation and is a versatile genetic system that can easily incorporate other techniques and adapt to a wide range of applications.

We describe a technology for the creation of inducible and reversible gene inactivation in mice. It combines two genetically modified mouse lines: a knock-out line with a tetracycline transactivator replacing the endogenous target gene, and a line in which a tetracycline-inducible cDNA of the target gene has been inserted into a specific genomic locus. A critical component of this system is the unique chromosomal loci we have identified and engineered that offer a platform for easy insertion of any gene of interest for tightly controlled expression. Because of its simple binary nature, allowing independent modification of each of the two components and possibility of use in a high-throughput mode, we believe that our system will be useful for multiple applications, such as introducing mutant or humanized form of the target gene as well as functional manipulating tools. We have applied this technology to the Apolipoprotein E (ApoE) gene and have demonstrated that: a) the expression of ApoE is strictly dependent on the presence of doxycycline, a tetracycline group antibiotic, in the mouse diet, b) in the absence of doxycycline (ApoE repressed) atherosclerotic plaques are formed, confirming the importance of ApoE in the process, and c) upon re-induction of ApoE in the animals with doxicyclin, atherosclerosis regressed.

In the post-genome era, a major challenge is deciphering the function of thousands of newly identified genes. One of the main approaches for studying gene function involves inactivation of genes in cells or animals using random (chemical or insertional) mutagenesis or gene targeting. A common problem with these methods stems from the fact that the gene of interest is usually mutated throughout the animal's life. As a result, 1) in many cases the mutation leads to embryonic or neonatal lethality, precluding the assessment of the gene's function in later life; 2) in viable mutants interpretation of observed phenotypes is often complicated by the inability to distinguish the direct effects of the gene loss at the time of observation from the results of developmental abnormalities caused by the gene loss earlier in life; 3) in still other cases, life-long absence of a gene product causes compensatory adjustments of activities of other genes precluding the elucidation of the function of the gene of interest. Conditional knockout and gene expression technologies, such as the Cre/lox-mediated tissue-specific knockout [1] and the tetracycline (Tet) regulated transcriptional activation system [2], can regulate gene expression in a more spatially and temporally controlled fashion. However, these technologies are often laborious to establish and the results are frequently variable. Here we report the development of a system that provides for the inducible and reversible gene inactivation in the mouse and can also be readily scaled up for high-throughput applications. The iKO system is a binary approach based on the Tet-dependent regulatory technology. It involves the combination of two mouse lines – a KO line that expresses the Tet-transactivator (tTA or rtTA) in place of the gene of interest, and a TIGRE (for tightly regulated) line that contains the gene of interest under the control of the Tet-responsive element (TRE) at a predetermined genomic locus. It has the advantage of, 1) ability to turn genes on or off at will by adding or removing doxycycline (Dox) at any time during the animal's life, thus minimizing embryonic lethality, developmental effects, and compensatory effects; 2) high degree of regulation to any gene inserted at the TIGRE locus, which has been selected to confer minimal basal expression and high inducibility, and to insert any gene of interest in a single step by Cre/loxP recombination; 3) efficiency; the design allowing streamlined production of both KO and TIGRE mice makes it possible to generate iKOs for a large number of genes in a cost-effective manner; 4) flexibility; KO and TIGRE lines can be engineered independently and combined in numerous ways, making a wide range of applications possible. As a proof of concept, we report the characterization of an iKO of the apolipoprotein E gene (ApoE iKO). ApoE plays a key role in regulating cholesterol metabolism and atherosclerosis progression. ApoE KO mice develop hypercholesterolemia and atherosclerosis that closely resemble the human conditions and are rapidly reversed when APOE protein is supplied [3],[4],[5],[6],[7],[8]. Thus, inducible and reversible regulation of ApoE expression could result in rapid physiological changes, which in turn can help assess the iKO technology. Furthermore, the phenotype of ApoE deficiency is quantifiable and very sensitive to leaky expression, allowing for the evaluation of the stringency of gene regulation by iKO technology [9]. Here we demonstrate that in the ApoE iKO mice, ApoE gene expression, as well as blood cholesterol levels, is tightly controlled by Dox. In the presence of Dox, ApoE is expressed and the cholesterol levels are low; in its absence, the reverse is observed. Furthermore, on examination of aortic atherosclerosis in the ApoE iKO mice we found that Dox treatment before the onset of atherosclerotic lesions completely prevented lesion formation and Dox treatment after extensive lesions had already formed resulted in regression of the lesions. These results demonstrate the reversibility of the iKO, leading to phenotype switching within the same animal. ApoE iKO is also useful in its own right as a novel model system for the study of molecular mechanisms underlying atherosclerosis progression and regression. As illustrated in Figure 1A, two genetically modified mouse strains are created. The first is a KO line in which a Tet-dependent transactivator (rtTA in this example) is inserted into the target gene (Gene X). The insertion inactivates Gene X, and places rtTA under the control of the endogenous promoter of Gene X. The KO line can be generated via either homologous recombination or insertional mutagenesis. The second line (TIGRE) contains an additional copy of Gene X cDNA (or genomic fragments) driven by TRE-promoter inserted in a specific locus in the genome (the TIGRE locus), which has been pre-selected for low basal transcriptional activity and high inducibility. When these two lines are crossed, rtTA protein produced from the KO allele can activate the TRE-Gene X in the TIGRE locus only in the presence of Dox. (Alternatively, tTA can be used, which works in the opposite way – Tet-off instead of Tet-on.) Figure 1B illustrates the breeding of KO and TIGRE lines to produce the iKO mouse, which is homozygous for the KO locus and carries one copy of the TIGRE allele. The status of the iKO mouse is regulated by Dox. In the absence of Dox, rtTA protein is produced, but is inactive. As a result, TRE-Gene X is silent, Gene X protein is not produced and the KO phenotype is manifested. In the presence of Dox, rtTA stimulates the synthesis of Gene X protein from the TIGRE locus in the same cells in which the endogenous gene would normally be expressed. Expression complements the missing endogenous Gene X activity and leads to phenotypically normal animals. Thus, one can switch between wild type and KO state of animals by simply adding or removing Dox (e.g. with food). To screen for TIGRE loci, we constructed a Moloney murine leukemia virus (MoMLV)-based retroviral vector, pRTonZ (Figure 2A), in which the TRE-controlled lacZ gene was used as a reporter for gene regulation. Retroviral transduction at low multiplicity of infection ensures integration of a single copy of the TRE-lacZ unit into the genome. pRTonZ contains a modified neomycin phosphotransferase gene, loxneo, in which initiating AUG has been placed upstream of a loxP site in frame with the neo coding region. Once optimal locus is selected, it is utilized as a target site for transgene-integration by the scheme shown in Figures 2B and 2C. The TRE-lacZ unit is removed by Cre-mediated recombination of flanking loxP sites, leaving one loxP site and the neo gene in the genome (Figure 2B). Since the promoter of the loxneo gene as well as the initiating AUG are also removed, ES cells become G418-sensitive. In this configuration, any gene of interest can be introduced into the same locus by Cre/loxP recombination (Figure 2C). Recombinant ES clones can be selected by G418-resistance because the neo expression unit is reconstituted. PCR screening showed that >90% of these G418-resistant clones had correct insertion of the new gene. Optimal loci were initially screened in ES cells. The ES cell line used was derived from CJ7 [10], of 129/Sv background. Following infection with the pRTonZ retroviral vector, G418-resistant (G418r) clones were stained with X-gal (Figure 3A). Most clones showed mosaic staining pattern. By the percentage of the X-gal-stained cell population, ES clones were classified into 4 categories (Figure 3A): class I (<1% of X-gal-stained cells) to class IV (>50% of X-gal-stained cells). From 242 ES clones analyzed, 55 clones were classified into class I. Inducibility was examined by β-galactosidase (β-gal) activity after transfecting 43 class I clones with a tTA expression vector (Figure 3B). We set the cut-off value for high induction level at 1500 µunits/mg of protein and nine clones belonged to this category. Those loci were further examined in mice. Three independent TRE-lacZ mouse lines were generated from class I ES clones T1, T2 and T3 (Figure 3C). Heterozygous TRE-lacZ mice were crossed to MMTV-tTA mice, which has been reported to be transcriptionally active in a wide variety of cell types [11], allowing for the examination of lacZ induction in various tissues. Three genotypes of mice (lacZ(−)tTA(−), lacZ(+)tTA(−), and lacZ(+)tTA(+)) were analyzed for β-gal activity (Figure 3C). Activity of lacZ(−)tTA(−) represents endogenous eukaryotic β-gal activity. Difference between lacZ(−)tTA(−) and lacZ(+)tTA(−) indicates basal activity of lacZ gene in the absence of tTA, and comparison of lacZ(+)tTA(−) and lacZ(+)tTA(+) reveals induction levels in the presence of tTA. Overall, the three mouse lines showed similar pattern of β-gal activity, although some differences were also seen. Basal activities were low but detectable in many tissues, and they were comparable to the values measured in the parental ES clones. β-gal activity was inducible in almost every tissue, and overall induction levels correlated with the expression levels of tTA (Figure 3C, bottom panel). Although the three loci T1, T2 and T3 showed tight regulation of gene expression, basal activity was still detectable in many tissues. This could result from enhancers in the vicinity of the integration sites. To solve this problem, we flanked the TRE-lacZ reporter by the insulator sequence derived from the chicken β-globin locus [12],[13]. The insulator sequences were introduced into all three loci (T1, T2, T3) by Cre-mediated recombination according to the scheme shown in Figures 2B and 2C (see Figure S1 for details) and regulation of the lacZ gene was examined by transient expression of tTA in ES cells (Figure 4A). The insulators reduced basal β-gal activity to levels indistinguishable from wild type ES cells. In contrast, inducibility was not impaired by the insulator sequence, indicating its effectiveness for increasing stringency of gene regulation. The insulators were also introduced into class II, III and IV ES clones, which showed high basal activity (Figure 3A). Although the insulator sequences were effective, basal activity was still clearly detectable in every clone (Figure 4B). To test whether tight regulation is achieved by using other genes, we replaced the lacZ gene with a luciferase gene (Figure S2). With this reporter, insulators reduced basal activity by 14, 20 and 11 fold in T1, T2 and T3 loci, respectively, bringing it close to the instrument detection limit (Figure 4C). Calculation of the number of luciferase molecule per ES cell (Supporting Methods) revealed that on average only 1, 0.7 and 0.3 luciferase molecule were expressed per cell in T1, T2 and T3 loci, respectively (Figure 4D). Importantly, induced levels were not impaired by the insulator (Figure 4C), leading to high induction ratio of luciferase activity. Similar basal activity levels could also be achieved by expressing transrepressor (Figure S3). However, our system is simpler because no additional protein expression is required. To evaluate the insulator effect in vivo, we also generated mice containing TRE-lacZ gene and insulators, or TRE-luciferase (Luc) gene and insulators at the TIGRE locus (T1), and bred them with mice containing tTA under the control of various promoters. Basal expression was examined in multiple tissues of TRE-Luc mice by RT-PCR (Figure 4E). Luc mRNA was undetectable in all tissues examined except testis, indicating very low basal levels of expression throughout the body. To examine if the Luc transcript detected in testis can produce functional Luc proteins, we conducted luciferase activity assays from protein extracts of all these tissues. The luciferase activity in testis was at the similar low basal level as all other tissues examined (data not shown) as well as the original TIGRE ES clone containing TRE-Luc with insulators that had been shown to express approximately one Luc molecule per cell (Figure 4D), suggesting that the RT-PCR band detected in testis resulted from aberrant transcription that did not generate functional protein. To examine induced expression, we used lines of mice with tTA under the control of two brain-specific promoters: α-CaMKII (calcium/calmodulin-dependent protein kinase II) [14] or NSE (neuron-specific enolase) [15]. Sagital sections of 50 µm thickness across the entire brain from TRE-lacZ, PCAMKII-tTA/TRE-lacZ or PNSE-tTA/TRE-lacZ mice were stained for β-galactosidase (gal) activity and representative sections are shown in Figure 4F. In TRE-lacZ mice, no β-gal staining was seen in any parts of the brain. In PCAMKII-tTA/TRE-lacZ or PNSE-tTA/TRE-lacZ mice, intense β-gal staining was observed in specific regions of the brain defined by the two promoters respectively. These results demonstrate tight control of the TIGRE locus in animals. Using the splinkerette PCR method [16], we obtained genomic fragments covering either 5′ or 3′ junctions of the TIGRE vector insertion site in the T1 ES cell line. After sequencing the fragments, we determined the precise integration site of the viral TIGRE vector in the T1 TIGRE locus, which is located on chromosome 9. Genomic sequences surrounding the T1 TIGRE locus are shown in Figure 5A. Characteristic of retrovirus insertions, four nucleotides immediately adjacent to the insertion were duplicated and the viral TIGRE vector was inserted exactly in between the duplication. BLAT search of the UCSC Mouse Genome Browser with genomic sequences surrounding T1 revealed the localization of T1 locus to chr9 qA3 (Figure 5B). The insertion site is flanked by two genes: AB124611 and Carm1. The insertion site is located 3′ to the hypothetical gene AB124611 with unknown function and undetermined polyA site. The insertion site is also ∼1.5 kb upstream of the transcriptional start of Carm1. Carm1 is ubiquitously expressed and Carm1−/− mice are embryonic lethal [17]. However, we have not observed overt developmental or other abnormalities in heterozygous or homozygous T1 TIGRE reporter lines TRE-lacZ or TRE-Luc (both with insulators), indicating that the viral insertion did not disrupt the nearby Carm1 gene. KO lines with target genes replaced by rtTA or tTA can be produced by any gene targeting or insertional mutagenic methods. To implement high-throughput production of iKO mice, we have utilized a large-scale insertional mutagenesis ES cell library developed in house for the KO production [18]. Figure 6A upper panel illustrates the structure of the retroviral vector used. In the particular case of ApoE, the virus is inserted in the third intron. The vector contains splice acceptor, stop codons, polyA signal and transcriptional terminator to ensure gene inactivation, which we confirmed to be the case by showing that ∼90% of isolated gene-specific ES clones were null alleles [18]. (The remaining were mostly knock-downs, and in nearly all these cases the retrovirus was inserted into 5′ UTR (exon or intron), suggesting that retroviral insertions upstream of the coding regions of genes should be avoided.) The vector also includes the rtTA gene immediately downstream of the splice acceptor, stop codons and internal ribosome entry site (IRES), so that rtTA protein can be synthesized from the ApoE-IRES-rtTA hybrid transcript. To examine if random insertion of the retroviral vector containing rtTA into a gene results in rtTA expression reflecting the expression patterns of the inactivated gene, we compared rtTA transcription with the transcription of the endogenous gene by RT-PCR in different tissues for 26 different G protein coupled receptor (GPCR) KO lines generated from the library. Heterozygous mice carrying one allele of the intact endogenous gene and one allele interrupted by the rtTA-bearing retroviral vector were used to prepare total RNA samples from different tissues and amplify gene-specific transcript and rtTA transcript from the same RNA preps for side-by-side comparison. Figure 6B shows three examples of such comparison for genes P2Y6, RE2 and LGR6 respectively. The retroviral vector was inserted into a different part of each of the three genes – a 5′UTR intron of P2Y6, an intron within the coding region of RE2 and the 15th coding exon of LGR6. In all three cases, rtTA expression profiles closely resemble those of the endogenous genes to be inactivated. Real-time qPCR for 16 tissues (Table S1) also showed high degree of correlation between rtTA and endogenous gene from tissue to tissue. Overall, out of the 26 lines examined, 19 lines showed good correlation between rtTA and the endogenous gene's expression (Table S2), i.e. rtTA is expressed in the tissues where the corresponding endogenous gene is expressed and the relative ratio across different tissues for each transcript also appears similar between rtTA and endogenous gene. In the remaining 7 lines in which rtTA was not expressed well, 5 lines had the retroviral vector inserted upstream of the first (and usually fairly large) intron of each gene. It has been recognized that the first intron (especially if it is large) could contain essential transcriptional regulatory elements, and thus insertions in this region might disrupt the transcriptional regulation and so should be avoided. Otherwise, our data showed that in the majority of insertional sites, rtTA could be expressed through the endogenous promoter upon integration via the retroviral vector. To prove the concept of iKO system, we generated ApoE iKO mice by creating ApoE KO line and ApoE TIGRE line separately and breeding them together (Figure 6A), and attempted to model human conditions such as hypercholesterolemia and atherosclerosis. ApoE KO line was created by screening the mutant ES cell library as mentioned above. ApoE TIGRE line was created by inserting a TRE-ApoE transgene, flanked with 4 copies of chicken β-globin insulators, into the T1 TIGRE locus. RT-PCR analysis of heterozygous mice (ApoE+/−; TRE-ApoE) showed that expression of TRE-ApoE was strictly dependent on Dox (Figure 6C). Real-time qPCR using primers specific for the endogenous ApoE or TRE-ApoE (2 sets of primers for each gene) showed that endogenous ApoE mRNA level is very high (compared to 18s rRNA), and the TRE-ApoE mRNA level in the presence of Dox is ∼6.7 fold lower than the endogenous ApoE (ΔCt = 2.7 between TRE-ApoE and endogenous ApoE), while in the absence of Dox TRE-ApoE is >55,000 fold lower than the endogenous ApoE (ΔCt>15.8). This shows the TRE-ApoE induction by Dox is >8,000 fold. Next, we carried out phenotypic analysis of homozygous ApoE iKO mice (ApoE−/−; TRE-ApoE), i.e. mice having both endogenous ApoE alleles inactivated and carrying TRE-ApoE in the TIGRE locus. We analyzed blood cholesterol levels in these mice in the absence and presence of Dox. Constitutive KO (i.e. ApoE−/− without TRE-ApoE in the TIGRE locus) and WT group mice that were littermates of iKO were used as controls. As shown in Figure 7A, in the absence of Dox the iKO mice had high cholesterol levels similar to that of the KO mice; in the presence of Dox, the iKO showed normal cholesterol levels, demonstrating that inducible expression of ApoE can lead to the reversion of the KO phenotype of hypercholesterolemia. When Dox was withdrawn, the cholesterol levels in the iKO mice rose again. These on/off switches occurred rapidly, within a few days after Dox administration or withdrawal. We examined the atherosclerotic lesion formation in the aortas of the ApoE iKO mice. As controls we used KO and WT group mice that were littermates of iKO. In the absence of ApoE protein, aortic atherosclerotic lesions start to form around 3–4 months of age and progress continually with time. One set of iKO, KO and WT group mice were treated with Dox-containing food throughout their life, and were compared with mice fed with normal food. Figure 7B shows the atherosclerotic lesions formed around the arch region of the aorta, as visualized by Sudan IV staining. By 7 months of age, extensive lesions had formed in KO mice, regardless of whether they were treated with Dox or not, whereas WT mice did not have any lesions in the absence (Figure 7B) or presence (data not shown) of Dox. The iKO mice developed extensive lesions in the absence of Dox (similar to KO mice), whereas in the presence of Dox no lesions had formed. We further investigated what happens if ApoE protein expression is turned on after the atherosclerotic lesions have already formed. ApoE iKO and KO mice of 5 months of age were switched from normal food to Dox-containing food for the next 4 months. Aortic atherosclerotic lesions were examined before (at 5 months) and after (at 9 months) the Dox treatment. As shown in Figure 7C, at 5 months of age, both iKO and KO had developed similar levels of lesions. After 4 months of Dox treatment, the lesions in KO mice continued to grow, whereas in the iKO mice, the lesions had regressed nearly completely with only scar-like tissues remaining, suggesting that the lipid-containing foam cells have disappeared from the lesions. The results were verified by quantification and statistical analysis of the lesions by one-way ANOVA followed by Neuman-Keul's post hoc test (Figure 7D). We have developed a reversible and inducible rescue system for gene KO in mice and have applied this method into the ApoE gene. Our system complements existing inducible gene expression approaches and provides certain advantages. The tamoxifen-dependent Cre-ERT2 [19],[20] recombination can drive inducible knockout of the endogenous gene, however, it is irreversible and the efficiency of tamoxifen inducibility throughout the body is yet to be demonstrated. The usefulness of the Tet-inducible system [21],[22] is critically dependent on the tightness of transcription induction and suppression. When either the Tet-transactivator or the target gene is randomly integrated into the genome, they are subjects to positional effects [23]. It is often necessary to screen through multiple transgenic lines every time a new transgene is introduced. Attempts to target both tTA and TRE together into an endogenous gene locus to achieve inducible activation and inactivation were successful in a few cases [24],[25],[26], but this method is not applicable to most other genes as the TRE is easily subject to activation from nearby enhancers independent from tTA. Our iKO technique utilizes each gene's own promoter to direct the expression of transcriptional activators, e.g. rtTA and tTA. Therefore it is not limited to certain tissues and available tissue-specific promoter driven transgenic lines. It could be applicable to any gene in any tissue. Dox-regulated expression can be turned on and off rapidly, i.e. within a few days, and at any time in development or adult. It also allows analysis of the effects of gene inactivation in the same animal. The TIGRE locus has been selected to confer little or no basal transgene expression throughout the body while maintaining high inducibility, enabling stringent control of the on/off switching of the target gene. Our system also has the flexibility to allow for further improvements. For example, in some cases it may be more desirable to put the genomic copy of the gene under TRE promoter into the TIGRE locus instead of the cDNA to more precisely mimic the expression of the endogenous gene. Also, the IRES we have used to drive rtTA translation may not work uniformly well in all tissues, and can be replaced by other approaches such as using the viral 2A-like sequences for bicistronic translation [27] or direct targeting of tTA/rtTA into the ATG start codon of the endogenous gene. Unique chromosomal loci for predicted gene expression provide fundamental tools for genetic studies. The most widely used is the ROSA26 locus [28] in which ubiquitous gene expression is achieved by endogenous promoter activity of this locus. The TIGRE loci identified in this study allow a different mode of gene regulation – they were selected from hundreds of insertion sites for tight gene regulation by an exogenous promoter. Therefore, the TIGRE loci offer a platform for easy insertion of any gene in a tightly regulated locus, applicable to not only the tetracycline system but also other gene expression systems utilizing exogenous promoters such as constitutively active promoters, tissue specific promoters, or other inducible promoters regulated by reagents such as ecdysone [29], mifepristone [30] and streptogramin [31]. Recently a similar approach was also applied to a human fibrosarcoma cell line to pre-screen optimal integration sites for transgenes [32]. In addition, here we demonstrated that the stringency of the regulation at TIGRE loci is further enhanced by the incorporation of insulators, i.e. basal expression level was reduced to less than one luciferase molecule per cell without impairing inducibility (Figure 4C, D). Insulators were also shown previously to improve inducibility of randomly integrated TRE-reporters [33]. It should be noted that in our study the insulator effect was limited in many other loci (Figure 4A versus 4B), demonstrating the uniqueness of the TIGRE loci. The genomic location of the T1 TIGRE locus, which was most extensively used in our study including the ApoE iKO mice, was determined (Figure 5). Further manipulation of this locus would be possible to expand its application. The stringent gene regulation is demonstrated in the ApoE iKO mice. It is known that regulation of blood cholesterol levels is very sensitive to the plasma ApoE protein levels. Even with the production of 3% of wild-type level of ApoE protein, the hypercholesterolemia and atherosclerosis phenotypes of the ApoE KO mice can be reversed [9]. Given this high sensitivity, the fact that our ApoE iKO mice in the uninduced state (i.e. in the absence of Dox) exhibit similarly high levels of cholesterol compared to ApoE KO mice indicates that there is hardly any expression of functional APOE. Reversing the KO at will is a particularly powerful approach in atherosclerotic regression studies. Dox-treatment of ApoE iKO mice results in expression of ApoE, marked reduction of plasma cholesterol levels, and regression of aortic atherosclerotic lesions. These findings are consistent with previous studies showing that aggressive lipid lowering or expression of ApoE can induce regression of pre-existing atherosclerotic lesion [7],[8],[34],[35]. It is becoming increasingly clear that lesion regression is regulated by a complex interplay between lipids, inflammation and the immune system [35]. The ApoE iKO mice will allow detailed studies on the roles of specific genes in these complex interactions. The binary nature of the iKO system is inherently simple, with the KO line serving dual roles: it could be used as a constitutive KO or combined with the TIGRE line to produce an inducible and reversible iKO. The 10-million clone ES cell library we utilized [18] has been estimated to contain insertional mutations for >90% of genes, and individual ES clones with retroviral insertions in a specific target gene can be rapidly identified through a PCR pooling strategy and subsequently isolated in a streamlined process. The identification and modification of the TIGRE locus allows rapid insertion of any gene of interest via co-transfection with Cre. Therefore, each component of the binary system, the KO or the TIGRE line, is amenable for high-throughput production to generate inducible and reversible KOs for a large number of genes. It should be noted that the iKO system may not be applicable in certain situations where highly stringent gene regulation is required. For example, even though the system had enough stringency in low basal activity and high induction, the induced gene expression level is usually not exactly the same as the endogenous gene's level and this could be a problem for genes that are highly sensitive to gene dosage effect or show haploid deficiency. The kinetics of the system (days) may be also too slow for some developmental problems where transient expression of developmental genes are critical, although it should be noted that it could still work well in carefully thought-out developmental studies (as in [24]). In addition, it has been reported that rtTA often can not induce sufficient gene expression in the brain, at least partially due to developmental inactivation of the TRE promoter in neurons [36]. It appears that the tTA (Tet-off) system is better suited for the use in brain [14], as substantial β-gal induction was observed in brain with tTA (Figure 4F). Our results of the ApoE iKO mice and the quantitative data using lacZ and luciferase reporters suggest that the iKO system could be a useful tool in addressing a variety of biological questions. The two components of the iKO system can be independently modified, and pairing of their different forms can generate numerous combinations. In KO lines, genes with unique expression patterns are tagged with a transcription transactivator, which can control, in an inducible fashion, the expression of a variety of genes derived from TIGRE lines, enabling a number of additional applications in specific types of tissues or cells. Those include: 1) introducing mutant forms of the target gene into the TIGRE locus for better functional probing of different domains, splice variants, post-transcriptional modifications (e.g. phosphorylation), or modeling human diseases; 2) humanizing target genes by placing a human ortholog of the mouse gene under TRE control, which can facilitate drug efficacy studies; 3) placing a cytotoxic gene under TRE control to allow inducible cell-type specific ablation; 4) introducing a marker gene such as GFP, protein interacting probe, or transneuronal tracer, into the TIGRE locus for cell-type specific tagging, functional analysis, isolation of specific cell population, or mapping neuronal networks; 5) combining with recently developed RNAi techniques [37],[38],[39] to down-regulate any genes of interest in a tissue-specific and inducible manner; 6) re-engineering the TIGRE locus to place a TRE-driven Cre, and combining it with a tissue-specific rtTA or tTA line and floxed target genes to achieve inducible region-specific gene knockout [40]. See Text S1 online for a more detailed description of the methods. Construction of ES cell library infected with the retrovirus and screening and isolation of ES clones with genes of interest inactivated by the viral insertion is described previously [18]. Specifically for ApoE, insertions were found by nested PCR analysis of the ES DNA using two vector-specific and gene-specific primer pairs. The ApoE-specific primers are antisense and located in the fourth exon. Several independent retroviral insertions in the ApoE gene were identified. PCR fragments were sequenced to confirm insertion into the gene. A library tube with a clone of interest identified by the PCR also contains a few hundred other ES cell clones. The sole desired clone was isolated from the mixture by three rounds of cell sorting and growing followed by PCR using the same pair of primers to identify positive clones. Additional PCRs using ApoE primers located in the third intron flanking the viral insertion site were conducted to confirm the precision of the insertion and integrity of the genomic sequence of ApoE. Full length cDNAs for the coding sequences of target genes were cloned into the TIGRE-targeting vector containing TRE, insulators, a PGK promoter and a pair of loxP sites (Figure 5A lower panel). The TIGRE-targeting vectors were co-transfected with a Cre-expressing plasmid into the neo-sensitive ES cells carrying the minimal TIGRE locus with a single loxP site and the promoterless, ATGless loxneo marker. When the TIGRE-targeting vector is integrated into the TIGRE locus through Cre/lox-mediated recombination, neo-resistance is restored to the ES cells by the addition of PGK promoter and in-frame fusion of ATG to the loxneo marker. Correctly integrated ES clones were identified and confirmed by PCR screening and southern blot analysis. ES cell clones were injected into blastocysts of C57BL/6J mice following standard techniques. Chimeric mice were bred with C57BL/6J mice to test germline transmission and generate heterozygous mice. Mice from the corresponding KO lines and TIGRE lines were crossed with each other to produce inducible KO mice according to the scheme shown in Figure 1B. All mice used for studies in this paper were in a mixed genetic background of 50% 129S1/SvImJ and 50% C57BL/6. For the ApoE KO line, southern blot using rtTA coding sequence as probe confirmed the correct insertion of the retroviral vector into the endogenous ApoE gene. It also revealed an additional retroviral insertion somewhere else in the genome. The additional insertion was selectively bred out, and the ApoE iKO colony was maintained with a single retroviral insertion at the ApoE locus. MMTV-tTA, PCAMKII-tTA and PNSE-tTA mice were purchased from The Jackson Laboratory (Bar Harbor, ME). All experimental procedures were approved by the Institutional Animal Care and Use Committee of NCI and Nura, Inc. in accordance with NIH guidelines. Dox was administered to the mice through Dox-containing food, which was custom made by Bio-Serv (Frenchtown, NJ) to contain 2 g Dox per kilogram of food. The nutrition content of the Dox food (e.g. 19% protein and 8.6% fat) was very similar to the regular diet (20% protein, 9% fat) used in our colony. Therefore the switching of food types did not result in change of cholesterol levels in mice.

10.1371/journal.pcbi.1005924

Mathematical modeling identifies optimum lapatinib dosing schedules for the treatment of glioblastoma patients

Human primary glioblastomas (GBM) often harbor mutations within the epidermal growth factor receptor (EGFR). Treatment of EGFR-mutant GBM cell lines with the EGFR/HER2 tyrosine kinase inhibitor lapatinib can effectively induce cell death in these models. However, EGFR inhibitors have shown little efficacy in the clinic, partly because of inappropriate dosing. Here, we developed a computational approach to model the in vitro cellular dynamics of the EGFR-mutant cell line SF268 in response to different lapatinib concentrations and dosing schedules. We then used this approach to identify an effective treatment strategy within the clinical toxicity limits of lapatinib, and developed a partial differential equation modeling approach to study the in vivo GBM treatment response by taking into account the heterogeneous and diffusive nature of the disease. Despite the inability of lapatinib to induce tumor regressions with a continuous daily schedule, our modeling approach consistently predicts that continuous dosing remains the best clinically feasible strategy for slowing down tumor growth and lowering overall tumor burden, compared to pulsatile schedules currently known to be tolerated, even when considering drug resistance, reduced lapatinib tumor concentrations due to the blood brain barrier, and the phenotypic switch from proliferative to migratory cell phenotypes that occurs in hypoxic microenvironments. Our mathematical modeling and statistical analysis platform provides a rational method for comparing treatment schedules in search for optimal dosing strategies for glioblastoma and other cancer types.

In vivo inhibition of tumor expansion requires a sufficient amount of therapeutic agent to be present in the tumor tissue. A number of factors affect drug concentrations including the maximum tolerated dose, pharmacokinetics and pharmacodynamics profiles. We present a computational modeling platform incorporating both in vitro data and published clinical trial data to investigate the efficacy of lapatinib as a function of different dosing schedules for inhibiting glioblastoma tumor cell growth. The goal of our method is to find the best dosing schedule balancing both toxicity and efficacy. Our modeling approach identifies continuous dosing as the best clinically feasible strategy for slowing down tumor growth even when taking into consideration intratumor heterogeneity, drug resistance and reduced lapatinib concentrations in the tumor due to the blood brain barrier.

Glioblastoma is the most common and aggressive form of brain tumors in adults, characterized by short survival and poor treatment response [1]. Currently, the standard of care for glioblastoma patients includes surgery followed by radiotherapy and adjuvant chemotherapy with temozolomide [2]. However, the addition of chemotherapy only modestly prolongs survival (median 14.6 months) compared to radiation alone (median 12.1 months). Thus, there remains a pressing unmet medical need for more effective therapeutic agents. Unfortunately, since the introduction of temozolomide, no other compound has been able to significantly prolong patient survival in clinical trials. For orally administered drugs, most trials have only explored daily continuous dosing schedules (Table 1). However, there is increasing evidence that for some targeted agents, intermittent schedules can deliver equal or potentially even superior therapeutic benefit with less toxicity [3, 4]. In the last decade, several molecularly targeted agents that inhibit recurrently mutated proteins have been investigated as a therapeutic strategy in glioblastoma. These have included several inhibitors of the epidermal growth factor receptor (EFGR), which is mutationally activated in approximately 50% of adult GBMs. Lapatinib is a small molecule tyrosine kinase inhibitor of human epidermal receptor 2 (HER2) and EGFR, which currently has regulatory approval for the treatment of HER2-positive advanced or metastatic breast cancer [5]. Additional indications for lapatinib [6–14] including glioblastoma, are currently being explored [15, 16]. In GBM, clinical trials of lapatinib have failed to show efficacy using continuous dosing [15, 16]. Interestingly, a study that evaluated EGFR inhibition as a means to prime tumor vasculature for efficient delivery of chemotherapy showed that in glioblastoma patients, a 2-day pulse of high dose (5,250 mg/day) lapatinib given through twice daily dosing was well tolerated [17]. However, the question remains whether an altered dosing strategy might increase the efficacy of this agent. Due to ethical concerns of testing several dosing strategies in the absence of preclinical data suggesting their benefit, as well as to speed up discovery, mathematical modeling of treatment response can be used to identify predicted best administration schedules. Here, we explored continuous and pulsatile dosing strategies in an EGFR-mutant GBM cell line, and used mathematical and statistical modeling to determine optimal lapatinib dosing schedules for inhibiting tumor growth. The use of mathematical models for treatment optimization is part of a growing effort to improve clinical trial design for cancer patients. Many models have been developed to investigate the relationship between toxicity and dose in an attempt to better characterize the toxicity profile of the therapeutic products under investigation. For instance, Huitema et al. reviewed several differential equation-based models of chemotherapy-induced myelosuppression, cardiovascular events and other ordinal adverse events [18]. Amantea et al. developed a model incorporating exposure, biomarkers, efficacy endpoints and adverse events in models describing the treatment response of patients with gastrointestinal stromal tumors [19]. Their analysis uncovered a correlation between adverse events and efficacy: drug-induced increases in diastolic blood pressure were positively associated with overall survival. Another study by Huitema et al. examined the effects of the anti-angiogenic drug E7080 and its drug-induced adverse events in an attempt to evaluate dosing regimens with regard to their reduction of adverse events and improvement of efficacy [20]. They found that proteinuria could be best described by a discrete-time Markov transition model. Similarly, Fuhr et al. constructed a continuous-time Markov model to investigate erlotinib-induced adverse events in non-small cell lung cancer patients; their simulation results provided support for the use of high-dose pulses as an alternative dosing strategy for addressing acquired resistance often found when using a low continuous dose [21]. In these examples, a delicate balance between toxicity and efficacy determines the most efficacious dose. In addition to understanding toxicity profiles, mathematical models have also been used to model pre-clinical data from animal or in vitro systems in order to bypass the difficulty associated with directly observing tumor progression in humans. Translational models incorporating in vitro or animal data have also be developed to investigate the relationship between efficacy endpoints and exposure. Mould et al. provided a comprehensive review of how the use of mathematical models can aid early development of anti-cancer therapy, in particular for describing tumor volume as a function of drug exposure [22]. Most papers these authors discussed describe changing tumor volumes using differential equation models with two terms: net tumor growth in the absence of therapeutic candidates and a drug-induced shrinkage effect [23–27]. These models suggest that the best strategy for inhibiting tumor growth is to maximize the effect of drug exposure, which often corresponds to higher drug concentrations. However, given toxicity constraints, high drug concentrations usually cannot be achieved for long time periods. An alternative strategy is to administer drugs in pulsatile doses in order to circumvent the toxicity limit. A small trial using weekly pulsatile high-dose erlotinib showed promising results for controlling central nervous system metastases from epidermal growth factor receptor-mutant non-small cell lung cancer [28]. Crooke et al. also performed computational analyses comparing continuous and pulsatile dosing and found that pulsed therapy is more effective than continuous therapy [29]. A clinical trial [30] based on mathematical modeling and preclinical [31] experiments demonstrated that a combined low-dose continuous and high-dose pulsed erlotinib schedule was successful at preventing progression in patients with CNS metastases but did not show significantly delayed emergence of resistance due to the T790M EGFR mutation. Here we present a computational modeling approach incorporating both toxicity data from phase I clinical trials and efficacy data from pre-clinical research to examine the effects of various lapatinib dosing schedules on tumor growth (Fig 1). The goal of our work is to determine the best dosing strategy given the observed toxicity limit. Specifically, we are interested in answering the question whether alternative dosing strategies can be applied to circumvent the toxicity limit while exceeding the level of efficacy observed in continuous dosing of lapatinib while taking into account known characteristics of GBM growth and treatment response such as diffusivity, intratumor heterogeneity, and the blood brain barrier. SF268 cells were obtained from the National Cancer Institute (NIH) and routinely grown in DMEM supplemented with 10% fetal bovine serum. To assess the effects of drug treatment, 150,000 cells were seeded on 6-cm dishes and allowed to attach overnight. Cells were then switched to growth media supplemented with 1% fetal bovine serum and the indicated concentrations of lapatinib or DMSO (vehicle). Each treatment condition was done in triplicate. For treatment discontinuation, cells were washed 3 times with media containing vehicle, and then allowed to continue growth in media containing vehicle. Vehicle-treated cells were also washed 3 times to control for the effects of washing. Cell viability and cell death were evaluated by the trypan blue exclusion assay using a ViCell cell viability analyzer (Beckman Coulter). We first modeled the in vitro cell pharmacodynamics during lapatinib treatment using a logistic ordinary differential equation (ODE) model (Eq 1). The choice of a logistic model over simpler exponential models was made based on exploratory analysis of viable cell dynamics over time (Fig 2A and 2B). During treatment with low concentrations of lapatinib, the growth rate of cells leveled off after 5 days; this change in the growth rate as a function of the cell number is better captured by a logistic than an exponential model. Note that the birth, death, and clearance rates and carrying capacity are restricted to be non-negative real numbers. The carrying capacity, K, is assumed to be invariant to lapatinib concentrations and is estimated as the maximum observed number of viable cells on day 5 in the absence of lapatinib. The birth, death, and clearance rate parameters b, d, and c, respectively, are functions of the lapatinib concentration; however, the exact functional form describing the relationship between these rates and the lapatinib dose is unknown. To estimate the birth, death, and clearance rates that best capture cell growth patterns, we implemented a grid search algorithm (Eq 2) to minimize the squared distance between observed and predicted cell numbers for each experimental setting in order to determine the parameters that best reproduce observed cell numbers. The patterns of birth, death, and clearance rates against lapatinib concentration are shown in S1A–S1C Fig. We observed that, when allowing the clearance rates to vary for different concentrations of lapatinib, only the death rates show a discernible pattern with increasing lapatinib dose (S1B Fig). However, if the clearance rate is constrained to a constant, we observed that both birth and death rates demonstrated clear dose-dependent correlations (S1D–S1F Fig). We thus adopted the latter approach, leading to birth rates exponentially decreasing and death rates exponentially increasing with escalating lapatinib concentrations. To further investigate the effects of a constant clearance rate, we studied a wide range of values for the clearance rate (0.0, 0.1, 0.2, 0.3, 0.4, and 0.5), and found only minor differences in the birth and death rates estimated for different clearance rates. We therefore selected the simplest assumption with a clearance rate of 0, which reduces the final model to Eq (1), with c = 0 in Eq (1b). We then modeled the relationship between birth and death rates and lapatinib dose as exponential distributions (Eq 3). We estimated the distribution coefficients for birth and death rates using nonlinear least squares regression, which are shown in Table 2. We found that birth and death rates are both significantly associated with lapatinib dose (pbirth = 1.204 × 10−5, pdeath = 3.351 × 10−6). Once we established the relationship between the birth and death rates and the lapatinib concentration, we modeled the growth trajectory for different lapatinib concentrations over time (Fig 2C and 2D) using the model described in Eq 1. We found that the number of viable cells predicted by our model agrees with the observed numbers of viable cells on day 5 (Fig 2E). The model also predicts a clear negative correlation between the lapatinib concentration and the number of viable cells at day 5; however, our model agrees less well with the observed number of dead cells on day 5 (Fig 2F). One possible reason for the discrepancy is that the number of observed dead cells in much smaller than the number of observed viable cells, and hence birth and death rates were determined predominantly using the observed viable cells. Note that our approach does not capture the peak in the number of dead cells on day 3; we believe that this peak does not represent true underlying biology but rather represents a technical artifact since, if we were to include the peak in our analysis, we would estimate growth and death rates that are inconsistent with the know action of anti-cancer agents (i.e. increasing doses lead to lower proliferation and/or enhanced death). Designing a model to capture this peak would likely lead to overfitting and hence decreased predictive abilities. For these reasons we decided to use the approach outlined above to describe the cell growth trajectories in the presence of lapatinib treatment. To identify optimum treatment schedules for the delivery of lapatinib to patients, our approach needs to take into account clinically determined toxicity constraints—the maximally tolerated dose (MTD) for a particular time interval that does not lead to dose-limiting side effects. The maximum allowable oral dose as a function of the number of treated days was constructed from two clinical studies: Thiessen et al. [16] and Chien et al. [17]; a third point was selected based on clinical expertise in order to estimate the shape of this function (Fig 3A). The three toxicity limits we then fitted to both a linear (Eq 4) and an exponential (Eq 5) model (Table 2). We found that the exponential model better explains the observed limiting toxicity as a function of the number of treatment days (AIC values: linear 49.48 vs. exponential 5.712). Limiting doses were extrapolated based on the exponential model for 1 to 5 treatment days in a 5-day treatment cycle. In addition to modeling tumor cell dynamics in the presence of lapatinib treatment, our approach also requires a model of lapatinib pharmacokinetics—the dynamics of lapatinib absorption into the blood stream, absorption into the tumor, and clearance from the tumor. We first modeled lapatinib absorption into the blood stream using pharmacokinetic parameters from clinical trials (Table 4 in Burris et al [37] and Table 4B in Chien et al. [17]). We then fit a logarithmic function (Eq 6) to this data to describe the relationship between the limiting oral doses identified above to the maximum plasma concentration (Fig 3B, Table 2). To convert the plasma concentration to tumor concentration, we used patient pharmacokinetic data reported in Table S4 in Vivanco et al. [32]. In this study, the authors found the mean tumor/serum concentration to be 0.61 in biopsied GBMs following 1 week of daily lapatinib treatment. The resulting serum to tumor conversion function is described in Eq 7. Finally, we constructed a time-dependent pharmacokinetic model using parameters Cmax = 4 hours, i.e. the time to reach the maximum serum concentration after drug administration, and T1/2 = 24 hours, i.e. the time to reach half of the maximum concentration after Cmax [38]. A linear function was used to model lapatinib uptake from 0 to 4 hours, and an exponential function was used to model the subsequent decay pattern, with a half-life of 24 hours (Fig 3C). Combining tumor cell dynamics with patient derived lapatinib pharmacokinetics provides the in vivo tumor growth trajectory—the basis for comparing different dosing strategies. In particular, we compared five treatment schedules that provide the MTD per day according to the identified toxicity constraints, including one continuous dosing strategy and four pulsatile dosing strategies (Table 3). Here we have presented a computational analysis platform of human GBM growth and treatment response, parameterized using in vitro tumor cell data measured based on the SF268 GBM cell line during lapatinib exposure. Our data demonstrate that lapatinib concentrations are negatively correlated with cell birth rates, but positively correlated with death rates. However, the necessary lapatinib concentration needed to arrest further tumor expansion is calculated to be 2,180nM in vitro, a concentration that cannot be reached in GBM patients based on current toxicity limits and oral absorption profiles. To achieve this concentration, our absorption model predicts that an oral dose of 34,000mg is required—5 times greater than the highest dose ever tested. The therapeutic potential seems to be restricted by the transport of lapatinib from the gastrointestinal tract to tumor sites. Due to the logarithmic relationship between oral intake and serum concentration, a higher than tolerated dose is required to achieve tumor reduction. Other methods of delivery, such as intravenous injection, may need to be considered to bypass the current delivery problem. This result agrees with the conclusions from early clinical trials that lapatinib failed to reduce tumor sizes [15], [16]. Despite these negative results, our model suggests that continuous lapatinib dosing does have positive effects. Continuous dosing was able to reduce long-term tumor burden significantly more than pulsatile treatment schedules. These findings were consistent with incorporating intratumor heterogeneity due to the presence of drug-resistance cells, variable penetration of the blood brain barrier leading to a reduction of lapatinib concentration in the tumor, and tumor cell motility. Optimization of dosing strategies for cancer treatments is a complex problem. Our mathematical model provides a quantitative relationship bridging in vitro tumor dynamics and in vivo lapatinib dose and schedule, taking into account the blood brain barrier, GBM motility, heterogeneity, and phenotypic changes that occur in hypoxic microenvironments. Our approach allows for the systematic comparison between different treatment strategies and their effects on tumor growth. Although none of the tolerable doses and schedules evaluated were found capable of arresting tumor expansion, our approach predicts that continuous treatment within these parameters is likely to be more successful at slowing down tumor growth. Further optimization of effective dosing strategies using our computational platform will benefit from additional dose-limiting toxicity studies that incorporate multiple pulsatile schedules.

10.1371/journal.ppat.1005753

Rabies Internalizes into Primary Peripheral Neurons via Clathrin Coated Pits and Requires Fusion at the Cell Body

The single glycoprotein (G) of rabies virus (RABV) dictates all viral entry steps from receptor engagement to membrane fusion. To study the uptake of RABV into primary neuronal cells in culture, we generated a recombinant vesicular stomatitis virus in which the G protein was replaced with that of the neurotropic RABV CVS-11 strain (rVSV CVS G). Using microfluidic compartmentalized culture, we examined the uptake of single virions into the termini of primary neurons of the dorsal root ganglion and ventral spinal cord. By pharmacologically disrupting endocytosis at the distal neurites, we demonstrate that rVSV CVS G uptake and infection are dependent on dynamin. Imaging of single virion uptake with fluorescent endocytic markers further identifies endocytosis via clathrin-coated pits as the predominant internalization mechanism. Transmission electron micrographs also reveal the presence of viral particles in vesicular structures consistent with incompletely coated clathrin pits. This work extends our previous findings of clathrin-mediated uptake of RABV into epithelial cells to two neuronal subtypes involved in rabies infection in vivo. Chemical perturbation of endosomal acidification in the neurite or somal compartment further shows that establishment of infection requires pH-dependent fusion of virions at the cell body. These findings correlate infectivity to existing single particle evidence of long-range endosomal transport of RABV and clathrin dependent uptake at the plasma membrane.

Rabies virus is the causative agent of a generally fatal and incurable disease of the central nervous system (CNS). Rabies lethality requires that the virus invade the brain, a feat accomplished by neuronal transmission from the site of infection to the CNS. Using cultures of peripheral neurons and chemicals that perturb specific cellular entry pathways we characterize the mechanism of rabies uptake. Using high resolution confocal microscopy, we visualize individual viral particles in the process of internalization and the establishment of infection by expression of a genetically encoded marker for infection. We show that clathrin-coated pits mediate internalization of the virus into endocytic vesicles that transport the virus to the cell body. We further demonstrate that release of the viral genomic core at the cell body is required to efficiently establish infection, and provide evidence that a subset of incoming virus particles fuse at non-productive sites prior to arrival at this site. This study extends the prior knowledge by identifying the entry mechanism and the site of fusion required for effective establishment of infection in neurons.

Rabies virus (RABV), a member of the Rhabdoviridae family, is a neurotropic pathogen that causes fatal encephalitis in animals and humans. The neurotropism of RABV is conferred by its single attachment and fusion glycoprotein (G) [1]. Virulence of specific RABV strains correlates with the neuroinvasiveness of their G proteins [2], such that exchange of G of an attenuated strain with that of a pathogenic strain and vice versa confers the corresponding level of pathogenicity [1,3–5]. Although differential glycosylation [6,7], dysregulation of G expression levels [8,9], and increased induction of apoptosis [8] all contribute to G-dependent attenuation of RABV strains, it is apparent that a predominant mechanism by which G modulates rabies virulence is by dictating affinity for and spread between neurons. Following the bite of a rabid animal, peripheral neurons serve as conduits of the virus to the CNS. Although both sensory and motor neurons can be infected [10–14], retrograde transmission of RABV dictates that motor neurons serve as the primary gateway for CNS invasion [15]. The predominant route of rabies virus entry into cells appears to be clathrin-mediated endocytosis (CME) [16–19]. Electron microscopic examination of chick embryo fibroblasts [18] and hippocampal neurons [20] show the presence of virions in coated pits. The relationship of those internalization events to infection, however, is not well established and existing studies that correlate the route of entry to eventual infection are restricted to non-neuronal cells [19]. Such studies also utilize vaccine RABV strains which may behave differently than their neurotropic counterparts. Available evidence suggests that RABV exploits existing cellular mechanisms that relay molecular signals from distal synapses to the somatodendritic compartment [21–23]. Long-range microtubule (MT) networks connect neuronal termini to the perinuclear region and mediate bidirectional axonal transport of proteins [24], mRNAs, organelles and endosomes [24–26]. Other neurotropic viruses exploit these routes to invade the CNS, but differ in directionality of transport and mode of MT engagement [27,28]. For example, polio- [29] and adeno- [30] viruses are transported within endosomes tethered to MTs via host proteins, whereas alpha herpesviruses [31] interact with cellular motors directly via capsid and tegument proteins. For RABV, single viral particles incorporating fluorescently-tagged transmembrane and RNP proteins appear to translocate intact within axons [21]. Consistent with this, receptors recruited to virions at the plasma membrane appear to remain associated during long-range axonal transport [22,23]. Collectively these studies provide evidence that rabies viruses are transported intact within endosomes, but the significance of this transport for productive infection has not been examined [21–23]. In the present study, we combine infectivity and single particle imaging approaches to study rabies internalization and fusion from the termini of neurons of the dorsal root ganglion (DRG) and motor neuron-rich ventral spinal cords (V SC) mediated by the neurotropic Challenge Virus Strain 11 (CVS) G protein. To model natural RABV infection at neuronal termini, we adapt a polydimethylsiloxane (PDMS) microfluidic culturing platform [32,33] which physically separates neuronal cell bodies from their neurites. We demonstrate that CVS G-mediated infection is reliant on dynamin-dependent uptake processes and that the predominant endocytic route is clathrin-mediated. We provide evidence that productive infection requires viral fusion at the somatodendritic compartment following long-range transport from the neuronal termini. This work extends previous findings from epithelial cells providing evidence that infection of neuronal cells by rabies virus occurs through a clathrin-dependent entry pathway with subsequent membrane fusion occurring at the cell body. To study uptake of RABV in neurons, we generated a recombinant VSV (rVSV) in which we replaced the glycoprotein (G) gene with that of the neurotropic rabies strain, CVS (rVSV CVS G; Fig 1A). This virus expresses eGFP as a marker of infection. Transmission electron microscopy (TEM) of purified rVSV CVS G showed characteristic bullet-shaped particles with readily discernible glycoprotein spikes (Fig 1B) consistent with efficient incorporation of CVS G. SDS-PAGE analysis of the protein composition of purified virions demonstrates comparable incorporation of CVS G into rVSV compared to RABV vaccine strain SAD B19 or VSV G (Fig 1C). CVS G migrates as a doublet consistent with the reported existence of two glycosylation variants of G [34] (Fig 1C). Neuroinvasiveness is a defining feature of pathogenic RABV G, and incorporation of CVS G results in a corresponding shift in rVSV tropism. BSR T7/5 monolayers were less susceptible to rVSV CVS G than their non-neuroinvasive counterparts, rVSV and rVSV SAD B19 G, as evident from a reduction in the number and size of the foci (Fig 1D). In addition, rVSV CVS G displayed greater capacity for spread resulting in larger plaque sizes than its non-neurotropic counterparts in N2a cells. A viral dose equivalent to MOI = 3 in N2a cells results in a calculated effective MOI < 0.05 in BSR T7/5 assuming the standard Poisson model of infection (Fig 1E). We exclude the possibility that differential lipid composition of virus grown in N2a cells is responsible for the difference in tropism by comparing the infectivity of rVSV grown in BSR T7/5 (rVSVBSRT7) or N2a cells (rVSVN2a; Fig 1E). Cumulatively, these observations are consistent with a restricted tropism conferred by the CVS G. We next investigated RABV uptake in neurons of the dorsal root ganglia (DRG) and ventral spinal cord (V SC) that project into muscle tissue (Fig 2A). Dissociated cultures, obtained from embryonic rats, yield neuronal populations with extensive projections as determined by immunofluorescence against the phosphorylated neurofilament H (NF) and the neuronal marker, Neuronal Nuclei (NeuN; Fig 2B). We selectively manipulated neuronal termini by culturing neurons in the S compartments of microfluidic devices, which allow spatial isolation of neurites from cell bodies (Fig 2C). By 10 days in vitro we detect significant outgrowth of projections into the distal, neurite (N) compartment (Fig 2D), and calcein staining demonstrates that neurites are contiguous across the channel (S1 Fig). Furthermore, we demonstrate by staining against phosphorylated NF that the majority of projections into the N compartment are axonal (Fig 2E). We restrict diffusion of molecules between the N and S compartments by controlling liquid levels and, therefore, hydrostatic pressure across the microchannels. Consistent with effective fluidic isolation, transferrin (Tfn) tagged with AlexaFluor 488 added to the N compartment is retained with no detectable diffusion across the channels (Fig 2F). We next evaluated the effect of endocytosis inhibitors on viral infection from the neuronal termini of cultured DRG and V SC neurons (Fig 3). For this purpose, we use dynasore and 5-(N-ethyl-N-isopropyl)amiloride (EIPA) to determine the requirement for dynamin- and macropinosome-mediated uptake respectively, and measure eGFP expression as a marker of viral infection. Consistent with dynamin-dependent endocytosis of rVSV CVS G, dynasore results in a near total block of infection in either neuronal population (Fig 3A and 3B). Treatment with EIPA also results in a reduction of infection; however, inhibition was not as pronounced as with dynasore and not significant in DRG neurons. Because neurite projections cannot be assigned to specific neurons in the somal chamber, the magnitude of the inhibitory effects of dynasore and EIPA cannot be precisely quantified. This is particularly apparent for V SC cultures which exhibit limited N compartment projection relative to the number of cultured neurons. For DRG cultures which exhibit a greater efficiency of neurite outgrowth (Fig 3C), dynasore reduces the percentage infected neurons from 43% to 2%. The inhibitory effect of EIPA in DRG infection was approximately two-fold with significant variability across experiments. This differed qualitatively from the more pronounced and reproducible inhibition observed in V SC cultures (Fig 3B). Reductions in infection by either inhibitor were not due to off-target effects on viral replication since addition of inhibitor at 2 hpi had limited effect on infection. We also exclude chemical degradation of projecting neurites as a contributing factor: intact neurite structures were retained at the experimental endpoint in both treated and untreated cultures (Fig 3D). We previously demonstrated that epithelial uptake of rVSV incorporating G from a vaccine strain of rabies (rVSV SAD B19 G) is clathrin-dependent [19]. To identify differences in uptake between epithelial and neuronal cells, we tested the effect of dynasore and EIPA on rVSV SAD B19 G uptake into DRG neurons (Fig 4). Treatment with dynasore resulted in an almost complete block of infection whereas EIPA treatment resulted in a more modest reduction (Fig 4B). These results suggest a shared mechanism for initial uptake at the plasma membrane during SAD B19 and CSV G-mediated entry. To determine whether chemical inhibition of infection was due to a reduction in the number of virions delivered to the soma, we next examined the cellular distribution of rVSV CVS G virions labeled with AF647 (rVSV CVS G AF647) by confocal microscopy (Fig 5). By 26 hpi neurons (Fig 5A and 5B) demonstrate efficient uptake and transport of virus: cell-associated particles are readily detected in all compartments including the cell bodies of some uninfected cells. In contrast, dynasore treatment restricts viral localization to the N compartment (Fig 5A) whereas EIPA has a limited effect on virus transport to the soma with viral particles present in both eGFP-positive and -negative neurons (Fig 5B). Irrespective of the inhibitor used, viruses efficiently bound N compartment neurites indicating that association with the plasma membrane is unaffected. At 26 hpi, remaining intact viral particles represent a population that did not contribute to infection. We, therefore, examined rVSV CVS G AF647 uptake disruption at an earlier timepoint, 6 hpi (Fig 5C and 5D). Here, we did not detect chemical disruption of viral association with the neuronal membrane (Fig 5C) but find differential viral accumulation within microchannel neurites. Since extracellular diffusion into the microchannels is restricted, viruses must access this compartment only via axoplasmic transport and are therefore intracellular. Both dynasore and EIPA administered at the time of inoculation abrogate viral accumulation in the microchannels (Fig 5D). Dynasore also impacts viral accumulation when added at 2 hpi, although the effect was less pronounced. A strong positive correlation between the number of cell-associated viruses at the opening to and the number within individual microchannels is observed in untreated controls. Dynasore decouples N-compartment and microchannel accumulation of virus consistent with a block to particle uptake. By contrast, following EIPA treatment that correlation remains positive suggesting that EIPA delays, rather than blocks, endocytosis and transport of virus. Fluorescent transferrin (Tfn) and dextran (Dex) are commonly used markers of clathrin-mediated and fluid phase endocytosis respectively. To corroborate our inhibitor studies, we next determined whether rVSV CVS G AF647 are colocalized with Tfn or Dex during entry (Fig 6A and 6B). In both neuronal populations, we observe a limited colocalization of virions with Dex in the N compartment and channel neurites (Fig 6A–6C). By contrast, a third of incoming rVSV CVS G associate with Tfn in the N compartment (Fig 6B), and this colocalization is enriched in the microchannels. This result suggests that internalization via clathrin shunts virus into long-range axoplasmic transport (Fig 6B and 6C). The transportation bias for particles originating from clathrin-coated pits is particularly pronounced within DRG neurons where 87% of virions within the channels colocalize with Tfn versus 33% in the N compartment (Fig 6C). Although fewer viral particles associate with Tfn in V SC channel neurites (Fig 6B and 6C) we note a similar enrichment from 31% in the N compartment to 58% in the channels. This observation is consistent with results showing greater sensitivity of infection to EIPA in the V SC neurons (Fig 3B). However, transmission electron micrographs of rVSV CVS G uptake in V SC displays viruses exclusively associated with endocytic structures consistent with clathrin mediated endocytosis (Fig 6D). This may suggest an off-target effect of EIPA in neurons which, by dysregulating Na+/H+ exchangers, may indirectly impact neuronal endocytic dynamics via disruption of the excitation properties of the plasma membrane [35]. Alternatively, 2hpi may be too early for the unambiguous identifications of macropinocytic structures in these cells. Furthermore, a caveat of the Tfn uptake within V SC neurons is the reported absence of transferrin receptors (TfnR) on the axons of mature motor neurons [36]. Axonal TfnR expression can be maintained by culturing of neurons in the presence of Tfn [37] and during axonal regeneration such as during outgrowth following axotomization inherent to the dissociation process [38]. Although the observation of Tfn uptake suggests that these primary cultures retain the receptor at some level, it is possible that a low expression of TfnR may result in an under-reporting of clathrin-mediated uptake in these neurons. To ascertain that viral particles are being transported, we performed live imaging studies in DRG neurons (S1 Video and S2 Video; Fig 7). Within the channels, we recorded single viral particles in the process of long-range axoplasmic transport. Many of these were transported concomitantly with fluorescent Tfn (S1 Video). Within highly polarized neurons, fusion can occur locally at the site of uptake or following endosomal transport at intermediate locations along the axon or at the cell body. Live imaging data of cotransport with Tfn corroborates previously published findings that RABV virions are sorted into a long-range vesicular transport route with delayed acidification and release at the cell body [21–23]. To investigate the role of delayed fusion on productive infection we administered Bafilomycin A1 (BAF A1), an inhibitor of vesicular H+ ion pumps, to the N or S compartment to respectively block localized or delayed fusion. We monitored the impact of this selective treatment on infection (Fig 8A–8C). Cells whose neurites were treated with BAF A1 displayed widespread infection in the S compartment despite presence of the inhibitor (Fig 8A and 8B). Under these conditions, we also observed enhanced viral spread relative to untreated counterparts in both DRG and V SC neurons. In contrast, somal treatment with BAF A1 resulted in robust inhibition of infection (Fig 8A–8C). This effect depended on administration of the drug early in infection: we observed lesser inhibition following treatment at 9–12 hpi. At this point incoming viruses from the N compartment have likely already fused at the cell body. We excluded potential cytotoxicity or off-target effects on the viral life cycle by assessing the effect of BAF A1 in non-compartmentalized culture. Post-entry administration at 2 hpi did not interfere with infection (Fig 8D). We corroborated our inhibitor experiments by demonstrating that incoming fluorescently-labeled viral particles colocalize with LysoTracker in the cell bodies of DRG neurons (Fig 9). LysoTracker accumulates in acidified endosomes where the luminal pH is low enough to trigger fusion. No colocalization was observed between LysoTracker and virions within microchannel neurites or neurites proximal to the somatodendritic compartment (Fig 9; S3 Video). We did not investigate colocalization of virions with LysoTracker at the growth cones or in distal neurites. Although we were unable to conclusively determine whether viruses enter acidified organelles prior to arrival in the cell body, the presence of acidified endosomes within the neurites is consistent with the possibility of fusion at earlier timepoints (Fig 9). Together, these results demonstrate that the majority of incoming virions are transported within endosomes to the cell body where acidification occurs. Fusion at the cell body is then a prerequisite for efficient infection. Here we provide evidence for a model of RABV entry into peripheral neurons that begins with clathrin- and dynamin-mediated uptake at the plasma membrane (Fig 10A). Endocytosed viruses are then transported intact and within endosomes from the distal neurites to the site of fusion at the cell body (Fig 10B). Furthermore, we show that somal fusion is required for efficient infection. This work extends the current understanding of RABV uptake by identifying the predominant internalization mechanism at the plasma membrane of two neuronal populations involved in early neuroinvasion in vivo. In addition, by combining single particle imaging and infectivity studies, we correlate single virion behavior with productive infection. In this study, we provide corroborating evidence of long-range axoplasmic transport of rhabdoviruses incorporating RABV G. Co-transport of virions with Tfn is consistent with trafficking within endosomes. Furthermore, exclusive sensitivity of infection to disruption of endosomal acidification in the S compartment demonstrates that viral fusion events leading to productive infection occur in the vicinity of the soma. Previous studies have reported long-range endosomal transport of intact vaccine strain RABV virions within differentiated neuroblastoma cells and DRG neurons [21,23]. Additionally, lentiviral vectors pseudotyped with G from the attenuated CVS-B2C strain have been tracked during retrograde endosomal transport in compartmentalized motor neuron culture [22]. We extend those earlier observations to the pathogenic CVS-11 G, and demonstrate that subsequent efficient establishment of infection is facilitated by pH-dependent fusion at the perikaryon. Previous work with the SAD vaccine strain of RABV provided evidence for transport within acidified endosomes [23]. This is in contrast to our observation that rVSV CVS G is exposed to acidic endosomal pH primarily at the cell body of neurons (Fig 9; S3 Video), with a minority of particles in acidified endosomes in the microchannel neurites. Vaccine strains of RABV are characterized by their reduced neurotropism and neuroinvasiveness, which are dictated by their respective glycoproteins and may account for the difference between our observations. The site of imaging might also be a contributory factor in the apparent difference because studies of acidified endosome distribution in neurons suggest a high density of these structures in areas proximal to the growth cone, but a significant reduction at intermediate locations on the axon [39]. The increase in infection and spread following inhibition of endosomal acidification at the neurites in both DRG and V SC neurons suggests that some virus-containing endosomes engaging long-range axoplasmic transport may undergo acidification before delivery to the soma (Fig 8). The presence of RABV in acidified organelles at the neuronal termini or within distal neurites has been previously suggested based on colocalization between RABV antigens and LysoTracker at NMJs [20], and cotransport of vaccine strain RABV and LysoTracker within DRG neurons [23]. Studies of tetanus neurotoxin trafficking suggest that sorting of certain cargoes into axonal carriers involves passage through intermediate acidified endosomes [40]. Although we did not independently verify the presence of infectious particles within acidified organelles in the distal neurites and growth cones, we cannot exclude that a similar sorting step occurs during transportation of RABV virions. If rhabdoviral particles traffic through analogous structures, exposure to the acidified environment may result in early fusion of a subset of incoming virions. Our results suggest that such early triggering of fusion leads to a delay or possibly abrogation of infection. Accordingly, if this early acidification event is circumvented–for example by BAF A1 inhibition of the vacuolar ATPase–viruses that would otherwise have fused prematurely now proceed along the transportation route and are subject to subsequent acidification events leading to infection. Although this observation is provocative it is important to emphasize that here we are studying infection initiated by delivery of the VSV RNP core and we cannot ignore the possibility of a RABV specific RNP transit to the cell body. The reported interaction between RABV P, and dynein LC8 [41,42] could facilitate RNP engagement of microtubules for retrograde transport. Mutation of the motif in P responsible for this interaction does not, however, abolish neuroinvasiveness of RABV in vivo [43,44] suggesting that direct transport of RNPs is not the primary mechanism of RABV axoplasmic transport. Previous studies investigated RABV uptake into non-neuronal [18,19], neuroblastoma [17] or hippocampal neurons [16]. Our work extends these studies into DRG and V SC neurons, the first populations of neurons invaded by rabies in vivo. We identify clathrin-mediated endocytosis (CME) as the primary mechanism of productive RABV uptake. We base our conclusion on three observations: (i) susceptibility of infection and single particle uptake to disruption of dynamin; (ii) co-packaging and -transport of incoming virions with transferrin; and (iii) detection of rhabdoviral particles within coated pits by electron microscopy. Because co-transport with transferrin was assessed during long-range axoplasmic transport, we cannot exclude that rhabdovirus-containing endosomes fuse or coalesce with Tfn-positive endosomes following internalization at the plasma membrane. However, transmission electron micrographs of uptake into V SC neurons provide direct evidence for clathrin-dependent uptake as viruses associated with coated endocytic structures that resemble clathrin-coated pits. Ultrastructural data from transmission electron micrographs also allows us to infer additional structural and mechanistic similarities between rhabdovirus-containing endosomes in neuronal and epithelial cells. The clathrin-coated pits share the elongated profile characteristic of their incompletely coated counterparts observed in BS-C-1 cells for both VSV and RABV (Fig 6D) [19,45,46]. It is likely, therefore, that actin polymerization is a requirement for completion of envelopment and scission also from the neuronal plasma membrane. Sensory DRG neurons and motor neurons are both susceptible to RABV infection in the host. Due to the morphological and functional differences between these neuronal populations, we explored the possibility of non-identical uptake mechanisms for RABV based on the neuronal subtype. Our infectivity experiments using pharmacological perturbation of endocytic processes reveal that productive infection in either neuronal population is dynamin-dependent. Single particle experiments further implicate clathrin-mediated uptake as a major route of RABV endocytosis in both cell types. Accordingly, 90% of particles in DRG microchannel neurites, and 55% in V SC neurites are co-packaged with transferrin. We also identified some differences in uptake between the two neuronal populations. In V SC culture, a significant fraction of incoming particles do not colocalize with transferrin, or fluorescent dextran. This suggests that a fraction of particles enter in a manner independent of clathrin or macropinocytic uptake mechanisms. Such differences in uptake appear to be cell-type dependent and may be dictated by differential engagement of particular receptors. Lentiviral vectors expressing CVS-B2C G undergoing axoplasmic transport in motor neurons were found to colocalize with all three known RABV receptors: p75 neurotrophin receptor, neural cell adhesion molecule, and nicotinic acetylcholine receptor [22]. Studies of uptake of putative rabies receptors following engagement with endogenous ligands, crosslinking antibodies or toxins indicate that p75NTR, NCAM and nAChR internalize by different cellular mechanisms. p75NTR bound to neurotropic factors internalizes via clathrin-mediated endocytosis [47–49]. Antibody crosslinking of NCAM induces its internalization primarily by clathrin-mediated endocytosis with caveolae playing a secondary role [48,49]. Finally, nAChR uptake into filamentous invaginations from the plasma membrane is clathrin-independent [50,51]. These observations raise the possibility that RABV interaction with specific receptors may contribute to the endocytic sorting of the virus. It will therefore be of interest to identify which receptors are internalized at the plasma membrane in complex with RABV and to relate these interactions to establishment of infection in these two neuronal populations. Device masters were manufactured by two-layer soft photolithography onto 3 inch mechanical grade silicon wafers for spin coating (UW3MEC, University Wafers) utilizing established methods [52]. Two negative photoresists were used: SU-8 2002.5 (MicroChem) for the 3 μm microchannel layer followed by SU-8 2050 (MicroChem) for the 100 μm culturing compartment layer. Photoresist was patterned by UV-crosslinking through custom 20,000 dpi transparency masks (CAD/Art Services), processed and cured according to supplier’s instructions. Following a final hard cure at 150°C for 15 min, masters were treated with (tridecafluoro-1,1,2,2-tetrahydrooctyl)trichlorosilane for 45 min to facilitate removal of cured polydimethylsiloxane (PDMS) following moulding. Devices were cast by applying a 10:1 prepolymer:curing agent mixture of Sylgard 184 (Dow Corning) to the master and curing at 65°C for a minimum of 1 h. After curing and release from the master, PDMS devices were cut, and wells punched out with round biopsy punches. We irreversibly bonded devices to acid cleaned glass coverslips by oxygen plasma bonding in a 500-II Plasma Etcher (Technics). Bonded devices were sterilized under UV in a biosafety cabinet for 10 minutes prior to consecutive overnight coatings with 300 μg ml-1 poly-D-lysine (P7886, Sigma-Aldrich) dissolved in 2X borate buffer solution (28341, Thermo Scientific) and 10 μg ml-1 laminin (L2020, Sigma-Aldrich) in sterile water. Pregnant Spraque-Dawley rats (Taconic) were a kind gift from C. Cepko. All animal work included in this study was approved by the Harvard Medical Area Standing Committee on Animals under protocol 428-R98 of the Institutional Animal Care and Use Committee (IACUC) of Harvard Medical School. Animals were housed and handled in accordance with the Guide for the Care and Use of Laboratory Animals. Euthanasia of pregnant Sprague-Dawley rats was performed by controlled exposure to carbon dioxide from compressed gas cylinders such that suffering and distress was minimized. Whenever possible, animals were kept in their housing cages during the procedure. Non-responsive animals were further subjected to cervical dislocation to exclude any possibility of accidental revival. Only once this sequence of procedures was completed did we remove the uterus and embryos from the carcass. E15 embryos were euthanized by removal from the amniotic sac and subsequent decapitation. Neuronal tissues were dissected from E14.5-E15.5 embryonic Sprague-Dawley rats (Taconic). Dorsal root ganglia (DRG) were dissected, dissociated by trypsinization and cultured in Neurobasal media (Gibco) supplemented with B27 (1:50; 17504–044, Gibco), β-nerve growth factor (100 ng ml-1; 450–01, Peprotech), 5% fetal bovine serum (FBS, Tissue Culture Biologicals), 2 mM glutamine (101806, MP Biochemicals), 25 mM HEPES (0511, AMRESCO) pH 7.4 and 25 μg ml-1 β-D-arabinofuranoside (AraC; C1768, Sigma-Aldrich) [53]. Ventral spinal cord neurons were dissected by adapting the strategy outlined for the harvest and culture of dorsal spinal cord commissural neurons [54]. Instead of harvesting the dorsal portion of the spinal cord, we retained the ventral portion for our cultures. Dissociation of the neurons was also carried out as detailed with the exclusion of the Opti-Prep purification step. V SC neurons were cultured in Neurobasal media (Gibco) supplemented as outlined by Leach et al. [55]. AraC treatment for selective kill-off of dividing non-neuronal cells was included and maintained in DRG media from first plating. For V SC culture AraC was applied after 48 h in culture. For preparation of compartmentalized cultures, we seeded dissociated neurons into the S compartment: 1.5 × 105 DRG or 1 × 105 V SC neurons were dispensed per device. A half-media swap was performed for both DRG and V SC culture following 2–3 days in culture; at this point, 25 μg ml-1 AraC (Sigma-Aldrich) was administered to the V SC cultures and maintained. Subsequently, media was periodically supplemented to counteract evaporation in culture. Experimental infections were carried out once adequate neurite outgrowth was observed in the N compartment: typically, from day 7 in culture for DRG and day 10 for V SC neurons. Devices were discarded following 12 days in culture. Neurons, mouse neuroblastoma Neuro-2a cells (N2a; ATCC CCL-131; American Type Culture Collection, Manassas, VA), baby hamster kidney BSR T7/5 cells (gift of U.J. Buchholz) [56], and African green monkey kidney BS-C-1 cells (ATCC CCL-26; American Type Culture Collection, Manassas, VA) were maintained at 37°C and 5% CO2. Non-neuronal cells were cultured in Dulbecco’s modified Eagle medium (DMEM; Corning) supplemented 10% fetal bovine serum (Tissue Culture Biologicals). N2a media was further supplemented with 2mM glutamine (Sigma) and 25mM HEPES pH 7.4. rVSV eGFP and rVSV eGFP SAD B19 G were amplified, purified and maintained as previously described [45,46]. rVSV eGFP CVS G (rVSV CVS G) was generated by insertion of the CVS-11 glycoprotein coding region into MluI and NotI restriction sites in a modified rVSV eGFP ΔG backbone [57–59]. A pUC57 plasmid containing the cDNA of CVS-11 G (Genbank: GQ918139.1) with flanking MluI and NotI sites was commercially synthesized by GenScript. A P0 stock of the virus was recovered by standard methods in BSR T7/5 monolayers [45]. Individual viral clones from the P0 stock were isolated by fluorescent focus assay in N2a culture, and further amplified in N2a monolayers. rVSV CVS G stocks were passaged and expanded in N2a monolayers. Infections were carried out according to standard technique [45]. At 24hpi supernatant and infected N2a cells were collected and subjected to 2 min sonication in a Branson 1510 ultrasonic cleaner (Branson, Richmond, VA) followed by 30 s vortex to release cell bound viruses. Cell debris was pelleted by centrifugation and the resultant virus supernatant purified by ultracentrifugation. Virus pellets were resuspended overnight in phosphate buffered saline (PBS) + 25 mM HEPES pH 7.4 + 50 mM EDTA. We sonicated the virus resuspension for an additional 2 min, followed by 30s vortex, immediately prior to a final gradient purification over a 15%-45% (wt/vol) sucrose gradient in PBS as previously described [45]. Viral titers were determined by fluorescent focus assay in N2a monolayers. We used established methods to label gradient-purified viral particles with 40 μg ml−1 Alexa Fluor (AF)-conjugated succinyl esters (Molecular Probes, Invitrogen) [45]. Titration of mock- or AF647-labeled virus preparations showed that dye conjugation had a negligible effect on infectivity: mock-labeled rVSV CVS G had a titer of 2.1 x 1010 ffu/mL compared to 1.5 1010 ffu/mL titer of the AF647-labeled virus. Viral proteins of gradient purified virions were separated by SDS-PAGE in a 10% polyacrylamide (wt/vol) and 0.13% (wt/vol) bis-acrylamide gel. Protein bands were visualized with SimplyBlue SafeStain according to manufacturer’s instructions. Viral protein amounts relative to N protein were determined using ImageJ (U.S. National Institutes of Health, Bethesda, Maryland; http://rsb.info.nih.gov/ij/). Neuronal cytoplasms in the N and S compartments were stained with calcein (diluted 1:1000; C3099, Molecular Probes) or CellTracker (diluted 1:500; C34552 Molecular Probes), and nuclei with NucBlue Live Cell Stain (diluted 1:50; R37605, Molecular Probes) in Neurobasal for 30 min prior to inoculation with virus. Stains were washed once with Neurobasal following removal, prior to infection or imaging by direct fluorescent microscopy. The following chemicals were administered at the listed concentrations: 0.1 μM bafilomycin A1 (BAF A1; 196000, Calbiochem, EMD Chemicals); 150 μM dynasore (Sigma-Aldrich); and 25 μM 5-(N-ethyl-N-isopropyl)amiloride (EIPA; A3085, Sigma-Aldrich). Infections were carried out exclusively in the N compartment. For N-compartment inhibitor treatment, culturing media in the S compartment was supplemented to replace evaporated liquid volume, and maintained throughout the experiment. For BAF A1 experiments where inhibitor treatment was also carried out in the S compartment, S compartment media was replaced with Neurobasal alone or BAF A1 diluted in Neurobasal. Neuronal culturing media in the N compartment was replaced with 30 μL of inoculation media. Inoculation media consisted of Neurobasal media alone or with inhibitor as indicated. 106 fluorescent foci forming units of virus were administered to the N compartment in 25 μL of inoculation media. Liquid volume differential between the S compartment and N compartment was optimized to eliminate diffusion of molecules across the channels. Inoculum was maintained for 2 h and then replaced with 55 μL treated or untreated Neurobasal media. For BAF A1 experiments where inhibitor was administered at 9–12 hpi, an additional media swap was carried out. Expression of eGFP was assessed at 26 hpi either by direct live fluorescence microscopy for DRG neurons or following fixation and immunofluorescence for V SC neurons. This end point was determined by monitoring appearance of eGFP and selecting a timepoint at which robust eGFP expression first occurs. In addition, we verified infection of neurons solely in the vicinity of the microchannel openings, consistent with primary infection without spread to secondary infection sites. For all co-uptake experiments neurons were first prestained with calcein or CellTracker as indicated. Infections were carried out, as described, in the N compartment with inocula containing Tfn conjugated to AF594 (50 μg ml-1; Molecular Probes), AF488-conjugated Dextran (MW 10 000, 1 μg ml-1; Molecular Probes) or LysoTracker DND-26 (75 nM; Molecular Probes). Tfn and Dextran results were analyzed in both fixed and live samples. All LysoTracker images were collected by live confocal microscopy. For fixed experiments, N compartments were washed twice at 2 or 5 hpi with Neurobasal, and fixed with 2% (wt/vol) paraformaldehyde in PBS + 5% (wt/vol) sucrose. For live imaging, uptake experiments were carried out in devices bonded to FluoroDish glass bottomed culture dishes (FD35-100, World Precision Instruments, Inc.) and imaged by high resolution spinning disk confocal microscopy. Neurons were fixed with 2% (wt/vol) paraformaldehyde in PBS + 5% (wt/vol) sucrose. Cell membranes were permeabilized with 0.2% Triton-X in PBS. Cells were consecutively stained with mouse monoclonal antibody against phosphorylated neurofilament H, SMI-31 (1:1000; NE1022, Calbiochem), and AF-conjugated anti-mouse secondary antibody (Molecular Probes, Invitrogen). When indicated, neuronal cells were detected by staining against Neuronal Nuclei (NeuN) with rabbit polyclonal antibody (1:500; ab104225, Abcam) and AF-conjugated anti-rabbit secondary antibody (Molecular Probes). Nuclei were stained with DAPI (1:10,000; Molecular Probes). Devices processed for immunofluorescence underwent a final wash with PBS and were imaged in solution. Non-compartmentalized neurons, cultured on coverslips, were mounted onto slides with ProLong Diamond (Molecular Probes). Devices were illuminated with a Mercury-100W mercury lamp (Chu Technical Corporation) and imaged using a Nikon Eclipse TE300 inverted microscope, outfitted with 4× Plan Fluor, 10× and 20× Plan Fluor objective lenses (Nikon). Images were collected using a SPOT RT Monochrome camera (Spot Imaging Solutions, Diagnostic Instruments Inc.) and recorded with the manufacturer’s Spot 3.5 Advanced software. Devices were imaged using a Marianas system (Intelligent Imaging Innovations) based on a Zeiss observer microscope (Carl Zeiss MicroImaging) outfitted with a CSU-22 spinning-disk confocal unit (Yokogawa Electric Corporation) and a 63× (Plan-Apochromat, NA 1.4; Carl Zeiss Microimaging) objective lens. Excitation wavelengths were 491 nm for AF488, 561 nm for AF594, and 660 nm for AF647. For three-dimensional acquisitions, the vertical position was manipulated in 0.3 μm increments using a PZ-2000 automated stage (Applied Scientific Instrumentation). Live imaging experiments were carried out on a temperature controlled sample holder (20/20 Technology Inc.; Wilmington, NC) maintained at 37°C and 5% CO2. Images were collected using a Photometrics Cascade II electron multiplication camera (Photometrics). SlideBook 5.0 (Intelligent Imaging Innovations) was used to command the hardware devices, and visualize and export the acquired data. Subsequent image manipulation was conducted using ImageJ (U.S. National Institutes of Health, http://rsb.info.nih.gov/ij/). To visualize viral morphology, we deposited gradient purified rVSV CVS G virions onto carbon-coated copper grids and stained them with 2% phosphotungstic acid (wt/vol) in H2O (pH 7.5). To visualize viral uptake, we inoculated V SC neurons cultured on Aclar with rVSV CVS G at a multiplicity of infection (MOI) exceeding 1,000 for 2 h at 37°C. rVSV uptake samples were prepared by inoculating BS-C-1 cells with rVSV at an MOI of 1,000 for 15 min at 37°C. Samples were fixed and processed for ultrathin sectioning as previously described [45,60]. Virus particles and ultrathin sections of cells were viewed using a Tecnai G2 Spirit BioTWIN transmission electron microscope (FEI).

10.1371/journal.pgen.1003711

Neuronal Reprograming of Protein Homeostasis by Calcium-Dependent Regulation of the Heat Shock Response

Protein quality control requires constant surveillance to prevent misfolding, aggregation, and loss of cellular function. There is increasing evidence in metazoans that communication between cells has an important role to ensure organismal health and to prevent stressed cells and tissues from compromising lifespan. Here, we show in C. elegans that a moderate increase in physiological cholinergic signaling at the neuromuscular junction (NMJ) induces the calcium (Ca2+)-dependent activation of HSF-1 in post-synaptic muscle cells, resulting in suppression of protein misfolding. This protective effect on muscle cell protein homeostasis was identified in an unbiased genome-wide screening for modifiers of protein aggregation, and is triggered by downregulation of gei-11, a Myb-family factor and proposed regulator of the L-type acetylcholine receptor (AChR). This, in-turn, activates the voltage-gated Ca2+ channel, EGL-19, and the sarcoplasmic reticulum ryanodine receptor in response to acetylcholine signaling. The release of calcium into the cytoplasm of muscle cells activates Ca2+-dependent kinases and induces HSF-1-dependent expression of cytoplasmic chaperones, which suppress misfolding of metastable proteins and stabilize the folding environment of muscle cells. This demonstrates that the heat shock response (HSR) can be activated in muscle cells by neuronal signaling across the NMJ to protect proteome health.

The protein quality control machinery is responsible for preventing the accumulation of misfolded and damaged proteins and loss of cellular function. The capacity of cellular surveillance is limited however, leading to increased appearance of protein aggregates and risk for age-associated diseases. Here, we show that upregulation of acetylcholine receptors and moderate increased cholinergic activity leads to a calcium-dependent stress response that suppresses protein misfolding and restores homeostasis in C. elegans muscle cells. This involves gei-11 knockdown-dependent upregulation of acetylcholine receptors, and the release of calcium into the cytoplasm of muscle cells through cell membrane and sarcoplasmic reticulum specific channels. Subsequently, activation of the heat shock factor 1 (HSF-1) leads to the expression of cytoplasmic chaperones that suppress misfolding of metastable and aggregating proteins, restoring folding and muscle function. This reveals a new non-canonical mechanism for the cell non-autonomous regulation of the heat shock response to ensure balance between cells in a metazoan.

Cellular health and organismal lifespan are critically dependent upon the fidelity of the proteome and the proteostasis network [1]. What are the molecular events that control proteostasis across tissues to activate protective responses at the cellular level to ensure organismal health? At the cellular level the heat shock response (HSR) and the unfolded protein response (UPR) respond to acute forms of proteotoxic stress with precise and rapid activation to restore homeostasis [2], [3]. In contrast to transient extreme stress, the chronic forms of protein damage and toxicity challenge the quality control machinery by their persistence and amplification effects on cumulative protein damage [4], [5]. How proteostasis monitors and responds to physiological stress is an area of active research [2], [6]–[8]. Yet, we know little about the regulation of stress responses under physiological conditions and at the organismal level. Elucidating these regulatory mechanisms is essential for a better understanding of diseases of altered protein conformation and age-related decline in cellular function [1], [9]–[11]. Much of our understanding of the HSR and the proteostasis network has come from studies using cultured cells and model organisms. The invertebrate animals C. elegans and D. melanogaster have been particularly amenable genetic models for identification of proteostasis components and modifiers of protein misfolding and toxicity [12]–[22]. These modifiers include cell autonomous factors such as molecular chaperones, proteasome subunits, components of the autophagy machinery, and the FOXO and heat shock factor 1 (HSF-1) transcriptomes that promote protein folding and clearance within the cell [20]. At the organismal level, the cell non-autonomous role of neuroendocrine signaling pathways and trans-cellular chaperone signaling has been shown to be important for lifespan, stress resistance, innate immunity and proteostasis [8], [9], . Moreover, tissue-specific regulation of mitochondrial function, including the electron transport chain and the mitochondrial UPR, was shown to affect the rate of aging [27]. Efforts to understand how cell autonomous and non-autonomous processes are integrated and co-regulated at the organismal level offer new genetic and pharmacological strategies to enhance the maintenance of proteostasis and health span. In this study, we describe a new pathway for the heat shock response involving calcium signaling, in which gei-11 knockdown-dependent upregulation of cholinergic receptor levels at the neuromuscular junction (NMJ) triggers activation of HSF-1. This reveals that under normal physiological conditions, the balance between cholinergic and GABAergic signaling at the NMJ regulates protein homeostasis in body-wall muscle (BWM) cells [28]. In contrast to the situation where complete inhibition of GABAergic signaling leads to overstimulation of muscle cells and dysfunction of post-synaptic cell proteostasis [29], we show here that a moderate increase in ACh receptors (AChR) at the NMJ, attained by down-regulation of gei-11, is beneficial to muscle cells. We demonstrate that there is a critical range of cholinergic activity at the NMJ leading to the Ca2+-dependent activation of HSF-1 and expression of molecular chaperones that results in an enhanced protective state in post-synaptic muscle cells. A genome-wide RNA interference (RNAi) screen performed in C. elegans for genetic modifiers of muscle proteostasis identified gei-11 RNAi as a potent suppressor of polyglutamine (polyQ) aggregation in BWM cells (Figure 1A: II, IV, VI) [30]. Knockdown of gei-11 also cross-protected against other aggregation-prone proteins, including polyQ37 that presents an earlier onset and more dramatic foci pattern (Figure S1A: I–IV) and mutant SOD1G93A (Figure S1A: V–VIII). Suppression of Q35 aggregation by gei-11 knockdown was achieved by maintaining the polyQ protein in a diffuse soluble state (Figure 1A–VI), as determined by fluorescence recovery after photobleaching (FRAP) (Figure 1B), and not by reducing the expression of Q35 mRNA or protein (Figure S1:B–D). Moreover, gei-11-mediated suppression of Q35 aggregation also led to the rescue of cellular toxicity as the motility of Q35 animals was restored to 100%, similar to knockdown of the polyQ transgene by yfp-RNAi (Figure 1C), without affecting the motility of wild-type (wt) animals. These results reveal that gei-11 knockdown has potent suppressor activity of polyQ aggregation and toxicity. The gene gei-11 encodes the GEX-3-interacting protein 11 (GEI-11), a member of the Myb superfamily of transcription factors that is homologous to mammalian SNAPC4 (snRNA-activating protein complex subunit 4) [31]. gei-11(tm6548) is the only mutant allele available for this gene, and has a lethal phenotype (National Bioresource Project of Japan). gei-11 has been proposed to have a negative regulatory effect on AChRs in C. elegans BWM cells, and is also expressed in head neurons, germ cells, somatic gonad, and intestine [32], [33]. In C. elegans, two types of ACh receptors, each with distinct subunit composition and pharmacology, are expressed at the NMJ [34]–[36]. To establish the specificity of gei-11 downregulation on the expression of NMJ AChR subtypes, we monitored the expression of the L-type (levamisole-sensitive) AChR subunits unc-29, unc-38, unc-63 and lev-1, and the N-type (nicotine-sensitive) homomeric AChR, acr-16. The effect of gei-11 knockdown increased the expression of only the three essential subunits of the L-AChR (unc-29, unc-38, unc-63) by approximately 3-fold, and had no effect on the expression of N-AChR acr-16 (Figure 2A). Likewise, gei-11 RNAi did not affect the expression of the NMJ GABA receptor (GABAR unc-49, Figure 2A). These results suggest that gei-11 has a highly selective effect on the regulation of the L-type of cholinergic receptors. To establish whether gei-11 RNAi-mediated suppression of polyQ aggregation was dependent on elevated expression of the AChR, we performed double RNAi knockdown experiments and downregulated gei-11 together with each of three L-AChR essential subunits (unc-38, unc-63 or unc-29). The results showed that polyQ aggregation was unaffected when an essential L-AChR subunit was co-downregulated with gei-11 (Figure 2B, Table S1). This was further confirmed using an L-AChR mutation (unc-38(e264)), that corroborated the results obtained with RNAi (Figure S2A). Moreover, gei-11 co-downregulation with the non-essential L-AChR subunit lev-1, that only reduces receptor function [37]–[40], only weakened the gei-11 effect on polyQ aggregation (Figure 2B). In contrast, double knockdown of gei-11 with the N-AChR subunit acr-16, still suppressed Q35 aggregation (Figure 2B). If the elevated expression of three essential L-AChR subunits results in increased L-AChR activity at the NMJ, this would predict increased sensitivity to levamisole, the cholinergic agonist that selectively activates L-AChR and causes hyper-contraction and paralysis [37]. Indeed, gei-11 RNAi-treated wt animals showed a more rapid paralysis on levamisole plates relative to vector RNAi-treated animals (Figure 2C), consistent with the enhanced L-AChR activity at the NMJ. The specificity of levamisole sensitivity to L-AChR activity was confirmed using loss-of-function receptor mutations (unc-38(e264), unc-63(x26), unc-29(e1072)) [28], [35]. These animals were resistant to levamisole upon gei-11 RNAi treatment (Figure 2C), supporting a gei-11 effect dependent on AChR function. Similarly, mutant Q35;unc-38(e264) animals were resistant to levamisole compared to Q35 animals upon gei-11 RNAi (Figure S2B). Consistent with specificity of gei-11 to the L-AChR sub-type, RNAi-treated animals exposed to the agonist nicotine that targets N-AChR [34]–[36], did not show altered sensitivity compared to either wt or N-AChR mutant animals (acr-16(ok789)) (Figure S2C). Additional support for cholinergic-mediated effect on proteostasis was obtained using (+)-Tubocurarine chloride (dTBC), a potent inhibitor of AChR activity [28]. Consistent with our hypothesis, dTBC inhibited gei-11 RNAi suppression of Q35 aggregation in a dose-dependent manner (Figure 2D, ); the specificity of this effect on L-AChR function was consistent with results obtained using the Q35;unc-38(e264) animals (Figure 2D). The expression of GEI-11 is not restricted to muscle cells [33], therefore we determined whether the effect of gei-11 RNAi and increased cholinergic receptor expression on muscle polyQ aggregation was a direct consequence of gei-11 knockdown in muscle cells, rather than a downstream effect from another tissue. For example, gei-11 RNAi did not have an effect on aggregation of polyQ expressed in the intestine (iQ44, Figure S1E). We next examined whether the enhancement of muscle proteostasis was a direct consequence of gei-11 downregulation in muscle by using a C. elegans mutant strain in which the effects of RNAi are restricted to muscle cells: Q35;rde-1(ne219);myo-3p-RDE-1 (here referred to as rde-1(ne219);mRDE-1) [41], [42]. As a negative control, we employed a mutant strain impaired for RNAi in all cells, Q35;rde-1(ne219) (Figure S2E). Knockdown of gei-11 in rde-1(ne219);mRDE-1 animals increased the expression of essential L-AChR subunits (>3-fold, Figure 2E), comparable to organism-wide gei-11 RNAi (compare Figure 2A and C to Figure 2E and F). Likewise, aggregation in Q35;rde-1(ne219);mRDE-1 animals was suppressed by 50% upon gei-11 RNAi relative to vector control, with no effects observed in Q35;rde-1(ne219) animals (Figure 2G). Taken together, these results show that suppression of protein aggregation in muscle cells is due to gei-11 down-regulation in muscle cells and the consequent upregulation of AChR at the NMJ. To address whether the effect of gei-11 knockdown-mediated suppression of polyQ aggregation was selective for this class of highly aggregation-prone species or reflected a more general enhancement of proteostasis in the BWM cells, we examined the effects of gei-11 knockdown on the folding stability of four endogenous metastable proteins that function as folding sensors in muscle cells [4]. These metastable proteins exhibit temperature-sensitive (TS) phenotypes and harbor missense mutations in the paramyosin ortholog UNC-15, the basement-membrane protein perlecan UNC-52, the myosin-assembly assisting protein UNC-45, and myosin heavy chain UNC-54 [4], [11]. Thus, in animals held at the permissive temperature (15°C), all four TS-proteins are fully functional, whereas at the restrictive temperatures (23° and 25°C) these sensors misfold and each mutation results in an 80–100% loss of muscle function phenotype (Figure 3A control). However, when gei-11 was downregulated at the restrictive temperature, the loss-of-function phenotype of each TS protein was decreased by 50–60% (Figure 3A). These results establish that downregulation of gei-11 has general protective effects on the stability of multiple muscle cell proteins. To establish whether gei-11 RNAi suppression of protein misfolding and aggregation in muscle cells is due to the expression of chaperones, we performed experiments in which gei-11 was knocked-down together with either hsf-1 or hsp-70 (C12C8.1), to reveal that Q35 aggregation was no longer suppressed (Figure 3B). We monitored the expression of cytoplasmic chaperones of the HSP-70 family (C12C8.1, F44E5.4 and C30C11.4) and small heat shock protein family (sHSPs hsp-16.1, hsp-12.6 and hsp-16.49) in wt animals, and show that gei-11 RNAi enhanced the expression of each chaperone gene from 2-to-10-fold (Figure 3C), a level that while substantial, is nevertheless much lower than observed when animals are exposed to acute heat shock treatment (>50-fold, Figure S3A). Moreover, the expression of these chaperones was not induced in the AChR mutant unc-29(e1072), or in HSF-1 mutant hsf-1(sy441) animals (Figure 3C). Therefore, upregulation of the proteostasis machinery by gei-11 RNAi was absolutely dependent upon both cholinergic receptor activity and HSF-1. These results were further corroborated with the cholinergic antagonist dTBC and the L-AChR unc-63(x26) mutation (Figure 3D) that also prevented the upregulation of chaperones upon gei-11 RNAi. The levels of hsps were also upregulated when gei-11 was knocked-down specifically in muscle cells (rde-1(ne219);mRDE-1 animals, Figure 3E), providing additional support that the regulation of cholinergic receptors at the NMJ enhances muscle proteostasis. To demonstrate directly that knockdown of gei-11 induced the HSR, we monitored the activity of the heat shock transcription factor, HSF-1, using the electrophoresis mobility shift assay (EMSA) [43]. Knockdown of gei-11 induced HSF-1 DNA-binding activity to the heat shock element (HSE) (Figure 3F: lane 6), similar to what was observed in animals exposed to heat shock (Figure 3F: lane 3). The specificity of HSF-1 binding was established using an in vitro competition reaction with excess unlabeled HSE (Figure 3F: lanes 4 and 7) and by using a mutant HSE radiolabeled oligonucleotide incapable of binding by HSF-1 (Figure 3F: lanes 5 and 8). These results demonstrate that gei-11 knockdown activates HSF-1 transcriptional activity. We further established the downstream regulatory effects of gei-11 knockdown by examining the expression of other components of the proteostasis network, including the expression of the UPR-regulated ER chaperones (hsp-3, hsp-4, dnj-7 and ero-1), metabolic stress FOXO/DAF-16 regulated genes (sod-3 and mtl-1), and oxidative stress regulated genes (hsp-6, gst-4 and gcs-1). As shown in Figure S3B, the expression of none of these other stress responsive genes was induced by gei-11 knockdown in wt animals. Taken together, these results demonstrate that modulation of cholinergic receptors at the NMJ reprograms post-synaptic proteostasis through the activation of HSF-1 and the selective induction of cytoplasmic chaperones. While upregulation of AChR at the NMJ induced the heat shock response and suppressed protein misfolding and toxicity in post-synaptic muscle cells, in previous studies we had observed that a null mutation in unc-30, the transcription factor that regulates the GABA operon, had the opposite result leading to enhanced aggregation in BWM cells [29]. This suggested that extreme cholinergic overstimulation was deleterious to muscle cell proteostasis [29]. Taken together with the results presented here for gei-11, we posit that there is a critical balance between the levels and activities of AChR and GABAR (Figure 4A), with a physiological enhancement of AChR activity being proteo-protective and extreme overstimulation having proteotoxic consequences. To address this hypothesis, we treated Q35 animals with the L-AChR agonist levamisole over a wide range of concentrations, and monitored aggregation in the post-synaptic muscle cells. At a low concentration of levamisole (5 µM) Q35 aggregation was suppressed by >40% (Figure 4B), whereas at higher levels (50 µM) that caused hyper-contraction, we observed the opposite result of 50–60% enhancement of aggregation (Figure 4B and S4). No effect on aggregation was observed in AChR mutant animals (Q35;unc-38(e264)) (Figure 4B). These results reveal that the beneficial effect on the folding environment is the consequence of a specific physiological range of cholinergic stimulation, and supports our conclusion that overstimulation has deleterious consequences on the folding environment of muscle cells. We further explored how the imbalance between AChR and GABAR activation at the NMJ affects muscle proteostasis by combining both genetic and small molecule agonists and antagonists probes. Reducing GABAergic activity by exposure to low concentrations (25 µM) of the GABAR antagonist Lindane [29], [44]–[46] led to the suppression of Q35 aggregation (Figure 4C), comparable to a moderate increase of cholinergic signaling. Consistent with the genetic observations after cholinergic overstimulation [29], exposure to the highest concentrations of Lindane (1 mM), that cause BWM cells overstimulation, enhanced Q35 aggregation (Figure 4C). These results provide additional support that shifting the balance at the NMJ can enhance or harm proteostasis in the post-synaptic cell, in a magnitude of signal-dependent manner. To further test this hypothesis, we titrated GABAR expression and GABA release at the synapse by using a combination of RNAi and specific loss-of-function mutations in the GABA pathway: unc-30 (GABA synthesis), unc-47 (GABA transport) and unc-49 (GABAR). Knockdown of unc-47 and unc-49 suppressed Q35 aggregation, and this effect was less penetrant upon dilution of RNAi (1∶1) with vector RNAi (Figure 4D, Table S1). Conversely, eliminating GABA signaling in the unc-30(e191) and unc-47(gk192) mutants had the opposite effect (Figure 4D), consistent with previous results [29]. Finally, we altered the balance between cholinergic and GABAergic signaling by exposing gei-11 RNAi-treated Q35 animals to increasing concentrations of GABA. At low concentrations (≤50 mM), GABA compensated for the gei-11-mediated increase in cholinergic signaling and prevented the suppression of Q35 aggregation, in a dose-dependent manner (Figure 4E). However, at very high GABA concentrations (50 mM–200 mM) this equilibrium shifted to the opposite direction and resulted in suppression of polyQ aggregation by GABAergic signaling (Figure 4E), also consistent with previous results [29]. The highest GABA concentrations tested (>200 mM, not shown) were very toxic and lethal to the animals. These results provide additional support to the importance of the magnitude of cholinergic and GABAergic signaling: an imbalance by either higher AChR or GABAR transmits a signal that is interpreted by the muscle cell to activate a proteo-protective response; however, when this balance is severely disrupted, the consequence is proteotoxic. Altering the balance between AChR and GABAR, also affected hsp-70 expression. Exposure of wt animals to low levels of the AChR agonist levamisole (5 µM), or to the GABAR antagonist Lindane (25 µM), resulted in elevated expression of the cytoplasmic hsp70 family of chaperone genes (Figure 4F: C12C8.1 and F44E5.4), consistent with enhanced proteostasis. Likewise, and as expected, genetic reduction of GABA signaling using unc-47 or unc-49 RNAi (Table S1), which is equivalent to a moderate increase in cholinergic signaling, also upregulated a low level of hsp-70 (<9-fold, Figure 4F) that restored the folding environment (Figure 4D). Consistent with our previous results, extreme overstimulation in the mutants unc-47(gk192), unc-49(e407) or unc-30(e191) that no longer express GABA or GABAR, led to a massive upregulation of hsp-70 (>50-fold, Figure 4F) that was not proteo-protective and resulted in elevated aggregation (Figure 4D), similar to the deleterious effects of acute heat shock treatment (Figure S3A). Taken together, these results provide strong support for the importance of the regulation of the equilibrium between cholinergic and GABAergic signaling for optimal proteostasis. Within a critical physiological range, we show that increased AChR activity was beneficial and led to HSF-1-dependent moderate upregulation of cytosolic chaperones in muscle cells to establish a proteo-protective state. In contrast, extreme cholinergic overstimulation, whether obtained by genetics or small molecules, resulted in a dysfunctional proteostasis network that was deleterious. The binding of ACh to receptors in BWM cells initiates a cascade of events that lead to the release of Ca2+ into the cytoplasm for muscle contraction [47] (Figure 5A). We therefore investigated whether the improvement in proteostasis through AChR-mediated HSF-1 activation was dependent on Ca2+ influx. Initially, we utilized the cell permeant Ca2+ chelator BAPTA [48], that alone (15 µM) had no effect on Q35 aggregation, but prevented gei-11 RNAi induction of hsp chaperones and the subsequent suppression of polyQ aggregation (Figure S5A, B). Activation of the HSR by increased levels of cytoplasmic Ca2+ also prompted us to examine the role of Ca2+-dependent kinases, as previous studies from our laboratory and others had identified serine residues of HSF-1 that are stress-inducibly phosphorylated by Ca2+-dependent kinases [49]–[55]. Therefore, we performed a candidate kinase screen to identify the kinases required for the gei-11 RNAi suppression of Q35 aggregation (Figure 5B, Figure S5C) and induction of the HSR in wt animals (Figure 5C). This candidate screen identified calmodulins cal-2 and cal-4; unc-43/CaMKII ortholog; pkc-1, pkc-3; and gsk-3; these genes correspond to the same mammalian kinases previously shown to regulate HSF-1. These results support that AChR upregulation initiates a cascade of Ca2+-signaling events leading to activation of HSF-1. Cholinergic activity at the NMJ leads to activation of the muscle voltage-gated Ca2+ channel (VGCC), EGL-19, and flux of Ca2+ into the cytoplasm [47] (Figure 5A). We therefore determined the role of EGL-19 activity on muscle proteostasis and polyQ aggregation using a partial (30% reduction) loss-of-function mutant (egl-19(n582)), a weak hypermorphic mutant (egl-19(n582ad952)), and a stronger hypermorphic mutant (egl-19(ad695)) in the background of Q35 [47], [56]. Our results showed that the magnitude of EGL-19 activity, that regulates Ca2+ flux into the muscle, had opposing effects: the strongest hypermorphic mutant ad695 enhanced Q35 aggregation, whereas the weak hypermorphic n582ad952 and weak hypomorphic n582 mutants both suppressed Q35 aggregation (Figure 6A) [29]. Consistent with these effects on aggregation, moderate levels of chaperone upregulation (3-fold) were detected in animals where Ca2+ suppressed aggregation (egl-19(n582) and egl-19(n582ad952); Figure 6B). By comparison, much higher levels of hsp-70 (C12C8.1, F44E5.4; 15-fold) were observed in animals with Ca2+-mediated enhanced aggregation (egl-19(ad695), Figure 6B). These results reveal a consistency between modulation of cholinergic signaling and Ca2+ influx on muscle cell folding environment, reflected by the effect on aggregation. Whereas a mild imbalance in Ca2+ influx, achieved with the weak hypermorph and hypomorph mutants, activated a protective HSR (corresponding to a moderate upregulation of hsp-70), the EGL-19 strong hypermorphic mutation, resulted in a much larger imbalance in signaling and accentuated stress response (corresponding to higher levels of hsp-70). These results provide further support for the importance of a critical physiological stimulation range to establish proteostasis. To determine whether the enhanced folding capacity regulated by gei-11 knockdown was dependent on the VGCC, we treated Q35;egl-19 mutant animals with gei-11 RNAi (Figure 6A). Knockdown of gei-11 in hypomorphic egl-19(n582) mutant animals had no effect on Q35 aggregation (Figure 6A) or hsp-70 expression (Figure 6B). This revealed that Ca2+ flux through EGL-19 was required for the beneficial effects of gei-11 knockdown on protein homeostasis. These results were further supported by chemical-genetic approaches using egl-19 RNAi and the EGL-19 inhibitor Nemadipine A [57] that block the gei-11 RNAi effect on muscle proteostasis (Figure 6A, B, S6A, Table S1). Consistent with these observations, the effect of the weak hypermorphic mutant egl-19(n582ad952) was additive to the beneficial effects of gei-11 RNAi on folding (Figure 6A, B), whereas the stronger hypermorphic mutant egl-19(ad695) effect on Ca2+ levels was deleterious (Figure 6A, B). These results establish the role of EGL-19 and Ca2+ influx function downstream of AChR upregulation to the rescue of post-synaptic proteostasis. Activation of the VGCC, and the flux of Ca2+ into the cytoplasm of muscle cells, triggers the opening of the ryanodine receptor (RYR) at the sarcoplasmic reticulum (SR), releasing additional Ca2+ into the cytosol (Figure 5A) [58], [59]. We examined the contribution of RYR to induction of the HSR by stimulating RYR activity to mimic the effect of enhanced cholinergic signaling at the NMJ. Ryanodine (Ryr), a plant alkaloid with high affinity to the RYR, is a pharmacological agent widely used to study intracellular Ca2+ signaling in muscle cells [58], [59]. At low (nM) concentrations, Ryr acts as an agonist and sensitizes RYR channels to activation by Ca2+ [60], [61]. Treatment of Q35 animals with Ryr (50 nM) caused suppression of aggregation (Figure 6C) and upregulation of hsp-70 (Figure 6D), and this effect was reduced in the background of the hypomorphic mutant egl-19(n582) (Figure 6C), supporting the model where enhancing both Ca2+ channels has beneficial effects on proteostasis. As observed for the natural agonist ryanodine, the clinically-used small molecule RYR activator 4-Chloro-m-cresol (4-CmC) [62] also up-regulated hsp-70 levels and significantly reduced Q35 aggregation by more than 60% (Figure 6C, D, S6B). At high concentrations of 4-CmC (>1 mM), we only observed toxicity and no effect on aggregation (Figure S6B). These results establish that Ca2+ release by the RYR is involved in the enhancement of folding in muscle cells. Recognizing that Ca2+ regulates many signaling cascades, we examined the specificity of gei-11 RNAi-dependent Ca2+ release by the RYR on induction of hsp-70 and suppression of polyQ aggregation, by testing a RYR-specific antagonist and a RYR mutant. We employed dantrolene sodium (DS), a clinically relevant muscle relaxant that selectively targets RYR and blocks Ca2+ release from the SR during muscle contraction [62], [63]. This antagonist prevented induction of hsp-70 and suppression of Q35 aggregation by gei-11 RNAi (Figure 6C, D, S6C). Similar results were obtained with the RYR mutant unc-68(kh30) (Figure 6D) [58]. From these observations, we conclude that Ca2+ flux from the RYR-SR is an important component of this new signaling pathway regulating muscle HSR. Finally, treatment of egl-19 hypomorphic mutant animals (Q35;egl-19(n582)) with the RYR modulators had no significant effect on aggregation (Figure 6C) supporting the epistatic relationship of the two Ca2+ channels (Figure 6E). Taken together, these results demonstrate that the downstream events of AChR upregulation involve Ca2+-dependent activation of the HSR, and establish a new proteo-protective state in BWM cells with enhanced folding capacity (Figure 6E). We demonstrate that cholinergic-dependent calcium signaling across the synaptic junction induces an atypical heat shock response that promotes protein homeostasis and suppresses misfolding and aggregation in post-synaptic muscle cells. These molecular events are dependent upon HSF1, but are distinct from the classical HSR regulated by transient exposures to acute heat shock stress. The key feature of this novel neuromuscular stress signaling mechanism, centers around the balance between cholinergic and GABAergic signaling at the NMJ. The importance of balanced signaling is highlighted by observations that muscle overexcitation caused by the complete absence of GABA is deleterious to the folding environment and results in a proteotoxic cellular environment [29]. Our present studies provide additional support for the importance of neuronal signaling in the control of somatic cell protein homeostasis, demonstrating that signaling balance at the NMJ can be perturbed to have either beneficial or detrimental consequences on the HSR-dependent proteome stability. The identification of gei-11 as a new genetic modifier of protein folding reveals a new strategy by which metazoans ensure proteostasis across tissues. Our biological observations suggest that the regulation of receptor expression in muscle cells can initiate a protective mechanism against stress and degeneration, such as age-dependent sarcopenia [11], [64], [65]. These results also suggest that neuronal signaling control of post-synaptic receptor function can achieve the same outcome. This may be highly relevant for complex pathologies, including neurodegenerative diseases and other neuromuscular disorders, where scenarios of protein misfolding initiated at one tissue have both autonomous (cell or tissue specific) and non-autonomous (inter-cellular) consequences on cellular function and organismal health. For example, neurodegeneration leads to muscle weakness and paralysis in motor neuron disorders such as ALS, hereditary spastic paraplegia, and spinal muscular atrophy [66]. Consequently, an understanding of the regulatory signaling cascades that trigger protective responses across tissues is of fundamental importance to delay or prevent the organismal collapse of proteostasis [22]. Modulation of signaling events at the NMJ to rescue muscle function, as described here upon knockdown of the gene gei-11, could suggest novel therapeutic targets for proteostasis maintenance with possible benefit for the patients suffering from somatic wasting diseases. Overall, it emphasizes the importance of dissecting neuronal signaling pathways that affect organismal stress responses and cellular function. We propose that induction of the HSR by physiological regulation of cholinergic receptors reveals a new class of regulatory pathways of HSF-1 and chaperone networks that is distinct from the classical activation of the HSR. gei-11 was identified from a genetic screen for proteostasis regulators that enhanced the cellular folding environment [30], and found to modulate AChR levels. The levels of ACh and GABA that activate the HSR are in contrast with the extreme imbalance and overstimulation of muscle cells caused by the absence of GABA, leading to proteotoxicity in the muscle cells [29]. An intermediate increase in cholinergic signaling at the NMJ, whether by genetic or small molecule upregulation of AChR activity or downregulation of GABAergic signaling, led to selective activation of the HSR and rescued folding capacity in muscle cells. As for cholinergic signaling, the folding rescue effect of Ca2+ influx in muscle cells homeostasis also occurs at a critical range. The activation of both EGL-19 and RYR channels by cholinergic signaling (Figure 6E) led to increased levels of cytoplasmic Ca2+ and activation of HSF-1 and chaperones that were physiologically beneficial, and well below the deregulated levels of signaling of these pathways that cause proteotoxicity. These results shift the emphasis from extreme environmental forms of stress to a new view on the roles of physiologically relevant in vivo stress signaling pathways regulation. Aging and chronic stress challenge the cellular quality control systems by the accumulation of misfolded toxic proteins. Our findings strongly suggest that control of HSR and proteostasis, at the cellular level and at cell-non-autonomous level through neuronal signaling, are critical mechanisms in the cellular challenge to activate proteo-protective pathways and maintain homeostasis at the level of the organism. “Fine-tuning” of post-synaptic receptor expression, and therefore regulation of neuronal cholinergic signaling within physiologically relevant levels, may provide a potential strategy to enhance the functional properties of the proteostasis network. Our results contribute to the growing understanding of the properties of stress response networks, as an integrated organismal response to diverse challenges to the health and lifespan of the organism. Animals were maintained according to standard methods, at 20°C on nematode growth media (NGM) with OP50 E. coli [67]. The strains utilized in this work were previously described: wild-type (wt) Bristol strain N2; polyQ strains Q24 AM138 (rmIs130[Punc-54::q24::yfp]II), Q35 AM140 (rmIs132[Punc-54::q35::yfp]I), Q37 AM470 (rmIs225[Punc-54::q37::yfp]II) [16], [17]; human SOD1G93A AM265 (rmIs177[Punc-54::sod1G93A::yfp]) [5]; intestinal Q44 GF80 dgEx80[pAMS66 Pvha-6::q44::yfp + rol-6(su1006) +pBluescript II] [68]; temperature sensitive (TS) mutant strains CB1402 [unc-15(e1402)], CB1157 [unc-54(e1157)], HE250 [unc-52(e669su250)] and CB286 [unc-45(e286)] [4]; WM27 [rde-1(ne219)] and WM118 [rde-1(ne219);neIs9[myo-3::HA::RDE-1+pRF4(rol-6(su1006))]] [41], [42]; CB1072 [unc-29(e1072)], CB904 [unc-38(e264)], ZZ26 [unc-63(x26)], RB918 [acr-16(ok789)], VC311 [unc-47(gk192)], CB845 [unc-30(e191)], CB407 [unc-49(e407)] [29], MT1212 [egl-19(n582)], DA952 [egl-19(n582ad952)], DA695 [egl-19(ad695)] [47], [56], HK30 [unc-68(kh30)], PS3551 [hsf-1(sy441)]. Where indicated, genetic crosses between mutant animals and Q35 animals were generated. RNAi gene knockdown in C. elegans was performed as described previously, using the commercial RNAi library (GeneService, USA) [30], [69]. Briefly, animals were added to RNAi bacteria (in liquid or RNAi-seeded NGM plates) at the L1 stage (first larval, day 1), incubated at 20°C for 5 days and scored for number of aggregates at 6 days old (which corresponds to 3 days after the onset of Q35 aggregation) [30], using the stereomicroscope Leica MZ16FA (Leica Microsystems, Switzerland). Q35 aggregation was scored in at least 50 animals, for each condition (n≥3). As a negative control, animals were fed bacteria carrying the L4440 empty vector (vector). Liquid RNAi treatment was performed in 96-well plates, with a total volume of 60 µl per well, consisting of 15–20 worms and RNAi bacteria. Bacteria was grown overnight (∼16 h), induced with isopropyl β-D-thiogalatoside (IPTG Sigma, 1 mM for 3 h at 37°C), pelleted and resuspended in S-medium complete (S-Basal supplemented with 3 mM MgSO4, 3 mM CaCl2, 10 mM Potassium Citrate, 100 mg/ml Ampicillin and 1 mM IPTG) so that the final OD595 nm was 0.9 in the well. RNAi assays on plates were performed as described previously [30], and for double knockdown experiments, equal volumes of each RNAi bacteria were mixed (1∶1 ratio) prior to plate seeding. Fluorescent microscopy images were taken using an Axiovert 200 microscope with a Hamamatsu digital camera C4742-98 (Carl Zeiss, Germany). All RNAi plasmids were sequenced to confirm correct and specific gene-target identity. Gene knockdown by RNAi was confirmed by PCR analysis. Animals were mounted on a 3% (w/v) agar pad on a glass slide, immobilized with 2 mM levamisole (Sigma), and subjected to FRAP analysis using the Zeiss LSM510 confocal microscope (Carl Zeiss, Germany) as previously described [30], [70]. The movement of 6 day old animals grown on RNAi-seeded NGM plates (>75 animals per experiment, n≥3) were digitally recorded using a Leica M205 FA microscope with a Hamamatsu digital camera C10600-10B (Orca-R2, Leica Microsystems, Switzerland), and the Hamamatsu Simple PCI Imaging software. Animals were tracked using a custom ImageJ plugin wrMTrck [30]. The average speed of each animal was calculated by dividing the distance of each track, corrected for body length, by the duration (in seconds) of the track (body length per second BLPS) (n≥3). Results are shown relative to wt animals' speed on L4440-RNAi vector control plates (100%). RNA from ∼50 animals was extracted with Trizol (Invitrogen), followed by DNase treatment (Applied Biosystems #AM1906). mRNA was reverse transcribed using the iScript cDNA Synthesis Kit (Bio-Rad #170-8891). cDNA real-time PCR amplification was done using the iQ-SYBR Green Supermix (Bio-Rad #170-8880) and the iCycler system (Bio-Rad) (see Protocol S1). Expression levels of each gene were determined using the Comparative CT Method (Real-Time PCR Applications Guide, Bio-Rad), normalized to actin (act-1) in the same sample, and relative to the non-treated or vector control sample. Measurements were performed for ≥3 biological samples. Five day old animals grown on RNAi-seeded NGM plates at 20°C (≤40) were transferred onto NGM plates, equilibrated at 20°C, containing 1 mM Levamisole (Sigma), 30 mM Nicotine (Sigma) or the solvent (water or ethanol, respectively). Sensitivity to the drugs was followed by visual inspection every 2 to 5 min and defined as paralysis, or lack of movement in response to prodding on the nose and tail of the animal (n = 3). Compound stock solutions: 800 mM levamisole (Sigma) in water, 300 mM nicotine (Sigma) in ethanol. Compound assays were performed in liquid culture as described previously [71], with 60 µl of final volume per well, 15–20 animals (in S-Basal complete), compound at the appropriate concentration and bacteria (OP50 or RNAi) at a final OD595 nm of 0.9 (resuspended in S-Basal complete). Replicates of each condition were included in each assay/plate. Animals were incubated with each drug, at the concentrations indicated in the respective Figures, from L1 stage (levamisole, GABA), L2 stage (Lindane, Ryanodine, Nemadipine A, Dantrolene Sodium, 4-CmC) or L4 stage (dTBC, BAPTA), until day 6 of age, at 20°C (n≥3) (see Protocol S1). At this time animals were transferred from liquid culture onto NGM plates for aggregate quantification (Leica MZ16FA) and collected for real-time qPCR analysis. Temperature sensitive (TS) mutant animals were age-synchronized to L1 stage, grown on RNAi-seeded NGM plates (∼20 animals per plate) from day 1 at a sensitized temperature of 23°C (to maintain the RNAi suppressor effect on aggregation), or at the control restrictive (25°C) and permissive (15°C) temperatures, and scored for phenotypes on day 5, as previously described [30]. >50 animals were scored for each TS phenotype, per assay: slow movement/paralysis assay for unc-15(e1402) and unc-54(e1157), stiff paralysis for unc-52(e669su250), and egg-laying phenotype for unc-45(e286) (partially paralyzed animals with a large belly of accumulated eggs) (n = 3) [4], [29], [30]. Native nuclear protein extracts were prepared from 200 µl of pelleted worms (grown on NGM RNAi-seeded plates), with the commercial Thermo Scientific NE-PER Nuclear and Cytoplasmic Extraction Kit (# 78835), as described previously [72]. Electrophoretic mobility shift analysis (EMSA) was performed as before [43] using a [32P]-labeled probe containing the proximal heat shock element (HSE) from the C. elegans hsp-70 (C12C8.1) gene promoter. Nuclear extracts (40 µg) were incubated with the [32P]-labeled probe (HSE or mutant) for 20 min at room temperature. For heat shock treatment (HS) the samples were pre-incubated at 35°C for 30 min. For competition experiments, a 100-fold molar excess of the same unlabeled oligonucleotide was added to the mixture. The samples were analyzed by electrophoresis on a 4% (w/v) polyacrylamide native gel that was dried and scanned using a PhosphorImager (Molecular Dynamics, Sunnyvale, CA). Oligonucleotide probes: HSE-F: taaattgtagaaggttctagaagatgccaga; HSE-R: tctggcatcttctagaaccttctacaattta; HSEmut-F: taaattgtaaaaggaaataaaagatgccaga; HSEmut-R: tctggcatcttttatttccttttacaattta.

10.1371/journal.pcbi.1000878

Lysine120 Interactions with p53 Response Elements can Allosterically Direct p53 Organization

p53 can serve as a paradigm in studies aiming to figure out how allosteric perturbations in transcription factors (TFs) triggered by small changes in DNA response element (RE) sequences, can spell selectivity in co-factor recruitment. p53-REs are 20-base pair (bp) DNA segments specifying diverse functions. They may be located near the transcription start sites or thousands of bps away in the genome. Their number has been estimated to be in the thousands, and they all share a common motif. A key question is then how does the p53 protein recognize a particular p53-RE sequence among all the similar ones? Here, representative p53-REs regulating diverse functions including cell cycle arrest, DNA repair, and apoptosis were simulated in explicit solvent. Among the major interactions between p53 and its REs involving Lys120, Arg280 and Arg248, the bps interacting with Lys120 vary while the interacting partners of other residues are less so. We observe that each p53-RE quarter site sequence has a unique pattern of interactions with p53 Lys120. The allosteric, DNA sequence-induced conformational and dynamic changes of the altered Lys120 interactions are amplified by the perturbation of other p53-DNA interactions. The combined subtle RE sequence-specific allosteric effects propagate in the p53 and in the DNA. The resulting amplified allosteric effects far away are reflected in changes in the overall p53 organization and in the p53 surface topology and residue fluctuations which play key roles in selective co-factor recruitment. As such, these observations suggest how similar p53-RE sequences can spell the preferred co-factor binding, which is the key to the selective gene transactivation and consequently different functional effects.

p53-response elements (p53-REs) are 20 base pairs (bps) DNA segments recognized by the p53 transcription factor (TF). They are found in promoters and enhancers across the genome and are associated with genes that have diverse functions. Because the DNA sequences of p53-REs can be very similar to each other, differing by as little as one or two bps, it is challenging to understand how p53 distinguishes between these to activate a specific function. Here we show that even a slight RE sequence change can be sufficient to elicit allosteric structural and dynamic perturbations in the p53 which propagate to other binding sites, and as such are expected to affect co-regulator recruitment. Among the major interactions between p53 and its REs involving Lys120, Arg280, and Arg248, the Lys120 interaction partners vary less than interactions between other residues. The outcome of our simulations of six p53-RE complexes shows that the variance of the interaction patterns triggers changes in the organization of tetrameric p53 and of residues away from the interaction sites. Subsequent events can depend on the level and post-translational states of co-regulators that are able to bind the unique p53 surface caused by the specific p53-RE binding.

p53-response elements (p53-REs) are two 10-bp palindromic DNA segments with the consensus sequence of 5′-Pu1Pu2Pu3C4(A/T)5(A/T)5′G4′Py3′Py2′Py1′-3′ for each of the two half sites, where Pu and Py stand for purine and pyrimidine bases, respectively [1], [2]. The two half sites can be separated by as many as 20 bps [1]–[6]. Hundreds of p53-REs have been identified [2], [5], and the numbers continue to grow [7]. Many of these are known to be related to regulation of genes involved in cellular pathways such as apoptosis, cell cycle arrest and senescence [8], [9]. However, upon stimulation only a small subset are selectively activated for transcriptional activation or repression through sequence-specific binding to tumor suppressor p53. Understanding the factors that determine the selective activation is crucial for deciphering the complex gene regulation by p53 [7], [10]–[14]. Binding affinities of functionally-diverse p53-REs showed that apoptosis-related p53-REs have higher affinities than cell cycle arrest-related p53-REs; however, at the same time, the affinities do not always correlate with functional effects [7], [12], [15], [16]. Spacer sizes also affect affinities: in spacers consisting of three or more bps, the two 10-bp half-sites are on opposite faces of the DNA [17], suggesting specific p53-RE interactions only with a single half-site, which results in lower affinity [7], [17]. Although several structures are available [9], [18]–[23], they involve a few engineered p53-REs and do not explain the in vivo selectivity. In vivo, p53-RE binding is affected by chromatin packaging epigenetic events known to be a key factor in RE occupancy [24], [25]. Nonetheless, even assuming genomic p53-REs availability, the question of the selective recognition by p53 still remains [12], [13]. Allostery is key to cellular signal transduction [26]–[30]. Mechanistically [12], [13], allostery can play a role either via protein co-factors binding to p53 prior to RE binding as could be in HIF-1 regulation of p53 and p300 [31], or ASPP family binding [32]; or via allostery-induced by RE sequences [33]–[37], or spacer sizes as in the pituitary-specific POU domain factor Pit-1 [38], in both cases through preferential interactions with certain side chain conformations [34]. In p53, RE bp changes were observed to relate to transactivation [39]. In the glucocorticoid receptor (GR) [40], [41], single bp changes were shown to allosterically affect GR conformational changes. These were amplified by ligand binding and propagated to the co-regulator binding site. Allosteric effects can shift the population toward co-factor binding-favored states. DNA methylation can lead to packing of the genome, making the REs unavailable; but it was also proposed to change the affinities of the REs [42], [43] either via direct interactions, or through allosteric effects on the DNA or the protein. In proteins, covalent modifications such as phosphorylation, glycosylation, and acetylation are well established to be allosteric effectors. The tetrameric p53 DNA-binding domains (DBD) are responsible for specific RE binding. However, the impact of the DNA sequence on the binding patterns, specificities and complex conformation has been studied only for the central 4 bps [44], [45]. Computational studies revealed that variation of the central four bps in the half site which contained the C(A/T)(T/A)G, conserved in most REs, resulted in conformational changes in the DNA and the DBD [45]. However, the impact of RE sequence variation in other bps on the complex organization and its dynamic properties is largely unknown due to the sparseness of available crystal structures. Here, using molecular dynamics (MD) simulations we study the conformational and dynamic consequences of p53 binding to six diverse p53-REs. We focus on the impact of specific interactions of Lys120, Arg280 and Arg248 with DNA as these are the most crucial for binding. We find that p53 Lys120-DNA interactions can change dramatically depending on the bp at positions 1-3 of the quarter site, which in turn affects the Arg280 binding. We find that such binding pattern changes at the DNA-protein interface have allosteric effects in terms of the p53 tetrameric organization and the fluctuations of residues on the p53 surface away from the DNA binding site. We propose that this combined allosteric effect could hold the key to selective transcriptional activation by the degenerate p53-REs and can serve as a paradigm for selective activation of transcription factors [13]. Six naturally-occurring p53-REs were selected, two each from the cell cycle arrest, DNA repair and apoptosis functional groups (Table 1). These REs differ from the consensus sequences by 1–3 bps (Table 1). To analyze the impact of the sequences on p53 binding, conformations and organization, hydrogen bond (HB) distances for p53 residues Lys120, Arg280, Arg248 and Arg273, DNA conformational differences, residue deviation and fluctuations in each quarter site (denoted as Q1, Q2, Q3 and Q4) and overall complex organizations were monitored. In the crystal structure Lys120 and Arg280 form HB with DNA bases in the major groove, while Arg248 anchors in the minor groove through electrostatic interactions (Fig 1a). The salt bridge network among Arg280, Glu281, and Arg273 (interacting with the DNA backbone) enhances the specific protein-DNA interactions (Fig 1b). Lys120 can interact with bps at three positions (positions 1–3 in a quarter site) (Fig 1a). However, the interaction patterns can vary, depending on the base identity. With a G base, Lys120 can make three center HBs (Fig 1c). For C, Lys120 can make the same interactions with the G on the other chain, but the protein has to adjust its relative position. For an A or T, Lys120 can only make one HB with either base but not both because the two HB acceptors are 6–7 Å apart in a Watson-Crick bp (Fig 1d). The methyl group next to the T O4 atom can also influence the interactions. All six potential HB distances for the three bps were monitored (Fig. S1) and the percentage of distances less than 3.5 Å are summarized in Table 1. Fig 2 highlights the average local conformation of Lys120 and Arg280 for selected binding sites. The results show that (a) with a quarter site whose sequence conforms to the consensus, Lys120 interacted mainly with the central G or A base, as in the crystal structures (Table 1: 14-3-3σ Q1 and Q4, Gadd45 Q2, Noxa Q1 and Q2, p21-5 Q1 and Q2, p53R2 Q2, Q3 and Q4, puma Q2 and Q4); the representative structure in Fig 2A shows that all four hydrogen bonds are well maintained. The simulations showed that Lys120 also interacted with G or A at positions 1 or 3 in these cases; the only exception is Gadd45 Q1 where Lys120 mainly interacted with G1 (Table 1 and Fig 2B), suggesting that G is preferred for HB; this was not observed in Gadd45 Q3 and p21-5 Q1, suggesting that geometrically the central position is more favorable for Lys120 interactions. (b) When there is a single base mutation, the mutation is at position 1 and the mutated base is C, Lys120 interacted with the central A or G (Noxa Q4, p21-5 Q3 and Q4, Puma Q3) or with both bases at the 2nd and 3rd positions (Gadd45 Q4, Noxa Q3); this is expected since Lys120 is unlikely to interact with G on the other chain at the 1st position. A typical structure is shown in Fig 2C. The interaction with the central base is usually weak if the base is A (Gadd45 Q4, Noxa Q4, p21-5 Q4); however, if T, the interaction is either abolished (p53R2 Q1) or weakened even when G is at the 2nd position (Puma Q3 in Fig 2D); the extra methyl group of T hampered the favorable Lys120 interaction with the 2nd G. (c) If the mutation is at the 2nd position (14-3-3σ Q2), Lys120 interacted with G at the 1st position (Fig 2E); although in this case Lys120 could interact with the A at the 3rd position, the fact that it did not suggests that Lys120 preferred G over A. Reaching the base at the 3rd position is also more difficult due to steric hindrance, requiring the movement of the whole protein. (d) When there were two mutations in a quarter site, Lys120 interacted weakly with the unmutated base (14-3-3σ Q3 and Puma Q1); in the case of 14-3-3σ Q3 the result is expected since both mutated bases were C which does not have HB acceptors; in the case of Puma Q1, the 2nd mutated base was T which was able to form HB; however, there was very little interaction with this base due to the presence of the protruding methyl functional group on T. The only option is the G at the 3rd position, which was also weak for reasons discussed earlier. More dramatic conformational adjustment is needed for better interactions between Lys120 and bases at the 2nd or 3rd positions. These results indicate that both base position and identity are important for specific binding. Lys120 is able to interact with bases at all three positions, depending on the environment; however, unless more significant conformational adjustment is involved, the binding of Lys120 to bases on the opposite DNA strand is not likely as it was only observed in a quarter site with a small population. The outcome is a unique binding pattern which can lead to a shift of the p53 organization and DNA conformation. The C at the 4th position is absolutely conserved in all the REs studied here and in most other known p53-REs. The importance of this bp for specificity and affinity has been shown (39,44). In addition, Arg280 formed a salt bridge with Glu281 as part of the HB network in Fig 1b. Arg280 distance fluctuation details are shown in Fig S2 and the HB percentages are summarized in Table 2. Unexpectedly, in many cases the Arg280-C HBs were disrupted for at least two of the four quarter sites for each of the six REs and the salt bridges were also very dynamic (Table 2 and Fig S2), suggesting HB sensitivity to environmental changes, possibly influenced by Lys120-DNA interactions. For example, in the complex of RE 14-3-3σ, Arg280 HB with DNA was intact for Q1 (Fig 2A) and 4, where Lys120 maintained its HB with the 2nd bp (Tables 2 and 3). This was also the case for Noxa Q1 where Lys120-DNA had good interactions at the 2nd and 3rd positions and Arg280 specific interactions were reasonably maintained as well, showing a good correlation between Lys120 and Arg280 interactions. In Q2 of the 14-3-3σ complex, Lys120 interacted with the base at the 1st position, which loosened the p53 from its original position and reduced the tightness of the Arg280 interaction with the G (Fig 2E, Tables 2 and 3). When Lys120 flipped out of the binding site, as in Q1 of the p53r2 complex, Arg280 also lost both HBs (Fig 2G). Similarly in Noxa Q3, Lys120 interacted with G3, which pushed Arg280 away from its original position, resulting in a conformation in which Arg280 interacted with the DNA backbone (Fig 2C). These results indicate cooperativity between the Arg280 and Lys120 interactions. Interestingly, in the case of Noxa Q4, Lys120 also flipped out of the major groove, yet the Arg280 interactions were still present (Fig 2H). However, such interactions without the concurrent HB of Lys120 nearby are expected to be vulnerable to environmental perturbations. There are also cases where Lys120 interacted with the 2nd base (G or A) but the Arg280 interactions were disrupted. Such changes were observed in the RE p21, Q1 and Q2 complexes. In both cases, Arg280 only partially maintained HBs with the bases (Fig 2I). These results indicate that specific HBs of Lys120 and Arg280 not only affect each other, but are also influenced by other interactions, such as the dynamic Arg248 interactions (Fig S3) and the Arg280, Glu281 and Arg273 salt bridge network (Table 2, Fig S4). However, the major factor in determining the conformational changes of the p53-DNA complex is the RE sequence at the Lys120 interaction site, which forces p53 to adjust its conformation locally and consequently the overall organization with respect to the DNA. Interactions at other sites such as those involving Arg280 and Arg248 also adjust their interactions even if the DNA sequences are unchanged. Thus, even very similar REs, which vary only by a single or a few bps, elicit different patterns of p53-RE interactions perturbing the p53, the DNA and their organization in different ways. The conformation with Arg248 inserted into the DNA minor groove was captured only in one crystal structure [46]. In others, Arg248 docked only at the edge/surface of the DNA backbone [20], [21], [47]. Arg248 was inside the minor groove at the beginning of our simulations. Once the simulations started, the residue was “ejected” in several complexes and then interacted with the backbone from the outside (Fig S3). As a result, Arg248 shifted away and adopted a conformation similar to those observed in some of the crystal structures. The change in Arg248 interaction patterns would affect the p53 conformation and cause conformational differences among the complexes. In order to further confirm the relationship between the sequence and the resulting complex conformations, the simulations of 14-3-3σ 1st half site, Gadd45 1st half site, and the Puma 2nd half site were repeated. In 14-3-3σ Q1 (Fig S5A) where Lys120 was expected to interact with the 2nd G base, these HBs were well maintained. In the Gadd45 Q1 (Fig S5B), the respective DNA sequence G1A2A3C4A5 suggests that Lys120 may prefer to interact with the G1 base as observed previously. These interactions were retained reasonably well, with Lys120 positioned within distance capable of HB formation. Because the DNA sequence in Puma Q3 is T1G2A3C4T5, it is expected that the presence of the methyl group on the T base at the 1st position would disrupt the Lys120 HB with the 2nd G base, which was indeed observed (Fig S5C). Comparison of these HB patterns for Lys120 and Arg280 with the corresponding panels in Fig 2A, B and D illustrates consistent and reproducible conformational preferences for a given DNA sequence. The other quarter sites for each of the three complexes were also analyzed and the results were consistent as well. Above, depending on bp identity in each RE the interactions were different. These subtle differences can allosterically propagate in both DNA and p53. To characterize these features, conformational changes for both the p53 and DNA were calculated. For p53, the RMS deviation (RMSD) of selected residues and RMS fluctuations (RMSF) of all residues were calculated (Figs 3 and 4). We focused on residues near Lys120 and Arg280. For 14-3-3σ, large RMSDs were observed for Lys120 in Q3 (Fig 3A); correspondingly, larger RMSF were observed for residues 96–100 and 125–135 next to Lys120 (Fig 4A). For Gadd45, Lys120 shifted significantly away in Q3 (Fig 3B), resulting in its large fluctuations and in nearby residues 115–140; although Lys120 in Q1 also had large RMSD, its interactions with the DNA backbone stabilized (Fig 3b). Noxa has a large RMSD for Lys120 in Q4 (Fig 3c). However, the RMSF was small, similar to Q1 in Gadd45. In p21, Q2 and Q4 had large Lys120 deviations (Fig 3d), slight increase in RMSF nearby in Q2, and large RMSF increase in nearby residues (100–110) in Q4 (Fig 4d). The RMSD for Arg248 were large in Q3 and Q4. Although the RMSF increase for Arg248 was not significant, it was higher for nearby residues 225 and 244. In the case of p53r2, large RMSDs of Lys120 in Q1 and of Arg248 in Q3 were observed (Fig 3e); the RMSF of residues 114–136 in the 1st and of residues 230–250 in Q3 also increased correspondingly (Fig 4e). For Puma, the RMSD of Lys120 in Q1 and Q3 were relatively large (Fig 3f), resulting in neighboring residues 111 and 125–132 in the 1st and 115–125 in Q3 fluctuating more (Fig 4f). While the RMSD for Arg248 in Q3 was also large, the RMSF of nearby residues changed little, although the pattern of the fluctuation magnitude was somewhat different from the other quarter sites. For the DNA, Table 3 summarizes the bending extent from the last 5 ns of each trajectory, illustrating the allosteric impact on the interactions. Thus, adjustments of specific interactions lead to larger fluctuations of nearby residues. In some cases these residues extended to the other side of the protein, suggesting amplified allosteric effect of the DNA on p53, which is likely to be important for selective co-regulator recruitment. To characterize the conformational changes of the complex elicited by the specific interactions, an angle and a dihedral angle were defined with two atoms from the protein (Cα of S269 and G112) and two from the DNA (C3′ at positions 0 and 4′) (see Fig 1B). These two geometrical parameters were expected to reflect the organizational change of the p53 core domain with respect to the DNA because the two protein atoms are located at the centers of the β-sheet secondary structures and the two DNA atoms belong to the base pairs that are in close contact with the corresponding p53. The calculated results (Table 4) show that the organizations of the p53 monomer-DNA varied to a large extent, ranging from 96 to 112 and from 14 to 44 degrees for the angle and dihedral angle, respectively (Table 4). In the context of the tetrameric p53-DNA complex, such orientation changes for each p53 core domain with respect to the DNA will propagate to the p53 surface away from the DNA binding site. The two examples shown in Figs 5 and 6 illustrate the conformational adjustments between p53 and the DNA. In the 14-3-3σ complex, the RMSDs of both p53 core domains were small (2.5 Å for all atoms) (Figs 5a and 5b). However, when the systems were superimposed with the DNA as the pivot, the p53 orientation changes significantly (Figs 5c and 5d). A major reason for such a change is the interaction pattern. Fig 5e shows that when Lys120 interacts with the G at the 1st position, Lys120, Arg280 and the whole molecule shifted significantly. The significant change of the helix orientation highlights this organizational difference (Fig 5d) which is also reflected in the small dihedral angle (17°) (Table 4). Although no large conformational changes were observed in the p53 itself in this case, allostery can be at play even with minor conformational changes [28]. In the p53 core domain, allosteric fluctuations were observed at locations distant from the allosteric perturbation site [48]. In the case of the p53r2 complex, the flip-out of the Lys120 in one core domain resulted in large protein backbone change (Fig 6a) relative to the other p53 (Fig 6b), leading to a conformational change on the surface of p53 away from the DNA binding site. Both p53 core domains shifted significantly in their orientation with respect to their corresponding DNA quarter sites (Figs 6c, 6d), an outcome of the amplified allosteric effect between the protein and DNA. Lys120 and Arg280 are the two major factors that determine the binding specificity to the p53-REs. While Arg280 mostly interacts with the G base at the 4th position within a quarter site, the adjustment of Lys120 interaction may affect the Arg280 interaction since these two residues are next to each other. To see if the two interactions are correlated, covariance map (Fig S6), interaction energy between the two residues (Fig S7), and the correlation between the HB distances of the two residues with DNA bases (Fig 7) were calculated. The covariance map revealed that the movements of residues 115–125 were negatively correlated with different portions of the p53 core domain, depending on the DNA sequence. One common negatively correlated portion was residues from 175–185, suggesting that the movement of the residues near Lys120 will affect the residues at the dimerization interface. Since these correlations were quarter-site specific, it is difficult to draw a general rule regarding the correlation between the conformational change and the RE type. The interaction energies between the two residues showed near zero net interaction energy (e.g. 14-3-3σ Q1, Q2, Q4) when Lys120 and Arg280 assumed near crystal structure conformation. When Lys120 popped out of the binding pocket, the interaction energies became either more favorable (14-3-3 σ Q3, Noxa Q4, Puma Q1) (Fig S7A, C, F), or less favorable (Gadd45 Q1, p21-5 Q2, Q4), or mostly changed little when Lys120 did not flip out. These results suggest that the altered packing of Lys120 triggers the readjustment of the Arg280 interactions with the new environment. Such a relationship is also reflected in the HB distances. Fig S8 shows that when the Lys120 HB broke, those of Arg280 also quickly disrupted (14-3-3σ Q2, Q3; Gadd45 Q3, Q4; p53R2 Q1; Puma Q3). Although in some cases the Lys120 HB disruption did not necessarily result in the disappearance of Arg280 HBs within the limited simulation time (Noxa Q4; p21-5 Q4; Puma Q1), their stability in the long run is likely to be compromised due to the lack of tight packing. To further demonstrate the correlation between the movement of Lys120 and Arg280, we present snapshots from two trajectories. Fig 7 shows that the conformational changes happened very early in the trajectories. For 14-3-3σ Q2 (Fig 7A), the distance between Lys120 and the C base at the 2nd position of the quarter site was too close (1.63 Å) and too far (3.66 Å) to interact with the G base at the same position on the complementary chain in the initial structure. After 0.01 ns, Lys120 shifted away from the 2nd bp moving toward the 1st bp, causing the weakening of the neighboring Arg280 HB (Fig 7A) with subsequent adjustment of the interactions of both residues with the DNA. While Lys120 was settling with the G1 base from 0.01 to 1 ns, Arg280 continued to lose contact with G4 base, shown by the longer interaction distances. In the p53R2 Q1 trajectory, both Lys120 and Arg280 HBs were nicely organized in the starting structure (0 ns) (Fig 7B). Because of the protruding methyl group of the T base at the 1st position of the quarter site, Lys pulled away from the G base at the 2nd position to avoid steric clash (0.1 ns) and drifted further away from the starting point (0.5 ns). While Lys120 was searching for favorable positions after pulling away from the major groove, Arg280 started to fray and the HB distance from the G base became longer and out of range from 1 to 1.5 ns. The final settled conformation is similar to that at 2 ns (Fig 7B). When compared with structures where both Lys120 and Arg280 maintained their HBs with the 2nd and 4th bases, these two examples clearly demonstrate that the movement of Arg280 or the loss of Arg280 HBs was the outcome of the Lys120 movement. In each quarter site, the p53-REs largely conform to the consensus sequence and are highly similar to each other. This raises a key question that has been largely overlooked [12], [13]: how does the small, often minor sequence variation of a single or few bps, translate into vastly different functional consequences, spelling transcription activation or repression? The in vitro, or cell-based affinity experiments do not necessarily correlate with the functional consequences [8], [9] and the sparseness of available experimental structures makes such an investigation highly challenging [49]. Our computational results provide insight into this crucial question, illustrating how minor DNA sequence changes can impact subsequent recognition events which in turn determine the functional outcome. We show that subtle conformational changes elicited by DNA sequences which can differ by as little as a single bp can result in altered p53 core domain organization and protein surface dynamics. The DNA is an allosteric effector; slightly different RE sequences lead to minor alterations in the core domain-DNA interactions. The core domain conformational changes may propagate and thus allosterically impact the full protein including the N- and C-terminal domains, providing preferred surfaces for recruitment of specific co-regulators such as STAGA [50], [51], CBP/p300 and HDM2 [52]. The amplified allosteric changes at the p53 surface can select different co-regulators [13]. Conformational selection and population shift have been proposed to play a key role in biomolecular recognition [26]–[28], [53], [54]. Cofactor binding can also affect RE selectivity by transcription factors through an alternative allosteric mechanism [12], [13]. In this case, the prior binding of the co-regulator will shift the population of the transcription factor leading to altered DNA-binding site conformation. ASPPs (apoptosis-stimulating proteins of p53) for example, when bound to p53 core domain, can shift the p53 ensemble enhancing a conformation that favors binding to specific p53-REs [12], [13], [55]. In light of the findings from this work, it is likely that the ASPP binding changes the loop L1 conformation of the p53 core domain, which has been demonstrated to be of crucial importance to the specificity of RE binding. The structured L1 loop could govern the allosteric pathway mediating these binding sites. The features captured here are only part of the story. DNA sequence variation can also code for the differential binding of p53 family proteins. For example, RE2 of the target gene GDF15 contains sequence variations that allow only p53 but not p63 and p73 binding [56]. This may explain why DNA sequences GGG, GGA or AGG all have similar binding patterns and affinities with p53 [20] but in combination can exclude the binding of other proteins. We further note that although our results clearly show that the p53-DNA interaction patterns and conformational and residue fluctuations vary with DNA sequence, allostery may not be saliently evident in some cases. The allosteric structural perturbations observed in experiments or simulations are the sum of multiple, major and minor pathways [57] and these may not be detected in the current analysis. The transmission of the signal over long distances may be difficult to observe in short MD simulations, and conformations that are relevant for cofactor binding may have high barriers to go through or higher energy, i.e. be less populated [58] and difficult to observe in simulations [59] and in experiment [58], [60]. However, recently a series of crystal structures coupled with biochemical and cell-based assays have shown how the glucocorticoid (GR) REs that vary by even a single bp can lead to different GR conformations at a cofactor binding site, thus affecting GR regulatory activity [13], [14], [40]. The cellular network, which reflects the environment, contributes critically to transactivation selectivity [12], [13] and p53 acetylation was shown to be related to the differential activation of apoptosis or cell cycle arrest [61], [62]. Methylation of cofactors such as the heterogeneous nuclear ribonucleoproteins hnRNP K can hamper the recruitment of p53 to the REs [63]. Similarly, arginine methylation in p53 may also control target gene selectivity [64]. Post-translational modifications of p53, including phosphorylation and acetylation [65], allosterically alter its activity. Covalent modifications provide an added level of cellular network regulation, in addition to protein co-regulator availability which is also regulated by the network in response to changes in the cellular environment. Although not addressed here, sequences flanking the REs are important for the overall organization of the complex, likely also via allosteric effects, combinatorial assembly of other transcription factors binding in these regions [13] and chromatin remodeling. Flanking segments assist in co-regulator transcription recruitment, as shown for the human BAX promoter [66] which can allosterically trigger conformational changes in p53 and neighboring DNA sequences, rendering the binding surface that is specific for cofactor binding. Further, the p53 core domain dimers interactions with DNA and with each other are primary factors responsible for specific cooperative DNA binding, with the interactions enhanced in the full-length protein [16]. The C-terminal domain is also involved in the interactions. While not included here, allosteric effects observed in this work further implicate the conformations of other p53 domains. p53-REs can have spacers with sizes ranging between 1–20 bps. p53-REs with 5- or 6-bp insertions have the weakest binding even with full fledged p53 [67]. p53 dimer-dimer cooperative interactions are important for function [17], and such cooperative interactions are unlikely for systems with 3–6 (and probably more) bp spacers [17]. In some cases, there is only one RE half site and there can still be significant transcriptional activity [68]. In these cases, the allosterically amplified p53 conformational changes induced by half-site DNA could still be large enough for specific recruitment of transcription co-regulators, while the second p53 dimer may bind DNA non-specifically. The notion that even when there is one bp change allosteric effects can still specify biomolecular recognition and hence determine function supports the likelihood that specificity of the 10-bp half site p53-REs is sufficient. Selective p53-related gene expression requires p53 binding to DNA and pre- and post-DNA binding regulatory events such as modifications of both p53 protein and DNA [69], the recruitment of transcriptional cofactors and RE availability. In a recent example [70], there exists an identical transcriptional target in apoptosis promoters such as BAX and Puma that was selectively blocked by SMAR1 expressed under mild DNA damage conditions. Under severe DNA damage, other factors displace the SMAR1 protein to allow the initiation of apoptotic processes. The actual repression of the relevant genes might involve direct p53 binding onto the target sites [71]. While selective transcription mechanisms are still unclear [12]–[14], our findings here on the p53-RE binding-induced selectivity and future developments are expected to provide further insight into the mechanisms of RE selectivity and the regulation of the first step in transcription initiation. To conclude, here we describe a molecular dynamics study of the p53-DNA interaction, particularly focusing on amino acids that make direct contact with DNA bases. We found that the side chain of Lys120 was able to make a number of alternative contacts with DNA bases at positions 1–3. This observation is consistent with low experimentally observed sequence specificity for p53 binding. We further observed that the conserved interaction of Arg280 with its cognate base pair may be broken in some cases, and that Arg248 is more likely to interact with the DNA backbone than make specific contact with DNA. We show that variant Lys120 interactions with bases at different positions can shift the overall p53-DNA interaction patterns, and how the conformation adopted by Lys120 influences the conformation adopted by other DNA-interacting residues. Most interestingly, the relative orientation of the p53 core domain and DNA changes depending on the sequence of the response element. This leads us to conclude that different response elements will result in different organization of p53-DNA complexes, potentially exposing different surfaces. This, in turn, could result in recruitment of different co-factors and explain the different functionality of response elements whose sequence differs by only a few nucleotides. MD simulations were performed on 12 p53 dimer-DNA half site complexes constructed based on the p53-DNA crystal structure with the PDB code 1tsr [46]. The detail construction methods of the models were described in the next section. Each system was solvated with a rectangular TIP3P water box [72] with a margin of at least 10 Å from any edge of the box to any protein or DNA atom. Solvent molecules within 1.6 Å of the DNA or within 2.5 Å of the protein were removed. The systems were then neutralized by adding sodium ions. The resulting systems were energy minimized for 1000 steps before the dynamic run using the CHARMm program [73] and the CHARMm 22 and 27 force field for the protein and nucleic acid, respectively [74]. The production MD simulations were performed at temperatures of 300 degrees Kelvin using the NAMD program [75] and the CHARMm force field. Periodic boundary conditions were applied and the non-bonded lists were updated every 20 steps. The NPT ensemble was applied and the pressure kept at 1 atom using Langevin-Nose-Hoover coupling. SHAKE constraints on all hydrogen atoms and a time step of 2 fs and a nonbonded cutoff of 12 Å were used in the trajectory production. The sizes of the systems were about 110,000 atoms and the duration for each simulation was 30 ns. The p53 core domain dimer-half site DNA complex was generated based on the crystal structure template (PDB code: 1tsr) [46], as described earlier [44], [45]. Briefly, we used two copies of the p53 monomer-DNA complex crystal structure and then superimposed the 10 consensus base pairs from the two copies of the extracted p53-DNA complex in reverse order so that the two copies of p53 were bound to two consecutive quarter sites of the DNA. The resulting p53 dimer-DNA complex structure ensures specific DNA-p53 binding and that the two copies of p53 have a C2 symmetry, with formation of the two salt bridges between Arg180 and Glu181 from the H1 helices of the p53 core domains. The DNA sequences that capped the 5′ and 3′ ends were 5′-ATAATT-3′ and 5′-ATTAA-3′, respectively. Each base pair that was different from the target sequence was mutated by removing the atoms in the base motif and these atoms were regenerated with GENERATE module in the CHARMm program. The systems were then minimized for 2000 steps with SD algorithm, the mutated base pairs were allowed to move with the NOE restrictions that all the distances between hydrogen bond partners (heavy atoms) were within 2.6 and 3.0 Å. The rest of the system was not allowed to move by applying a force constant of 2 kcal/mol/å during the minimization. The obtained structures were then further minimized for 1000 steps with the ABNR algorithm without any restriction. The models obtained in such a manner yielded reasonable local and overall conformations and served as the starting structure for the MD simulations. For the three duplicate simulations for the purpose to ensure the reliability of the results, additional 1000 steps with the ABNR algorithm was applied before the start of MD trajectories.

10.1371/journal.pcbi.1006557

Representations of regular and irregular shapes by deep Convolutional Neural Networks, monkey inferotemporal neurons and human judgments

Recent studies suggest that deep Convolutional Neural Network (CNN) models show higher representational similarity, compared to any other existing object recognition models, with macaque inferior temporal (IT) cortical responses, human ventral stream fMRI activations and human object recognition. These studies employed natural images of objects. A long research tradition employed abstract shapes to probe the selectivity of IT neurons. If CNN models provide a realistic model of IT responses, then they should capture the IT selectivity for such shapes. Here, we compare the activations of CNN units to a stimulus set of 2D regular and irregular shapes with the response selectivity of macaque IT neurons and with human similarity judgements. The shape set consisted of regular shapes that differed in nonaccidental properties, and irregular, asymmetrical shapes with curved or straight boundaries. We found that deep CNNs (Alexnet, VGG-16 and VGG-19) that were trained to classify natural images show response modulations to these shapes that were similar to those of IT neurons. Untrained CNNs with the same architecture than trained CNNs, but with random weights, demonstrated a poorer similarity than CNNs trained in classification. The difference between the trained and untrained CNNs emerged at the deep convolutional layers, where the similarity between the shape-related response modulations of IT neurons and the trained CNNs was high. Unlike IT neurons, human similarity judgements of the same shapes correlated best with the last layers of the trained CNNs. In particular, these deepest layers showed an enhanced sensitivity for straight versus curved irregular shapes, similar to that shown in human shape judgments. In conclusion, the representations of abstract shape similarity are highly comparable between macaque IT neurons and deep convolutional layers of CNNs that were trained to classify natural images, while human shape similarity judgments correlate better with the deepest layers.

The primate inferior temporal (IT) cortex is considered to be the final stage of visual processing that allows for object recognition, identification and categorization of objects. Electrophysiology studies suggest that an object’s shape is a strong determinant of the neuronal response patterns in IT. Here we examine whether deep Convolutional Neural Networks (CNNs), that were trained to classify natural images of objects, show response modulations for abstract shapes similar to those of macaque IT neurons. For trained and untrained versions of three state-of-the-art CNNs, we assessed the response modulations for a set of 2D shapes at each of their stages and compared these to those of a population of macaque IT neurons and human shape similarity judgements. We show that an IT-like representation of similarity amongst 2D abstract shapes develops in the deep convolutional CNN layers when these are trained to classify natural images. Our results reveal a high correspondence between the representation of shape similarity of deep trained CNN stages and macaque IT neurons and an analogous correspondence of the last trained CNN stages with shape similarity as judged by humans.

Recently, several studies compared the representations of visual images in deep Convolutional Neural Networks (CNN) with those of biological systems, such as the primate ventral visual stream [1–4]. These studies showed that the representation of visual objects in macaque inferior temporal (IT) cortex corresponds better with the representations of these images in deep CNN layers than with representations of older computational models such as HMAX [5]. Similar findings were obtained with human fMRI data [6–10]. The images used in these studies were those of real objects in cluttered scenes, which are the same class of images as those employed to train the deep CNNs for classification. Other single unit studies of IT neurons employed two-dimensional (2D) shapes and observed highly selective responses to such stimuli (for review see [11]). If deep CNNs provide a realistic model of IT responses, then the CNNs should capture also the selectivity observed for such two-dimensional shapes in IT. To our knowledge, thus far there has been no comparison between the 2D-shape representation of IT neurons, measured with such reduced stimuli, and that of deep CNN models. It is impossible to predict from existing studies that compared deep CNN activations and neurophysiology whether the deep CNNs, which are trained with natural images, can faithfully model the selectivity of IT neurons for two-dimensional abstract shapes. Nonetheless, such correspondence between CNN models and single unit selectivity for abstract shapes is critical for assessing the generalizability of CNN models to stimuli that differ markedly from those of the trained task but have been shown to drive selectively IT neurons. Previously, we showed that a linear combination of units of deep convolutional layers of CNNs trained with natural images could predict reasonably well the shape selectivity of single neurons recorded from an fMRI-defined body patch [4]. However, in that study, we adapted for each single unit the shapes to the shape preference of that neuron, precluding a comparison between the shape representation of the population of IT neurons and deep CNNs. To perform such a comparison, one should measure the responses of IT neurons to the same set of shapes. Furthermore, the shape set should include variations in shape properties IT neurons were shown to be sensitive to. Also, the IT response selectivities for such shapes should not trivially be explainable by physical image similarities, such as pixel-based differences in graylevels. Kayaert et al. [12] measured the responses of single IT neurons to a set of shapes that varied in regularity and the presence of curved versus straight boundaries (Fig 1). The first group of stimuli of [12] was composed of regular geometric shapes (shown in the first two rows of Fig 1 and denoted as Regular (R)) that all have at least one axis of symmetry. These shapes are simple, i.e., have low medial axis complexity [13]. The stimulus pairs in each column of these two rows (denoted by a and b) differed in a non-accidental property (NAP). NAPs are stimulus properties that are relatively invariant with orientation in depth, such as whether a contour is straight or curved or whether a pair of edges is parallel or not. These properties can allow efficient object recognition at different orientations in depth not previously experienced [14–16]. NAPs can be contrasted with metric properties (MPs), which vary with orientation in depth, such as aspect ratio or the degree of curvature. The three other groups are all ‘Irregular’. They differed from the Regular shapes in that they do not have a single axis of symmetry. The two shapes in each row of the three Irregular groups differed in the configuration of their concavities and convexities or corners. The shapes in the Irregular Simple Curved (ISC) set all had curved contours. The Irregular Simple Straight (ISS) shapes were derived from the ISC shapes by replacing the curved contours with straight lines. Thus, the corresponding stimuli in the ISS and ISC shapes differed in a NAP. Last, the Irregular Complex (IC) group was more complex in that the shapes in that group had a greater number of contours. Kayaert et al. [12] found that anterior IT neurons distinguished the four groups of shapes. Importantly, the differences in IT responses amongst the shapes could not be explained by pixel-based gray level differences, nor by HMAX C2 unit differences. In fact, none of the tested quantitative models of object processing could explain the IT response modulations. Furthermore, the IT response modulations were greater for the Regular shapes and when comparing the curved and straight Irregular Simple shapes than within the 3 Irregular shape groups, suggesting a greater sensitivity for NAPs than for MPs (see also [17,18]). We reasoned that this shape set and corresponding IT responses was useful to examine to what degree different layers of deep CNNs and IT neurons represent abstract shapes similarly. We employed deep CNNs that were pretrained to classify ImageNet data [19], consisting of images of natural objects in scenes. Hence, the CNNs were not exposed during training to silhouette shapes shown to the IT neurons. Deep CNNs have a particular architecture with early units having small receptive fields, nonlinear pooling of units of the previous layer, etc. Such a serial, hierarchical network architecture with increasing receptive field size across layers may result in itself, i.e. without training, in changes in the representational similarity across layers. To assess whether potential correlations between IT and CNN layer response modulations resulted from classification training or from the CNN architecture per se, we also compared the activations of untrained CNNs with the IT response modulations. Kayaert et al. [12] had also human subjects sort the same shapes based on similarity and found that human subjects had a pronounced higher sensitivity to the difference between the curved and straight simple irregular shapes (relative to the regular shapes) than the IT neurons. We examined whether a similar difference in response pattern between macaque IT neurons and human similarity judgements would emerge in the deep CNNs. We expected that deeper layers would resemble the human response patterns while the IT response pattern would peak at less deep layers. Kayaert et al [12] recorded the responses of 119 IT neurons to the 64 shapes shown in Fig 1. The 64 shapes are divided in four groups based on their regularity, complexity and whether they differed in NAPs. We presented the same shapes to 3 deep CNNs: Alexnet [20], VGG-16, VGG-19 [21] and measured the activations of the units in each layer of the deep nets. These deep nets differ in their number of layers, the number of units in each layer and the presence of a normalization stage, but each have rectifying non-linearity (RELU) and max pooling stages (Fig 2). We employed deep nets that were pre-trained in classification of a database of natural images, which were very different in nature from the abstract shape stimuli that we employ here to test the models and neurons. The aim was to compare the representations of the shapes between IT neurons and each layer of the deep nets. To do this, we employed representational similarity analyses [22,23], following the logic of second order isomorphism [24,25], and examined the correlation between the neural IT-based similarities and CNN-based similarities in responses to shapes. We are not trying to reconstruct the shapes based on IT neuron or CNN unit outputs but we are examining whether shapes that are represented close to each other in the neural IT space are also represented close to each other in the CNN layer space. In a first analysis, we computed the pairwise dissimilarity between all 64 stimuli using the responses of the IT neurons and the activations in each of the CNN layers. We employed two dissimilarity metrics: Euclidean distance and 1 –Spearman rank correlation ρ. The dissimilarity matrices computed with the Euclidean distance metric for the IT neurons and for 5 layers of the trained CNNs are illustrated in Fig 3B and 3C, respectively. In this and the next figures, we will show only the data for Alexnet and VGG-19, since VGG-16 and VGG19 produced similar results. In addition, Fig 3A shows the pixel-based dissimilarities for all image pairs. Visual inspection of the dissimilarity matrices suggests that (1) the pattern of dissimilarities changes from the superficial to deep layers in a relatively similar way in the CNNs, (2) the dissimilarity matrix of the first layer (e.g. conv1.1) resembles the pixel-based similarities (Fig 3A) and (3) the deeper layers resemble more the IT neural data (Fig 3B). We quantified the similarity between the IT shape representation and that of each layer by computing the Spearman Rank correlation between the corresponding pairwise dissimilarities of IT and each layer. Thus, we could assess to what degree stimuli that produce a very different (similar) pattern of responses in IT also show a different (similar) pattern of activations in a CNN layer. We found that for both dissimilarity metrics the similarity between IT neuronal responses and trained CNN layer activations increased significantly with the depth of the layer. This is shown using the Euclidean distance metric for Alexnet and VGG-19 in Fig 4 (see S1 Fig for the data of both distance metrics and the 3 networks). In the VGG nets, the similarity peaked at the deepest convolutional layers (Fig 4) and then decreased for the deepest layers. In fact, the Spearman correlations for the last two fully connected layers did not differ significantly from that of the first convolutional layer in each CNN (Fig 4). The decrease in similarity for the deepest layers was weaker in Alexnet. The peak similarity was similar amongst the 3 nets, with ρ hovering around 0.60, and were larger for the correlation (mean peak ρ = 0.64) compared with Euclidean distance metric (mean peak ρ = 0.58). To assess the degree to which the models explained the neural data, we computed the reliability of the neural-based distances giving the finite sampling of the IT neuron population. This noise ceiling was computed by randomly splitting the neurons into two groups, computing the dissimilarities for each group, followed by computation of the Spearman rank correlation between the dissimilarities of the two groups. This split-half reliability computation was performed for 10000 random splits. Fig 4 shows the 2.5, 50 (median) and 97.5 percentiles of the Spearman-Brown corrected correlations between the two groups. The correlations between (some) CNN layers and neural responses were close but still below the estimated noise of the neural dissimilarities. In order to assess to what degree the similarity between neural data and the CNN layers reflects the architecture of the CNNs versus image classification training, we computed also the similarity for untrained networks with random weights. Fig 3C illustrates dissimilarity matrices computed using Euclidean distances for 5 untrained layers of two CNNs. Visual inspection suggests little change in the dissimilarity matrices of the different layers of the CNNs, except for fc8. Furthermore, the pattern of dissimilarities resembled the pixel-based dissimilarities shown in Fig 3A. Both observations were confirmed by the quantitative analysis. The Spearman correlations of the neural data and untrained CNNs increased only weakly with depth, except for a marked decrease in correlation for the last two fully connected layers. Except for the deep convolutional and the last two layers, the trained and untrained networks showed similar Spearman correlations of the neural and CNN distances (Fig 4). This suggests that overall the similarity between the IT data and the shallow CNN layers are unrelated to classification training but reflect merely the CNN architecture. Significant differences between trained and untrained CNNs were observed for the deeper convolutional layers (Fig 4), suggesting that the similarity between IT and the deep convolutional layers depends on classification training. The similarities for the first fully connected layer (fc6 and relu6 in Fig 4) did not differ significantly between the trained and untrained layers (except for the correlation metric in AlexNet (S1 Fig). The deepest two (fully connected) layers showed again a significantly greater similarity for the trained compared with the untrained networks. However, this can be the result of the sharp drop in correlations for these layers in the untrained network. Overall, these data suggest that the shape representations of the trained deep convolutional layers, but not of the deepest layers, shows the highest similarity with shape representations in macaque IT. Receptive field (RF) size increases along the layers of the CNNs, allowing deeper layer units to integrate information from larger spatial regions. The difference in IT-CNN similarity between untrained and trained layers shows that the increase in RF size cannot by itself explain the increased IT-CNN similarity in deeper layers, since untrained CNN also increase their RFs along the layer hierarchy. Also, the decrease in similarity between IT responses and the fully connected layers argues against RF size being the mere factor. Nonetheless, although not the only contributing factor, RF size is expected to matter since arguably small RFs cannot capture overall shape when the shape is relatively large. Hence, it is possible that the degree of IT-CNN similarity for different layers depends on shape size, with smaller shapes showing a greater IT-CNN similarity at earlier layers. We tested this by computing the activations to shapes that were reduced in size by a factor of two in all layers of each of the 3 trained CNNs. Fig 5 compares the correlations between dissimilarities of the trained Alexnet and VGG-19 networks and IT dissimilarities for the original and reduced sizes, with dissimilarities computed using Euclidean distances. The stars indicate significant differences between the similarities for the two sizes (tested with a FDR corrected randomization test; same procedure as in Fig 4 when comparing trained and untrained correlations). In each of the CNNs (S2 Fig), the IT-CNN similarity increased at more superficial layers for the smaller shape. The overall peak IT-CNN similarity was highly similar for the two sizes in the VGG networks and occurred at the deep convolutional layers. For Alexnet, the overall similarity was significantly higher for the smaller shapes in the deep layers. This analysis indicates that shape size is a contributing factor that determines at which layer the IT-CNN similarity increases, but that for the VGG networks, peak similarity in the deep layers does not depend on size (at least not for the twofold variation in size employed here). Note that also for the smaller size the IT-CNN similarity drops markedly for the fully connected layers in the VGG networks. Thus, the overall trends are independent of a twofold change in shape size. In the preceding analyses, we included all units of each CNN layer. To examine whether the similarity between the CNN layers and the IT responses depends on a relatively small number of CNN units or is distributed amongst many units, we reran the representational similarity analysis of deep CNN layers and IT neurons for the whole shape set for smaller samples of CNN units. We took for each network the layer showing the peak IT-CNN similarity and for that layer sampled 10000 times at random a fixed percentage of units. We restricted the population of units to those that showed a differential activation (standard deviation of activation across stimuli greater than 0) since only those can contribute to the Euclidean distance. Fig 6A plots the median and 95% range of Spearman rank correlation coefficients between IT and CNN layer dissimilarities for the whole shape set as a function of the percent of sampled units for two CNNs. We found that the IT-CNN similarity was quite robust to the number of sampled units. For instance, for Alexnet, the IT-CNN similarity for the original and the 95% range of the 10% samples overlap, indicating that 315 Alexnet units can produce the same IT-CNN similarity as the full population of units. Note also that the lower bound of the 95% range is still above the IT-CNN similarities observed for the untrained network (median Spearman rho about 0.40; see Fig 4). This indicates that the IT-CNN similarity does not depend on a small subset of units, since otherwise the range of similarities (Spearman rho correlations) for the 10% samples would be much greater. The same holds for the other CNNs (S3 Fig), except that these tolerated even smaller percent sample size (for VGG19 even 0.1%, which corresponds to 100 units). The above analysis appears to suggest that the activations of the CNN units to the shapes are highly correlated with each other. To address this directly, we performed Principal Component Analysis (PCA) of unit activations of the same peak CNN layers as in Fig 6 and computed Euclidean distance based dissimilarities between all stimulus pairs for the first, first two, etc. principal components (PCs), followed by correlation with the neural dissimilarities as done before for the distances computed across all units of a CNN layer. For both the Alexnet and VGG-19 layer, the first 10 PCs explained about 70% of the variance in CNN unit activations to the 64 stimuli (Fig 7B). Only the first 3 (Alexnet) or 5 (VGG-19) PCs were required to obtain a similar correlation between the model and neural distances as observed when using all model units of the layer (Fig 7A; about 7 PCs were required for VGG-16; see S4 Fig). This analysis shows that the neural distances between the abstract shapes relate to a relatively low dimensional shape representation in the CNN layer, with a high redundancy between the CNN units. In the above analyses, we compared the overall similarity of the shape representations in IT and CNN layers. However, a more stringent comparison between the shape representations in IT and the CNNs involves response modulations for the shape pairs for which Kayaert et al [12] observed striking differences between predictions of pixel-based models or computational models like HMAX and the neural responses. The average response modulations (quantified by pairwise Euclidean distances) for the different group pairs comparisons are shown in Fig 8 for the IT neural data, the HMAX C2 layer and the pixel differences. Kayaert et al [12] showed that the mean response modulation in IT (Fig 8A)was significantly greater for the regular shape pairs (1–8 in Fig 1) than for the 3 irregular shape group pairs, despite the pixel differences between members of a pair being, on average, lower or similar for the regular group than for the 3 irregular groups (Fig 8D). In addition, the response modulation to ISC vs. ISS was significantly greater than the modulations within IC, ISC and ISS, although the average pixel-difference within the ISC vs. ISS-pairs was much lower than the pixel-differences within the other pairs. This differential neural response modulation to ISC vs ISS was present for both members of the ISC and ISS pairs (a and b members: “ISCa vs ISSa” and “ISCb vs ISSb”) and thus was highly reliable. Note that the difference between ISC vs. ISS and the IC and ISS shape groups that are present in the neural data is not present for the HMAX C2 distances (Fig 8C). Kayaert et al. [12] reported also a relatively higher sensitivity to the straight vs. curved contrast of the ISC vs. ISS comparison compared with the regular shapes in human similarity ratings (Fig 8B), compared with the IT neural data. In other words, human subjects appear to be more sensitive to the curved versus straight NAP difference than macaque IT neurons. In a second analysis, we determined whether the marked differences in IT response modulations and human judgements shown in Fig 8 are present in the dissimilarities for the different layers of the deep CNNs. Fig 9 illustrates the results for 8 layers of VGG-19. The left column of the figure plots the distances for the trained network. The dissimilarities for the first convolutional layer fits the pixel-based distances amongst the shape pairs (Fig 8D; Pearson correlation between pixel-based distances and first layer distances = 0.966), but differ from those observed in IT and for human judgements. Similar trends are present until the very deep convolutional layers where the dissimilarities became strikingly similar to those observed in macaque IT (e.g. compare trained conv5.4 or pool5 of Fig 9 with Fig 8A). The dissimilarities for the last two layers (e.g. trained relu7 and fc8 in Fig 9) are strikingly similar to those observed for the human judgements (Fig 8B), and differ from the pattern seen in macaque IT neurons. Indeed, as noted above, the human judgements differ from the IT responses in their sensitivity for the ISC vs ISS comparison relative to that for the regular shape pairs: for the human judgement distances, the ISC vs ISS distances are greater than for the regular shape distances while for the neural distances both are statistically indistinguishable (Kayaert et al. [12]). Therefore, we tested statistically for which CNN layer the ISC vs ISS distances were significantly greater than the regular shape distances (Wilcoxon test), thus mimicking the human distances. We found a significant difference for the very deep VGG19 layer fc8 (p = 0.039) and VGG16 layers fc7 (p = 0.039), relu7 (p = 0.023), and fc8 (p = 0.023). Although the deepest Alexnet (fully connected) layers showed the same trend, this failed to reach significance. These tests showed that only the very deep CNN layers mimicked the human judgements. None of the untrained CNN layers showed a dissimilarity profile similar to that observed in monkey IT or in human judgements (Fig 9, right column). In fact, the untrained data resembled more the pixel-based distances (see Fig 8D). Indeed, the Pearson correlation between the pixel-based distances and the conv1.1 distances was 0.999 for the untrained VGG-19. We quantified the correspondence between the neural response dissimilarities of Fig 8A and the CNN layer dissimilarities (as in Fig 9) by computing the Pearson correlation coefficient between the dissimilarity profiles (n = 6 dissimilarity pairs). The same quantification was performed for the human judgements (Fig 8B) and the CNN dissimilarities (n = 5 pairs). These correlations are plotted in Fig 10A and 10B as a function of layer for two CNNs, trained and untrained. For the neural data, the correlations are negative for the shallow layers and highly similar for the trained and untrained CNNs. The negative correlations are a result of the nearly inverse relationship between neural and low-level (pixel) differences between the shapes (Fig 8D). This was not accidental, but by design: when creating the stimuli, Kayaert et al [12] ensured that the NAP differences (e.g. between ISC and ISS) were smaller than MP differences. For both VGG networks (S5 Fig; Fig 10B), there was a sharp increase in correlations at the trained deep 5.1 convolutional layer, followed by a decrease of the correlations for the fully connected layers. This trend was similar, although more abrupt, to that observed for the global dissimilarities of Fig 4. For Alexnet, the increase of the correlations with increasing depth of the trained convolutional layers was more gradual, but like the VGG networks, high correlations were observed for the deeper trained convolutional layers. For the human judgement data, the correlations were already higher for the trained compared with the untrained CNNs at the shallow layers, although still negative. Like the neural data, there was a marked increase in correlation at the very deep trained layers. Contrary to the neural data, the correlations for the human judgements continued to increase along the trained fully connected layers, approaching a correlation of 1 at the last layer. These data show that the average response modulations of IT neurons for the shape groups of Fig 1 correspond nearly perfectly with those of the deeper layers of CNNs, while the differences in human similarity judgements between the groups are captured by the later fully connected layers of the CNNs. This holds for Alexnet and VGG nets. Note that the deep CNN layers performed better at predicting the neural IT and human perceptual dissimilarities than the HMAX C2 layer output (Fig 10C). As for the representational similarity analysis for all shapes (Fig 6A), we computed also the Pearson correlation coefficients between the dissimilarity profiles (n = 6 dissimilarity pairs) of the same peak CNN layers and the IT distances for the 6 shape groups (as in Fig 10) for smaller samples of units. As shown in Fig 6B, we observed similar IT-CNN correlations for the within-group distances up to the 1% and 0.1% samples compared with the full population of units for Alexnet and VGG, respectively. Again, this suggests that IT-CNN similarity does not depend on a small number of outlier CNN units. The greater tolerance for percent sample size for the VGG units is because the VGG layers consisted of a larger number of units per se (total number of units are indicated in the legend of Fig 6). In addition, we computed the mean distances for the same layers and their correlation with the mean neural modulations as a function of retained PCs (Fig 7B). Up to 30 PCs were required to obtain a similar correlation between neural and CNN layer distances for the six groups of shapes as when including all units of the layer (Fig 7B). This suggests that the close to perfect modeling of the mean response modulations across the 6 shape groups required a relatively high dimensional representation of the shapes within the CNN layer. The particular set of shapes that we employed in the present study was designed originally to test the idea that the shape selectivity of IT neurons reflects the computational challenges posed when differentiating objects at different orientations in depth [12,14]. Here, we show that deep CNNs that were trained to classify a large set of natural images show response modulations to these shapes that are similar to those observed in macaque IT neurons. We show that untrained CNNs with the same architecture than the trained CNNs, but with random weights, demonstrate a poorer IT-CNN similarity than the CNNs trained in classification. The difference between the trained and untrained CNNs emerged at the deep convolutional layers, where the similarity between the shape-related response modulations of IT neurons and the trained CNNs was high. Unlike macaque IT neurons, human similarity judgements of the same shapes correlated best with the deepest layers of the trained CNNs. Early and many later studies of IT neurons employed shapes as stimuli (e.g. [26–31,22,32–37]), in keeping with shape being an essential object property for identification and categorization. Deep CNNs are trained with natural images of objects in cluttered scenes. If deep CNNs are useful models of biological object recognition [38], their shape representations should mimic those of the biological system, although the CNNs were not trained with such isolated shapes. We show here that indeed the representation of the response modulations by rather abstract, unnatural shapes is highly similar for deep CNN layers and macaque IT neurons. Note that the parameters of these CNN models are set via supervised machine learning methods to do a task (i.e. classify objects) rather than to replicate the properties of the neural responses, as done in classic computational modeling of neural selectivities. Thus, the same CNN model that fits neural responses to natural images [1–4] also simulates the selectivity of IT neurons for abstract shapes, demonstrating that these models show generalization across highly different stimulus families. Of course, the high similarity between deep CNN layers and IT neurons activation patterns we show here may not generalize for (perhaps less fundamental) shape properties that we did not vary in our study. Kubilius et al. [39] showed that deep nets captured shape sensitivities of human observers. They showed that deep Nets, in particular their deeper layers, show a NAP advantage for objects (“geons”), as does human perception (and macaque IT [18]). Although we also manipulated NAPs, our shapes differed in addition in other properties such as regularity and complexity. Furthermore, our shapes are unlike real objects and more abstract than the shaded 3D objects employed by Kubilius et al. [39] when manipulating NAPs. In both the representational similarity analysis and the response modulations comparisons amongst shape groups, we found that the correspondence between IT and deep CNN layers peaked at the deep convolutional layers and then decreased for the deeper layers. Recently, we observed a similar pattern when using deep CNN activations of individual layers to model the shape selectivity of single neurons of the middle Superior Temporal Sulcus body patch [4], a fMRI-defined region of IT that is located posterior with respect to the present recordings. The increase with deeper layers of the fit between CNN activations and neural responses has also been observed when predicting with CNN layers macaque IT multi-unit selectivity [40], voxel activations in human LO [9] and the representational similarity of macaque and human (putative) IT [8,10] using natural images. However, the decrease in correlation between CNNs and neural data that we observed for the deepest layers was not found in fMRI studies that examined human putative IT [8,10], although such a trend was present in [6] when predicting CNN features from fMRI activations. The deepest layers are close to or at the categorization stage of the CNN and hence strongly dependent on the classifications the network was trained on. The relatively poor performance of the last layers is in line with previous findings that IT neurons show little invariance across exemplars of the same semantic category [41,42], unlike the deepest CNN units [43]. The question of what the different layers in the various CNN models with different depths represent neurally remains basically unanswered. Shallow CNN layers can be related to early visual areas (e.g. V1; V4) and deeper layers to late areas (e.g. IT). However, different laminae within the same visual area (e.g. input and output layers) may also correspond to different layers of CNNs. Furthermore, units of a single CNN layer may not correspond to a single area, but the mapping might be more mixed with some units of different CNN layers being mapped to area 1, while other units of partially overlapping CNN layers to area 2, etc. Finally, different CNN layers may represent different temporal processing stages within an area, although this may map partially to the different laminae within an area. Further research in which recordings in different laminae of several areas will be obtained for the same stimulus sets, followed by mapping these to units of different layers in various CNNs, may start to answer this complex issue. In contrast with IT neurons, human similarity judgements of our shapes matched to a greater extent the last rather than the less deep convolutional layers. In particular, the deepest layers showed a similar enhanced sensitivity for straight versus curved irregular shapes. The untrained CNNs did not show such straight versus irregular bias for the irregular shapes. Thus, it appears that a system, be it artificial like the CNNs or a biological system like humans, that is required to classify natural images of objects develops such bias for curved versus straight contours, indicating that this shape property must be highly informative for object categorization. Whether this relates to straight versus curved being a NAP [14] is unclear. Kayaert et al. [12] employed a sorting task to rate shape similarity. In this task, subjects were required to sort the shapes into groups based on their similarity. Although this is not the same as labeling an object, the task for which the CNNs were trained, higher order classification can intrude the sorting task judgements. This may explain why the human sortings of Kayaert et al. [12] resembled that closely the activation pattern seen at the deepest CNN layers, which are strongly category label driven. Interestingly, even for the shallow convolutional layers, the correlations between the human judgements and the CNN activations were higher for the trained compared with the untrained CNNs. This contrasted with the equal correlations for trained and untrained shallow layers for the IT data. This suggests that the trained shallow CNN layers show already some, albeit weak, bias for higher order category-related information. Previous studies that compared deep CNNs and neural responses rarely included untrained CNNs as control (e.g. [8], [40]). We found the untrained CNNs helpful in interpreting our data. The comparison with untrained CNNs can inform to what extent neural responses reflect features that can be picked up by untrained CNNs (because of CNN architectural properties such as tiling of local RFs in shallow layers and non-linear pooling). Indeed, we found that most layers of the untrained CNNs represented rather closely the pixel-based graylevel differences between the shape groups, which assisted to interpret the representational similarity of the trained CNNs at shallow layers. Thus, we advise that future studies use untrained CNNs as control or benchmark. Currently, deep CNNs are the best models we have of primate object recognition, providing the best quantitative fits of ventral stream stimulus selectivities and primate recognition behavior [38]. However, recent studies show that CNNs have their limitations, especially when stimuli are noisy or partially occluded. For instance, the commonly used deep CNNs tolerate less image degradation than humans [44], can be fooled by unrecognizable images [45] or show a sensitivity to imperceptible stimulus perturbations (“adversarial examples”; [46]). Our data show that training CNNs in object categorization produces at least some shape selectivities (that are thought to reflect fundamental aspects of shape processing [14]) similar to those that are observed in neural IT data and human similarity judgements. This does not imply that CNNs can explain all shape or stimulus selectivity in IT and there is still considerable room for model improvement (e.g. recurrent connectivity etc.). In conclusion, deep CNN layers that were trained to classify complex natural images represented differences among relatively simple abstract 2D-shapes similar to macaque IT neurons. Human sorting of the same shapes corresponded better with the deepest layers of the CNNs. The similarity between IT neurons and the deeper convolutional layers is greater for trained compared to untrained CNNs, suggesting a role of image classification in shaping the shape selectivity of macaque IT neurons. The latter likely occurs during ontogenetic development, but may not result from the same supervised learning algorithm as employed to train the CNNs. Indeed, independent of the particular training protocol (e.g. supervised versus unsupervised), any biological object classification system may have similar shape representation biases that are inherently useful for performing invariant object classification. Two male rhesus monkeys served as subjects. The animals were housed individually with visual and auditory contact with conspecifics. During the recording weeks, they had controlled access to fluids but food was available at libitum. All procedures were in accordance with the Weatherall report on “The use of non-human primates in research” and were approved by the Animal Ethics committee of the KU Leuven (protocol number: P631/2002). The 64 shapes were identical to the first stimulus set employed by Kayaert et al. [12] and are shown in Fig 1. The Regular shapes R were created with Studio MAX, release 2.5, while the Irregular shapes were made with Fourier Boundary Descriptors, using MATLAB, release 5. The Irregular Simple Straight (ISS) stimuli were made by replacing the curves of the Irregular Simple Curved (ISC) shapes by straight lines while preserving the overall shape. The increase in complexity of the Irregular Complex (IC) shapes compared to the simpler ISC shapes was produced by increasing the number and frequency of the Fourier Boundary Descriptors. Each group contains 8 pairs of stimuli (one stimulus in row a and one in row b in Fig 1). The columns of Fig 1 comprise a set of 4 pairs (one for each group) that were matched in overall size and aspect ratio, both within and between groups. The averaged pixel-based graylevel differences between the members of the pairs were balanced across groups (see [12] for more details). The members of the pairs within the Regular shapes differ in a NAP, such as parallel vs. nonparallel sides, or straight vs. curved contours. The differences among the members of an irregular pair were created by varying the phase, frequency or amplitude of the Fourier Boundary Descriptors. For the single unit recordings and the human behavioral study, all stimuli were filled with the same random dot texture pattern. The number of black and white dots was required to be equal for 2*2 squares of pixels, so the texture patterns were highly uniform. Stimuli were presented on a gray background. In the single unit study, they extended approximately 7 degrees and were shown at the center of the screen. We employed the identical shapes for the CNN modeling, except that the noise pattern was replaced by a uniform white surface (see Fig 1 for the actual stimuli presented to the CNNs). The single unit data have been published before [12] and the procedures have been described in detail in that paper. Therefore, we will summarize here only briefly the experimental procedures. The IT recordings were made while the two monkeys performed a passive fixation task. Eye movements were measured with the scleral search coil technique or with a noninvasive eye tracker (ISCAN). During the recordings, their head was fixed by means of an implanted head post. We employed the standard dorsal approach to IT and recording sites were verified with MRI and CT scans with the guiding tube in situ. We lowered a tungsten microelectrode through the guiding tube that was fixed in a Crist grid, which was positioned within the plastic recording chamber. The signals of the electrode were amplified and filtered using standard single-cell recording equipment. Single units were isolated on line and their timing was stored together with stimulus and behavioral events for later offline analysis. The stimuli were presented during fixation for 200 ms in a randomly interleaved fashion. In the present study, the response of a neuron was defined as the average firing rate in spikes/s during a time interval of 250 ms, starting from 50 to 150 ms after stimulus onset. The starting point of this time interval was chosen for each neuron to best capture its response, by inspection of the peristimulus time histograms averaged across the stimuli. Responses were averaged across presentations per stimulus. The minimum number of presentations per stimulus was 5 (median = 10). The data set consisted of 119 anterior IT neurons (76 in monkey 1 and 43 in monkey 2) that showed significant response selectivity to the stimuli of the set (ANOVA, p<0.05). The data were pooled across animals. The neurons were located in the lower bank of the Superior Temporal Sulcus and the lateral convexity of anterior IT (TEad). As described by Kayaert et al. [12], printed versions of the shapes were given to 23 naive adult human subjects who were asked to sort the stimuli in groups based on shape similarity. No further definition of similarity was given and they could freely choose the number of groups. This is a classical task to measure image similarities [47]. Dissimilarity values between pairs of stimuli were computed by counting the number of subjects that put the two members in different groups. In order to compare the shape representation of the IT neurons’ population with deep CNN layers, we extracted stimulus features for each processing stage (layer) of three deep models: Alexnet [20], VGG-16 and VGG-19 [21]. We used the pretrained networks, which are available through the MatConvNet toolbox [48] in MATLAB, and their untrained versions. The pretrained CNNs were trained on ~1.2 million natural images divided in 1,000 classes for the ImageNet Large Scale Visual Recognition Challenge 2012 (ILSVRC2012). The untrained versions of these networks have the same architecture, but did not undergo any training, thus no update of their weights took place after initialization. Their layer weights were initialized by sampling randomly from a normal distribution, using the opts.weightInitMethod = 'gaussian' setting in the cnn_imagenet.m function of the MatConvNet toolbox. The stimuli shown to the CNNs were black and white images with pixel values ranging from 0–255 (0 for black and 255 for white). Before feature extraction, the mean of the ILSVRC2012 training images was subtracted from each stimulus, since this was also part of the preprocessing stage of the networks’ training procedure. In addition, the stimuli were rescaled accordingly to match each network’s input requirements (227x227 pixels for Alexnet and 224x224 pixels for VGG-16 & VGG-19). In all analyses, we employed as distance metric the normalized Euclidean distance between the neuronal responses or deep CNN unit activations: (Σin(Ri1−Ri2)2n)12, where Ri1 is the response of neuron or deep CNN unit i, to stimulus 1, and n is the number of neurons or the number of deep CNN units in a specific layer. For the representational similarity analyses, we computed also a second distance metric: 1- Spearman’s correlation coefficient. The Spearman rank correlation coefficient ρ was computed between the neural responses or CNN units’ activations for all stimulus pairs. To compare neuronal data to CNN layers, we performed representational similarity analysis [49], using both distance metrics. We constructed representational dissimilarity matrices (RDMs) for the whole stimulus set (n = 64 stimuli) for both IT neurons and each deep CNN layer (trained and untrained separately; for examples see Fig 3), by arranging all possible pairwise distances in 64x64 RDMs. We extracted all values above the diagonal (upper triangle of the RDM, excluding the diagonal) of the symmetrical RDMs, and computed for each layer the Spearman rank correlation coefficient between the distances of the corresponding pairs of the neural and CNN matrices. We computed 95% confidence intervals of the Spearman correlation coefficient between neural and CNN distances by resampling with replacement 10,000 times 119 neurons out of our pool of IT neurons and correlating each time the resulting neural distance matrix with each deep CNN layer for the trained and untrained versions of the same network. The confidence intervals corresponded to the 2.5 and 97.5 percentiles of the bootstrapped correlation coefficient distributions. To assess whether the trained deep CNN layers significantly differed from the untrained, we computed for each CNN layer the distribution of the paired differences of trained minus untrained layer correlations across the 10,000 iterations (one difference per bootstrapped neuronal sample). For each layer, we computed the percentile in the corresponding distribution of the zero difference value and these defined the p values of the test. For each of the 3 CNNs, we corrected the p values for multiple comparisons (n = number of CNN layers) using the Benjamini and Hochberg [50] False Discovery Rate (FDR) procedure. A difference between the trained and untrained CNNs was judged to be significant when FDR q < 0.05. The same procedure was used to assess the significance of the difference in IT-CNN correlations between the original and reduced shape size for each of the CNN layers (Fig 5). We employed a similar procedure to test the significance of the difference in Spearman rank correlation coefficients of the neural and CNN distances between the first layer and each subsequent layer. Thus, we computed the pairwise difference between the correlation for the first and a subsequent layer for each of the 10,000 bootstrapped neural samples and then obtained the percentile of the zero difference in that distribution of differences. The p values were corrected for multiple comparisons using the FDR procedure and significance was defined when q < 0.05. In the second analysis, using only the original, non-bootstrapped distances, we compared the pairwise Euclidean stimulus distances amongst the 4 stimulus groups R, IC, ISC, ISS. For each group, we included only the stimulus pairs numbered 1–8 in Fig 1, i.e. for each group the members of the a and b rows of Fig 1. In addition, we selected the distances for the “ISCa vs. ISSa” and “ISCb vs. ISSb” pairs of Fig 1, e.g. the column-wise distances between row a of ISC and row a of ISS in Fig 1 (likewise for the b rows). This produced twice 8 distances for the ISC versus ISS comparison, which we analyzed separately, unlike in Kayaert et al. [12]. For each of the 4 groups and the two ISC versus ISS comparisons, we computed the mean distance (and standard errors of the mean) across the 8 pairs per group or comparison. To quantify the relationship between the mean distances across groups for the neural data and each CNN layer, we computed the Pearson correlation coefficient between the mean neural distances and the mean distances of the CNN layers (n = 6 pairs of distances per layer). A similar analysis was performed comparing the CNN layer distances and the distances based on the human ratings. However, for this analysis, the available human rating data consisted of the distances that were computed by Kayaert et al [12], having an ISC versus ISS comparison of 8 stimulus pairs (for selection of those pairs, see [12]) instead of twice 8 pairs as above. We compared those distances with the average of the “ISCa vs. ISSa” and “ISCb vs. ISSb” pairs of the CNN layers. Note that our average neural distances for the “ISCa vs. ISSa” and “ISCb vs. ISSb” pairs were highly similar to those for the 8 “ISC vs. ISS” pairs selected by Kayaert et al. [12], justifying this procedure. We compared neural dissimilarities also with dissimilarities based on pixel graylevels and the HMAX model [5], employing the same procedures as in Kayaert et al.. We computed the Euclidean distance between the gray-level values of the pixels for all image pairs (Fig 3B). In addition, we computed the Euclidean distances between the outputs of C2-units of the HMAX model as described by Riesenhuber and Poggio [5] and presented in [12]. The HMAX C2 units were designed to extract moderately complex features from objects, irrespective of size, position and their relative geometry in the image. HMAX-based dissimilarities were computed as the Euclidean distance between the output of the 256 C2 units.

10.1371/journal.pntd.0004688

Tegument Glycoproteins and Cathepsins of Newly Excysted Juvenile Fasciola hepatica Carry Mannosidic and Paucimannosidic N-glycans

Recently, the prevalence of Fasciola hepatica in some areas has increased considerably and the availability of a vaccine to protect livestock from infection would represent a major advance in tools available for controlling this disease. To date, most vaccine-target discovery research on this parasite has concentrated on proteomic and transcriptomic approaches whereas little work has been carried out on glycosylation. As the F. hepatica tegument (Teg) may contain glycans potentially relevant to vaccine development and the Newly Excysted Juvenile (NEJ) is the first lifecycle stage in contact with the definitive host, our work has focused on assessing the glycosylation of the NEJTeg and identifying the NEJTeg glycoprotein repertoire. After in vitro excystation, NEJ were fixed and NEJTeg was extracted. Matrix-assisted laser desorption ionisation-time of flight-mass spectrometry (MALDI-TOF-MS) analysis of released N-glycans revealed that oligomannose and core-fucosylated truncated N-glycans were the most dominant glycan types. By lectin binding studies these glycans were identified mainly on the NEJ surface, together with the oral and ventral suckers. NEJTeg glycoproteins were affinity purified after targeted biotinylation of the glycans and identified using liquid chromatography and tandem mass spectrometry (LC-MS/MS). From the total set of proteins previously identified in NEJTeg, eighteen were also detected in the glycosylated fraction, including the F. hepatica Cathepsin B3 (FhCB3) and two of the Cathepsin L3 (FhCL3) proteins, among others. To confirm glycosylation of cathepsins, analysis at the glycopeptide level by LC-ESI-ion-trap-MS/MS with collision-induced dissociation (CID) and electron-transfer dissociation (ETD) was carried out. We established that cathepsin B1 (FhCB1) on position N80, and FhCL3 (BN1106_s10139B000014, scaffold10139) on position N153, carry unusual paucimannosidic Man2GlcNAc2 glycans. To our knowledge, this is the first description of F. hepatica NEJ glycosylation and the first report of N-glycosylation of F. hepatica cathepsins. The significance of these findings for immunological studies and vaccine development is discussed.

Fasciola hepatica is a parasite responsible for the zoonotic disease fasciolosis, prevalence of which has increased in recent years because of the emergence of triclabendazole-resistant strains as well as changing climatic conditions. A number of F. hepatica protein antigens are used for assessing the immune response of the definitive host for the development of recombinant vaccines but no such vaccine has been commercialised yet. Glycans that may contribute to the antigenic, immunological and protective properties in F. hepatica have not been characterised. Using a panel of plant lectins with defined sugar binding specificities alongside mass spectrometric analysis, we found that high mannose and oligomannose N-glycans are the most abundant in the surface of the juvenile fluke and identified eighteen proteins likely to be glycosylated. Additionally, we found that the proteases cathepsin B1 and L3 contain a glycosylation site occupied by an unusually short Man2GlcNAc2 N-glycan. This work is the starting point for understanding how F. hepatica glycans interact with the definitive host at the initiation of infection. Additionally, it provides useful information for including glycans in the design of new vaccine candidates.

The trematode Fasciola hepatica, commonly known as the liver fluke, is widely distributed across five continents and is responsible for fasciolosis in livestock and humans. It has a large economic impact on the livestock industry, causing production losses in terms of reduced milk yield, liver condemnation and problems with animal fertility among others [1–3]. The World Health Organisation (WHO) has estimated an increase of 11% in the prevalence of human fasciolosis over the last decade, with more than 2.6 million people infected globally [4]. To date, the chief method of control has been administration of flukicidal drugs, with triclabendazole the most frequently used. However, in recent decades, triclabendazole resistance in fluke populations has been reported globally [5]. This, together with the goal of reducing drug use in food-producing animals [6] has encouraged the scientific community to investigate new methods of control. The development and use of vaccines against F. hepatica would be a more sustainable and environmentally friendly future alternative to anthelmintic drugs. The tegument (Teg) and the excretory/secretory (ES) components are important sources of F. hepatica antigens with the highest potential as vaccine targets. Native F. hepatica molecules identified and isolated using proteomic approaches have in some cases been able to produce significant reductions in liver fluke burden and reduced pathology not only in small animal models but also in cattle and sheep. For example, the native members of the Cathepsin clade FhCL1, FhCL2 and FhCL3—some of the most intensively studied vaccine candidates—were able to induce protection in experimental models including rats [7], and in cattle [8] and sheep [9]. In addition, they were able to decrease liver fluke egg viability by as much as 98% in vaccinated animals [8,10]. Other important vaccine candidates investigated included peroxiredoxin (PRX), paramyosin, glutathione S-transferase (GST), fatty acid-binding protein [11] and leucine-aminopeptidase [9]. In many cases, the protective capacity of recombinant versions was lower than that of the native proteins, or more variable between one study and another. For example, the recombinant version of FhCL1 induced 48% protection in cattle [12] in a small-scale field trial but did not provide protection in small ruminants [13,14]. Although animal variability and differences in vaccine formulation are key factors that could explain these discrepancies [15], it is also likely that there are differences in terms of protein folding or post-translational modifications between native and recombinant proteins which influence protective capacity. Glycosylation is one of the main post-translational modifications that occurs following protein synthesis, and glycosylation pathways of prokaryotic and simple eukaryotic vectors used to produce recombinant vaccine candidates are substantially different from those of the parasite itself [16]. Glycomic studies of other trematodes, such as Schistosoma mansoni [17] and Opisthorchis viverrini [18] and most recently Echinostoma caproni [19,20] have led to a deeper understanding of the structural and functional aspects of glycans in this class of helminths. The immunomodulatory properties of F. hepatica glycans and glycoconjugates are now also being studied; for example, their apoptotic effect in peritoneal eosinophils and macrophages in vitro has been demonstrated [21,22] along with their induction of arginase 1, IL-10 and TGF-β transcription in peritoneal macrophages, indicators of M2a macrophages [23]. Recently, the requirement of F. hepatica glycans to influence dendritic cell (DC) maturation and to inhibit IFN-γ production by splenocytes from infected animals has been reported [24,25]. It has been shown that the adult stage of F. hepatica contains at least several types of glycans and glycan motifs such as fucosylated LacdiNAc (LDN-F) motifs [26], mucin-type O-glycosylated proteins [27], the glycolipid CD77 [28] and other glycoceramides [29,30]. Even though the full range of F. hepatica glycoconjugates has not been characterised, the presence of various glycans has been confirmed by lectin binding studies not only in adult stages [25,29,31] but also in the other lifecycle stages (miracidia, rediae, sporocysts) [32–34]. However, systematic studies of glycosylation in the NEJ have not yet been carried out. It is also important to characterise the protein backbones containing these glycans and to verify whether proteins used previously as vaccine candidates are glycosylated. Finally, the potential immunomodulatory properties of these glycans are important both for our understanding of the immunology of fasciolosis and for vaccine development. The goals of this work were to describe the glycosylation patterns of NEJ tegumental (NEJTeg) and somatic (NEJSom) fractions using lectin-affinity techniques and perform a deeper glycan characterisation of NEJTeg using mass spectrometry (MS). We also identified NEJ proteins that are glycosylated and assessed whether they have been previously described as antigenic/vaccine candidates. F. hepatica metacercariae with their outer cyst wall removed were obtained from Baldwin Aquatics, Inc. (Monmouth, Oregon) and stored at 4°C until use. Excystation of metacercariae was performed as described previously with minor modifications [35]. Approximately 15,000 excysted NEJ were used for Teg isolation as previously described [36]. Briefly, NEJ were washed three times in PBS followed by incubation for 30 min at room temperature with 1ml of 1% Nonidet P40 (NP40) in PBS. They were then centrifuged at 300 x g for 5 min and the supernatant collected. NP40 was removed from the supernatant with 0.3 g of Bio beads (Bio-Rad) according to the manufacturer’s recommendations. The pellet obtained after Teg removal, containing denuded NEJ, was suspended in 1 ml of RIPA buffer (Sigma-Aldrich) and used for somatic fraction isolation as previously described [37]. After incubation, both samples were centrifuged at 1000 x g for 5 min and supernatants, which correspond to the NEJTeg and NEJSom, respectively, were aliquoted and stored at -80°C. The protein content was determined by the bicinchoninic acid assay (Thermo Fisher Scientific) according to the manufacturer’s recommendations. The protocol for lectin fluorescence staining was adapted from that described [38]. After fixation in 10% buffered paraformaldehyde solution, NEJ were incubated overnight with 10 ml of 1% BSA, 0.1% sodium azide in PBS (blocking solution) at 4°C. Following extensive washing in PBS, they were suspended in 500 μl of 0.1% BSA, 0.1% sodium azide in PBS (ABD buffer) and aliquoted in batches (40 per batch). NEJ were then incubated overnight in the dark at RT with a panel of FITC conjugated-lectins (Vector Labs) diluted in ABD buffer in a ratio at 1:200. The lectins used and their nominal binding specificity are listed in Table 1. Specific lectin binding was confirmed by pre-incubation of each lectin with the appropriate sugar inhibitor for 2 h at the concentration provided in Table 1. After incubation and extensive washes with ABD solution, NEJ were placed in Vectashield mounting medium with DAPI (Vector Labs). Slides were viewed with a LEICA DM IL LED using 10x and 40x HI PLAN I objectives (Leica Microsystems) equipped with an epifluorescence source and filter system for FITC and DAPI fluorescence. Images were merged using Adobe Photoshop CC software version 14.0 x 64. NEJTeg and NEJSom (10 μg) were analysed by SDS–PAGE using 4–20% gradient Tris-glycine precast gels (Thermo Fisher Scientific) in a vertical electrophoresis system (ATTO) for 90 min at 40 mA. Gels were either subjected to silver staining or transferred to nitrocellulose membrane using the iBlot system (Thermo Fisher Scientific). After blocking with 5% BSA in PBS for 1 h, membranes were incubated with biotin-conjugated lectins (Vector Labs) for 1 h. For secondary detection, membranes were washed and incubated with IRDye-labelled streptavidin (LI-COR Biosciences) for 1 h. Specific lectin binding was confirmed as described previously. A membrane incubated with IRDye-labelled streptavidin only was performed as an additional negative control in order to confirm lack of reactivity between NEJ extracts and streptavidin. Images were acquired using an Odyssey infrared scanning imaging system (LI-COR Biosciences). NEJTeg (glyco)proteins (500 μg) were enzymatically fragmented by using trypsin-coupled beads (GE Healthcare Life Sciences) following the manufacturer’s instructions. Four μl of 1 U/μl N-glycosidase-F (PNG-F) (Roche) were added to the sample and incubated for 24 h at 37°C while shaking. The N-glycan mixture released by the PNG-F was purified with a C18 Reverse phase (RP) cartridge (500 mg; JT Baker) and a carbon cartridges (150 mg Carbograph; Grace) as previously described [39]. The (glyco)peptides remaining in the C18 RP cartridge were eluted, dried in a Speed-Vac (Thermo Fisher Scientific) and dissolved in 50 μl of 1 M sodium acetate solution (pH 4.5). Following pH adjustment to 4.5, 2 μl of N-glycosidase-A (PNG-A, 1 U/μl; Roche) were added to the sample and incubated for 48 h at 37°C while shaking. The N-glycan mixture released by the PNG-A treatment was purified and incubated overnight as described previously [39]. The purified N-glycans were subsequently labelled with the fluorophore 2-aminobenzoic acid (2-AA) (Sigma-Aldrich), as described elsewhere [40]. The labelled N-glycans were brought to 75% AcN and loaded on Biogel P10 (Bio-Rad) columns conditioned with 80% AcN. N-glycans were eluted with 400 μl of dH2O and dried down in a Speed-Vac. AA-labelled PNG-F released glycans (1 μl/treatment) were incubated in a volume of 10 μl for 24 h at 37°C in 250 mM sodium citrate buffer with one of the following exoglycosidases: α-mannosidase from jack bean (15 μU/μl; Sigma Aldrich), β-galactosidase from jack bean (5 μU/μl; Prozyme) or recombinant β-N-acetylglucosaminidase (4 μU/μl; New England Biolabs). PNG-F and PNG-A released N-glycans were dissolved in 50 and 25 μl of dH2O respectively. One μl of each sample was spotted on a matrix-assisted laser desorption ionisation (MALDI) polished steel targeted plate (Bruker Daltonics) using 2,5-dihydroxybenzoic acid (DHB; 20 mg/ml in 30% AcN) as matrix. Samples were analysed with an Ultraflex II MALDI–time-of-flight (MALDI-TOF) mass spectrometer (Bruker Daltonics, Bremen, Germany) operating in the negative-ion reflectron mode. For assessing the results of exoglycosidase treatments, digestion products were purified using a HILIC Ziptip column (Millipore) following the manufacturer’s instructions. Glycans were eluted directly to a MALDI target plate with 10 mg/ml DHB in 50% AcN containing 0.1% TFA and analysed in the negative-ion reflectron mode. Glycopeakfinder (http://www.glyco-peakfinder.org) was used to define glycan composition. The glycan components of NEJTeg-derived glycoproteins were biotinylated with EZ-link Hydrazide-Biotin (Thermo Fisher Scientific) according to the manufacturer’s instructions. The biotinylated glycoproteins (B-NEJTeg) were purified from the rest of the NEJTeg components by affinity chromatography using monomeric avidin agarose (Thermo Fisher Scientific). The flow through was collected as avidin-unbound fraction (A-UB-NEJTeg) and the avidin-bound fraction (A-BB-NEJTeg) separated from the resin following a previously reported protocol [41]. All fractions were concentrated and buffer exchanged in PBS using Pierce Concentrator 3K MWCO (Thermo Fisher Scientific). BCA assays were performed in order to quantify the protein content. SDS–PAGE, silver staining and streptavidin blots were used in order to assess the quality of the biotinylation reaction. Total NEJTeg (50 μg), B-NEJTeg and A-BB-NEJTeg, samples (10 μg each) were fractionated by 12% SDS-PAGE. The bands contained in B-NEJTeg and A-BB-NEJTeg fractions were stained with SYPRO Ruby gel stain (Bio-Rad) according to the manufacturer’s instructions. Bands were excised and a liquid handling station (MassPrep; Waters) was used with sequencing-grade modified trypsin (Promega) according to the manufacturer’s instructions, for in-gel protein digestion. Peptide extracts were then dried by evaporation in a Speed-Vac. Liquid Chromatography and tandem mass spectrometry (LC-MS/MS) was carried out with a linear trap quadrupole (LTQ) mass spectrometer connected to a Thermo Surveyor MS pump and equipped with a nano electrospray ionisation (ESI) source (Thermo Fisher Scientific) [42]. For total NEJTeg, bands were stained with a Coomassie blue kit (Thermo Fisher Scientific), excised and digested as described [43]. The proteomics analysis of the NEJTeg extract was performed by nano reverse-phase (RP) LC-ESI-ion trap MS/MS, consisting of an Ultimate 3000 RSLC nano LC system (Thermo Fisher Scientific) coupled to a Captive Spray nano Booster (Bruker Daltonics) according to previous protocol [43]. All MS/MS spectra were analysed with Mascot (version 2.4.0; Matrix Science), set up to search against two proprietary Fasciola hepatica databases, assuming digestion with trypsin with 1 miss cleavage allowed: (1) a database comprising the gene models identified from the F. hepatica genome (101,780; [44]) (2) a database comprising all available EST sequences (633,678 entries). Fragment and parent ion mass tolerance were set at 0.100 Da. Carbamidomethylation of cysteine was specified as a fixed modification. Dehydration of the N-terminus, Glu->pyro-Glu of the N-terminus, ammonia-loss of the N-terminus, Gln->pyro-Glu of the N-terminus, deamidation of Asn and Glu, oxidation of Met, Arg and Thr and biotinylation of Lys were specified as variable modifications. Scaffold v4.34 (Proteome Software Inc.) was used to validate MS/MS based peptide and protein identifications. Peptide identifications were accepted if they could be established at greater than 95% probability to achieve a 0% false discovery rate (FDR) by the Peptide Prophet algorithm [45] with Scaffold delta-mass correction. Protein identifications were accepted if they could be established at greater than 95% probability to achieve a 0% FDR and contained at least 2 identified peptides. Protein probabilities were assigned by the Protein Prophet algorithm [46]. Proteins that contained similar peptides and could not be differentiated based on MS/MS analysis alone were grouped to satisfy the principles of parsimony. Putative N-glycosylation sites of the glycoproteins detected in the A-BB-NEJTeg and from the cathepsins identified in total NEJTeg were searched using the NetNGlyc 1.0 Server (www.cbs.dtu.dk) [47]. Bands from the NEJTeg sample in which the presence of cathepsins were confirmed by MS/MS were selected for in-depth glycopeptide analysis. The amino acid sequence and the molecular weight of the N-glycosylated tryptic peptides of the cathepsins were predicted using the online tool http://web.expasy.org/peptide_mass [48] taking into account carbamidomethylation of cysteine, and methionine oxidation as a variable modification. For nano-RP-LC-ESI-ion trap-MS/MS the peptides and glycopeptides extracted from the cathepsin bands were loaded on a trap column (Acclaim PepMap100 C18 column, 100 μm × 2 cm, C18 particle size 5 μm, pore size 100 Å, Thermo Fisher Scientific) for concentration prior to separation on an Acclaim PepMap RSLCnano-column (75 μm × 15 cm, C18 particle size 2 μm, pore size 100 Å, Thermo Fisher Scientific). The column was equilibrated at RT with eluent A (0.1% formic acid in dH2O) at a flow rate of 700 nL/min. After injection of the sample, elution conditions were switched to 10% solvent B (95% AcN, 5% dH2O), followed by a gradient to 60% B in 45 min and a subsequent isocratic elution of 10 min. The eluate was monitored by absorption at 215 nm. MS was performed on an AmazonSpeed ion trap (Bruker Daltonics) containing an electron-transfer dissociation (ETD) module (PTM Discovery System). The MS instrument was operated in positive ion mode with a mass window of m/z 400–2000. The five most abundant ions in an MS spectrum were selected for MS/MS analysis by collision-induced dissociation (CID) using helium as the collision gas, with ion detection over m/z 300–1300. For the electrospray (1300 V), solvent evaporation was achieved at 180°C with N2 stream at a flow rate of 3 L/min. In manually selected cases glycopeptide sequence analysis was performed using ETD as described [49]. The selected glycopeptide ions were isolated in the ion trap and fluoranthene radical anions were formed by negative chemical ionisation (nCI) with methane as mediator. For the accumulation (typical accumulation time 5 ms) of fluoranthene reactant anions in the ion trap, the polarity was switched to negative mode. Glycopeptide cations and fluoranthene anions were incubated in the ion trap for 70 ms, allowing electron transfer, followed by the registration of the ETD fragment ion spectrum for m/z 150 to 3000. Selected MS/MS spectra of CID and ETD were interpreted manually using Bruker Daltonics Data Analysis software (Bruker Daltonics). Multiple sequence alignment of selected proteins was performed using Clustal Omega [50] using default parameters. Pairwise sequence alignment was carried out using EMBOSS matcher, based on the Bill Pearson's align application, version 2.0u4 [51]. BN1106_s6570B000051 BN1106_s4187B000061 Fh_Contig2249 BN1106_s1922B000122 BN1106_s7612B000030 BN1106_s1518B000071 BN1106_s7307B000022 BN1106_s25B000189 BN1106_s1612B000138 BN1106_s462B000766 BN1106_s2763B000063 BN1106_s3008B000074 BN1106_s1081B000242 BN1106_s5172B000090 BN1106_s666B000200 BN1106_s9461B000006 BN1106_s4565B000032 BN1106_s10139B000014 BN1106_s3227B000227 BN1106_s10667B000018 BN1106_s5100B000033 BN1106_s8462B000006 BN1106_s6570B000050 BN1106_s4482B000044 gi|27526823 The main components of the F. hepatica Teg are glycoproteins [52]. However, the nature of the different glycans and the identity of individual glycoproteins are still unknown. For that reason, a panel of 17 fluorescein-labelled plant lectins were employed in fluorescence microscopy experiments for glycan localisation (Fig 1). In parallel, lectin blots were used to identify and compare protein bands carrying glycans. The most dominant protein bands detected in NEJTeg by silver staining had an apparent molecular weight between 25–37, 50–75 and 100 kDa (Fig 1A). This was consistent with the broad range of bands detected in the blots (Fig 1B), indicating that the majority of the NEJTeg components were glycoproteins. Indeed, this recognition was probed with the mannose (Man)-binding lectins ConA and GNL, all Man and fucose (Fuc)-binding lectins (AAL, LCA and PSA), the Gal- binding lectins GSL-I and all the T-antigen-binding lectins (PNA and Jacalin) suggesting the abundance of these terminal carbohydrate motifs on these NEJTeg-derived glycoproteins. The exclusive 25 and 37 kDa pattern detected by the lectins PHA-L, PHA-E, GSL-II, SJA, DBA, ECL (Fig 1B), UEA and SJA incubated membranes (S1 Fig), suggested the presence of complex glycan mixture and complex glycosylation patterns in these protein bands. Lectin binding revealed that glycans were distributed uniformly on the parasite surface, but most intensely on the tegumental spines. Strong carbohydrate recognition at the oral and ventral suckers was detected by PNA, Jacalin, ConA, PSA and GSL-I (Fig 1B), suggesting the possible role of Man, Fuc and Gal/GalNAc decorated glycoproteins in adhesion and nutrient intake from the host. Another lectin (SBA) bound to a variety of high molecular sized glycoproteins (Fig 1B) with terminal Gal or GalNAc located at the ventral sucker and its periphery. To provide additional information regarding the distribution of F. hepatica glycoproteins, NEJTeg and NEJSom fractions were compared by silver staining and lectin blots. Similarities observed by silver staining in the migratory components between the NEJTeg (Fig 1A) and NEJSom (Fig 2A) and by eleven lectin blots indicated that there were some (glyco)proteins shared in both extracts. In addition, a small group of lectins did not recognise glycoproteins in any of the NEJ extracts, i.e. the Gal-binding lectin RCA-120, all NeuAc-binding lectins (SNA and MAL-II) and the majority of the GlcNAc-binding lectins (DSL, LEL and STL) (S1 Fig), Nevertheless, the lectins GSL-I, SBA, PNA and Jacalin revealed significant differences in glycosylation of NEJSom (Fig 2B). For example, in the NEJSom, PNA, which shows a similar pattern of recognition as Jacalin, bound the 37 kDa band as observed in the NEJTeg, but also two additional prominent (glyco)protein bands at 25 and 75 kDa. These results suggest the presence of T-antigen in the internal organs of the NEJ. As the NEJTeg showed high (glyco)protein abundance and complexity, we decided to perform an in-depth characterisation of the protein-derived glycans of this preparation by MALDI-TOF-MS. In order to characterise the tegumental N-glycans, these were enzymatically released and MALDI-TOF-MS spectra were recorded before and after specific exoglycosidase treatments (Fig 3). The spectrum obtained from PNG-F released N-glycans showed 5 dominant peaks with m/z values 1354.63, 1516.69, 1678.49, 1840.81 and 2002.87 [M−H]− (Fig 3A). These peaks corresponded to glycans with composition H5N2, H6N2, H7N2, H8N2 and H9N2 (hexose (H), N-acetylhexosamine (N)) respectively, suggesting, in line with the previous lectin binding affinity results, a clear dominance of high mannose N-glycans in NEJTeg. This composition was confirmed after exoglycosidase treatment as all these dominant peaks completely disappeared or drastically decreased after α-mannosidase treatment while peak m/z 1030.32 [M−H]−, which corresponded to the N-glycan with composition H3N2, became the most dominant ones in the spectrum (Fig 3B). Some compositions suggestive of complex type N-glycans were detected in the MALDI-TOF MS of the NEJTeg preparation, with lower relative abundance. The comparison of the PNG-F spectra before and after β-N-acetylglucosaminidase treatment confirmed the presence of terminal, unsubstituted GlcNAc on the antennae of glycans with m/z 1379.66 [M−H]− (H3N3F1) (deoxyhexose (F)) and m/z 1582.75 [M−H]− (H3N4F1) seen in Fig 3A but undetectable in Fig 3C. This observation also confirms that the Fuc residue found in these glycans is attached to the chitobiose core. We assume that the Fuc is α1-6-linked to GlcNAc-1 since α1-3-linkage would have rendered the glycans resistant to PNG-F. Both PNG-F glycan spectra before and after β-galactosidase treatment were very similar suggesting that no terminal β-Gal was present which was in line with the negative binding detection of RCA-120 lectin. The spectrum obtained from PNG-A released N-glycans (S2 Fig) showed a similar glycan pattern as the PNG-F release indicating that no specific PNG-F resistant core α1–3 fucosylated glycans are present in NEJTeg. All peaks identified and their corresponding glycan structures are described in Fig 4. To specifically identify which NEJTeg proteins are glycosylated, we biotinylated the glycan structures contained in NEJTeg glycoproteins and then selectively isolated the glycoprotein components by affinity monomeric avidin chromatography from the total NEJTeg protein pool. The efficiency of the glycoprotein biotin-labelling (B-NEJTeg) and isolation of the proteins (A-UB-NEJTeg) and glycoproteins (A-BB-NEJTeg) was confirmed by comparing the protein profile between fractions using silver staining and the streptavidin blot (S3 Fig). Proteomic MS analysis identified a total of 20 proteins in B-NEJTeg and A-BB-NEJTeg preparations whose molecular weights correlated with protein bands found by silver staining. A total of eighteen proteins were identified in the A-BB-NEJTeg (glycosylated) fraction (Table 2). There were potentially important F. hepatica molecules identified in this fraction including the asparaginyl endopeptidase legumain-1, the antioxidants GST and PRX and heat shock protein 70 (HSP70). It is noteworthy also to highlight the identification of the proteases FhCB3 and FhCL3 in the glycosylated fraction. Two isoforms of FhCL3 were identified in this extract. All the glycoproteins except GST and Histone H4 were predicted to contain potential N- glycosylation sites. GST and Histone H4 were predicted not to contain N-glycosylated sites, suggesting that they are O-glycosylated. Cathepsins were the most abundant proteins found in the glycosylated fraction; due to their pivotal role in parasite survival and since cathepsins from the same family have been employed as vaccine candidates, we investigated the total NEJTeg material focusing on evaluating cathepsin-specific glycosylation at the glycopeptide level. After selecting and excising bands found at the predicted molecular weight of the cathepsins (Fig 5A), tryptic in-gel digestion and protein identification was performed in order to select cathepsins to be examined in detail. We were able to identify several cathepsins, including FhCL3, FhCB1, FhCB2, FhCB3 and partial novel FhCB protein (Fig 5B). The NetNGlyc 1.0 server suggested the presence of one potential N-glycosylation site in the amino acid sequence of two of the FhCL3 members (BN1106_s3008B000074 and BN1106_s10139B000014), at the positions N151 and N255, respectively. FhCB1, FhCB2 and the partial novel FhCB were predicted to contain two N-glycosylation sites at positions N145 and N156 for partial novel FhCB, N80 and N315 for FhCB1 and N345 and N356 for FhCB2. FhCB3 was predicted to have four N-glycosylation sites at positions N180, N223, N336 and N453. However, one of the FhCL3 (BN1106_s4187B000061) and the partial FhCB (BN1106_s8462B000006) were predicted not to be N-glycosylated. When 0 missed cleavages were allowed, tryptic peptide masses containing potential N-glycosylation vary from 617.337 to 2658.15 kDa. Once the cathepsin members and the relevant tryptic peptides with potential N-glycosylation sites were identified, we wanted to determine whether these glycosylation sites were occupied by glycans and also the composition of the glycans involved. All tryptic glycopeptides from the four bands were applied to a RP-nano-LC column coupled to an ion-trap MS/MS system and fragmented in the auto-MS/MS mode. A parent-ion of m/z 829.50 [M+3H]3+ in the 20.1 min elution was detected in band 2, giving rise to oxonium glycan fragment ions at m/z 366.09 [M+H]+ (H1N1) and m/z 528.21 [M+H]+ (H2N1), which indicated the presence of the terminal di-mannosylated HexNAc element. Further inspection of the CID-MS/MS spectrum of the parent ion m/z 829.50 [M+3H]3+ indicated the presence of a glycan with the composition H2N2 linked to a peptide of 1755 Da (Fig 6A). [M+3H]3+ fragment ions were observed for this peptide carrying N2 (m/z 721.28) or a H1N2 trisaccharide (m/z 775.31) structure of an Asn-linked glycan. Furthermore, a [M+2H]2+ fragment ion at m/z 980.40 (N1-peptide) was fully in line with the presence of an N-glycan. The CID-MS/MS spectrum of another parent ion m/z 878.70 [M+3H]3+ in the 19.9 min elution indicated the presence of another sort of glycan with the composition F1H2N2 linked to a peptide with the same mass as the previous peptide described (Fig 6B). [M+3H]3+ fragment ions were observed for this peptide carrying a Fuc attached to the N2 (m/z 770.75) or a F1H1N2 tetrasaccharide (m/z 824.64) structure of an Asn-linked glycan. In addition, a series of [M+2H]2+ fragment ions at m/z 1243.62 (H2N2-peptide), 1235.44 (F1H1N2-peptide), 1162.51 (H1N2-peptide), 1154.60 (F1N2-peptide), 1081.49 (N2-peptide), 1052.96 (F1N1-peptide) and 979.94 (N1-peptide) were fully in line with the presence of the fucosylated version of the dimannosyl core N-glycan. The mass of the peptide containing the N2H2 and F1N2H2 N-glycans matched the mass of one of the tryptic peptides of FhCB1 (BN1106_s6570B000050). To obtain information on the glycopeptide sequence and identify the cathepsin that contained the N-glycan, the ETD-MS/MS spectra of the same parent ions m/z 829.83 [M+3H]3+ and 878.7 [M+3H]3+ were recorded in which peptide cleavages were predominantly observed while leaving the glycosidic linkages intact [53]. For the parent ion m/z 829.83 [M+3H]3+ (Fig 6C), the c’-type as well as z’-type ions arising from peptide bond cleavages provide partial sequences of the peptide backbone from the N- and C-terminal side, respectively. As annotated in Fig 6C, the VSENDLPESFDAR sequence can be read directly from the c2-c15 ion series, while the signals indicated with z2-z15 read the DFSEPLDNESVN(glycan)Y sequence, including the mass increment of 732 Da accountable to the glycan H2N2. This sequence is the FhCB1 tryptic peptide (no missed cleavage) Y79NVSENDLPESFDAR93, containing the N-glycosylation consensus sequence NVS. On the other hand, when the ETD-MS/MS spectrum of the parent ion m/z 878.7 [M+3H]3+ was recorded, the VSENDLPESFDAR sequence containing the glycan F1N2H2 could not be directly read. Taken together, the MS/MS data indicated that FhCB1 is modified at the Asn in position N80 with a Manα1-3/6Manα1-4GlcNAcβ1-4GlcNAcβ1-Asn N-glycan. Nevertheless, despite the strong evidence of the presence of a monofucosylated version of the same N-glycan attached to the FhCB1 in CID-MS analysis, its confirmation remained inconclusive by ETD-MS. For band 2, the presence of the H2N2 and F1H2N2 N-glycans attached to a peptide of 1741.2 kDa were detected in the chromatograms of the parent-ions of m/z 824.70 [M+3H]3+ in the 19.9 min elution and m/z 873.70 [M+3H]3+ in the 19.8 min elution respectively (Fig 7A and 7B). The annotation of the ETD-MS/MS spectrum of the same parent ion m/z 825.20 [M+3H]3+ showed the VSDNDLPESFDAR sequence from the c2-c15 ion series and the DFSEPLDNDSVN(glycan)Y sequence from the z2-z15 ion series, including the mass increment of 732 Da of the glycan H2N2 (Fig 7C). This sequence is slightly different to the sequence of the tryptic peptide identified for FhCB1 (no missed cleavage) as there is a replacement of the glutamic acid (E) in position 83 by an aspartic acid (D). As the new sequence was not identified in the list of tryptic peptide containing N-glycosylation, the sequence tag YNVSDNDLPESFDAR was analysed by BLAST (NCBI nr database). The result showed a 100% coverage and identity hit with FhProCB2 (gi|27526823) which contains the N-glycosylation consensus sequence NVS in position N70. Following identification of further cathepsin B class proteins for F. hepatica and the comparative analysis of those cathepsin sequences publically available, the FhProCB2 (gi|27526823) is consistent with the FhCB1 sequence (S1 File, S1 Table). The slight difference (represented as at most 2 SNPs resulting in the amino acid change from aspartic acid to glutamic acid) between the tryptic peptide YNVSDNDLPESFDAR and the sequence corresponding to FhCB1 sequence in the F. hepatica genome and that submitted to Genbank is most likely due to isolate differences between the various fluke samples used for these studies. As demonstrated for FhCB1, FhCL3 also showed firm evidence of the same glycosylation pattern. The fragmented ions confirming the glycan structures are described in Table 3. As with the analysis performed in tryptic peptides found in band 2, tryptic glycopeptides of the FhCL3 (BN1106_s10139B000014, scaffold10139) were detected in bands 1 and 4. The same glycan structures seen for FhCB1 were found to be linked to a peptide of 2113.9 kDa in the chromatograms of the parent-ions of m/z 949.20 and 997.80 [M+3H]3+ in the 22.2 and 22.1 min elution respectively. To investigate this further for the tryptic peptide derived from FhCL3, ETD-MS/MS for the same parent ion in the same band was carried out. The results showed that the ion c2-c17 and z4-z17 ion series have correspond to the sequences N(glycan)QTLFSEQQVLDCTR and VLQQESFLTQN(glycan)QF respectively, taking into consideration the mass increment of the H2N2 glycan. This sequence is the FhCL3 tryptic peptide (no missed cleavage) F151QNQTLFSEQQVLDCTR167, containing the N-glycosylation consensus sequence NQT. No definitive identification of its fucosylated version could be achieved in the ETD-MS/MS analysis. Sixty four additional N-glycopeptide species were found in the CID-MS/MS spectra whose parent ions were registered in the elution times from 2.5 to 29.3 min (S2 Table). The glycans attached to those peptides matched those observed in the NEJTeg N-glycan profile (Fig 4). Nevertheless, the peptide mass to which the glycans were linked did not match with the masses of the remaining tryptic peptides derived from the cathepsins. To date, this is the first study characterising the glycans attached to the proteins in the F. hepatica NEJTeg and NEJSom preparations by the systematic use of an extensive panel of lectins. In addition, a mass spectrometric glycan characterisation of the tegumental extract was performed. The lack of sialic acid and the wide presence and distribution of high mannose N-glycans and α1–6 Fuc attached to the chitobiose core of oligomannose and truncated N-glycans in the NEJTeg was firstly detected by the binding patterns of the lectins ConA, GNL, PSA and LCA, and confirmed later on by MALDI-TOF MS which represent an important percentage of the total of NEJ N-glycans. The wide distribution of Man-rich N-glycans and the presence of fucosylated glycans among the NEJ glycoproteins have been found in other helminths including O. viverrini [18] Schistosoma spp. [54,55], Haemonchus contortus [56,57] and Taenia solium [58]. These glycans in S. mansoni are recognised and internalised by DC via the C-lectin type receptor (CLR) DC-specific ICAM-3-grabbing non-integrin (DC-SIGN) [59], suppressing DC maturation and therefore decreasing the secretion of pro-inflammatory cytokines. We propose that glycoconjugates with terminal Man-N-glycans in the NEJTeg are likely to interact with DCs and macrophages via Man-specific CLRs at the early stages of infection, impairing the DC function and augmenting the M2a population in the peritoneal cavity, as seen with adult F. hepatica homogenate or ES fractions [23–25], contributing to the development of the biased Th2 that is elicited in the mammalian host [25,36]. In our study, lectin affinity techniques in some cases revealed different results than those using other glycomics techniques with respect to identification of terminal carbohydrates. For instance, β-N-acetylglucosaminidase and β-galactosidase treated N-glycans did not show evidence of terminal β-Gal- or GalNAc- on hybrid or complex glycans in the MS spectra but did show positive recognition of terminal Gal/GalNAc- residues on the parasite surface. This may reflect a low abundance and potential masking by high mannose N-glycans which are very abundant in NEJTeg. As seen in lectin blots, not many glycoproteins contain these residues. Alternatively, lectin fluorescence staining may also indicate the presence of glycolipids as described in adult F. hepatica [29,30], which would not have been detected in our MS spectra results, or the O-glycosylation pattern of NEJTeg. Similar PNA recognition has been described in adult F. hepatica [27,31] and in the nematode O. viverrini where 80% of the O-glycan profile is based on the T-antigen structure [18]. Tn antigen was also detected by the binding lectin VVL, which was also described in the basal membrane of the Teg and in the caeca of the adult fluke [27]. Although the same author also described the presence of sialyl-Tn antigen in the adult stage, we did not detect terminal sialic acid in the NEJTeg by MAL-II and SNA. Taking into account the presence of these glycans in the oral and ventral suckers of the parasite, we suggest that NEJ-derived glycoproteins decorated with terminal Man, Fuc and Gal/GalNAc are in close contact with host mucins and intestinal enterocytes. Based on PNA positive recognition at the parasite surface and around the suckers in F. hepatica, O-glycan motifs may also play a role in intestinal epithelial invasion by this parasite, as has been shown also in studies of protozoal infections [60–63]. Differences in NEJTeg and NEJSom glycosylation were confirmed with four lectins. However, we also observed similar protein profiles in silver stained gels and in most of the lectin blots. It is possible that some tegumental material remained in the somatic preparation or vice versa. Due to the size of the NEJ and the relatively large number of parasites needed to obtain enough material, it is very difficult to achieve a complete separation of NEJTeg and NEJSom. It is known that some proteins are present in both F. hepatica Teg and Som. For example, the protein PRX as well as paramyosin are found in both Teg and Som of adult F. hepatica [64–66], although these studies also may be compromised by the difficulties encountered in achieving complete separation of the two fractions. This is the first report that suggests that F. hepatica PRX and GST, which have well known antioxidant and immunomodulatory properties [67–70], are glycoproteins. Particularly for GST F. hepatica, it was not predicted to be potentially N-glycosylated, suggesting that it could be O-glycosylated. The facts that (1) the F. hepatica NEJTeg O-glycan profile could not be confirmed in our previous analysis and (2) the O-glycan consensus is not as specific as that for N-glycans limited our attempts to analyse potential O-glycosylation sites. It is reasonable to assume that mucin-type core-1 O-glycan would be the most likely to be attached to the GST due to the positive lectin binding of proteins at molecular weights of GST. Further studies will be required to assess whether O-glycan structures are present and involved in the detoxification and of immunoregulatory activities of these molecules. To our knowledge, this is the first report describing the specific glycosylation of the F. hepatica NEJ-specific cathepsin cysteine proteases. These molecules, which play an important role in tissue degradation, in parasite penetration in the host gut and in parasite feeding [37,71] are glycosylated by Manα1-3/6Manα1-4GlcNAcβ1-4GlcNAcβ1- N-glycan. Although there is some evidence of the presence of Fuc on the chitobiose core, our analysis was inconclusive due to restrictions on material. However, the presence of this glycan is shown in the PNG-F released glycan pool of NEJTeg. The confirmation of N-glycosylation of FhCL3 in our study differs from previous findings [37,72]. Five genes were found in the F. hepatica genome encoding FhCL3 [44], meaning that the proteins evaluated by previous workers and in the present study could be different. Methodological differences may also be responsible. Previously, glycosylation of FhCL3 was assessed by detecting differences in protein migration in SDS-PAGE gel after enzyme deglycosylation [37,72]. Because of the size and the mass of the N-glycan detected, which is lower than 1 kDa, it is very difficult to measure differences in protein migration in SDS-PAGE gels before and after deglycosylation. Therefore, this is not an accurate approach to verify protein glycosylation when the glycans present are small. The use of MS for glycan analysis is more sensitive and suitable for these purposes. The absence of N-glycosylation sites in other trematode CL-like proteases [73], as well the presence of glycans in other various cathepsin proteinases including those from Trypanosoma brucei Cathepsin B [74] and human Cathepsin V (also known as cathepsin L2) [75] has been reported. Crucially, GST, PRX and FhCL3 have been proposed and used as potential vaccine candidates. When these molecules were used as vaccines their maximum efficacy was observed when antigens were used in their native, rather than recombinant, versions [11,15]. As the putative glycans attached to GST and PRX have not been characterised we suggest that further in-depth molecular glycomic analyses are necessary in order to prove the existence of N- and/or O- glycosylation on these vaccine candidate molecules. There is only one report in the literature where glycosylation of FhCL3 was taken into account for measuring vaccine efficacy. This account involved rats vaccinated with recombinant FhCL3 expressed by Saccharomyces cerevisiae and recombinant baculovirus [76]. The formation of hyper high mannose-type N-glycans by yeasts [77] and mammalian-like post-translational modifications by baculovirus vector expression systems [78] differs significantly from the N-glycan structure identified for the native FhCL3 in the present work. This could have an effect on the final protein folding and/or the differences in the antigenic epitopes that could exist between the native protein and recombinant versions, explaining lack of efficacy of recombinant vaccine candidates. It is possible that the presence of this unusual N-glycan attached to the FhCL3 peptide backbone is required for proper processing and presentation. Correct glycosylation is an important issue for the production of recombinant antigens of F. hepatica and other helminths [79]. Further work in which alternative strategies for protein expression systems could be applied in order to express cathepsins with their correct glycosylation profiles and again, ascertain their importance for vaccine efficacy, is highly desirable. In conclusion, we have demonstrated the presence of carbohydrates in the F. hepatica NEJTeg and NEJSom, with differences in glycosylation between the two fractions. We also have performed a glycan characterisation of the NEJTeg extract showing that high mannose and oligomannose N-glycans are the most dominant glycan structures. Finally, this work has shown that a set of important immunomodulatory proteins from F. hepatica NEJ, including some of the cathepsin family clade, are glycosylated. Knowledge of the specific structures of the glycans decorating these target proteins is particularly valuable for enhancing our understanding of immunoevasion and parasite migration as well as optimising the efficacy of future vaccines.

10.1371/journal.ppat.1002397

The Splicing Factor Proline-Glutamine Rich (SFPQ/PSF) Is Involved in Influenza Virus Transcription

The influenza A virus RNA polymerase is a heterotrimeric complex responsible for viral genome transcription and replication in the nucleus of infected cells. We recently carried out a proteomic analysis of purified polymerase expressed in human cells and identified a number of polymerase-associated cellular proteins. Here we characterise the role of one such host factors, SFPQ/PSF, during virus infection. Down-regulation of SFPQ/PSF by silencing with two independent siRNAs reduced the virus yield by 2–5 log in low-multiplicity infections, while the replication of unrelated viruses as VSV or Adenovirus was almost unaffected. As the SFPQ/PSF protein is frequently associated to NonO/p54, we tested the potential implication of the latter in influenza virus replication. However, down-regulation of NonO/p54 by silencing with two independent siRNAs did not affect virus yields. Down-regulation of SFPQ/PSF by siRNA silencing led to a reduction and delay of influenza virus gene expression. Immunofluorescence analyses showed a good correlation between SFPQ/PSF and NP levels in infected cells. Analysis of virus RNA accumulation in silenced cells showed that production of mRNA, cRNA and vRNA is reduced by more than 5-fold but splicing is not affected. Likewise, the accumulation of viral mRNA in cicloheximide-treated cells was reduced by 3-fold. In contrast, down-regulation of SFPQ/PSF in a recombinant virus replicon system indicated that, while the accumulation of viral mRNA is reduced by 5-fold, vRNA levels are slightly increased. In vitro transcription of recombinant RNPs generated in SFPQ/PSF-silenced cells indicated a 4–5-fold reduction in polyadenylation but no alteration in cap snatching. These results indicate that SFPQ/PSF is a host factor essential for influenza virus transcription that increases the efficiency of viral mRNA polyadenylation and open the possibility to develop new antivirals targeting the accumulation of primary transcripts, a very early step during infection.

The influenza A viruses cause annual epidemics and occasional pandemics of respiratory infections that may be life threatening. The viral genome contains 8 RNA molecules forming ribonucleoproteins that replicate and transcribe in the nucleus of infected cells. Influenza viruses are intracellular parasites that need the host cell machinery to replicate. To better understand this virus-cell interplay we purified the viral RNA polymerase expressed in human cells and identified several specifically associated cellular proteins. Here we characterise the role of one of them, the proline-glutamine rich splicing factor (SFPQ/PSF). Down-regulation of SFPQ/PSF indicated that it is essential for virus multiplication. Specifically, the accumulation of messenger and genomic virus-specific RNAs was reduced by SFPQ/PSF silencing in infected cells. Furthermore, transcription of parental ribonucleoproteins was affected by SFPQ/PSF down-regulation. The consequences of silencing SFPQ/PSF on the transcription and replication of a viral recombinant replicon indicated that it is required for virus transcription but not for virus RNA replication. In vitro transcription experiments indicated that SFPQ/PSF increases the efficiency of virus mRNA polyadenylation. This is the first description of a cellular factor essential for influenza virus transcription and opens the possibility to identify inhibitors that target this host-virus interaction and block virus gene expression.

The influenza A viruses belong to the family Orthomyxoviridae and contain a segmented, single-stranded RNA genome of negative polarity (for a review see [1]. Each of the genomic RNA segments is encapsidated in a ribonucleoprotein particle (RNP) containing the polymerase complex and a number of nucleoprotein (NP) monomers, depending on their size [2], [3]. Contrary to many other RNA viruses, the influenza virus RNPs are transcribed and replicated in the nucleus of infected cells. The enzyme responsible for these activities is the viral polymerase, a heterotrimer that comprises the PB1, PB2 and PA subunits [4]–[6]. The PB1 subunit acts as polymerase [7], [8] while PB2 and PA are responsible for cap-binding and cap-snatching, respectively [9]–[12]. The heterotrimer has a compact structure [2], [13]–[15] and is required for both transcription and replication [7], [16]–[19]. The polymerase complex can be found associated to the RNP structure or in a soluble form [20], the latter being able to oligomerise in vivo [21], [22]. Along the years, a number of human cell factors have been described as interactors of influenza virus polymerase and in some specific cases their role in virus replication has been studied [23]–[36]. In one such studies, we identified the human SFPQ/PSF factor as associated in vivo to influenza virus polymerase by proteomic analysis of purified complexes [34]. Human SFPQ/PSF is a nuclear multifunctional protein that has been implicated in a series of steps in the human gene expression pathway (for a review, see [37]. It was first described as associated to the polypyrimidine tract-binding protein (PTB) [38] and contains regions rich in arginine/glycine and proline/glutamine close to its N-terminus as well as two RRMs located more C-terminal. SFPQ/PSF can be found as a heterodimer with p54nrb/NonO, a protein that is highly homologous to the SFPQ/PSF C-terminal half. The SFPQ/PSF-p54nrb/NonO heterodimer co-purifies with DNA topoisomerase and interacts with RAD1 recombinase, leading to the stimulation of nucleic acid strand transfer and the cleavage/religation steps [39]–[43]. In addition, several reports have shown that SFPQ/PSF and/or p54nrb/NonO can regulate cellular transcription in a variety of genes (reviewed in [37]. In agreement with the SFPQ/PSF association with PTB, its binding has been reported to elements in the splicing machinery, like the U4/U6-U5 tri-snRNP and many other splicing factors [44]–[46] and the RNA pol II CTD [47], probably in a RRM-dependent manner [48]. Consistent with these interactions, SFPQ/PSF has been shown to stimulate the splicing of mRNAs transcribed by strong transcriptional activators and to control alternative splicing [49], [50]. In spite of the above mentioned associations, most of SFPQ/PSF-p54nrb/NonO can be found in the “nuclear matrix” fraction [51], consistent with the proposed role for the heterodimer in the specific retention in the nuclear matrix of RNA which has been A>I hyperedited [37], [52]. In addition to these many potential functions assigned to SFPQ/PSF or the SFPQ/PSF-p54nrb/NonO heterodimer, the former has been shown to bind specifically to a defined stem-loop in hepatitis delta RNA [53] and the 3′-end of HCV genome [54], while the latter inhibits the transport and expression of HIV mRNAs containing the instability region (INS) [55]. Here we have analysed the role of SFPQ/PSF in the influenza virus infection. Silencing the SFPQ/PSF gene, but not p54nrb/NonO, strongly reduced virus multiplication. The accumulation of viral genomic vRNA and mRNA as well as viral proteins was reduced, probably as a consequence of the inhibition of primary and secondary transcription, but normal splicing of virus mRNA was observed. In vitro transcription of recombinant RNPs generated in SFPQ-silenced cells resulted in reduced levels of viral poly A+ RNA. These results are consistent with a role for SFPQ/PSF during the polyadenylation step in the synthesis of viral mRNAs from the parental RNP templates, the earliest nuclear step in virus replication cycle, as well as during secondary transcription. A proteomic analysis of the intracellular complexes formed by recombinant influenza virus polymerase revealed a series of human proteins that were stably associated to the viral enzyme [34]. One of such associated factors was SFPQ/PSF, a multifunctional nuclear protein involved in transcription, post-transcriptional processing of mRNAs and DNA rearrangements [37]. To study the role of SFPQ/PSF in influenza virus replication we first analysed the expression and localisation of the protein along the infection cycle. Cultures of A549 cells were infected with VIC influenza virus at high multiplicity and total cell extracts were prepared at various times thereafter. The accumulation of SFPQ/PSF was determined by Western-blot using specific antibodies and is shown in Figure 1A. No changes in the level of expression of the protein were observed when compared to GAPDH as a standard, whereas the progressive accumulation of virus proteins was apparent (see NP marker in Figure 1A). Similar experiments were obtained when the WSN virus strain was used (data not shown). The localisation of SFPQ/PSF was analysed by confocal immunofluorescence and is presented in Figure 1B. The protein was found in a punctuate pattern within the nucleus, slightly more intense in the nuclear periphery and excluded from the nucleoli. No significant change could be observed in such distribution along virus infection, although a small increase in cytoplasmic staining was apparent at late times (Figure 1B). Similar results were obtained using the HeLa cell line and the WSN virus strain (data not shown). In previous studies we analysed the localisation of SFPQ/PSF and NP by confocal immunofluorescence of cells infected with VIC influenza virus. A clear co-localisation was observed at 4–6 hours post-infection (hpi), particularly prominent at the periphery of the cell nucleus [34]. To further analyse this association in infected cells we carried out co-immunoprecipitation experiments. Cultures of A549 cells were infected at high moi with the VIC strain of influenza virus and at various times after infection cell extracts were prepared and used for immunoprecipitation with anti-SFPQ/PSF or control antibodies. The immunoprecipitates were analysed by Western-blotting with antibodies specific for SFPQ/PSF and NP. The results are presented in Figure 2 and a clear co-immunoprecipitation of NP was observed with the SFPQ/PSF-specific antibodies. In view of these results, the relevance of SFPQ/PSF for influenza virus multiplication was then studied by gene silencing. Cultures of A549 cells were transfected with SFPQ/PSF-specific siRNA, or a scrambled unspecific siRNA as a control, and then infected with VIC influenza virus at low multiplicity. Samples were withdrawn from the supernatant medium at various times after infection and the virus titre was determined by plaque-assay on MDCK cells. The results of kinetics experiments run in triplicate are presented in Figure 3A. A protracted virus growth kinetics was apparent in the SFPQ/PSF–silenced cultures as compared to cultures transfected with control siRNA and the final titre was reduced by 2–3 log units in various experiments similar to that presented in Figure 3A. The level of SFPQ/PSF down-regulation was verified at various times after siRNA transfection and virus infection and is presented in Figure 3D. These results suggested that SFPQ/PSF plays an important role during influenza virus infection. To verify that the virus growth inhibition is really due to SFPQ/PSF down-regulation and not to an spurious off-target effect of the particular siRNA used, similar experiments were performed with an independent SFPQ/PSF-specific siRNA and the results are presented in Figure 3B and 3C. Furthermore, two additional virus strains were used in these experiments, WSN and a VIC/WSN recombinant virus. Again, a strong reduction in the yield of virus production was obtained, thereby confirming that SFPQ/PSF in an important human host factor for the multiplication of influenza virus. As SFPQ/PSF is a multifunctional protein involved in many steps of cellular transcription and post-transcriptional RNA processing [37], it is conceivable that its down-regulation indirectly leads to reductions in influenza virus multiplication. For instance, it would be conceivable that SFPQ/PSF down-regulation inhibits cellular transcription and/or splicing to a level that makes influenza virus unable to replicate, as it depends on these processes for its own transcription and gene expression. Such a possibility would seem unlikely, as the general pattern of cellular protein synthesis is not altered by SFPQ/PSF silencing (see below), but we carried out controls to ascertain the specificity of SFPQ/PSF requirement for influenza virus multiplication. The multiplication of two additional viruses was studied in SFPQ/PSF-silenced cells: Vesicular stomatitis virus (VSV), as an additional example of negative-stranded RNA virus, and Adenovirus 5 (Ad5), as a nuclear virus that is strongly dependent on the cellular transcriptional and splicing machineries. Cultures of A549 cells were SFPQ/PSF- or control-silenced and infected with either VSV or Ad5 and the virus accumulated in the culture supernatant (VSV) or the infected cells (Ad5) was determined by plaque-assay on BHK21 (VSV) or HEK293T cells (Ad5). As presented in Figure 4A and 4B, the multiplication of neither virus was affected by the down-regulation of SFPQ/PSF, indicating that this human protein is a host factor specifically important for influenza virus multiplication. Since it has been shown that SFPQ/PSF associates to p54nrb/NonO (see above), it was important to ascertain whether influenza virus requires SFPQ/PSF or the SFPQ/PSF-p54nrb/NonO heterodimer for proper multiplication. Hence, we analysed the multiplication of influenza virus in A549 cells after silencing p54nrb/NonO by transfection of a p54nrb/NonO-specific siRNA. Importantly, silencing of p54nrb/NonO did not alter the accumulation of SFPQ/PSF or vice versa (data not shown). As indicated in Figure 5, no reduction in virus yield was observed when using the VIC virus strain (Figure 5A) and a small reduction in WSN virus amplification was observed by down-regulation of p54nrb/NonO, much more limited than that observed when silencing SFPQ/PSF (Figure 5B), in spite of an almost complete block of p54nrb/NonO expression (Figure 5C). Therefore, we conclude that it is SFPQ/PSF by itself what influenza virus requires carrying out a normal infection cycle. Once the relevance of SFPQ/PSF for influenza virus infection was verified, we analysed the role of this host factor in the virus cycle. First, the synthesis of viral proteins was studied in SFPQ/PSF- and control-silenced A549 cells. Cultures of SFPQ/PSF- or control-silenced A549 cells were infected at high multiplicity and pulse-labelled with 35S-met-cys at various times after infection. The labelled proteins were analysed by polyacrylamide gel electrophoresis and autoradiography and the results are presented in Figure 6. The synthesis of the major virus proteins was reduced and delayed in SFPQ/PSF-silenced cells, indicating that SFPQ/PSF down-regulation leads to a general reduction and delay in virus gene expression. However, it is worth mentioning that only a slight change was observed in the level and pattern of cellular protein synthesis upon SFPQ/PSF silencing (compare lanes 0 in siCRTL and S1 panels in Figure 6). Similar results were obtained when the accumulation of viral proteins was determined by Western-blot (Figure S1). To analyse further the phenotype of the infection cycle in SFPQ/PSF-silenced cells, the localisation of progeny RNPs was studied by confocal immunofluorescence with anti-NP antibodies. The results are presented in Figure 7. As expected, the level of SFPQ/PSF was decreased in SFPQ/PSF-silenced cells and a general reduction in NP signal was observed, as compared with infected, control-silenced cells (Figure 7A). Only cells with a level of SFPQ/PSF similar to that of control-silenced cells showed an accumulation of NP comparable to control infected cells, although the localisation was not normal (Figure 7A; arrow). Significantly, a linear correlation was observed between the immunofluorescence signals of SFPQ/PSF and NP in random fields of control- or SFPQ/PSF-silenced cells (Figure 7B). Thus, the low levels of virus protein synthesis (Figure 6) and the small virus production (Figure 3) observed in SFPQ/PSF-silenced cultures could be simply the consequence of the infection of a small number of cells not completely silenced upon transfection of SFPQ/PSF-specific siRNA. The reduction of virus protein synthesis under conditions of SFPQ/PSF down-regulation might be due to defects in genome amplification, virus transcription, splicing or translation of viral mRNA. As SFPQ/PSF has been described as a splicing factor, we first analysed whether its down-regulation would alter the splicing of virus mRNAs. The amounts of NS1 and NS2 mRNAs were determined in control and SFPQ/PSF silenced infected cells by a RT-qPCR procedure using TaqMan probes (Table S1). The proportion of NS1 versus total NS mRNA was indistinguishable in both experimental conditions (ratio control/silenced cells 1.016+/−0.035; n = 7 experiments). Next, the levels of accumulation of vRNA, cRNA and mRNA were determined in SFPQ/PSF-silenced and control-silenced cells. Total cell RNA was isolated at various times after high-multiplicity infections and the amounts of virus-specific RNAs corresponding to the NP, NS and M virus segments were determined by hybridisation with specific probes. The results of a representative experiment are presented in Figure 8. In control-silenced cells, the kinetics of accumulation of virus RNAs showed a pattern analogous to that previously described [56], [57], while in SFPQ/PSF-silenced cells a protracted kinetics was observed and 4-5 fold reductions in maximal accumulations of vRNA, cRNA and mRNA were determined. These results indicated that SFPQ/PSF is required for virus RNA replication but could not tell whether this was a direct effect and whether SFPQ/PSF also played a role in virus transcription. To clarify these questions, the accumulations of primary virus transcripts were determined after infection of SFPQ/PSF-silenced and control-silenced cells. The cells were infected at high multiplicity in the presence of cycloheximide to avoid the synthesis of viral proteins and hence the replication of viral RNA [58]. The accumulation of total NS transcripts was determined by RT-qPCR and demonstrated that silencing SFPQ/PSF leads to a 3-fold reduction in primary transcription (Figure 9A). To verify the inhibition of virus multiplication, virus mRNA was determined in infected cells treated or not with cycloheximide. The results are presented in Figure 9B and document around 50-fold reduction upon treatment with the drug. These results suggest that SFPQ/PSF might simply be required for virus primary transcription and the observed reduction in vRNA accumulation would be an indirect consequence, since viral protein expression is essential for viral RNA replication [58]. To test whether virus RNA replication is directly inhibited by SFPQ/PSF down-regulation, in addition to the observed inhibition of primary transcription, we made use of a recombinant replicon system to analyse RNA replication and secondary transcription. In this system, no primary transcription is required for RNA replication to take place, as the virus proteins are provided by plasmid expression in trans. Human HEK293T cells were transfected with SFPQ-specific or control siRNAs and later transfected with plasmids encoding the virus polymerase subunits, NP and a virus genomic plasmid encoding the cat gene in negative polarity. The down-regulation of SFPQ/PSF was ascertained by Western-blot (Figure 10D) and the CAT protein accumulation was determined as a reporter of total replicon activity. The results indicated that silencing SFPQ/PSF lead to a 5-fold reduction in CAT expression (Figure 10A). To determine whether this reduction was due to alterations in viral RNA replication, transcription or both, total cell RNA was isolated and used to determine negative-polarity and positive-polarity RNA accumulations by hybridisation with specific probes. As shown in Figure 10B, a 5-fold reduction in viral transcription was apparent, whereas a two-fold increase was observed between the accumulations of vRNA in control or SFPQ/PSF-silenced cells (Figure 10C). These results indicated that SFPQ/PSF is required for viral transcription, but not for virus RNA replication and suggest that, in the absence of SFPQ/PSF, the viral RNPs are preferentially devoted to RNA replication instead of transcribing their template. One possible mechanism to explain the effects of SFPQ/PSF down-regulation on influenza virus transcription would imply that the interaction of SFPQ/PSF with the viral polymerase present in the RNP increases the affinity of the enzyme for the cap structure. The apparent Kd for the interaction of the isolated PB2 cap-binding domain with 7mGTP is around 170 µM [10], in good agreement with the inhibition data reported for cap-RNA crosslinking to influenza virus RNPs [59], whereas the affinity of binding of eIF4E or CBC to cap-analogues is much higher [60], [61]. This is in contradistinction to the elution profiles of eIF4F and influenza polymerase-template complexes on a 7mGTP-Sepharose resin [14] which show a stronger binding of the latter. Hence, it is conceivable that some cellular factor(s), for instance SFPQ/PSF, interact with viral polymerase and enhances its affinity for binding to cap. To test such possibility we generated recombinant RNPs by transfection of HEK293T cells, which had previously been either control- or SFPQ/PSF-silenced. Total cell extracts of these cells were used for in vitro transcription using ß-globin mRNA as cap-donor. A wide concentration range of cap-donor was employed to test for a potential difference in cap-donor dose response when the reaction was performed in the presence or absence of SFPQ/PSF protein. The results are presented in Figure 11 and no change in the profile of virus RNA synthesis was apparent when SFPQ/PSF was silenced (Figure 11B). However, a clear change in the size distribution of RNA products was observed depending on the downregulation of SFPQ/PSF and irrespective of the cap-donor concentration (Figure 11A). In the presence of SFPQ/PSF the RNA profile was reminiscent of a polyadenylated virus mRNA, as the transcript size was slightly larger than the template (used in the gel as a marker) while downregulation of SFPQ/PSF led to a variety of RNA sizes, always smaller than the template. To further characterise the transcripts generated with SFPQ/PSF-silenced samples they were separated into poly A+ and poly A− fractions and analysed by denaturing polyacrylamide gel electrophoresis. The results are presented in Figure 12. A consistent reduction in the amount of polyadenylated RNA was observed when SFPQ/PSF was downregulated, with a corresponding increased in the poly A− fraction (Figure 12A). Quantification of 5 experiments indicated that the total amount of transcript was not affected by SFPQ/PSF downregulation but the fraction of poly A+ viral mRNA was reduced about 4–5 fold (Figure 12 B). Similar results were obtained when other SFPQ/PSF-specific siRNA was used (see Figure S2). A considerable fraction of the viral poly A− transcrips showed sizes smaller than the template, consistent with RNA degradation or premature termination but, interestingly, the profile of these poly A− transcripts was identical for control- or SFPQ/PSF-silenced samples, suggesting that the reduction in poly A+ transcripts is not due to a defect in transcript elongation but most probably to a deficiency in the polyadenylation step. The results presented here show that SFPQ/PSF is specifically required for influenza virus multiplication and indicate that this cellular factor is essential for the transcription of viral RNPs during both primary and secondary mRNA synthesis. Furthermore, the results shown suggest that SFPQ/PSF plays a role during the polyadenylation step in virus transcription. With the evidence presented, the following picture could be envisaged upon SFPQ down-regulation in infected cells: The primary viral mRNAs lacking poly A would be unstable and hence their accumulation diminished as compared to infections performed in normal cells. In addition, normal recycling of the transcribing polymerase could be affected. As a consequence of these effects, viral RNA replication would be indirectly inhibited, since little polymerase and NP would be synthesised. Moreover, a similar inhibition could be predicted for secondary transcription of the small amount of viral progeny RNA, with the final result of very low viral gene expression and virus production. At present it is not clear how SFPQ/PSF participates in the polyadenylation step of viral transcription. The available evidence indicates that polyadenylation of viral mRNAs is carried out by the polymerase by reiterative copy of the oligo U signal located close to the 5′-terminus of the vRNA template [62]–[65] and is mechanistically distinct from the cellular cleavage and polyadenylation process. The SFPQ/PSF protein is an RNA-binding protein that has been described as essential for the formation of the spliceosome and can be cross-linked to the pre-mRNA in the spliceosome [46]. Furthermore, purified SFPQ/PSF can be specifically cross-linked to poly U, but not to poly C, A or G, showing the same sequence specificity than PTB [38]. Therefore, it is tempting to speculate that SFPQ/PSF could interact both with the viral polymerase [34] and with the viral polyadenylation signal within the RNP to promote polymerase stuttering at the site. Further experiments will be required to test this and other possible alternatives. Down-regulation of SFPQ/PSF leads to a dramatic decrease in the yield of virus infection and hence it could be considered as a new target for antiviral design, a particularly interesting one as it is involved in a very early stage of the infection. Silencing of mouse SFPQ/PSF leads to chromosome instability [66] and we have verified that down-regulation of SFPQ/PSF in some human cells strongly reduces their growth kinetics (unpublished results). Therefore, one should aim at blocking the association of SFPQ/PSF with virus polymerase/RNP for the design of potential new antivirals. The HEK 293T cell line [67] was obtained from J.C. de la Torre and the A549 human cell line [68] was obtained from J.A. Melero. The MDCK and BHK21 cell lines were purchased from ATCC. Cell culture was carried out as described [69]. The influenza virus strains A/Victoria/3/75 (H3N2) (VIC), WSN (H1N1) and a recombinant of both strains (VIC/WSN) was used in these experiments. Titrations were done in MDCK cells as described [70]. Vesicular stomatitis virus (VSV) was provided by R. Alfonso and titrated in BHK21 cells. Adenovirus 5 was provided by P. Fortes. Virus stocks were prepared and titrated in HEK 293T cells as described [71]. Plasmids pCMVPB1, pCMVPB2, pCMVPA and pCMVNP, expressing the influenza virus polymerase subunits and the NP have been described [57]. Plasmid pHHCAT, that transcribes a virus-like cat gene in negative polarity, was provided by A. Rodríguez. The monoclonal antibodies specific for PA have been described [72], [73]. A monoclonal antibody specific for the N-terminal region of M1 and M2 proteins [74] was provided by A. García-Sastre. Antisera specific for PB1 and NP were generated by immunisation of rabbits with recombinant proteins [2], [14], [31]. A monoclonal antibody specific for SFPQ/PSF (ab11825) and polyclonal sera specific for GAPDH (ab9485) and actin (ab8226) were purchased to Abcam. The secondary antibodies used for Western-blot and immunofluorescence was purchased to Sigma and Invitrogen, respectively. Total mouse IgG, with no known specificity, was used as immunoprecipitation control. Protein samples were separated by polyacrylamide-SDS gel electrophoresis and transferred to Immobilon filters. Western-blotting was carried out essentially as described [75]: The membranes were saturated with 3% BSA for 1 h and then incubated with the primary antibodies for 1 h at room temperature. The filters were washed with PBS containing 0.25% Tween 20 and incubated with the appropriate secondary antibody conjugated to horseradish peroxidase. After further washing as above the filters were developed by enhanced chemiluminiscence. For immunoprecipitations, extracts from mock-infected or infected cells were incubated with goat-antimouse-agarose beads (A6531, Sigma) loaded with a monoclonal antibody specific for SFPQ/PSF or an equal amount of total mouse IgG as negative control. After extensive washing with a buffer containing 50 mM Tris HCl pH 7,5, 150 mM NaCl, 0,5% NP-40, 1,5 mM MgCl2, 1 mM DTT, 1 u/µl HPRI and protease inhibitors containing EDTA, the beads were extracted by boiling with Laemmli loading buffer and the samples were analysed by Western-blot with antibodies specific for SFPQ/PSF and NP. For immunofluorescence, cells were fixed with 3% paraformaldehyde. The cultures were permeabilised with 0.5% Triton X100 and processed for indirect immunofluorescence as described before [31]. Images were collected on a Leica SP5 confocal microscope (Leica Microsystems) and processed with the LAS AF Software (Leica Microsystems). For quantisation of cellular staining with anti-SFPQ/PSF and NP antibodies, the average intensities of 50 random images (1024×1024 pixels) of each preparation were determined using the LAS AF Software. The procedures for protein labelling in vivo have been described [76]. Cultures were washed, incubated for 1 h in a DMEM medium lacking met-cys and labelled with 35S-met-cys to a final concentration of 200 µCi/ml. After incubation for 1 h, total extracts were prepared in Laemmli sample buffer and processed by polyacrylamide gel electrophoresis and autoradiography. Quantisation of CAT protein in total cell extracts was done by ELISA (GE Healthcare). The amplification in vivo of recombinant RNPs was performed essentially as described [77]. In brief, cultures of HEK293T cells (2.5 106 cells) were transfected with a mixture of plasmids expressing the polymerase subunits (pCMVPB1, 12.5 ng; pCMVPB2, 12.5 ng; pCMVPA, 2.5 ng), NP (pCMVNP, 500 ng) and a genomic plasmid expressing a vRNA-like cat gene (pHHCAT, 500 ng), using the calcium phosphate technique [78]. At 24 hours post-transfection, total cell extracts were prepared for CAT determination or total cell RNA was extracted. For RNA extraction cell pellets were resuspended in 1 ml of TRIZOL reagent (Invitrogen) and the RNA was purified as recommended by the manufacturer. The RNA was digested with RNAse-free DNAse (1 u/µg) for 1 h at 37°C, extracted with phenol-chloroform-isoamylalcohol and precipitated with ethanol. For alternative splicing studies, poly A+ RNA was isolated by two rounds of chromatography on oligo-dT-cellulose as described previously [79]. The purified RNA was resuspended in nuclease-free water and the absorbance was measured at 260 nm (NanoDrop ND-1000). For dot-blot hybridisation, aliquots of purified RNAs were denatured for 15 min at 55°C in 10SSC, 7.5% formaldehyde and fixed on nylon filters by UV cross-linking. As controls, total yeast RNA or various amounts of plasmids containing cDNAs of the full-length virus genomic segments or the corresponding viral mRNAs, were fixed on the hybridisation filters. Hybridisation was carried out overnight at 40°C in 6xSSC, 0.5% SDS, 5× Denhardt's mixture 26–47% formamide, depending on the probe, and 100 µg/ml single-stranded DNA. After washing with 0.5xSSC-0.5% SDS at 40°C, the filters were quantified in a phosphorimager. As probes, synthetic oligonucleotides specific to detect the various RNA species of each RNA segment analysed were used (Table S1). They were labelled with gamma-32P-ATP and polynucleotide kinase. Additionally, specific riboprobes were used to specifically detect cRNAs. They were transcribed by T3 RNA polymerase using as templates synthetic DNAs containing T3 promoters fused to the viral sequences (Table S1). For siRNA transfection, cultured A549 cells were incubated independently with 5 nM of siSFPQ 1 (107613), siSFPQ 2 (15923), specific for SFPQ, or siNonO 1 (s9614), siNonO 2 (s9612), specific for NonO, or an irrelevant siRNA (AM4611) from Ambion, using Lipofectamine (Invitrogen) as recommended by the manufacturer. Transfection was carried out twice on consecutive days to increase the silencing efficiency before infection. Quantification of virus-specific RNAs for splicing and primary transcription analyses was carried out by RT-qPCR as follows: The RT reaction was performed by addition of 100 ng of RNA resuspended in 10 µl of nuclease-free water and 10 µl of Reaction Mix 2x (Applied Biosystems) as recommended by the manufacturer. From each 20-µl reaction, 2 µl of cDNA was transferred directly to 96-well PCR plates and 12,5 µl of TaqMan universal master mixture (Applied Biosystems) and 1,25 µl of Custom TaqMan assay (designed by Applied BioSystems) were added. PCR was carried out in a PRISM 7000 Sequence detection system (Applied Biosystems), with 1 cycle of 50°C for 2 min followed by 1 cycle of 95°C for 10 min, 40 cycles of 95°C for 15 s and 60°C for 1 min. The cycle threshold (Ct) was determined with analytical software (SDS; Applied Biosystems). Serial dilutions of cDNA were used to ensure amplification in the lineal range. To construct calibration curves for quantification, we generated PCR products whose sequences were identical to the spliced or unspliced mRNAs of NS segment. The sequences of TaqMan probes and primers are presented in Table S1. For in vitro transcription, extracts from control- or SFPQ/PSF-silenced HEK293T cells in which a mini-RNP was reconstituted [2] were incubated for 60 min at 30°C in 20 µl reactions containing 50 mM Tris.HCl, 100 mM KCl, 2 mM MgCl2, 1 mM DTT, 1 mM each of ATP, CTP and UTP, 1 µM alpha-32P-GTP (0.5 mCi/µmol), 10 µg/ml actinomycin D, 1 u/µl HPRI, pH 8.0 and µg/ml (or various amounts, depending on the experiment) ß-globin mRNA. The reaction products were phenol-extracted, ethanol-precipitated and analysed by electrophoresis on 6% polyacrylamide-7 M urea gels. The purified in vitro transcripts were separated into poly A+ and poly A− fractions by chromatography on oligo dT-cellulose as described [80]. To monitor the recovery of RNAs during extraction and fractionation two labelled synthetic oligonucleotides were added, a 50 nt oligonucleotide lacking poly T sequences and a 70 nt long containing a 39T tract at its 3′-terminus.

10.1371/journal.pcbi.1002282

Collective Animal Behavior from Bayesian Estimation and Probability Matching

Animals living in groups make movement decisions that depend, among other factors, on social interactions with other group members. Our present understanding of social rules in animal collectives is mainly based on empirical fits to observations, with less emphasis in obtaining first-principles approaches that allow their derivation. Here we show that patterns of collective decisions can be derived from the basic ability of animals to make probabilistic estimations in the presence of uncertainty. We build a decision-making model with two stages: Bayesian estimation and probabilistic matching. In the first stage, each animal makes a Bayesian estimation of which behavior is best to perform taking into account personal information about the environment and social information collected by observing the behaviors of other animals. In the probability matching stage, each animal chooses a behavior with a probability equal to the Bayesian-estimated probability that this behavior is the most appropriate one. This model derives very simple rules of interaction in animal collectives that depend only on two types of reliability parameters, one that each animal assigns to the other animals and another given by the quality of the non-social information. We test our model by obtaining theoretically a rich set of observed collective patterns of decisions in three-spined sticklebacks, Gasterosteus aculeatus, a shoaling fish species. The quantitative link shown between probabilistic estimation and collective rules of behavior allows a better contact with other fields such as foraging, mate selection, neurobiology and psychology, and gives predictions for experiments directly testing the relationship between estimation and collective behavior.

Animals need to act on uncertain data and with limited cognitive abilities to survive. It is well known that our sensory and sensorimotor processing uses probabilistic estimation as a means to counteract these limitations. Indeed, the way animals learn, forage or select mates is well explained by probabilistic estimation. Social animals have an interesting new opportunity since the behavior of other members of the group provides a continuous flow of indirect information about the environment. This information can be used to improve their estimations of environmental factors. Here we show that this simple idea can derive basic interaction rules that animals use for decisions in social contexts. In particular, we show that the patterns of choice of Gasterosteus aculeatus correspond very well to probabilistic estimation using the social information. The link found between estimation and collective behavior should help to design experiments of collective behavior testing for the importance of estimation as a basic property of how brains work.

Animals need to make decisions without certainty in which option is best. This uncertainty is due to the ambiguity of sensory data but also to limited processing capabilities, and is an intrinsic and general property of the representation that animals can build about the world. A general way to make decisions in uncertain situations is to make probabilistic estimations [1], [2]. There is evidence that animals use probabilistic estimations, for example in the early stages of sensory perception [3]–[11], sensory-motor transformations [12]–[14], learning [15]–[17] and behaviors in an ecological context such as strategies for food patch exploitation [18]–[20] and mate selection [21], among others [13], [17], [21], [22]. An additional source of information about the environment may come from the behavior of other animals (social information) [23]–[28]. This information can have different degrees of ambiguity. In particular cases, the behavior of conspecifics directly reveals environmental characteristics (for example, food encountered by another individual informs about the quality of a food patch). Cases in which social information correlates well with the environmental characteristic of interest have been very well studied [29]–[37]. But in most cases social information is ambiguous and potentially misleading [26], [38]. In spite of this ambiguity, there is evidence that in some cases such as predator avoidance [39], [40] and mate choice [41], animals use this kind of information. Social animals have a continuous flow of information about the environment coming from the behaviours of other animals. It is therefore possible that social animals use it at all times, making probabilistic estimations to counteract its ambiguity. If this is the case, estimation of the environment using both non-social and social information might be a major determinant of the structure of animal collectives. In order to test this hypothesis, we have developed a Bayesian decision-making model that includes both personal and social information, that naturally weights them according to their reliability in order to get a better estimate of the environment. All members of the group can then use these improved estimations to make better decisions, and collective patterns of decisions then emerge from these individuals interacting through their perceptual systems. We show that this model derives social rules that economically explain detailed experiments of decision-making in animal groups [42], [43]. This approach should complement the empirical approach used in the study of animal groups [42]–[47], finding which mathematical functions should correspond to each experimental problem and to propose experiments relating estimation and collective motion. The Bayesian structure of our model also builds a bridge between the field of collective behavior and other fields of animal behavior, such as optimal foraging theory [18]–[22] and others [21], [22]. Further, it explicitly includes in a natural way different cognitive abilities, making more direct contact with neurobiology and psychology [3]–[10],[17]. We derived a model in which each individual decides from an estimation of which behavior is best to perform. These behaviors can be to go to one of several different places, to choose among some behaviors like forage, explore or run away, or any other set of options. For clarity, here we particularize to the case of choosing the best of two spatial locations, and (see Text S1 for more than two options). ‘Best’ may correspond to the safest, the one with highest food density or most interesting for any other reasons. We assume that each decision maker uses in the estimation of the best location both non-social and social information. Non-social information may include sensory information about the environment (i.e. shelter properties, potential predators, food items), memory of previous experiences and internal states. Social information consists of the behaviors performed by other decision-makers. Each individual estimates the probability that each location, say , is the best one, using its non-social information () and the behavior of the other individuals (),(1)where stands for ‘ is the best location’. , because there are only two locations to choose from. We can compute the probability in Eq. 1 using Bayes' theorem,(2)By simply dividing numerator and denominator by the numerator we find an interesting structure,(3)where(4)and(5)Note that does not contain any social information so it can be understood as the “non-social term” of the estimation. We can also understand as the “social term” because it contains all the social information, although is also depends on the non-social information . The non-social term is the likelihood ratio for the two options given only the non-social information. This kind of likelihood ratio is the basis of Bayesian decision-making in the absence of social information [5], [11]–[14]. Eq. 3 now tells us that this well known term interacts with the social term simply through multiplication. We are seeking a model based on probabilistic estimation that can simultaneously give us insight into social decision-making and fit experimental data. For this reason we simplify the model by assuming that the focal individual does not make use of the correlations among the behaviour of others, but instead assumes their behaviours to be independent of each other. This is a strong hypothesis but allows us to derive simple explicit expressions with important insights. The section ‘Model including dependencies’ at the end of Results shows that this assumption gives a very good approximation to a more complete model that takes into account these correlations. The assumption of independence translates in that the probability of a given set of behaviors is just the product of the probabilities of the individual behaviors. We apply it to the probabilities needed to compute in Eq. 5, getting(6)where is the set of all the behaviors of the other animals at the time the focal individual chooses, , and denotes the behavior of one of them, individual . is a combinatorial term counting the number of possible decision sequences that lead to the set of behaviors , that will cancel out in the next step. Substituting Eq. 6 and the corresponding expression for into Eq. 5, we get(7)Instead of an expression in terms of as many behaviors as individuals, it may be more useful to consider a discrete set of behavioral classes. For example, in our two-choice example, these behavioral classes may be ‘go to ’ (denoted ), ‘go to ’ () and ‘remain undecided’ (). Frequently, these behavioral classes (or simply ‘behaviors’) will be directly related to the choices, so that each behavior will consist of choosing one option. For example, behaviors and are directly related to choices and , respectively. But there may be behaviors not related to any option as the case of indecision, , or related to choices in an indirect way. These behaviors can still be informative because they may be more consistent with one of the options being better than the other (for example, indecision may increase when there is a predator, so the presence of undecided individuals may bias the decision against the place where the non-social information suggests the presence of a predator). Let us consider different behavioral classes, . We do not here consider individual differences for animals performing the same behavior (say, behavior ), so they have the same probabilities and . Thus, if for example the first individuals are performing behavior , we have that . We can then write Eq. 7 as(8)where is the number of individuals performing behavior , and(9)The term is the probability that an individual performs behavior when is the best option, over the probability that it performs the same behavior when is the best choice. The higher the more reliably behavior indicates that is better than , so we can understand as the reliability parameter of behavior . If , observing behavior indicates with complete certainty that is the best option, while for behavior gives no information. For , observing behavior favors as the best option, and more so the closer it is to 0. Note that and are not the actual probabilities of performing behavior , but estimates of these probabilities that the deciding animal uses to assess the reliability of the other decision-makers. These estimates may be ‘hard-wired’ as a result of evolutionary adaptation, but may also be subject to change due to learning. To summarize, using Eqs. 3 and 8, the probability that is the best choice, given both social and non-social information is(10)with in Eq. 4 and in Eq. 9. We have so far only considered the perceptual stage of decision-making, in which the deciding individual estimates the probability that each behavior is the best one. Now it must decide according to this estimation. A simple decision rule would be to go to when is above a certain threshold. This rule maximizes the amount of correct choices when the probabilities do not change [48], but is not consistent with the experimental data considered in this paper. Applying this deterministic rule strictly, without any noise sources, one would obtain that all individuals behave exactly in the same way when facing the same stimuli, but in the experiments considered here this is not the case. Instead, we used a different decision rule called probability matching, that has been experimentally observed in many species, from insects to humans [49]–[55]. According to this rule an individual chooses each option with a probability that is equal to the probability that it is the best choice. Therefore, in our case the probability of going to (), is the same as the estimated probability that is the best location (), so(11)Probability matching does not maximize the amount of right choices if we assume that the probabilities stay always the same, but in many circumstances it can be the optimal behavior, such as when there is competition for resources [56], [57], when the estimated probabilities are expected to change due to learning [53], [55], or for other reasons [53], [58]. Finally, using Eqs. 10 and 11 we have that the probability that the deciding individual goes to is(12)The assumption of probability matching has the advantage that the final expression for the decision in Eq. 12 is identical to the one given by Bayesian estimation in Eq. 10, with no extra parameters. Alternative decision rules could be noisy versions of the threshold rule, but at the price of adding at least one extra parameter to describe the noise. Also, decision rules might not depend on estimation alone, but also on other factors or constraints. These more complicated rules fall beyond the scope of this paper. In the following sections, we particularize Eq. 12 to different experimental settings to test its results against existing rich experimental data sets that have previously been fitted to different mathematical expressions [42], [43]. We first considered the simple case of two identical equidistant sites, and , Fig. 1A. For a set-up made symmetric by experimental design there is no true best option. But deciding individuals must act, like for any other case, using only their incomplete sensory data to make the best possible decision. Even when non-social sensory data indicates no relevant difference between the two sites, the social information can bias the estimation of the best option to one of the two sites. Using Eq. 12 and that the three possible behaviors are ‘go to ’ (), ‘go to ’ () and ‘remain undecided’ (), we obtain(13)where and are the number of individuals that have already chosen and , respectively, and is the size of the group containing our focal individual and other animals. As the set-up is symmetric, the sensory information available to the deciding individual is the same for both options so and then according to Eq. 4. Also, since indecision is not related to any particular choice, symmetry imposes , so indecision is not informative, (Eq. 9). For the other two behaviors, going to () and going to (), Eq. 9 gives(14) and are the estimated probabilities of making the right choice, that is, going to when is the best option, or going to when is the best option. Since in this case the sensory information is identical for both options, the probability of making the correct choice must be the same for both options, . An analogous argument holds for the incorrect choices, , giving(15)In cases in which , we find it convenient to express reliability more generally as(16)which is the ratio of the probability of making the correct choice and the probability of making a mistake, for both behaviors. Using this definition and given that , Eq. 13 reduces to(17)with the variable . Eq. 17 describes a sigmoidal function that is steeper the higher the higher the value of (Fig. 1B). Therefore, for very reliable behaviors (high , meaning individuals that are much more likely to make correct choices than erroneous ones), grows fast with and the deciding individual then goes to with high probability when taking into account the behaviors of only very few individuals. The behavior of the group is obtained by applying the decision rule in Eq. 17 sequentially to each individual (see Methods). After each behavioural choice, we update the number of individuals at and , using the new and for the next deciding individual (Fig. 1C, bottom). Repeating this procedure for all the individuals in the group, we can compute the probability for each possible final outcome of the experiment (Fig. 1C, top). The relevance of the symmetric case is that the model has a single parameter and a single variable, enabling a powerful comparison against experimental data. We tested the model using an existing rich data set of collective decisions in three-spined sticklebacks [42], a shoaling fish species. This data set was obtained using a group of fish choosing between two identical refugia, one on their left and another one on their right (Fig. 2A), equivalent to locations and in the model (Fig. 1A). At the start of the experiment, () replica fish made of resin were moved along lines on the left (right) towards the refugia (Fig. 2A). The experimental results consisted on the statistics of collective decisions between the two refugia for 19 different cases using different group sizes  = 2, 4 or 8 and different numbers of replicas going left and right,  = {1∶1, 2∶2, 0∶1, 1∶2, 0∶2, 1∶3, 0∶3} (Fig. 2B, blue histograms). To compare against these experimental data, we calculated the probability of finding a collective pattern applying the individual behavioural rule in Eq. 17 iteratively over each fish for the 19 experimental settings. We found a good fit of the model to the experimental data using for the 19 graphs the same value (Fig. 2B, red line). The model is robust, with good fits in the interval (Fig. 3, red line). Despite the simplicity of the behavioral rule in Eq. 17, it reproduces the experimental results, including the dependence on the total number of fish , even though the rule is independent of this parameter, except for determining the range of possible values of . The dependence of the final distributions on emerges from the application of the rule to the individuals in the group, as is illustrated in Fig. 4. Each small box represents a state of the system in which fish have already decided to go to and , respectively. The lines connecting each box with another two boxes on top represent the decision made by the next deciding individual, that takes the system to the next state. The width of the lines is proportional to the probability of the decision. As more individuals decide, the central states become less likely simply because they accumulate more unlikely decisions. Therefore, the U-shape or J-shape becomes more pronounced for larger groups, even though the individual decision rule in Eq. 17 is independent of the total number of individuals . Group decision-making in three-spined sticklebacks shows a single type of distribution in which probability is minimum at the center and increases monotonically towards the edges, denoted here as U-shaped distribution (or J-shaped when there is a bias to one of the two options). However, the model in Eq. 17 also gives two other types of distributions, Fig. 5A. For non-social behavior () the histogram is bell-shaped due to combinatorial effects. However, a bell-shape is also compatible with social animals for a certain range of and group size (white region on the bottom-left of Fig. 5A). For higher values of , the histograms are M-shaped, with two maxima located between the center and the sides (region coloured in black and blue in Fig. 5A). However, the M shape becomes clear only with enough number of bins because the drop in probability near the edge or at the center of the distribution disappears when binning is too coarse, producing a bell-shaped or U-shaped histogram, Fig. 5B. This is an important practical issue, because the amount of data that can be collected rarely allows for more than 5 bins. The colorscale in Fig. 5A reflects the number of bins needed to observe the M shape (black has been reserved for exactly 5 bins). For high values of , the histograms are U-shaped (white region on the top of Fig. 5A). Also, all the M-region above the black zone becomes of type U when the binning is too coarse. An interesting prediction of our model is that, for a given number of bins, the shape of the distribution of choices changes with the number of decided individuals, and the dynamics of this change depends on . For high values of , the probability is U-shaped from the beginning and becomes steeper as more individuals decide (as is the case for the stickleback dataset), Fig. 5C. For lower values of , we observe M-shaped distributions for the first individuals and then U-shaped ones when more individuals decide, Fig. 5D. For even lower values of , we observe bell-shaped distributions for the first individuals, then M-shaped and finally U-shaped, Fig. 5E,F. An interesting modification of the experimental set-up consists in using replicas of the animals that we can modify to potentially alter their reliability estimated by the animals. We considered the particular case, motivated by experiments in [43], of two types of modified replicas with different characteristics (for example, fat or thin), Fig. 6A. We considered 7 behaviors: ‘animal goes to ’ (), ‘animal goes to ’ (), ‘most attractive replica goes to ’ (), ‘most attractive replica goes to ’ () ‘least attractive replica goes to ’ (), ‘least attractive replica goes to ’ (), and ‘animal remains undecided’ (). The probability of going to in Eq. 12 then reduces to(18)where subindex ‘f’ refers to real fish and ‘R’ (‘r’) to replicas of the most (least) attractive type. As in the previous section, symmetry imposes that and . It also imposes the following relations between the reliability parameters, , , . Therefore,(19)where , and . In the particular case of only two different replicas, one going to and the other to and for notational simplicity taking the convention that the most (least) attractive replica goes to (), we have and . Therefore,(20)Note that the probability in Eq. 20 does not depend on and separately, but only on their ratio. Therefore, in this case the model uses only two parameters ( and ). We compared the model with the stickleback data set from [43], Fig. 6. The data in Fig. 6B has a different type of replica pair in each row, so in principle we would fit a different ratio for each row. But note that the first three rows correspond to experiments with the same three replicas (large, medium and small), combined in different pairs. The same can be said for the second and third threesomes of rows. Therefore, there are only two free parameters for each three rows. On the other hand, should have the same value for all cases. The model again reproduces the experimental results reported in reference [43], obtaining the best fit for (Fig. 6B). The result is robust, with good fits for (Fig. 3, green line) in accord with the value obtained for the case shown in Fig. 2B. We finally considered the case in which sites and are different and the three behaviors are ‘go to ’ (), ‘go to ’ () and ‘remain undecided’ (). Eq. 12 reduces to(21)The term represents the non-social information and in general because the set-up is asymmetric by design. This asymmetry might also affect how a deciding animal takes into account the behaviours of other animals depending on which side they chose, making in general . Also, indecision might be informative. For example, if non-social information indicates the possible presence of a predator at , the indecision of other animals might confirm this to the deciding individual, further biasing the decision towards . Therefore, we may have . But it may also be the case that the set-up's asymmetry does not affect the social terms, so we also tested a simpler model in which and , giving(22) The stickleback dataset reported in reference [42] is ideally suited to test the asymmetric model for the experiments that were performed with a replica predator at the right arm (Fig. 7A). The model in Eq. 22 fits best the data with (Fig. 7B) and it is robust with a good fit in (Fig. 3, blue line). The more complex model in Eq. 21 gives fits very similar to those of simpler model. Specifically, parameter was rejected by the Bayes Information Criterion [59], [60], suggesting that fish do not rely on undecided individuals. The fact that fish rely differently on other fish depending on the option they have taken could not be ruled out by the Bayes Information Criterion, but in any case the impact of this difference on the data is small. In the experiments in Fig. 2 and Fig. 7, we have assumed that the replicas are perceived by fish as real animals. However, it is reasonable to think that fish might perceive the difference, and rely differently on replicas and real fish. To test this, we considered different behaviors for fish and replicas, such as ‘fish goes to ’ and ‘replica goes to ’. Making that distinction, we get that Eq. 12 reduces to(23)The Bayes Information Criterion rejects only parameter . However, the addition of the new parameters that distinguish replica from real fish give very small improvements in the fits compared to results of the simpler models in Eq. 17 and Eq. 22 (see Fig. S1 and S3), suggesting that fish follow replicas as much as they follow real fish. In this section we will remove the hypothesis of independence among the behaviors of the other individuals (Eq. 6). We now consider that the focal individual not only takes into account the behaviors of the other animals at the time of decision but the specific sequence of decisions that has taken place before, , being the number of individuals that have decided before the focal one. For example, the sequence may give different information to the focal individual than the sequence . This is illustrated in Fig. 8A, where there are two possible paths leading to states labeled as 1∶1, but these two states are in different branches of the tree (in contrast with Fig. 4, in which these two states were collapsed in a single one). To calculate the probability of the observed sequence of behaviors provided that is the correct choice,(24)one can apply repeatedly to obtain(25)This expression substitutes the assumption of independence in Eq. 6. Each of the terms in the product is simply the probability that the individual makes its decision, given the previous decisions, and also given that is the correct choice. This result was expected since if we look at the tree in Fig. 8A we see that the probability of reaching a given state is simply the product of the probabilities of choosing the adequate branches in each step. So the problem reduces to computing the individual decision probabilities . We assume in the following that these probabilities are calculated by the focal individual by assuming that all animals use the same rules to make a decision. The rule for the focal individual is, as in previous sections,(26)where the non-social and social terms are(27)and(28)respectively, and where we have added subscript to , and to reflect that they apply to the focal individual, that makes its decision in the place. The assumption that all animals apply the same rules translates into the following. To apply an equation like Eq. 26 but on a different individual (say, individual ) it is necessary to know the non-social information . Remember that all these computations are made from the point of view of the focal individual, and obviously the focal individual does not have access to the non-social information of the other individuals. It may seem reasonable for the focal animal to assume that all the other individuals have the same non-social information (), but this would result in no social behavior at all (if the other individuals have the same non-social information, their behaviors will not give any extra information). Instead, one can assume that the other individuals may have a different non-social information, . Furthermore, this non-social information depends on which is the best choice, because if for example is the best choice the other individuals have some probability of detecting it, and therefore their non-social information will be on average biased towards . We approximate this average bias by assuming that, if () is the best choice, all the other individuals will have non-social information () that will bias the decision towards (). It is therefore the same to assume that () is the best option as to assume that all the other individuals have non-social information (). Therefore, for the probabilities of individual behaviors in Eq. 25, we have that(29)where now applies to the individual, so we can compute this probability simply by applying Eq. 26 to the individual,(30)where(31)Then, if we denote , we have that(32)These are the individual probabilities needed in Eq. 25, that takes into account the correlations among the other individuals. So we can already calculate using Eq. 28,(33)Eqs. 30 and 33 have a recursive relation, because we need the probabilities up to step to compute , and then we need to compute the probabilities in step . At the beginning no individual has made any choices, so we start with and work recursively from there until we obtain the probabilities for individual , that allow to compute . Then, we can already use Eq. 26 to compute the decision probability of the focal individual, this time using its actual non-social term (which is 1 for the symmetric cases, and fitted to the data in the non-symmetric case). The equations above constitute the model taking into account dependencies. The new parameters of this model are and , which substitute and in the previous models, so the number of parameters is exactly the same. In the symmetrical case we must have that , so the model has a single parameter. For the non-symmetrical case these parameters may be independent of each other, but we find good results even assuming that they are not, as was the case for the simplified model. So for simplicity we always assume that(34)For the case with different replicas at each side, each of them has a different value of , thus making one replica more attractive than the other. The new model also matches very well with the experimental data discussed in this paper. Results for the case of two different replicas are shown in Fig. 9, for the symmetric case in Fig. S4 and for the case with predator in Fig. S5. Fits are robust, and all cases are well explained by the model with the same value of , Fig. S6. See Figs. S1, S2, S3 for a comparison of all models. We now ask how different is the model including dependencies from the model that neglects them. To compare the two models, we plot the probability of going to as a function of for the new model, as we did in Fig. 1B for the old one. The inclusion of dependencies has the consequence that the probability of going to does not depend only on , since now different states with the same may have different probabilities. Therefore, when we plot the probability of going to as a function of we obtain different values of the probability for each value of . This is shown by the black dots in Fig. 8B, where the size of the dots is proportional to the probability of observing each state when starting from 0∶0. The red line shows the average probability for each , taking into account the probability of each state. Both the dots and this line correspond to , which is the one that fits best the data. The green line corresponds to the probability for the simplest model neglecting dependencies, with the value that best fits to the data (). This line is close to the mean probability for the new model and to the values with highest probability of occurrence, so the simple model is as a good approximation to the model with dependencies. We find an interesting prediction of the new model: There are some states in which the most likely option is to choose the option chosen by fewer individuals (for example, note in Fig. 8D that some points with are above 0.5). This surprising result comes from the fact that, as more fish accumulate at one side, their choices become less and less informative (because it is very likely that they are simply following the others). If then one fish goes to the opposite side, its behavior is very informative, because it is contradicting its social information. This effect can be so strong that it may beat the effect of all the other individuals, resulting in a higher probability of following this last individual than all the individuals that decided before. We have shown that probabilistic estimation in the presence of uncertainty can explain collective animal decisions. This approach generated a new expression for each experimental manipulation, Eq. 17–22, and was naturally extended to test for more refined cognitive capacities, Eq. 23. The model was found to have a good correspondence with the data in three experimental settings (Figs. 2, 6 and 7), always giving a good fit with the social reliability parameter in the interval 2–4. Indeed, all the data have a very good fit with (Figs. 2, 6 and 7, green lines). According to Eq. 9, this value for has the interpretation that, for the behaviors relevant for these experiments, the fish assume that their conspecifics make the right choice 2.5 times more often than the wrong choice. For the data used in this paper, previous empirical fits used more parameters [42] (Figs. S1, S2, S3, blue line), and added more complex behavioral rules when the basic model failed [43] (Fig. S2, blue line). Our approach thus gains in simplicity. It also finds an expression for each set-up with expressions for complex set-ups obtained with add-ons to those of simpler set-ups, making the model scalable and easier to understand in terms of simpler experiments. Also, taking the models as fits to experimental data, the bayesian information criterion finds our models to be better than those in [42] and [43] (see captions in Figs. S1, S2, S3 for details). Collective animal behavior has been subject to a particularly careful quantitative analysis. Previous studies have given descriptions led by the powerful idea that complex collective behaviors can emerge from simple individual rules. In fact, some systems have been found empirically to obey rules that are mathematically similar or the same as some of the ones presented in this paper, further supporting the idea that probabilistic estimation might underlie collective decision rules in many species. For example, a function like the one in Eq. 17 has been used to describe the behavior of Pharaoh's ant [61], a function like Eq. 22 for mosquito fish [62], and a function like the one in the right-hand-side of Eq. 22 for meerkats [63]. But despite the importance of group decisions in animals, little is known about the origin of such simple individual rules. This paper argues that probabilistic estimation can be an underlying substrate for the rules explaining collective decisions, thus helping in their evolutionary explanation. Also, this connection between patterns in animal collectives and a cognitive process helps to explain the similarities that exist between decision-making processes at the level of the brain and at the level of animal collectives [64], [65]. Our model is naturally compatible with other theories that use a Bayesian formalism to study different aspects of behavior and neurobiology, thus contributing to a unified approach of information processing in animals. For example, it may be combined with the formalism of Bayesian foraging theory [18], through an expansion of the non-social reliability . Related to this case, a very well studied example of use of social information is the one in which one individual can observe directly the food collected by another individual [29]–[33]. In this case the social information is as unambiguous as the non-social one, so in this case both types of information should have a similar mathematical form [29]–[33]. This is consistent with our model, that in this case will give a similar expression for and . Other kinds of social information (such as another individual's decision to leave a food patch or choices of females in mating [41]) would enter naturally in our reliability terms . In discussing these and similar problems, it has been proposed that animals should use social information when their personal information is poor, and ignore it otherwise [25], [26], [41]. Our model provides a quantitative framework for this problem, predicting that social information is always used, only with different weights with respect to other sources of information. Bayesian estimation is also a prominent approach to study decisions in neurobiology and psychology [3]–[17] and it would be of interest to explore the mechanisms and role played by the multiplicative relation between non-social and social terms. Our approach also makes a number of predictions. For example, it derives the probability of choosing among options (see Eq. S16 of the Text S1), that for the symmetric case reduces to(35)predicted also to fit the data for cases with options. We also predict a quantitative link between estimation and collective behavior. The parameters and in our model are in fact not merely fitting parameters, but true experimental variables. Manipulations of and should allow to test that changes in collective behavior follow the predictions of the model. A counterintuitive prediction about the manipulation of is that external factors unrelated to the social component can nevertheless modify it. For example, a fish that usually finds food in a given environment should interpret a sudden turn of one of his mates as an indication that it has found food, and therefore will follow it. In contrast, another fish that is not expected to find food in that environment will not interpret the sudden turn as indicative of food, and will not follow. Thus, the model predicts that the a priori probability of finding food (to which each fish can be trained in isolation) will modify its propensity to follow conspecifics. An alternative approach that would not need manipulation of the reliabilities would consist in showing that the probability of copying a behavior increases with how reliably the behavior informs about the environment. We can also extend the estimation model to use, instead of the location of animals, their predicted location. We would then find expressions like the ones in this paper but for the number or density of individuals estimated for a later time. Consider for example the case without non-social information, described in Eq. 17 for two options and in Eq. 35 for more options. We can rewrite these equations as with one of the options and is the normalization, , where is the number of options. Then, we would have for the continuous case using prediction. Future positions at times (where does not need to be constant) in terms of variables at present time would be given by for animals moving at constant velocity . Consider then a simple case of an animal located at and estimating the future position of a compact group at and moving with velocity . The deciding animal would be predicted to move with a high probability in the direction . Estimation of future locations thus naturally predicts in this simple case a particular form of ‘attraction’ and ‘alignment’ forces of dynamical empirical models [46], [66] as attraction to future positions, but in the general also deviations from these simple rules. The estimation rules presented in this paper refer to a single individual. To simulate the behavior of a group, we use the following algorithm: The current individual decides between and . After the decision, we recompute the relevant parameters of the model and use the new values for the next deciding individual. The undecided individuals are only those that are waiting for their turn to decide. We tested an alternative implementation in which individuals may remain undecided or in which two individuals can decide simultaneously, obtaining no relevant differences. For the case of the model including dependencies, the model always starts at state 0∶0, with . Most experiments have initial conditions in which several replicas are already going to either side, and the fish have no information about the path followed to reach this state. In these cases, we average the probabilities of all the paths that might have possibly led to the initial state to compute the initial value of . Protocol S1 and Protocol S2, contain Matlab functions that run the models (extensions of the files must be changed from .txt to .m to make them operative). Protocol S1 corresponds to the model without dependencies, and Protocol S2 corresponds to the model with dependencies. These functions have been used to generate all the theoretical results presented in this paper. We computed log likelihood as the logarithm of the probability that the histograms come from the model. We searched for the model parameters giving a higher value of log likelihood, corresponding to a better fit. This search was performed by optimizing each parameter separately (keeping the rest constant) and iterating through all parameters until convergence. In all cases convergence was rapidly achieved. We performed multiple searches for best fitting parameters starting from random initial conditions and always found convergence to the same values, suggesting there are no local maxima. Indeed, we observed that log-likelihood is smooth and with a single maximum in all the cases with 1 or 2 parameters (see Fig. 3 for an example). For model comparison we used the Bayesian Information Criterion (BIC) [59], [60], which takes into account both goodness of fit and the number of parameters. According to this criterion, among several models that have been fitted to maximize log likelihood, one should select the one for which(36)is largest, where is the logarithm of the probability that the data comes from the model once its parameters have been optimized to maximize this probability, is its number of parameters of the model and is the number of measurements (which in our case is the same for all models). More intuitive than the direct values in Eq. 36 are the BIC weights, defined as [60](37)when we assume that all models are a priori equally likely. Roughly speaking, can be interpreted as the probability that model is the most correct one [60]. We used BIC to compare different versions of our model, and also to compare our model with those of references [42], [43] (see Figs. S1, S2, S3). The models of refs. [42], [43] were originally fitted by minimizing the mean squared error instead of by maximizing logprob. For this reason, they score very poorly in BIC with their reported parameters. For this reason, we re-optimized for maximum logprob all their model parameters (these parameters are, using the notation of refs. [42], [43], , , , and , with only applicable in the case of predator present). For the case of different replicas going to each side, parameter takes a different value for each row in the figure, adding up to 10 parameters. The model in ref. [43] is computationally expensive, so it is not feasible to re-optimize these many parameters. Therefore, we treated them as if they were independently measured: we fixed in each case so that the results of the trials with a single individual matched exactly the model's prediction (as reported in [43]). We also followed this procedure with the ratios of our model without dependencies, and the pairs in our model with dependencies. Then, we performed BIC taking into account neither these parameters ( the ratios and the pairs ) nor the data from trials using single individuals.

10.1371/journal.pgen.1005731

dbl-1/TGF-β and daf-12/NHR Signaling Mediate Cell-Nonautonomous Effects of daf-16/FOXO on Starvation-Induced Developmental Arrest

Nutrient availability has profound influence on development. In the nematode C. elegans, nutrient availability governs post-embryonic development. L1-stage larvae remain in a state of developmental arrest after hatching until they feed. This “L1 arrest” (or "L1 diapause") is associated with increased stress resistance, supporting starvation survival. Loss of the transcription factor daf-16/FOXO, an effector of insulin/IGF signaling, results in arrest-defective and starvation-sensitive phenotypes. We show that daf-16/FOXO regulates L1 arrest cell-nonautonomously, suggesting that insulin/IGF signaling regulates at least one additional signaling pathway. We used mRNA-seq to identify candidate signaling molecules affected by daf-16/FOXO during L1 arrest. dbl-1/TGF-β, a ligand for the Sma/Mab pathway, daf-12/NHR and daf-36/oxygenase, an upstream component of the daf-12 steroid hormone signaling pathway, were up-regulated during L1 arrest in a daf-16/FOXO mutant. Using genetic epistasis analysis, we show that dbl-1/TGF-β and daf-12/NHR steroid hormone signaling pathways are required for the daf-16/FOXO arrest-defective phenotype, suggesting that daf-16/FOXO represses dbl-1/TGF-β, daf-12/NHR and daf-36/oxygenase. The dbl-1/TGF-β and daf-12/NHR pathways have not previously been shown to affect L1 development, but we found that disruption of these pathways delayed L1 development in fed larvae, consistent with these pathways promoting development in starved daf-16/FOXO mutants. Though the dbl-1/TGF-β and daf-12/NHR pathways are epistatic to daf-16/FOXO for the arrest-defective phenotype, disruption of these pathways does not suppress starvation sensitivity of daf-16/FOXO mutants. This observation uncouples starvation survival from developmental arrest, indicating that DAF-16/FOXO targets distinct effectors for each phenotype and revealing that inappropriate development during starvation does not cause the early demise of daf-16/FOXO mutants. Overall, this study shows that daf-16/FOXO promotes developmental arrest cell-nonautonomously by repressing pathways that promote larval development.

Animals must cope with feast and famine in the wild. Environmental fluctuations require a balancing act between development in favorable conditions and survival during starvation. Disruption of the pathways that govern this balance can lead to cancer, where cells proliferate when they should not, and metabolic diseases, where nutrient sensing is impaired. In the roundworm Caenorhabditis elegans, larval development is controlled by nutrient availability. Larvae are able to survive starvation by stopping development and starting again after feeding. Stopping and starting development in this multicellular animal requires signaling to coordinate development across tissues and organs. How such coordination is accomplished is poorly understood. Insulin/insulin-like growth factor (IGF) signaling governs larval development in response to nutrient availability. Here we show that insulin/IGF signaling activity in one tissue can affect the development of other tissues, suggesting regulation of additional signaling pathways. We identified two pathways that promote development in fed larvae and are repressed by lack of insulin/IGF signaling in starved larvae. Repression of these pathways is crucial to stopping development throughout the animal during starvation. These three pathways are widely conserved and associated with disease, suggesting the nutrient-dependent regulatory network they comprise is important to human health.

C. elegans L1-stage larvae must feed upon hatching in order to exit developmental arrest. Larvae in L1 arrest have increased stress resistance and can survive for weeks, initiating postembryonic development when food is available [1]. C. elegans larvae can also arrest development during the dauer stage, an alternative to the third larval stage that forms in response to crowding, nutrient stress or high temperature [2]. Unlike dauer formation, L1 arrest is an acute starvation response without an alternative developmental program, making it an excellent model for nutritional control of development. The insulin/insulin-like growth factor (IGF) signaling pathway is a key regulator of L1 arrest, mediating the systemic response to nutrient availability [3,4]. In fed larvae, insulin-like peptides act through the insulin/IGF receptor DAF-2/InsR, activating a conserved PI3K cascade to repress function of the forkhead-type transcription factor DAF-16/FOXO [5–8]. In starved larvae, DAF-16/FOXO is active and promotes L1 arrest [3]. A variety of genome-wide gene expression analyses have been published for daf-16/FOXO, and a meta-analysis of them identified thousands of genes whose expression is positively or negatively affected by daf-16/FOXO activity [9]. This study also identified the transcription factor PQM-1 as a mediator of daf-16-dependent effects on gene expression. However, these experiments were done in young adult animals and with a daf-2/InsR mutant background to activate DAF-16/FOXO as opposed to starvation. DAF-16/FOXO target genes that promote L1 arrest are largely unknown. daf-16/FOXO activates the cyclin-dependent kinase inhibitor cki-1/p27 and represses the developmental timing microRNA lin-4 [3], but whether such regulation is direct is unclear. The insulin/IGF pathway is pleiotropic, serving as a key regulator in dauer formation, lifespan, associative learning, and stress resistance [5,10–15]. daf-2/InsR and daf-16/FOXO affect lifespan and dauer formation cell-nonautonomously [16–20]; that is, daf-16/FOXO activity in limited tissues affects the phenotype of the entire organism. The insulin/IGF pathway is highly conserved, and it also affects lifespan cell-nonautonomously in Drosophila and mice [21–24]. daf-16/FOXO activity in one tissue has been shown to affect daf-16/FOXO activity in other tissues through feedback regulation, termed FOXO-to-FOXO signaling [25–27]. Insulin/IGF receptor cell-nonautonomy could result from FOXO-to-FOXO signaling, since FOXO is still present in the affected tissues. However, FOXO cell-nonautonomy is inconsistent with FOXO-to-FOXO signaling because in this experimental scenario FOXO is not present in the affected tissues. Instead, FOXO cell-nonautonomy suggests FOXO regulates an additional signaling pathway to affect cells where it is not present. The daf-12/nuclear hormone receptor (NHR) signaling pathway is an attractive candidate for a signaling pathway mediating daf-16/FOXO cell-nonautonomy. daf-12/NHR signaling is known to play a key role in coordinating a variety of systemic effects, including dauer formation, aging, and developmental timing [28]. The Rieske oxygenase DAF-36 and the cytochrome P450 DAF-9 are necessary for the production of dafachronic acid, which is a ligand for the nuclear hormone receptor DAF-12 [29–33]. Ligand-bound DAF-12 promotes dauer bypass and reproductive development, and ligand-free DAF-12, together with its co-repressor DIN-1/SHARP, promotes dauer formation [34,35]. Another potential candidate for a signaling pathway mediating the effects of FOXO cell-nonautonomy is the Transforming Growth Factor-β (TGF-β) Sma/Mab pathway. This pathway was first identified by mutations causing small body size and male tail abnormalities [36–38]. Core pathway components include the TGF-β ligand DBL-1, the TGF-β receptor subunits DAF-4 and SMA-6, the SMADs SMA-2, SMA-3, and SMA-4, and the SMAD co-factor SMA-9 [38–43]. The dbl-1/TGF-β pathway has also been shown to regulate reproductive aging, aversive olfactory learning, and mesodermal patterning [44–46]. Here we show that daf-16/FOXO promotes L1 arrest cell-nonautonomously. mRNA-seq reveals that daf-16/FOXO has overlapping but distinct effects on gene expression in starved L1 larvae compared to young adults with reduced insulin/IGF signaling. Our mRNA-seq analysis identified Sma/Mab TGF-β and daf-12/NHR pathways as candidate mediators of FOXO cell-nonautonomy. We show that daf-16/FOXO promotes developmental arrest throughout the animal by inhibiting expression of Sma/Mab TGF-β and daf-12/NHR steroid hormone signaling pathway components. These pathways promote development in fed L1 larvae and in starved daf-16/FOXO mutants, but their activation does not cause the early demise of daf-16/FOXO mutants during starvation. Loss of daf-16/FOXO reduces starvation survival during L1 arrest [3,47]. To determine if daf-16/FOXO expression in a specific tissue is sufficient to rescue starvation survival, tissue-specific promoters were used to express GFP-tagged DAF-16 in a daf-16null background [18,48]. Visible DAF-16::GFP expression patterns were restricted to the tissues targeted by each promoter. Expression from the native daf-16 promoter resulted in complete rescue of the starvation survival defect while expression from neuronal (Punc-119), intestinal (Pges-1), and epidermal (Pcol-12) promoters resulted in significant partial rescue (Fig 1A, S1 Table). Additional transgenic lines with the same promoters provided comparable results (S1 Fig). These data show that DAF-16::GFP expression in various individual tissues is sufficient to affect a whole-animal phenotype. Though it is plausible that individual tissues could have an additive effect on survival, we were unable to detect differences between single and double promoter rescues (S1 Fig). These results suggest that daf-16 functions cell-nonautonomously during L1 arrest as it does for adult lifespan [16,18–20]. daf-16 mutants inappropriately initiate somatic postembryonic development during starvation [3]. To determine if daf-16/FOXO regulates developmental arrest cell-nonautonomously, a variety of markers were examined in tissue-specific DAF-16::GFP rescue strains. The M cell is a mesoblast that undergoes a series of divisions to produce 18 cells during L1 development [49]. The M cell lineage was visualized using a Phlh-8::GFP reporter [50]. The M cell of wild-type larvae did not divide during L1 arrest, though a significant proportion of daf-16null larvae had at least one M lineage division (Fig 1B). Expression of DAF-16::GFP from the native promoter (Pdaf-16) resulted in complete rescue while expression from the intestinal (Pges-1), neuronal (Punc-119) and epidermal (Pcol-12) promoters resulted in significant partial rescue of the daf-16null phenotype. The muscle promoter (Pmyo-3) had no effect. Pairs of tissue-specific promoters also provided significant rescue but were not significantly different from rescue with single promoters, failing to provide evidence of additive effects (S2 Table). There appeared to be variation between independent transgenic lines for each promoter and promoter pair, suggesting allele-specific effects can confound effects of promoter/site of expression (S2 Fig). Rescue of the M cell lineage division defect by expression of DAF-16::GFP in other cells implies that daf-16/FOXO can function cell-nonautonomously during L1 arrest. We investigated cells from different lineages with different developmental fates and behaviors to determine if the cell-nonautonomous effect of daf-16/FOXO extends to development of the entire animal. The intestinal rescue line had the strongest effect of the integrated lines tested, so we used it for these assays. P cells migrate ventrally and divide, and some of their descendants differentiate into the VB motor neurons near the L1 molt [49]. The Pdel-1::GFP reporter is expressed specifically in differentiated VB motor neurons in late L1 larvae [51]. Expression was not detectable in wild type during L1 arrest, but the reporter was active in a significant proportion of daf-16null larvae, reflecting inappropriate VB differentiation (Fig 1D and 1E). Expression of DAF-16::GFP from the intestinal promoter suppressed inappropriate differentiation (Fig 1D and 1E). The lateral epidermal seam cells divide about five hours after hatching in fed larvae [49,52]. AJM-1::GFP marks the adherens junctions of the seam cells, enabling progression of seam cell development to be visualized during L1 development (S3 Fig) AJM-1::GFP was used to determine if seam cells v1–6 divide in starved larvae [53]. daf-16null larvae had significantly more divisions than wild type when starved (Fig 1F and 1G), as expected [3]. Intestinal expression of DAF-16::GFP suppressed the division of seam cells in otherwise daf-16null animals (Fig 1F and 1G). These results show that daf-16 can function cell-nonautonomously to promote developmental arrest in a variety of tissues throughout the body, consistent with systemic regulation. Cell-nonautonomous function of daf-16/FOXO suggests that it regulates at least one additional signaling pathway. FOXO-to-FOXO signaling cannot account for daf-16/FOXO cell-nonautonomy, since daf-16/FOXO is not available as an effector of insulin/IGF signaling in the affected cells. That is, when expressing daf-16/FOXO in specific tissues in a daf-16null background daf-16 is absent from the cells being assayed for division. However, feedback regulation of insulin/IGF signaling could nonetheless be responsible if an alternative effector were involved (i.e. "FOXO-to-effector X" signaling). Double mutant analysis of daf-2/InsR and daf-16/FOXO showed that daf-2/InsR was not epistatic to daf-16/FOXO for M cell division (S4 Fig), arguing against this possibility. Insulin-like peptides ins-4, ins-6, and daf-28 function redundantly to promote L1 development [54]. Simultaneous disruption of ins-4, ins-5, ins-6, and daf-28 along with daf-16/FOXO showed that these insulin-like peptides were also not epistatic to daf-16/FOXO (S4 Fig). The daf-2 allele used is not null (null is lethal) and other insulin-like peptides could be involved, but these results suggest that the cell-nonautonomous effects of daf-16/FOXO on M cell division are not mediated through insulin/IGF signaling. We used mRNA-seq for expression analysis during L1 arrest to identify candidate signaling molecules regulated by daf-16/FOXO activity. A variety of studies have investigated the effects of daf-16 on gene expression in the context of aging. These studies used fed daf-2/InsR adults as a background within which the effect of daf-16 mutation was examined. We suspected that the effects of daf-16/FOXO on gene expression during L1 arrest were sufficiently different from its effects in young adult daf-2/InsR mutants to warrant analysis of L1 arrest in a wild-type background. mRNA-seq analysis of wild type and daf-16null worms on the first day of L1 starvation identified 1,353 genes with reduced expression and 558 genes with increased expression in the mutant with a false-discovery rate (FDR) of 5% (S1 Dataset). This analysis revealed widespread effects of DAF-16/FOXO on gene expression in starved L1 larvae, consistent with the phenotypic consequences of its disruption. Gene Ontology (GO) term enrichment analysis revealed a variety of effects on metabolic and immune system gene expression, with distinct patterns of enrichment for genes up- and down-regulated in the mutant and no overlapping terms between the two gene sets (Fig 2A, S1 Dataset). Notably, the term "determination of adult lifespan" was prominently enriched among genes down-regulated in the mutant, consistent with the known role of daf-16/FOXO in promoting lifespan. These results suggest that we effectively captured the effects of daf-16/FOXO on gene expression underlying the phenotypic consequences of its disruption during L1 arrest. Our mRNA-seq analysis revealed similar but different effects of daf-16/FOXO on gene expression during L1 arrest compared to other contexts. A meta-analysis of multiple different genome-wide analyses of daf-16/FOXO-dependent gene expression in young adult animals with a daf-2/InsR mutant background has been published [9]. We compared the genes identified in our analysis to the "Class I" and "Class II" genes defined in the meta-analysis (Class I genes have reduced expression in the daf-16/FOXO mutant, as if activated by daf-16, and Class II genes have increased expression, as if repressed) to validate our mRNA-seq results. We found significant overlap between Class I genes and those with decreased expression in the daf-16/FOXO mutant during L1 arrest ("Down"; Fig 2B; hypergeometric p-value = 2.8e-88). We also found significant overlap between Class II genes and those with increased expression in the daf-16/FOXO mutant during L1 arrest ("Up"; hypergeometric p-value = 5.3e-87). These observations corroborate our results in that there is significant overlap with the effects of activating daf-16/FOXO via starvation in L1 larvae compared to activating daf-16/FOXO with reduction of daf-2/InsR function in adults. However, because of differences in experimental design, this comparison also allowed us to determine the extent to which the genes affected by daf-16/FOXO activity depends on stage (L1 larvae vs. young adults) and condition (wild type vs. daf-2 mutant and starved vs. fed). Despite the significance of the overlap, the majority of the Up and Down genes we identified were not identified as Class II or Class I genes, respectively (Fig 2B). This result shows that daf-16/FOXO-dependent effects on gene expression are sensitive to stage and/or condition. We examined the Up and Down genes for candidate signaling molecules that could mediate cell-nonautonomous effects of daf-16/FOXO. dbl-1/TGF-β expression was increased 2.1-fold in the mutant (FDR = 0.8%; Fig 2C; S1 Dataset), though other components of the Sma/Mab pathway were not significantly affected. In addition, expression of the Rieske oxygenase daf-36 was increased 3.6-fold (FDR = 3%). DAF-36 is necessary to produce the ligand for DAF-12/NHR, and daf-12/NHR expression was increased 1.7-fold (FDR = 11%), though only marginally significant. Other components of the daf-12/NHR pathway were not significantly affected. The NanoString nCounter platform was used to measure transcript abundance and validate the mRNA-seq results for dbl-1, daf-36, and daf-12. dbl-1 and daf-12 were expressed significantly higher in daf-16null mutants (p<0.005; S5 Fig). Although daf-36 showed a marginally significant increase in expression (p = 0.04), it was below the limits of reliable detection for this assay so we performed qRT-PCR [55]. qRT-PCR of daf-36 showed an increase of 1.7-fold in the mutant (SEM = 0.32; p = 0.01). These results suggest daf-16/FOXO activity during L1 arrest leads to repression of dbl-1/TGF-β, daf-36 and daf-12/NHR. Given the effect of daf-16/FOXO on the dbl-1/TGF-β and daf-36/oxygenase pathways, we hypothesized that the transcription factors SMA-9/co-SMAD, the downstream effector of dbl-1 signaling, and DAF-12/NHR may contribute to the differential expression of genes in the daf-16null mutant. We performed transcription factor binding motif enrichment analysis for DAF-16 (DBE motif; positive-control), SMA-9/co-SMAD (DBD-1 and DBD-2 motifs of the homolog Schnurri), and DAF-12/NHR (M-2 motif) [56–58]. We also performed motif enrichment analysis for PQM-1 (DAE motif), a transcription factor known to be associated with both Class I and Class II genes but not implicated in young larvae [9]. As expected, we found enrichment for the DAF-16 motif (DBE) in the Down genes (p = 9.4x10-6). We also found enrichment for the PQM-1 motif (DAE) in both Up and Down genes (p = 3.1x10-47 and p = 1.0x10-7, respectively), suggesting a functional role of pqm-1 in L1 larvae. We failed to detect enrichment of the DAF-12/NHR motif in either set of genes. However, DAF-12/NHR may act on a small but functionally important set of genes, and DAF-16/FOXO is likely to regulate other genes that affect expression, both of which could limit our ability to detect enrichment of the DAF-12/NHR motif. We found enrichment of the SMA-9/co-SMAD motif (DBD-1) in the Down genes (p = 0.023). SMA-9/SMAD has been shown to act as both a transcriptional activator and repressor [59], and this result suggests it actively represses a significant number of the Down genes. This result suggests that a subset of DAF-16-regulated genes are direct targets of the dbl-1/TGF-β signaling pathway. We used genetic epistasis analysis to determine if the effects of daf-16/FOXO on gene expression identified by mRNA-seq are functionally significant in vivo. Specifically, we hypothesized that up-regulation of daf-12/NHR signaling in the daf-16null mutant contributes to its arrest-defective phenotype. Components of the pathway daf-36/oxygenase, daf-9/CYP450, and daf-12/NHR were epistatic to daf-16/FOXO with respect to M cell division (Fig 3A and 3D). daf-12(m20) is a null allele for all but the B isoform, which is a short isoform containing the ligand-binding domain (LBD) but not the DNA-binding domain (DBD). rh61rh411 is a null allele for all daf-12 isoforms. daf-12(rh273) contains a substitution in the LBD, which is hypothesized to interrupt ligand binding [34]. All three alleles suppressed the daf-16null arrest-defective phenotype. din-1/SHARP collaborates with ligand-free daf-12 as a co-repressor in promoting dauer formation [35]. din-1 was dispensable for the daf-16null arrest-defective phenotype (Fig 3A), suggesting that ligand-bound daf-12/NHR promotes M cell division and that it is normally inhibited by daf-16/FOXO activity during starvation. Pharmacological manipulation of the daf-12/NHR pathway corroborated genetic analysis. Dafadine is a small molecule inhibitor of daf-9 [60]. Wild-type and daf-16null worms were exposed to dafadine or dafachronic acid during starvation. Dafachronic acid did not cause an arrest-defective phenotype, suggesting that the steroid hormone pathway is not sufficient to promote development. Dafadine suppressed M cell lineage divisions in daf-16null worms (Fig 3B), phenocopying daf-9(m540). These results further support the conclusion that daf-16/FOXO activity inhibits daf-12/NHR signaling during L1 arrest, and that this inhibition is functionally significant. Given the mRNA-seq results, we also hypothesized that up-regulation of dbl-1/TGF-β signaling in the daf-16null mutant contributes to its arrest-defective phenotype. Mutations affecting dbl-1/TGF-β and its downstream effector sma-9/co-SMAD suppressed the daf-16null arrest-defective phenotype (Fig 3C and 3D). These results suggest that daf-16/FOXO also inhibits dbl-1/TGF-β signaling during L1 arrest, and that this inhibition has physiological consequences. daf-12/NHR and dbl-1/TGF-β signaling were also required for inappropriate seam cell divisions in starved daf-16null worms. Both daf-36 and dbl-1/TGF-β were epistatic to daf-16/FOXO for seam cell division (Fig 3E and 3F). Overall, these data suggest that daf-16/FOXO activity leads to repression of both Sma/Mab TGF-β and daf-12/NHR signaling during starvation to promote developmental arrest. We hypothesized that daf-12/NHR and dbl-1/TGF-β signaling promote L1 development in fed larvae, since they are required for the daf-16null arrest-defective phenotype, though such early larval function has not been shown. We used Phlh-8::GFP, AJM-1::GFP and molting to monitor the rate of L1 development in fed larvae with mutations in these pathways. dbl-1/TGF-β and sma-9/co-SMAD mutants clearly had delayed M cell lineage divisions (Fig 4A), consistent with our hypothesis. The daf-36 mutant also had significantly fewer M cell divisions on average than wild type (Fig 4B), suggesting that ligand-bound DAF-12 promotes L1 development. daf-12(rh273) and daf-12(rh274), which have a constitutive dauer-formation phenotype due to mutations in the LBD, also caused developmental delay (Fig 4B). These results suggest that ligand-free DAF-12 represses L1 development, analogous to its repression of reproductive development during dauer formation. However, the null alleles of daf-12 had no effect (Fig 4C), which we believe is due to the opposing effects of ligand-bound and -free forms of DAF-12. This interpretation is supported by the fact that these same alleles do not cause a constitutive dauer-formation phenotype [34]. The null allele for din-1, the co-repressor for ligand-free DAF-12, also had no effect on M cell lineage division rate, as expected based on our model and its role in regulation of dauer formation. daf-9(m540) is a suspected hypomorph [30], and residual activity appeared to be sufficient for normal developmental rate (Fig 4C). These results support the conclusion that daf-12/NHR and dbl-1/TGF-β signaling promote M cell lineage division in fed L1 larvae. We assessed seam cell divisions and molting progression in fed larvae to broaden our developmental analysis, examining developmental events that occur before and after M cell division, respectively. daf-36 and dbl-1/TGF-β mutants had significantly fewer seam cell divisions than wild type (Fig 4D). L1 larvae have a stage-specific cuticular ridge (alae) on their lateral midline that is absent from L2 larvae. The presence or absence of alae was scored, along with detached cuticles, to identify L1 or L2 larvae, and larvae undergoing the L1 molt, respectively. The L1 molt was significantly delayed in mutants of the daf-36 and dbl-1 pathways compared to wild type (Fig 4E). These data further support the conclusion that Sma/Mab TGF-β and daf-12/NHR signaling promote L1 development in fed larvae. Does initiation of post-embryonic development in daf-16null cause reduced starvation survival? dbl-1/TGF-β and sma-9/co-SMAD were not epistatic to daf-16/FOXO for starvation survival (Fig 5A; S3 Table). Mutations affecting steroid hormone pathway components daf-36, daf-9, and daf-12/NHR also did not affect starvation survival in a wild-type or daf-16null background (Fig 5B; S3 Table). din-1/SHARP and daf-12/NHR null alleles also had no effect on starvation survival in either background (Fig 5C; S3 Table). Together these results suggest that dbl-1/TGF-β and daf-12/NHR signaling do not affect starvation survival, implying that development during starvation is not the cause of daf-16null starvation sensitivity. This work shows that daf-16/FOXO functions cell-nonautonomously to regulate C. elegans L1 arrest. A recent study corroborates this finding, highlighting the need to determine the molecular basis of cell-nonautonomous function [61]. The dbl-1/TGF-β and daf-12/NHR signaling pathways have not been reported to affect early larval development, but our results reveal that these pathways promote L1 development in fed larvae. mRNA expression analysis together with genetic epistasis analysis suggest that daf-16/FOXO inhibits these pathways to promote developmental arrest. Taken together, this work shows that daf-16/FOXO promotes L1 arrest by inhibiting pathways that promote development (Fig 6). Our results suggest that daf-16/FOXO can function in the intestine, nervous system or epidermis to promote starvation survival and developmental arrest. We analyzed four independent transgenic lines per tissue-specific rescue construct. Apparent allele-specific effects confounded our ability to distinguish tissue-specific effects on rescue. Variation in promoter strength may also obscure tissue-specific effects. These technical complications limit our ability to conclude whether daf-16/FOXO activity in one tissue is functionally more important than another. Though we could not statistically distinguish the effects of rescue in intestine, nervous system or epidermis, we clearly found that muscle rescue had no effect on starvation survival or developmental arrest. Expression in the intestine, nervous system or epidermis was sufficient to partially rescue daf-16null, but rescue was not complete, though it was complete when the native daf-16 promoter was used. We considered that additive effects of multiple tissues might explain this finding, but we did not observe a significant change in penetrance when pairs of promoters were used for rescue. Our starvation survival assay may have lacked sufficient dynamic range to observe this. Alternatively, and particularly with respect to developmental arrest, different components of a single pathway may be regulated by DAF-16/FOXO in different tissues such that regulation in any single tissue is sufficient to disrupt the pathway. Consistent with this interpretation, dbl-1/TGF-β is expressed in neurons while its receptor and downstream SMADs are expressed broadly [40,41,43,62,63]. Likewise, daf-36 is expressed in the intestine, daf-9/CYP450 is expressed in the epidermis and XXX cells, and daf-12/NHR is expressed broadly [29,32,34,64]. We speculate that DAF-16/FOXO functions in several of these tissues to inhibit these genes directly or indirectly. Anatomical sites of action for the insulin/IGF signaling pathway have been characterized for regulation of dauer formation, aging, reproductive aging, developmental arrest and starvation survival. These analyses have produced largely overlapping results, implicating primarily the intestine, nervous system and epidermis. Rescue of daf-2/InsR or age-1/PI3K suggested the nervous system as a key site of action in regulation of aging, but rescue of daf-16/FOXO in a daf-2/InsR mutant background suggested the intestine [18,19]. However, because daf-2 and age-1 antagonize daf-16, these two experimental designs should not be expected to produce the same results. Nonetheless, both studies actually found some effect in each site. In addition, a subsequent study found that epidermal rescue of daf-16/FOXO in a daf-2/InsR mutant background can also rescue lifespan [20]. Curiously, daf-16/FOXO expression in the intestine was originally found not to affect dauer formation while neural expression did [18], but more recently it was reported that daf-16/FOXO expression in the intestine determines dauer formation without contribution from neurons [17]. It is unclear why there are such discrepancies among these results, though specific experimental conditions, promoters, and in some cases limited numbers of transgenic alleles could contribute. Muscle expression of daf-16/FOXO was not found to affect lifespan or dauer formation in any of these studies, but it does affect reproductive aging [65]. Overlapping but discordant results are also evident with respect to regulation of developmental arrest. For example, the daf-16/FOXO target mir-235, which robustly phenocopies daf-16null, was found to function in the epidermis and nervous system to promote developmental arrest in starved L1 larvae [66]. In contrast, transgenic rescue of daf-18/PTEN, which is thought to act through daf-16/FOXO for arrest of somatic cells, highlighted the epidermis as the primary site of action but also showed that it can function cell-autonomously as well [61]. Likewise, rescue of daf-16/FOXO in the context of starvation-induced post-dauer arrest also implicated the epidermis [48]. Though we found the epidermis to be an important site of daf-16/FOXO action in regulation of L1 arrest, our results also suggest that the nervous system and intestine are important. Neurons and intestine also appear to be important sites of action upstream of daf-2/InsR and daf-16/FOXO. That is, the insulin-like peptides ins-4, ins-6, and daf-28 have been shown to be regulators of L1 arrest and dauer formation with expression in chemosensory and motor neurons as well as intestine [17,54]. As the organismal signaling network mediating nutritional control of dauer formation, aging, and developmental arrest is characterized it is increasingly clear that it is complex, being distributed over several different tissues and likely involving feedback and crosstalk at a variety of levels. Our results reveal that dbl-1/TGF-β and daf-12/NHR signaling pathways are each required for the daf-16null arrest-defective phenotype. However, the dbl-1/TGF-β and daf-12/NHR signaling pathways do not affect starvation survival, uncoupling control of development and starvation resistance, as seen with mir-235 [66]. This finding is also consistent with previous work showing no effect of daf-12/NHR on starvation survival [67]. Such uncoupling suggests that daf-16/FOXO acts through different target genes to promote starvation resistance and developmental arrest, and that daf-16/FOXO mutants do not die rapidly during starvation as a result of inappropriate development. It has been suggested that DAF-16 functions as a transcriptional activator [9,68,69], implying that repression of dbl-1/TGF-β and daf-12/NHR signaling is indirect. Our mRNA-seq analysis presumably captures direct and indirect effects, but it identified 2.4 times more genes with decreased expression in daf-16null than with increased expression, consistent with daf-16/FOXO having more effect as an activator than a repressor. However, DAF-16/FOXO directly binds the promoter of daf-12/NHR as well as components of the dbl-1/TGF-β pathway, including the SMADs sma-2 and sma-3 and sma-9/co-SMAD, based on DamID analysis in adults [69]. Likewise, modENCODE data for DAF-16 ChIP-seq in L3-stage larvae suggests direct binding to daf-12/NHR and the upstream cytochrome P450 daf-9 as well as the TGF-β Sma/Mab receptor subunits sma-6 and daf-4, and sma-2 and sma-3 [9,70]. In Drosophila, dFOXO directly binds the promoter of the TGF-β ligand dawdle and the daf-12/NHR homolog dHR96 [71,72]. These observations suggest that daf-16/FOXO regulation could actually be direct. daf-16/FOXO activates the microRNA mir-235 during L1 arrest [66], which could function as an intermediate if regulation is indirect. However, mir-235 is not expressed in and does not function in the intestine, a site of daf-16/FOXO action based on our results. DAF-16/FOXO antagonizes the transcription factor PQM-1, which promotes expression of genes up-regulated in daf-16/FOXO mutants [9], suggesting it too could function as an intermediate. Consistent with this possibility, we found significant enrichment of the PQM-1 binding site motif (DAE) among genes up- and down-regulated in daf-16null, suggesting PQM-1 contributes to gene regulation during L1 arrest. In addition, insulin/IGF signaling and the Rag-TORC1 pathway crosstalk in regulation of L1 arrest [61], providing an additional possibility for indirect effects. It is also important to note that the changes in expression observed for daf-36, daf-12, and dbl-1 in daf-16null were relatively small, suggesting that post-transcriptional mechanisms could also contribute to regulation. FOXO regulates TGF-β and NHR signaling in other contexts. daf-16/FOXO functions upstream of daf-36/oxygenase, daf-9/CYP450, and daf-12/NHR in dauer formation [14,73]. daf-9 also functions downstream of daf-16/FOXO in progression through post-dauer developmental checkpoints, though daf-12/NHR was not implicated [48]. In addition, dbl-1/TGF-β and insulin/IGF signaling both participate in pathogen-associated olfactory learning [10,46]. Regulation of TGF-β signaling by insulin/IGF signaling has not been reported in C. elegans or mammals, but dFOXO represses the TGF-β ligand dawdle in Drosophila, affecting lifespan [71]. These observations suggest that FOXO regulation of TGF-β and steroid hormone signaling pathways is broadly significant and conserved among metazoa. Strains were maintained on agar plates containing standard nematode growth media (NGM) seeded with E. coli OP50 at 20°C. Strains containing the alleles daf-12(rh273) and daf-12(rh274) were maintained at 15°C. The wild-type strain N2 (Bristol) and the following mutants and transgenes were used: daf-2(e979), daf-2(e1370), daf-16(mgDf47), daf-16(mgDf50), daf-16(mu86), ins-4 ins-5 ins-6(hpDf761), daf-28(tm2308), daf-36(k114), daf-9(m540), daf-9(dh6), daf-12(m20), daf-12(rh273), daf-12(rh274), daf-12(rh61rh411), din-1(dh127), dbl-1(wk70), sma-9(wk55), ayIs7[Phlh-8::GFP], syIs78[Pajm-1::AJM-1::GFP], wdIs3[Pdel-1::GFP], qyEx264[Pmyo-3::GFP::DAF-16], qyIs288[Pdaf-16::GFP::DAF-16], qyIs290[Pcol-12::GFP::DAF-16], qyIs292[Pges-1::GFP::DAF-16], qyIs294[Punc-119::GFP::DAF-16], dukEx88–90[Phlh-8::GFP +unc-119(+) + Pcol-12::DAF-16::GFP], dukEx91,92,103[Phlh-8::GFP +unc-119(+) + Pges-1::DAF-16::GFP], dukEx93–95[Phlh-8::GFP +unc-119(+) + Pcol-12::DAF-16::GFP + Pges-1::DAF-16::GFP], dukEx96–98[Phlh-8::GFP +unc-119(+) + Punc-119::DAF-16::GFP + Pges-1::DAF-16::GFP], dukEx99[Phlh-8::GFP +unc-119(+)],dukEx100–102[Phlh-8::GFP +unc-119(+) + Pcol-12::DAF-16::GFP + Punc-119::DAF-16::GFP], dukEx104–106[Phlh-8::GFP +unc-119(+) + Punc-119::DAF-16::GFP], dukEx107–109[Phlh-8::GFP +unc-119(+) + Pdaf-16::DAF-16::GFP], and dukEx110–112[Phlh-8::GFP +unc-119(+) + Pmyo-3::DAF-16::GFP]. Standard genetic techniques were used to create different combinations of alleles. Mixed-stage cultures on 10 cm NGM plates were washed from the plates using S-basal and centrifuged. A hypochlorite solution (7:2:1 ddH2O, sodium hypochlorite (Sigma), 5 M KOH) was added to dissolve the animals. Worms were centrifuged after 1.5–2 minutes in the hypochlorite solution and fresh solution was added. Total time in the hypochlorite solution was 8–10 minutes. Embryos were washed three times in S-basal buffer (including 0.1% ethanol and 5 ng/μL cholesterol) before final suspension in 5 mL S-basal at a density of 1.5 worms/μL. Embryos were cultured in a 16 mm glass tube on a tissue culture roller drum at approximately 25 rpm and 21–22°C. For the M cell division and VB motor neuron differentiation assays during starvation, the larvae were starved for 7 days before 200 larvae per replicate were examined on a slide on a compound fluorescent microscope. For the dafadine and dafachronic acid experiments, 25 μM DMSO, 25 μM dafadine, or 50 nM Δ7-dafachronic acid was added to the final suspension in S-basal. For the seam cell division assay during starvation, the larvae were starved for 3 days and the v1–6 cells on one side of the animal were scored for 60–70 larvae per replicate. The data for wild type (N2) was published previously (GEO accession number GSE33023; "0 hr recovery") [74], and daf-16null data are first reported here. Two biological replicates were performed. Worms were cultured, RNA was prepared and sequencing and analysis were done as described [74]. The only exceptions to this pertain to the software used to count reads aligning to genes. Briefly, liquid culture was used and total RNA was prepared with TRIzol (Invitrogen). Strain GR1307 [daf-16(mgDf50)] was used. mRNA was isolated by polyA-selection. Sequencing libraries were prepared using the Solid Total RNA-seq Kit using the Whole Transcriptome Protocol (Applied Biosystems). Twelve PCR cycles were used to amplify the libraries. Fifty base pair single-end reads were sequenced on the Solid 4 system according to the manufacturer's protocols in the Genome Sequencing Shared Resource at Duke University. Sequencing reads were mapped using Bowtie v0.12.7 allowing two mismatches and requiring unique alignments [75] and genome coordinates for WS210. The script htseq-count [76] was used to count reads mapping to genes defined using WS220 annotation that had been mapped to WS210 coordinates. Reads were counted using the options “-m union–s yes”. DESeq v1.20.0 was used to analyze differential gene expression [77]. Cufflinks v2.1.1 was used to determine fragments per kilobase per million (FPKM; [78]). GO term enrichments were identified using GOrilla [79]. Enrichments were plotted in semantic space using REVIGO [80]. Specific parameters used with GOrilla and REVIGO are documented in S1 Dataset. GEO accession number for the daf-16null dataset is GSE69329. Enrichment for DBE (DAF-16) and DAE (PQM-1) [9], SMAD DBD-1/DBD-2 (SMA-9) [56] and M-2 (DAF-12) [57] motifs were calculated using the AME application of the MEME software package [58]. The primary sequences scanned were the 700 bp upstream of the most upstream translation start sites of the Up and Down genes (Fig 2B) from version WS220 of the C. elegans genome. The background control promoter sequences (N = 2350) was comprised of all genes well detected in the mRNA data (expression in both WT and DAF-16 being in the top 75% percentile) and a having a strong lack of evidence of differential expression between wild-type and the DAF-16 mutant (adjusted p-value > 0.9). For each motif the background model was set to uniform, and the remaining AME parameters were default. NanoString expression analysis was conducted as described with the following exceptions [54]. All five strains were cultured at 15°C since daf-2 mutants are temperature sensitive. Worms were cultured in liquid at 180 rpm and 4–5 worms/μL with 40 mg/ml HB101 as food. N2, daf-16(mgDf47), and daf-16(mgDf47); daf-2(e1370) were bleached after 5 d culture as young gravid adults. daf-2(e1370) and daf-2(e979) were bleached at 6 d and 7 d, respectively, as young gravid adults. For each strain, embryos were divided into three flasks of S-basal at 5 worms/μL and cultured at 15°C, 20°C or 25°C at 180 rpm. Arrested L1 larvae were collected 24 hr after bleaching for 20°C and 25°C, and after 48 hr for 15°C. Larvae were washed, pelleted and flash frozen. Two or three biological replicates were included for each strain. Total RNA was prepared with TRIzol, and 3 μg was used for each hybridization. Data were first normalized by spike-in controls and then normalized by three internal control genes included as targets in the codeset (grld-1, rnf-5 and T16G12.6). Internal controls were identified from genome-wide time-series analysis of fed and starved L1 larvae by virtue of moderate expression levels and invariant expression over time and between conditions [74,81]. RNA was collected from N2 and daf-16(mgDf50) after 1 day of starvation in S-basal. cDNA was synthesized from 1 μg of total RNA using oligonucleotide (dT) primers and SuperScript III Reverse Transcriptase (Thermo Fisher). qPCR was performed with Brilliant II QPCR Master Mix (Agilent) according to manufacturer’s protocol. The genes T16G12.6, rnf-5, and grld-1 were used as internal controls (also used as internal controls for NanoString analysis). Primers were PrimeTime qPCR primers from IDT as follows (in the order of forward, probe, reverse): daf-36, ATCACAGACTCATATTGCCCG, TGTCACGTACTACCCGTCCTCCAA, ACACATTTTCCAGTTTCTGCAC; T16G12.6, CACCACAGACACAAGAACACTA, AACCATACGGGACATCAGCCCTTG, CGGCCAAATTGAAGCGAATC; rnf-5, AACCACCACCGCAATCAT, ATGCACATTTGGTCCGCCGC, TCAACGGGAACAGACCATTC; grld-1, AAGCTGCAGGCGTTGTAA, TGGGAAGATGTAGAGAATGCCGCC, AAGAGCTCCGAGCAAGAATG. Three technical replicates were performed for each of three biological replicates. Standard curves were analyzed to determine reaction efficiency. daf-36 was normalized based on the internal controls and reaction efficiencies to calculate a fold-change between N2 and daf-16 null. An unpaired t-test was performed to determine significance. Following hypochlorite treatment, cultures were synchronized by overnight passage in L1 arrest at a density of 1.5 worms/μL in 5 mL S-basal. For recovery and development, 2,000 arrested L1 larvae were plated per NGM OP50 plate and placed at 20°C for 6 hours (AJM-1 marker assay), 12 hours (M cell marker assay), or 18 hours (molting progression assay). Larvae were then washed off the plate with S-basal, centrifuged, and mounted on an agarose pad. A compound fluorescent microscope was used to score cell divisions or cuticle structure in 200 worms (M cell assay) or 75 worms (seam cell and molting progression assays) per replicate. Animals were treated in hypochlorite solution and suspended in S-basal as described above. 100 μL aliquots were sampled on different days and placed around the edge of an OP50 lawn on NGM plates. Number of plated worms (Tp) was counted and the plates were incubated at 20°C. After two days the number of animals that survived (Ts) was counted. Survival was calculated as Ts/Tp. The daf-16 tissue-specific starvation survival experiments were done with an alternative protocol. Instead of S-basal, virgin S-basal (no ethanol or cholesterol) was used for both hypochlorite treatment and final suspension. Worms were suspended at 1 worm/μL. For the integrated strains, 100 μL aliquots were sampled every two days and spontaneous movement in liquid was used to score survival. For the strains with an extrachromosomal array, the culture was centrifuged after seven days and the pelleted worms were placed on NGM plates. Worms expressing GFP were then scored as alive or dead based on spontaneous movement. Data were handled in R and Excel. Graphs were plotted in the R package ggplot2 or Excel. Statistical tests were performed in R or Excel. Starvation survival analysis was performed on 50% survival times (thalf), which were obtained by fitting survival data for each trial with the function S=100−1001+e(thalf−t)/rate which we have modified slightly from [82]. Goodness of fit is reported as R2 in S1 and S3 Tables.

10.1371/journal.pgen.1006791

USP9X counteracts differential ubiquitination of NPHP5 by MARCH7 and BBS11 to regulate ciliogenesis

Ciliogenesis is a fundamental biological process central to human health. Precisely how this process is coordinated with the cell cycle remains an open question. We report that nephrocystin-5 (NPHP5/IQCB1), a positive regulator of ciliogenesis, is a stable and low turnover protein subjected to cycles of ubiquitination and deubiquitination. NPHP5 directly binds to a deubiquitinating enzyme USP9X/FAM and two E3 ubiquitin ligases BBS11/TRIM32 and MARCH7/axotrophin. NPHP5 undergoes K63 ubiquitination in a cell cycle dependent manner and K48/K63 ubiquitination upon USP9X depletion or inhibition. In the G0/G1/S phase, a pool of cytoplasmic USP9X recruited to the centrosome by NPHP5 protects NPHP5 from ubiquitination, thus favouring cilia assembly. In the G2/M phase, USP9X dissociation from the centrosome allows BBS11 to K63 ubiquitinate NPHP5 which triggers protein delocalization and loss of cilia. BBS11 is a resident centrosomal protein, whereas cytoplasmic USP9X sequesters the majority of MARCH7 away from the centrosome during interphase. Depletion or inhibition of USP9X leads to an accumulation of centrosomal MARCH7 which K48 ubiquitinates NPHP5, triggering protein degradation and cilia loss. At the same time, BBS11 K63 ubiquitinates NPHP5. Our data suggest that dynamic ubiquitination and deubiquitination of NPHP5 plays a crucial role in the regulation of ciliogenesis.

Centrosomes are non-membrane bound organelles that modulate a variety of cellular processes including cell division and formation of hair-like protrusions called primary cilia. Primary cilia function as cellular antennae to sense a wide variety of signals important for growth, development and differentiation. Defects in cilia formation or ciliogenesis can give rise to a bewildering array of human ciliary diseases collectively known as ciliopathies. Ciliogenesis is controlled in part by nephrocystin-5 (NPHP5/IQCB1), and NPHP5 dysfunction causes ciliopathies in humans, mice and dogs. We are interested in studying how the stability, localization and biological activity of NPHP5 are regulated at the molecular level. We present here that NPHP5 directly interacts with, and is a substrate of, one deubiquitinase (USP9X/FAM) and two ubiquitin ligases (BBS11/TRIM32 and MARCH7/axotrophin), enzymes involved in controlling protein stability, localization and activity. Our results suggest that timely ubiquitination and deubiquitination of NPHP5 is critical for the regulation of ciliogenesis.

Primary cilia, microtubule-based protrusions found on the surface of most eukaryotic cells, are derived from centrosomes and possess sensory function such as chemosensation and mechanosensation[1,2]. Formation of primary cilia is tightly regulated during the cell cycle: they assemble primarily during the G0 phase and undergo complete disassembly prior to entry into mitosis[3]. Defects in cilia formation (ciliogenesis) or function can give rise to a myriad of human genetic disorders collectively known as ciliopathies that are often pleiotropic, exhibiting clinical manifestations such as retinal degeneration, renal failure and neurological disorders[4]. In addition, cilia regulate several signalling pathways commonly perturbed in cancer and a loss of cilia is known to occur early in the development of several human cancers[5,6,7,8,9,10,11,12,13,14]. Although several hundred proteins are required for ciliogenesis[15,16,17], a critical step towards understanding their role in health and disease is to delineate their precise spatial and temporal regulation. Ciliogenesis is controlled in part by nephrocystin-5 (NPHP5/IQCB1). NPHP5 was originally identified as the causative gene of two human ciliopathies, Senior-Løken syndrome and Leber congenital amaurosis, typified by retinal degeneration with or without renal failure[18,19,20]. Murine and canine models of NPHP5 develop retinal degeneration[21,22]. NPHP5 might also be involved in tumorigenesis since its mRNA expression is up-regulated in gastrointestinal cancer[23]. We are others have shown that NPHP5 and its interacting partner Cep290 are essential for ciliogenesis[24,25]. Pathogenic mutations of NPHP5 lead to truncated products that become mis-localized and are unable to interact with Cep290[25]. NPHP5 localizes to the centrosome including the ciliary base during interphase[24,25] but disappears from the organelle during mitosis for reasons that are not understood[25]. Moreover, exactly how the stability or activity of this protein is controlled at the molecular level has not been studied. Ubiquitination is a post-translational modification crucial for controlling protein stability, localization and activity[26]. It is a multi-step process in which ubiquitin (Ub) is transferred onto a substrate via the action of three enzymes: an Ub-activating enzyme E1, an Ub-conjugating enzyme E2 and an Ub ligase E3 which is mainly responsible for substrate recognition. In humans, several hundred E3 ligases exist and they are grouped into three families based on the presence of characteristic domains and the mechanism of ubiquitin transfer[27]. A given substrate can be monoubiquitinated, multi-monoubiquitinated or polyubiquitinated. The most common types of polyubiquitination are the K48-linkage which targets a substrate for proteasomal degradation, and the K63-linkage which has non-proteasomal function. Substrate ubiquitination by E3 ligases can be reversed by the ~100 or so deubiquitinating enzymes or deubiquitinases (DUBs) that are divided into five families[28,29]. E3 ligases often work in concert with DUBs to control the ubiquitination status of a substrate, and deregulation of these enzymes is known to cause human disorders and cancer[30,31]. It is currently unknown if NPHP5 associates with E3 ligases and/or DUBs and undergoes ubiquitination and/or deubiquitination. In an effort to identify factors that regulate NPHP5 and hence ciliogenesis, we performed a proteomic screen for components of NPHP5-containining complexes. Three enzymes involved in substrate ubiquitination and deubiquitination[32,33,34], including one DUB, USP9X/FAM, and two E3 ligases, MARCH7/axotrophin and BBS11/TRIM32, were found (Fig 1A and S1 Table). We first explored whether NPHP5 can undergo ubiquitination by conducting in vivo ubiquitination assays. Polyubiquitination of NPHP5 was detected upon immunoprecipitating endogenous protein from HEK293 cells treated with a pan DUB inhibitor N-ethylmaleimide (NEM[35]) (Fig 1B). When Flag-NPHP5 was immunoprecipitated under denaturing conditions with 1% SDS to prevent co-immunoprecipitation of interacting proteins such as Cep290 and Usp9x (S1A Fig), the recombinant protein remained highly polyubiquitinated in the presence of NEM (S1B Fig), suggesting that NPHP5 itself undergoes ubiquitination. Polyubiquitination of NPHP5 could be observed without NEM when immunoprecipitations were performed from lysates of HEK293 cells expressing recombinant Flag-NPHP5 and HA-Ub (Fig 1C and S2A Fig). Of note, NPHP5 polyubiquitination occurred in multiple cell lines including HEK293, U2OS and normal diploid retinal pigmented epithelial cells (RPE-1), a well-established model for primary cilia formation (Fig 1C and S2A–S2C Fig). To identify ubiquitination sites on NPHP5 and characterize Ub chain linkages, ubiquitinated NPHP5 species purified from asynchronous cell extracts treated with NEM were analyzed by mass spectrometry. A double glycine was attached to several lysine residues of NPHP5, including K19, K31/K33, K205, K263, K396 and K528, suggesting that these residues are potential ubiquitination sites (Fig 1D). Indeed, a NPHP5 mutant in which six of these lysine residues were mutated to arginines (K19RK31RK33RK205RK263RK528R or hexamutant) exhibited diminished levels of ubiquitination compared to wild type (Fig 1E). In addition, about 80% and 20% of double glycines were linked to K48 and K63 residues of Ub, respectively (Fig 1F), indicating that NPHP5 is conjugated to both K48- and K63-Ub chains. To validate this finding, we observed robust ubiquitination of NPHP5 in the presence of NEM when this protein was co-expressed with wild type Ub or K63R mutant Ub, but not with K48RK63R Ub or KallR Ub in which all lysine residues were mutated to arginine (Fig 1G). The use of K48R mutant Ub did not completely abolish NPHP5 ubiquitination (Fig 1G), suggesting that a fraction of NPHP5 is K63-ubiquitinated. Taken together, these data indicate that NPHP5 is ubiquitinated at multiple sites and undergoes K48 and K63 ubiquitination upon DUB inhibition. Given that NPHP5 is ubiquitinated in cells and that a relatively large number of unique peptides corresponding to USP9X were identified in our proteomic screen (Fig 1A), we focused on verifying the interaction between NPHP5 and USP9X. We expressed Flag-NPHP5 in HEK293 cells, performed anti-Flag immunoprecipitations, and showed that recombinant NPHP5 and endogenous USP9X are co-precipitated (Fig 2A). Endogenous NPHP5 was detected in USP9X immunoprecipitates in HEK293 cells (Fig 2B) and similarly, endogenous USP9X was co-immunoprecipitated with an NPHP5 antibody against the C-terminus of the protein in HEK293, U2OS and RPE-1 cells (Fig 2B and see below). In sharp contrast, an NPHP5 antibody against the N-terminus of the protein did not efficiently immunoprecipitate USP9X (Fig 2B), suggesting that this antibody might compete with USP9X for binding to NPHP5. Indeed, mapping studies using a series of Flag-tagged NPHP5 truncates revealed that the N-terminal region spanning residues 1–19 is critical for USP9X binding (Fig 2C and 2D). The interaction between NPHP5 and USP9X is probably direct since purified NPHP5 bound to USP9X in vitro (Fig 2E). Both wild type and catalytically inactive mutant USP9X (C1566S) could bind to NPHP5 in cells and in vitro (Fig 2E and S3 Fig). To determine if NPHP5 is a substrate of USP9X, we performed in vitro deubiquitination experiments by incubating ubiquitinated NPHP5 with purified USP9X. Only wild type USP9X could robustly deubiquitinate NPHP5 (Fig 3A). Subsequently, we examined the effects of manipulating USP9X levels in cells, and showed that depletion of USP9X induces NPHP5 ubiquitination (Fig 3B), whereas expression of wild type but not mutant USP9X suppresses ubiquitination (Fig 3C). These results indicate that USP9X directly binds to, and catalyzes the deubiquitination of, NPHP5. Next, we sought to characterize the nature and impact of NPHP5 ubiquitination. NPHP5 underwent K48 and K63 ubiquitination upon USP9X depletion (Fig 3D), akin to our earlier findings with DUB inhibition (Fig 1F and 1G). Depletion of USP9X significantly reduces the intensity of NPHP5 immunofluorescence at the centrosome (Fig 4A and S4 Fig) and protein levels, endogenous (Fig 4B and S5A Fig) or recombinant (S5B Fig), without affecting mRNA levels (S5C and S5D Fig). The decrease in NPHP5 protein levels could be rescued by a proteasome inhibitor MG132 (Fig 4C and S5E Fig) and enhanced by a protein synthesis inhibitor cycloheximide (Fig 4D and S5F Fig). Conversely, over-expression of wild type but not mutant USP9X elevated the protein levels of NPHP5 (Fig 4E). Ablation of USP9X inhibited the formation of cilia in cycling and quiescent RPE-1 cells (Fig 4F and 4H), a phenotype reminiscent of NPHP5 loss, without interfering with cell cycle progression (Fig 4I), cell proliferation (Fig 4J) or cell cycle exit (Fig 4K). Interestingly, cilia formation in USP9X-depleted cells could be restored by enforced expression of NPHP5 (Fig 4L). These observations together suggest that depletion/inhibition of USP9X sensitizes NPHP5 for K48 and K63 ubiquitination and proteasomal degradation, thereby compromising ciliogenesis. To further explore the connection between NPHP5 and USP9X, we examined the consequences of ablating NPHP5 on USP9X protein levels and localization. Loss of NPHP5 did not impinge on USP9X levels (Fig 5A). Previously, it was shown that USP9X localizes predominately to the cytoplasm (where it controls the localization/activity of MARCH7[36,37,38]), but its targeting to the centrosome has not been documented. We confirmed that USP9X is indeed mostly cytoplasmic; nevertheless, this protein is also detected at the centrosome where it co-localizes with a centriolar marker centrin (Figs 5B and 6C). Depletion of NPHP5 specifically disrupted the centrosomal localization of USP9X (Fig 5B), suggesting that a pool of cytoplasmic USP9X is recruited to the centrosome where it might serve to deubiquitinate NPHP5. To better understand the role and requirements of USP9X in regulating NPHP5 ubiquitination, we examined the protein levels and localization of NPHP5 and USP9X, along with the interaction between the two proteins, across the cell cycle. Although NPHP5 becomes K48 and K63 ubiquitinated when USP9X is lost or inhibited, we found that endogenous NPHP5 levels remain relatively constant throughout the cell cycle (Fig 6A and 6B). Neither cycloheximide nor MG132 dramatically destabilized or stabilized NPHP5 (control lanes; Fig 4C and 4D), suggesting that this protein is highly stable and has a low turnover rate when USP9X is not perturbed. In terms of localization, NPHP5 was highly enriched at the centrosome during interphase (S6 Fig). Its centrosomal staining became fuzzy in late G2 and disappeared in mitosis (S6 Fig), in agreement with our previous study[25]. Thus, NPHP5 is delocalized but is not down-regulated in mitosis. For USP9X, we found that while its levels remain the same throughout the cell cycle (Fig 6A and 6B), the protein localizes to the centrosome in G0, G1 and S only (Fig 6C). The absence of centrosomal USP9X staining in G2 and M (Fig 6C) suggests that it is lost from the organelle at a much earlier point in the cell cycle than NPHP5. Consistent with this, the interaction between NPHP5 and USP9X was strong in G0, G1 and S and became substantially weaker in G2 and M (Fig 6A and 6B). Collectively, our data support the idea that USP9X associates with NPHP5 at the centrosome for most of the cell cycle. This enzyme then dissociates from the centrosome in the G2 and M phase, concomitant with its decreased affinity for NPHP5. During mitosis, NPHP5 also becomes delocalized. Because USP9X deubiquitinates NPHP5 (Fig 3), we hypothesize that NPHP5 is prone to ubiquitination when its association with USP9X is compromised in G2/M. The levels of NPHP5 ubiquitination were indeed elevated in mitosis (Fig 6D), and this increase was specifically attributed to the attachment of K63-Ub, but not K48-Ub chains, to NPHP5 (Fig 6E). Thus, our data suggest that K63 ubiquitination of NPHP5 is likely responsible for its delocalization from the centrosome in the mitotic phase of the cell cycle. To further prove that NPHP5 is prone to K63 ubiquitination when it is not binding to USP9X, we conducted a serious of experiments with 20–598, a NPHP5 mutant refractory to USP9X binding (Fig 2C and 2D). Curiously, 20–598 was more highly ubiquitinated compared to wild type NPHP5 (S7A Fig) and underwent predominantly K63 ubiquitination (S7A Fig). Despite possessing an intact centrosome localization domain[25] and expressing at levels similar to wild type NPHP5, 20–598 was poorly targeted to the centrosome (S7B and S7C Fig). Weak centrosomal staining of the mutant protein was seen in ~35% of transfected cells, while no centrosomal staining was observed in the remaining ~65%. We have thus far demonstrated that NPHP5 becomes ubiquitinated when it is not protected by USP9X: depletion/inhibition of USP9X triggers K48/K63 ubiquitination of NPHP5 and protein degradation/mis-localization, whereas impaired binding to USP9X provokes K63 ubiquitination and protein mis-localization. To reconcile these observations and to identify enzymes responsible for the differential ubiquitination of NPHP5, we examined the connection between NPHP5 and the two E3 ligases, MARCH7 and BBS11 (Fig 1A). We first validated the interaction between NPHP5 and MARCH7 (Fig 7A). This interaction is likely direct since the two proteins bound to each other in vitro (S8A Fig). Unlike the catalytically inactive mutant[32], wild type MARCH7 was able to trigger NPHP5 ubiquitination (Fig 7B). More specifically, MARCH7 induced K48 but not K63 ubiquitination of NPHP5 (Fig 7B), resulting in decreased protein levels (Fig 7C). Conversely, depletion of MARCH7 elevated NPHP5 protein levels (Fig 7D). Enforced expression of MARCH7 reduced the fluorescence intensity of NPHP5 at the centrosome (Fig 7E) and suppressed cilia formation in cycling and quiescent RPE-1 cells (Fig 7F and 7G). Although MARCH7 is predominately nuclear and cytoplasmic[32,39], a weak endogenous signal could be detected at the centrosome in a small percentage of interphase cells (Fig 7H and S8B Fig). This minute amount of MARCH7 was not sufficient to induce substantial K48 ubiquitination of NPHP5 (Fig 6E). During mitosis, a strong centrosomal MARCH7 signal could be seen in most cells (Fig 7H and S8B Fig). Remarkably, we found that depletion of USP9X leads to an accumulation of MARCH7 at the centrosome in the majority of cells irrespective of the cell cycle (Fig 7I and S8C Fig). The interaction between MARCH7 and NPHP5 became stronger upon USP9X depletion (Fig 7J). Our results suggest that MARCH7 has subtle effects on NPHP5 under normal, unperturbed conditions. Upon USP9X depletion or inhibition, however, NPHP5 is exposed to E3 ligases. MARCH7 becomes mis-targeted to the centrosome where it adds K48-Ub chains to NPHP5 (Figs 1F, 1G and 3D). Meanwhile, another enzyme might be responsible for putting K63-Ub chains onto NPHP5 (Figs 1F, 1G and 3D). Next, we validated the interaction between NPHP5 and BBS11 (Fig 8A), which was reported to be weak [40], and showed that these proteins directly interact in vitro (S9A Fig). Wild type but not catalytically inactive mutant BBS11[41] induced NPHP5 ubiquitination (Fig 8B). In striking contrast to MARCH7, BBS11 specifically provoked K63 rather than K48 ubiquitination of NPHP5 (Fig 8B). Endogenous BBS11 localized to the centrosome in all cell cycle phases examined (S9B Fig), suggesting that it might have the capacity to ubiquitinate NPHP5 throughout the cell cycle. Consistent with the notion that BBS11 K63-ubiquitinates NPHP5, neither ectopic expression nor depletion of BBS11 impinged on the protein levels of NPHP5 (Fig 8C and 8D). Ectopic expression of BBS11 triggered the delocalization of NPHP5 (Fig 8E) and suppressed cilia formation in cycling and quiescent cells (Fig 8F and 8G). Depletion of BBS11 enhanced the localization of the 20–598 mutant of NPHP5 to the centrosome (S7D Fig) and, unlike depletion of MARCH7, caused a dramatic accumulation of centrosomal NPHP5 in late G2 cells (S10 Fig). Furthermore, depletion of BBS11 induced late G2 arrest/delay (Fig 8H), in agreement with results reported earlier[42], and this phenotype could be reversed by co-depletion of NPHP5 (Fig 8H). In a subset of BBS11-depleted cells that have managed to slip into mitosis, persistent centrosomal NPHP5 staining could be observed (S9C Fig). Taken together, our data suggest that BBS11 catalyzes the addition of K63-Ub chains to NPHP5, leading to its delocalization from the centrosome without affecting protein levels. Furthermore, timely delocalization of NPHP5 appears to be crucial for proper mitotic entry. On the basis of our findings that NPHP5 is deubiquitinated by USP9X, K48-ubiquitinated by MARCH7 and K63-ubiquitinated by BBS11, we asked whether simultaneous ablation of MARCH7 and/or BBS11 might override the effects of USP9X loss on NPHP5. Depletion of USP9X induced down-regulation and mis-localization of NPHP5 (Fig 4A–4D and S4, S5A, S5E, S5F, S11A and S11B Figs). Co-depletion of MARCH7 reinstated protein levels but not the centrosomal localization of NPHP5 (S11A and S11B Fig), whereas co-depletion of BBS11 could barely restore localization, presumably because protein levels remained low (S11A and S11B Fig). Co-depletion of MARCH7 and BBS11, however, drastically restored the centrosomal localization and protein levels of NPHP5 (S11A and S11B Fig). We have thus established a functional connection between USP9X, MARCH7, BBS11 and NPHP5. We present here the spatial and temporal regulation of NPHP5. We show that there is a significant correlation between the sub-cellular localization of NPHP5, the biological activity of NPHP5, and the capacity of cells to possess cilia. Our data suggest that at G0/G1/S when cilia are present, NPHP5 directly recruits a fraction of cytoplasmic USP9X to the centrosome which in turn protects NPHP5 from ubiquitination (Fig 9). In the G2/M phase when cilia formation is not favourable, USP9X dissociates from the centrosome, making NPHP5 vulnerable to ubiquitination. We identify two E3 ligases, MARCH7 and BBS11 that exhibit distinct localization patterns and exert different effects on NPHP5. In contrast to MARCH7 which is mostly a nuclear/cytoplasmic protein, BBS11 is present at the centrosome throughout the cell cycle. We believe that BBS11 is the enzyme responsible for K63-ubiquitinating NPHP5 and triggering its delocalization at late G2/M (Fig 9). The molecular cause of USP9X dissociation from NPHP5 is currently under investigation. Why is it necessary for NPHP5 to move out of the centrosome in mitosis? NPHP5 is known to be a positive regulator of ciliogenesis[24,25]. Given that the presence of cilia is incompatible with spindle formation and/or function[43], failure to delocalize NPHP5 might interfere with the removal of downstream ciliogenesis events or cilia disassembly and affect cell cycle progression. In support of this idea, our studies show that a loss of BBS11 causes an accumulation of NPHP5 at the centrosome at late G2, which in turn compromises mitotic entry. We hereby propose that the delocalization of NPHP5 triggered by BBS11 represents a novel mechanism that acts in concert with the recently characterized Nek-Kif24 pathway[44] to favour cilia disassembly at G2/M. The localization/activity of MARCH7 is controlled by two different DUBs, USP7 in the nucleus and USP9X in the cytoplasm[32]. In particular, it is known that USP9X, which is predominantly cytoplasmic, interacts with and sequesters MARCH7 in this compartment (Fig 9 and [32]). Although we demonstrate that NPHP5 is a substrate of MARCH7, these two proteins are mostly found in distinct sub-cellular compartments. Indeed, MARCH7 is barely detectable at the centrosome during interphase. MARCH7 does become highly enriched at the centrosome during mitosis, but NPHP5 becomes K63 ubiquitinated by BBS11 and is delocalized at this juncture. Thus, under normal and unperturbed conditions, MARCH7 does not K48-ubiquitinate NPHP5 to a significant degree (Fig 9), which could explain the observed stability of NPHP5. A priori, MARCH7 could contribute to the low turnover of NPHP5 by targeting a small fraction of this protein for degradation. When USP9X is depleted or inhibited, however, the localization/activity of MARCH7 is dramatically altered. MARCH7 becomes aberrantly translocated to the centrosome where both MARCH7 and BBS11 can now ubiquitinate NPHP5 (Fig 9). MARCH7 induces K48 ubiquitination of NPHP5 and protein degradation, while BBS11 triggers K63 ubiquitination and protein delocalization. Previously, it has been reported that USP9X localizes to cilia and that depletion of USP9X does not impair ciliogenesis in human fibroblasts[45]. At first glance, these observations appear to contradict our findings. We speculate that these discrepancies could be attributed to the knockdown efficiency and the use of different cell lines (fibroblasts versus RPE-1 cells; see [32]) which give different USP9X sub-cellular localization patterns. In our hands, depletion of USP9X in RPE-1 cells was extremely robust. In summary, the ubiquitination status and hence the stability, localization and biological activity of NPHP5 is controlled by two E3 ligases (MARCH7 and BBS11) and a DUB (USP9X). These studies raise the intriguing possibility that targeting certain DUBs or E3 ligases might represent a novel strategy to manipulate cilia assembly pathways and treat cilia-related diseases. HEK293, U2OS and hTERT RPE-1 cells were grown in DMEM supplemented with 10% FBS at 37°C in a humidified 5% CO2 atmosphere. Plasmids expressing GFP-NPHP5, Flag-NPHP5 and Flag-NPHP5 truncates were described previously[25]. Mouse USP9X (S. Wood), mouse USP9X C1566S (S. Wood) and human BBS11 cDNAs were sub-cloned into mammalian vector pEGFP-C1 to generate pEGFP-C1-USP9X, pEGFP-C1-USP9X(C1566S) and pEGFP-C1-BBS11 respectively. Mutations in BBS11 (C20AC39AH41A) and NPHP5 (K19RK31RK33RK205RK263RK528R and 20–598) were introduced into full-length cDNAs by PCR mutagenesis and sub-cloned into pEGFP-C1, pEGFP-N1 or pCBF-Flag. All constructs were verified by DNA sequencing. Plasmids expressing the following proteins were obtained or purchased: HA-Ub (J. Archambault); Flag-Ub, Flag-Ub KallR, Flag-Ub K48R, Flag-Ub K63R and Flag-Ub K48RK63R (Division of Signal Transduction Therapy, University of Dundee); Myc-MARCH7 and Myc-MARCH7WI (P. Lehner). Antibodies used in this study included anti-HA, anti-NPHP5, anti-Myc, anti-MARCH7, anti-γ tubulin-FITC (Santa Cruz); anti-Flag, anti-GFP, anti-β actin, (Sigma); anti-CP110, anti-USP9X, anti-CEP290 (Bethyl Laboratories); anti-BBS11 (Thermo Fisher), anti-centrin, anti-K48 Ub (Millipore); anti-K63 Ub (Abcam); anti-Ub (Dako); anti-IFT88 (Proteintech); anti-glutamylated tubulin GT335, anti-acetylated tubulin and anti-Ki67 (Invitrogen). Two additional anti-NPHP5 antibodies (IRCM1 and IRCM2) were also used[25]. To identify NPHP5-interacting proteins, HEK293 cells expressing Flag-NPHP5 were lysed in a buffer containing 50 mM HEPES/pH 7.4, 250 mM NaCl, 5 mM EDTA/pH 8, 0.1% NP-40, 1 mM DTT, 0.2 mM AEBSF, 2 μg/ml leupeptin, 2 μg aprotinin, 10 mM NaF, 50 mM β-glycerophosphate and 10% glycerol for 30 minutes at 4°C. Lysates were immunoprecipitated with anti-Flag agarose beads (Sigma) for 2 hours at 4°C. Bounded proteins were eluted with Flag peptide for 30 minutes, and the resultant eluates were precipitated with trichloroacetic acid and fractionated by sodium dodecyl sulfate–polyacrylamide gel electrophoresis. Gel slices containing polypeptides were excised after Coomassie staining and subjected to proteolytic digestion mass spectrometric analysis. To identify ubiquitination sites on, and Ub chains linked to, NPHP5, the procedure was similar except that 5 mM NEM (Bioshop) was added to the lysis buffer. After immunoprecipitation, beads were washed three times with 500 mM NaCl to disrupt interacting proteins from co-immunoprecipitating with Flag-NPHP5. Gel slices containing polypeptides ≥75 kDa were excised for mass spectrometric analysis. Analyses were performed at the IRCM mass spectrometry core facility by micro-capillary liquid chromatography-mass spectrometry/mass spectrometry. LC-MS peptide quantitation was performed by manual integration of the extracted ion chromatogram of each peptide using Qual Browser/Xcalibur version 2.2 (ThermoFisher Scientific, Waltham, MA, USA). Immunoprecipitation, immunoblotting and immunofluorescence were performed as described previously[40,46]. To obtain cells synchronized in G1, S, G2 and M phases, cells were treated with 0.4 mM mimosine for 24 hours, 2 mM HU for 24 hours and released for 4 hours, 2 mM HU for 24 hours and released for 9 hours or 15 μM RO 3306 for 24 hours, and 40 ng/ml nocodazole for 24 hours, respectively. For RPE-1, cells grown in the presence of serum (cycling) or brought to quiescence by serum starvation for 48–72 hours (quiescent) were stained with Ki67, a cellular marker for proliferation. A high percentage of Ki67-positive cells is indicative of cell proliferation, whereas a low percentage is indicative of cell cycle exit. Cell cycle distribution was performed by fluorescence-activated cell sorting as described previously[25]. To prevent proteasomal degradation and protein synthesis, cells were treated with 10 μM MG132 and 10 mg/ml cycloheximide, respectively, for a maximum of 8 hours. Transfection of siRNA was performed using siImporter (Millipore) per manufacturer's instructions. Target siRNA sequences were: non-specific (NS) control: 5′-AATTCTCCGAACGTGTCACGT-3′; USP9X-2: 5′-ACACGATGCTTTAGAATTT-3′; USP9X-3: 5′-GTACGACGATGTATTCTCA-3′; NPHP5: 5′-ACCCAAGGATCTTATCTAT-3′. USP9X-2 oligo was used unless stated otherwise. ON-TARGETplus SMARTpool for MARCH7 and BBS11 were used. All siRNAs were purchased from Dharmacon. Total RNA was prepared from HEK293 cell culture using Trizol reagent (Invitrogen). Messenger RNA was reverse-transcribed to cDNA (42°C for 1 hour, 50°C for 1 hour and 90°C for 10 minutes) using random hexamers and Superscript II reverse transcriptase (Applied Biosystems, Carlsbad, USA). A negative control RT- reaction was carried out to establish that the target RNA was not contaminated with DNA. The cDNA product was used as a template for subsequent PCR amplification. Target sequences for NPHP5 were: 5′-GCTTACACAGATATTAAGC-3′, 3′-TCATCTTTTTCTTCAGCCTTA-5′; and for β actin were: 5′-AGAGCTACGTGCCTGAC-3′, 3′-AGCACTGTGTTGGCGTACAG-5′. β actin was used as a control. Flag-NPHP5 or GFP-USP9X expressed in HEK293 cells was immunoprecipitated with anti-Flag or anti-GFP antibody. To obtain poly (HA)-ubiquitinated NPHP5, Flag-NPHP5 co-expressed with HA-Ub in HEK293 cells was immunoprecipitated with anti-Flag antibody in the presence of 5 mM NEM to increase the amount of polyubiquitinated products. To obtain BBS11 or MARCH7, endogenous protein was immunoprecipitated from HEK293 cells with anti-BBS11 or anti-MARCH7 antibody. Beads were washed three times with lysis buffer containing 500 mM NaCl to remove interacting proteins from co-immunoprecipitating with the desired protein. For Flag-NPHP5 and poly (HA)-ubiquitinated NPHP5, proteins were eluted with Flag peptide and collected through poly-prep chromatography columns (BioRad). In vivo ubiquitination assays were performed as described previously[47]. 1 μg of purified Flag-NPHP5 was mixed with 1 μg of purified USP9X, BBS11 or MARCH7 bound to beads at 4°C for 2 hours. After extensive washing, bound proteins were analyzed by sodium dodecyl sulfate–polyacrylamide gel electrophoresis and immunoblotting. 1 μg purified, poly (HA)-ubiquitinated NPHP5 was incubated with 1 μg of purified USP9X bound to beads in 20 μl of reaction buffer (50 mM HEPES-KOH/pH 8, 150 mM KCl, 5% glycerol, 0.01% Triton X-100, and 2 mM DTT) for 2 hours at room temperature. The reaction was terminated by the addition of sodium dodecyl sulfate sample buffer and analyzed by gel electrophoresis and immunoblotting. Graphs were generated in Microsoft Excel. Error bars represent standard error of the mean. For Western blot quantitation, densitometric analyses were performed with ImageJ.

10.1371/journal.pcbi.1003101

PUMA: A Unified Framework for Penalized Multiple Regression Analysis of GWAS Data

Penalized Multiple Regression (PMR) can be used to discover novel disease associations in GWAS datasets. In practice, proposed PMR methods have not been able to identify well-supported associations in GWAS that are undetectable by standard association tests and thus these methods are not widely applied. Here, we present a combined algorithmic and heuristic framework for PUMA (Penalized Unified Multiple-locus Association) analysis that solves the problems of previously proposed methods including computational speed, poor performance on genome-scale simulated data, and identification of too many associations for real data to be biologically plausible. The framework includes a new minorize-maximization (MM) algorithm for generalized linear models (GLM) combined with heuristic model selection and testing methods for identification of robust associations. The PUMA framework implements the penalized maximum likelihood penalties previously proposed for GWAS analysis (i.e. Lasso, Adaptive Lasso, NEG, MCP), as well as a penalty that has not been previously applied to GWAS (i.e. LOG). Using simulations that closely mirror real GWAS data, we show that our framework has high performance and reliably increases power to detect weak associations, while existing PMR methods can perform worse than single marker testing in overall performance. To demonstrate the empirical value of PUMA, we analyzed GWAS data for type 1 diabetes, Crohns's disease, and rheumatoid arthritis, three autoimmune diseases from the original Wellcome Trust Case Control Consortium. Our analysis replicates known associations for these diseases and we discover novel etiologically relevant susceptibility loci that are invisible to standard single marker tests, including six novel associations implicating genes involved in pancreatic function, insulin pathways and immune-cell function in type 1 diabetes; three novel associations implicating genes in pro- and anti-inflammatory pathways in Crohn's disease; and one novel association implicating a gene involved in apoptosis pathways in rheumatoid arthritis. We provide software for applying our PUMA analysis framework.

Genome-wide association studies (GWAS) have identified hundreds of regions of the human genome that are associated with susceptibility to common diseases. Yet many lines of evidence indicate that many susceptibility loci, which cannot be detected by standard statistical methods, remain to be discovered. We have developed PUMA, a framework for applying a family of penalized regression methods that simultaneously consider multiple susceptibility loci in the same statistical model. We demonstrate through simulations that our framework has increased power to detect weak associations compared to both standard GWAS analysis methods and previous applications of penalized methods. We applied PUMA to identify novel susceptibility loci for type 1 diabetes, Crohn's disease and rheumatoid arthritis, where the novel disease loci we identified have been previously associated with similar diseases or are known to function in relevant biological pathways.

Genome-wide association studies (GWAS) have identified many susceptibility loci underlying the molecular etiology of complex diseases [1]. These studies have been responsible for the discovery of many individual genes that contribute to disease risk [2]–[10], for discoveries on the front line of personalized medicine [11], [12], and for discovering novel pathways important for the progression of complex heritable diseases [13]. The expense of each GWAS that is capable of finding well-supported disease loci is considerable and, as a consequence, each robust and interpretable association discovered in a GWAS is valuable, not only from the point of view of scientific discovery but also in terms of return on investment [14], [15]. A clear picture that has an important bearing on the investment-discovery tradeoff in GWAS experiments is that the associations identified to date generally explain only a small to moderate fraction of total heritability [16], [17]. Recent analyses have suggested that a considerable amount of this ‘missing’ heritability can be accounted for by rare variants or variants with weak effects [18]–[20]. This suggests that there is an opportunity to identify more risk loci through studies that require even greater investment, by including larger sample sizes and/or by incorporating higher genetic marker coverage of the genome by using next-generation sequencing (NGS). The novel associations discovered by large consortia GWAS studies support this supposition [7]–[10]. Another complementary strategy that leverages both the current and future investment in GWAS experiments is the application of new statistical analyses that can reliably identify weaker associations [21]–[25]. Although there has been an explosion of methods in this area [26], [27], few have produced robustly supported associations that are not detectable by single marker tests of association [1], [26]–[29]. Here, we report a general framework for applying a family of GWAS analysis methods that is extremely promising for detection of weak associations yet has not been widely applied to learn novel biology from GWAS datasets: penalized multiple regression (PMR) methods. PMR methods work by simultaneously incorporating tens to hundreds of thousands of genetic markers in a single statistical model where a penalty is incorporated to force most marker regression coefficients to be exactly zero, so that only a small fraction are estimated to make a contribution to disease risk [22], [30]–[39]. By jointly analyzing markers, PMR methods are able to consider the correlation of each marker with the phenotype, conditional on all other relevant markers. This can increase the power to detect weak associations compared to single marker methods due to the smaller residual variance and the fact that the conditional correlation of a marker with the phenotype can often be substantially higher than the marginal correlation [40]. The latter effect is a consequence of non-zero correlation structure between associated markers when the underlying genetic architecture is polygenic [41]. These methods therefore model the underlying biology more accurately than single marker tests, by explicitly modeling the polygenic architecture of complex phenotypes to account for the effects of multiple susceptibility loci. They also leverage the same type of statistical model used in single marker testing methods that have demonstrated reliability in the identification of strong associations [1], [28], [29]. Yet, despite theoretical power of PMR methods, the large body of statistical literature exploring their theoretical properties (see reviews [42], [43]), and the recent interest in the methods development community [22], [30]–[39], these methods have not been successful in GWAS analysis. This is due to a combination of limitations: 1) inability to scale for very large GWAS datasets [22], [32], [34], [35], 2) poor performance on simulated data [22], [31], 3) they often find too many ‘hits’ to be biologically plausible for a given GWAS sample size [22], and 4) they do not identify novel, well-supported associations that are not detectable by standard methods [22], [31]. In order to address these issues, we present a combined algorithmic and heuristic framework for PUMA (Penalized Unified Multiple-locus Association) analysis that optimizes these methods for reliable detection of weak associations when applied to large GWAS datasets. The complete PUMA framework includes an extremely efficient implementation of a new minorize-maximization (MM) algorithm [44] for generalized linear models (GLM) [45], a theoretically motivated data-adaptive heuristic approach to determine penalty strength and for model selection, and post hoc methods for assessing the rank of identified associations. Within PUMA, we implement all sparse feature selection, penalized regression approaches proposed for GWAS analysis to date, including four penalties implemented in a maximum likelihood framework (i.e. Lasso, Adaptive Lasso, NEG, MCP), as well as theoretically justified penalties that have not been previously applied to GWAS (i.e. LOG) (Figure 1). We demonstrate the power of our framework for detecting weaker associations that are invisible to individual marker testing through analysis of simulated GWAS data that mirror observations from analyses of real GWAS data. We also demonstrate that our approaches correct issues with all current PMR methods where software is available for GWAS analysis, where we find that all of these currently available PMR GWAS methods can perform worse than single marker testing for our simulation conditions. As an illustration of the value of PUMA for mining existing GWAS data for novel associations, we apply these methods to the original Wellcome Trust Case Control Consortium (WTCCC) [2] GWAS datasets for type 1 diabetes, Crohn's disease and rheumatoid arthritis. Our re-analysis identifies weak associations that implicate additional susceptibility loci for these autoimmune diseases, which did not appear significant by standard single marker tests of association in these datasets, yet were 1) identified in an independent GWAS of the same phenotype that did not include WTCCC data, 2) previously known to play a role in disease etiology, or 3) known to function in a relevant biological pathway. Our results demonstrate that appropriately tuned PMR methods can provide a complementary approach to large meta-analyses [4]–[10] to identify susceptibility loci with weak associations. We also provide a discussion concerning how the framework can be extended to perform penalized analysis of epistasis, to incorporate mixed model analysis, and to address challenges of genome-wide genotypes provided by whole-genome next-generation sequencing. The methods implemented in our PUMA framework are orders of magnitude faster than existing software when assigned identical computational tasks and no pre-screening of markers is performed (Table 1). This substantial boost in computational speed allows PUMA to perform a dense two-dimensional search of tuning parameter values for non-convex penalties (i.e. MCP, NEG, LOG) and examine upwards of 1 million total modes of the likelihood surface for simulated case/control dataset of 5,000 individuals and 650K genetic markers in less than 24 hours on a 6 core Intel® Xeon® W3690 @ 3.47 GHz with 12 Gb memory when a pre-screening p-value cutoff of 0.01 from single marker analysis is applied (Table 2). This is a huge improvement compared to existing software for non-convex PMR methods [22], [32] which only examine a single mode. While pre-screening markers based on a p-value cutoff may initially seem to detract from the purpose of a multiple-locus analysis, it is supported by statistical theory, is necessary for large scale analysis and has almost no impact on the set of markers identified as associated. In a seminal paper, Fan and Lv [46] demonstrate that pre-screening by ranking the marginal correlation of each variable with the response will retain the relevant variable asymptotically with probability tending to 1. Fan and Song [43] extend this result to generalized linear models. Moreover, Tibshirani, et al. [47] and El Ghaoui, et al. [48] establish exact pre-screening methods for linear and logistic Lasso models where relevant variables are guaranteed to be retained for finite sample sizes and demonstrate that the number of variables can be reduced by up to 3 orders of magnitude. Intuitively, both the asymptotic [43], [46] and exact pre-screening methods [47], [48] rely on the fact that a variable is unlikely to have a very small marginal correlation with the response but a large and very significant conditional correlation for a particular sample size when the relevant variables explain only a small fraction of the variation in the response. Moreover, pre-screening is often computationally necessary because storing 650 K markers for 5000 samples requires 26 Gb of memory. Finally, we note that pre-screening is used by previous applications of PMR methods to GWAS data [22], [31] in order to handle genome-scale data. We use a pre-screening p-value cutoff based on single marker analysis, because 1) it retains all relevant variables asymptotically [43], [46], 2) it approximates the exact methods proposed for Lasso [47], [48], which cannot be easily adapted to other penalties, 3) it reduces memory requirements so that very large datasets can be analyzed on a high-end desktop computer, 4) it substantially reduces the computational burden, 5) by using a p-value it is naturally calibrated to the sample size and the fraction of variation in the response being explained, and 6) it has very little empirical effect on the results. We demonstrate this final and most important point in two complementary simulation studies. First we consider a simple two-step forward regression method, which is known to approximate penalized multiple regression [49], [50] and, under a range of biologically motivated simulation conditions, demonstrate that variables that do not cross an initial p-value threshold have a very low probability of being significant in the second step (Figure S1). Second we demonstrate that the pre-screening has no noticeable effect on the performance of Lasso and MCP methods but substantially reduces the computational time (Figure S2). We analyzed 960 simulated GWAS datasets to assess the performance of our PUMA framework compared to other published methods for PMR GWAS analysis. We note that these simulations, while far more extensive than other published works on PMR GWAS analysis [22], [30]–[38] are not meant to be exhaustive or to capture all the possible complexities in a GWAS but rather to: 1) serve as a baseline for comparing GWAS analysis methods and 2) provide an estimate of the expected performance for these methods when applied to GWAS data under relatively ideal experimental conditions. Our goal therefore was not to attempt to model a broad spectrum of possible GWAS data complexities (e.g. stratified experimental sampling schemes, known or cryptic population structure effects on phenotype, relatedness among individuals, measured or latent covariates, etc.) but rather to simulate data that captured the most basic components of a GWAS experiment (see Methods for details). In simulated data a causal variant is defined as a variant whose coefficient value is nonzero, so that number of minor alleles at this marker contributes to the phenotype. In order to mimic the fact that true causal variants are not available from array-based genotyping, the simulated causal variants were removed from the dataset so that they are not considered by the tests of association. Therefore, just like in all array-based genotyping datasets, our simulations identify associations based on markers in linkage-disequilibrium with the (omitted) causal variant. We assessed the performance of PMR methods for which there is available software. We compared the performance of the Lasso penalty from Wu, et al. [31], the NEG penalty as implemented in the HyperLasso program [22], and a permutation-based approach to selecting tuning parameter values for the MCP penalty [32], [51] that we term perm-MCP. We note that we only considered PMR approaches that are designed to handle the specific challenges of GWAS data and that also perform feature selection, such that we do not consider ridge, elastic net, or group-penalties since they set many correlated markers to have nonzero coefficients and thus complicate the generation of interpretable p-values [30], [52]. We also did not consider Markov Chain Monte Carlo (MCMC) approaches [34], [35] since they could not efficiently scale to genome-wide data while exploring a range of tuning parameter values. We ran the HyperLasso program [22] with standard settings (see Text S1). We applied the method of Wu, et al. [31], setting the number of selected markers to the true number of causal markers in each simulation since Wu, et al. [31] do not specify a criterion for selecting the model size. As a benchmark, we also ran a single marker analysis implemented by applying a logistic regression model to each marker individually. We used a pre-screening p-value cutoff of 0.01 from single marker analysis for the PMR methods to make them computationally tractable. Simulations indicate that HyperLasso [22] and the Lasso of Wu, et al. [31] are generally less powerful than a standard single marker test (Figure 2, S3, S4, S5, S6, S7, S8, S9, S10). While Lasso is sometimes comparable or slightly more powerful than a single marker test for low FDR, the performance of the method benefits from the fact that the number of selected markers is set using information not available in real data. Setting the marker number to 10 (the default in the implementation of Wu, et al. [31]) or another arbitrary value results in poor performance and is not competitive with a single marker test (results not shown). The performance of HyperLasso is especially poor as is it suffers from the fact that the choice of tuning parameters has a huge effect on performance, but the method does not implement a search over tuning parameter values. Moreover, HyperLasso does not include a way to evaluate the significance of a selected marker, so we used their default approach of using coefficient values from selected markers to assess performance. Alternatively, perm-MCP was the most powerful in our simulations. We note that for perm-MCP, by setting the expected false positive rate (eFPR) and using permutations to obtain the value of the tuning parameter based on this rate, perm-MCP generates a single model with relatively few nonzero coefficients while explicitly addressing the multiple testing problem. Yet in practice this result indicates that perm-MCP may assign p-values to only a handful of markers so that the method may not identify any novel associations for a particular dataset. Since the number of nonzero coefficients is directly related to the specified eFPR and the pre-screening cutoff, we examined multiple eFPR values (, , , , ) and cutoff values (0.1, 0.01, 0.001), and selected the values the yielded the highest power (eFPR = , cutoff = 0.001) to present in Figure 2, where other cutoff combinations produce poorer performance (see Figure S11 for a representative plot showing the results for all cutoffs). We note that the eFPR value is based on the number of markers that pass the pre-screening cutoff, not the total number of markers. Therefore the performance of perm-MCP is sensitive to the eFPR and cutoff values, yet there is no clear method to optimally specify this value a priori. Furthermore, determining the appropriate cutoff for a desired eFPR for correlated high-dimensional data is the subject of current research [53], and its application to permutation methods for selecting a tuning parameter remains an open question. We also note that the performance achieved with PUMA methods does not require the optimal determination of eFPR and pre-screening cutoffs. In addition, we note that while Ayers and Cordell [32] have previously shown that penalized regression methods can perform well on simulated data, the datasets we address here are orders of magnitude larger. Ayers and Cordell [32] conducted two simulation studies, one with 4000 markers and the other with no more than 228. By considering such a small set of markers, which is not the product of a pre-screening step, they were able to use standard R packages and apply a permutation method to select tuning parameters on the full dataset. Moreover, the multiple testing problem is less severe in their analysis. For the HyperLasso program, Ayers and Cordell [32] selected the tuning parameter as described by Hoggart, et al. [22]. However, using these settings for the genome-scale datasets examined here caused the HyperLasso program to crash (Text S1) and so we use the default program settings. We note that the program worked as expected for smaller datasets. It is unclear whether this problem is an issue with the underlying algorithm or the specifics of the implementation. Thus the difference between the performance of methods in Ayers and Cordell [32] and the current study is the scale of the data, the large multiple-testing burden for genome-scale data and the necessity of a pre-screening step for genome-scale data. PUMA's statistical power is due to its data-adaptive properties. PUMA 1) performs a two dimensional search of the tuning parameter space 2) selects the number of nonzero coefficients based on both the fit to the data and the sample size, and 3) uses a heuristic methods to assess the significance of correlated markers. Conversely, perm-MCP fixes one of the tuning parameters, does not incorporate the sample size, and does not address the issues of testing the significance of correlated markers. Moreover, perm-MCP relies on setting the eFPR despite problem of determining an appropriate value a priori for high dimensional data. For the 960 simulated GWAS datasets we analyzed, almost all PMR GWAS approaches implemented in PUMA except NEG and adaptive Lasso outperformed single marker analysis under simulation conditions with sufficient sample size (Figure 3, Figures S3, S4, S5, S6, S7, S8, S9, S10). Quite critically, the performance is far greater even when using a conservative control of FDR that is commonly employed in GWAS studies. Moreover, the improvement of the PMR methods in PUMA is most noticeable for causal variants with intermediate marginal heritability. Overall, these simulations demonstrate that the advantage of PMR methods over a single marker test increases with sample size, but decreases with the number of susceptibility loci (Figure S3, S4). While the penalized methods implemented in our PUMA framework consistently had higher power than single marker analysis as a function of FDR under most simulation conditions, none of the penalties consistently stood out as the most powerful. However, our PUMA framework, which includes a fast novel algorithm for penalized maximum likelihood estimation in generalized linear models, data-adaptive tuning of tuning parameters, heuristics for model selection and novel method of assigning p-values (see Methods) increased the power of PMR methods compared to current approaches using the same penalties [22], [31]. We note that our implementation of the NEG penalty showed a substantial increase in power over the HyperLasso program [22] and indicates that our search over tuning parameter values and heuristic approach for model selection was successful. Moreover, our search of one or both tuning parameter values for MCP (termed 1D-MCP and 2D-MCP, respectively) showed that our approach to applying MCP (i.e. 2D-MCP) can be more powerful than that of Ayers and Cordell [32]. The fact that our implementation of Lasso had higher power than the version of Wu, et al. [31] confirms the usefulness of our data-adaptive approach for selecting penalty strength and our novel method for assigning p-values. We also note that for comparison we applied a conditional regression test and our previously published algorithm VBAY, a variational Bayes approach for fitting a mixture prior penalty [33]. We found that perm-MCP and VBAY had similar performance to our PMR methods and while the conditional test of association was sometimes more powerful than single marker analyses it was generally not as powerful as the PUMA PMR methods. In our re-analysis of type 1 diabetes, Crohn's disease and rheumatoid arthritis datasets, we applied a single-marker analysis and all PMR analysis approaches (Lasso, Adaptive Lasso, NEG, LOG, 1D-MCP, 2D-MCP, perm-MCP) using all the recommended components of our framework. We included sex and the first two principal components as unpenalized covariates, applied a pre-screening cutoff of 0.01 on the p-values from the single marker test, and ran 100 reorderings for the non-convex penalties. Quantile-Quantile (QQ) plots of p-values from a standard single marker analysis indicate that the effects of any remaining population structure is minimal. Moreover, including the subset of significantly associated markers identified by the PMR methods as covariates in a single marker analysis of remaining markers does not yield an inflation of the QQ plots and thus indicates that the PMR methods are not overfitting the data (Figure S12). We also note that due to the complex LD around the MHC on chromosome 6, while we included this region in our analysis, we omit this region from any post hoc analysis and discussion. Our single-marker re-analysis of type 1 diabetes, Crohn's disease and rheumatoid arthritis datasets reproduced the same associations as reported in the original analysis (Figure S13). Our PMR methods recapitulated almost all of the associations identified by single marker analysis, although there were differences among the methods. The PUMA Lasso and Adaptive Lasso identified almost no additional associations compared to single marker tests, and while LOG, NEG and 1D-MCP identified more, almost all of the associations found by these five methods (Lasso, Adaptive Lasso, LOG, NEG, 1D-MCP) were identified by 2D-MCP (Figure 4, S14). We note that perm-MCP identified very few associations (12 overall, across the three diseases), all but one of which was identified by a single marker test, and all were identified by 2D-MCP. We therefore discuss the associations found by 2D-MCP, where we consider three categories of interest (Table 3): those concordant with single marker tests, those that recapitulate associations identified in external GWAS studies but not by single marker analysis of the WTCCC, and novel associations, of which many were deemed to be biologically interpretable in terms of the current knowledge of disease etiology. In the absence of functional validation, the presence of a feasible biological interpretation lends more credibility to these novel findings. A critical point to note about the performance of our PUMA framework for PMR analysis of GWAS data is that these methods not only result in the correct identification of more loci than a single marker testing analysis (when controlling the false discovery rate at the same level), but also lead to re-orderings of the rank of markers that are considered the most significant when compared to a single marker analysis (Figure 5). As a consequence, we are able to identify etiologically relevant and replicated disease loci that are too weak to be detected by single marker analysis, yet show strong signals of association by PMR analysis. This means that our PMR GWAS analysis is not simply taking advantage of the lower residual variance to improve performance, but is also taking advantage of the fact that conditional correlation of a relevant marker with the phenotype is often more significant than the marginal correlation. When the coefficients for multiple markers, each tagging different susceptibility loci throughout the genome, have nonzero values in the PMR framework, their association with the phenotype becomes more significant. Our framework can therefore identify disease susceptibility loci in a GWAS with weak associations with phenotype, when they are invisible to a single marker testing approach (i.e. they have p-values in a single marker test that would never be considered significant). The associations identified by PUMA generally recapitulate associations identified by single marker analysis, and the PUMA hits have perfect concordance for strong associations. Overall 2D-MCP recapitulates the largest number of associations, while the union of the other PMR methods (considered here for illustrative purposes due to the high degree of concordance with each other, and the fact that 2D-MCP identifies almost all of the associations they find) had a lower degree of concordance with the single marker analysis (Figure 4, 6, S14, Table S1). Of the 6 associations identified by a single marker analysis that were missed by our methods, 5 were from type 1 diabetes and 1 was from Crohn's disease (Table S2). One of these associations was borderline significant by 2D-MCP with a p-value of 1.42×10−7. We compared associations identified by our PUMA methods that were not detected by single marker tests in the WTCCC dataset to markers identified by independent studies in the HuGE database of published GWAS [53] in order to find associations identified in both our analysis and an independent study that did not include WTCCC data. Such replications are considered the gold standard for validating a putative association [54]. In the ideal case the same marker would show an association in both the WTCCC dataset and those summarized in the HuGE database. However, given 1) the lack of overlap of marker-sets between genotyping platforms, 2) that the HuGE database reports only the most significant marker in an associated LD block, and 3) that PMR methods tend to select only a single marker within a LD block, we considered a marker to recapitulate a known association if the two are within 0.1 cM [6]. A representative example from Crohn's disease is shown in Figure 7 where only 2D-MCP is able to identify STAT3 as a susceptibility locus in the WTCCC dataset (Figure 7a). While this association has also been replicated in non-independent datasets [6], which included WTCCC data, the role of STAT3 in Crohn's and other autoimmune disease is well established [55], [56]. While all PUMA methods and a single marker test are able to replicate associations from independent studies, LOG, NEG and 1D-MCP, stood out in terms of identifying associations replicated by non-independent studies, but not detected in the WTCCC dataset by a single marker analysis. These counts reflect the results when the number of markers considered as ‘hits’ was set to be equal across methods so that they reflect the ordering of markers by PMR methods rather than the number of associations. When comparing the total number of significant hits from each method to associations identified in either independent studies or non-independent external studies that incorporated WTCCC data, 2D-MCP is the only PMR method to identify as many total replicated associations as a single marker test (Tables 4, S3, S4, Figures S15, S16, S17, S18, S19). However, 2D-MCP is able to replicate known associations that cannot be replicated by a standard single marker test in this dataset, thus demonstrating that PMR methods can extract biologically relevant information that is overlooked by standard analyses (Table S5). These results demonstrate that PMR methods overall are able to identify replicated associations in this dataset that are invisible to a standard single marker test. Moreover, our methods provide an opportunity to replicate previously unreplicated associations by re-analyzing existing GWAS datasets. Re-analysis of type 1 diabetes, Crohn's disease and rheumatoid arthritis datasets from the original WTCCC [2] with our PUMA methods revealed novel associations that have not been identified in previous GWAS of these diseases (Table 5, Figures S14, S20, S21, S22). These methods, most notably 2D-MCP, identify novel associations in or near genes which have been previously associated with etiologically related diseases or which are known to function in biologically relevant pathways based on public databases and disease literature (Tables 5,6). In addition, PUMA also identified associations without a clear biological link to the disease phenotype (Tables S6, S7, S8). PUMA methods identified novel susceptibility loci for type 1 diabetes involved in pancreatic function, insulin pathways and immune cell function and for Crohn's disease that are involved in pro- and anti-inflammatory pathways (Table 6). 2D-MCP identified a gene functioning in apoptosis as a susceptibility locus for rheumatoid arthritis (Table 6). These genes are known to function in relevant pathways or have been previously implicated in the etiology of the disease but have not been found by previous GWAS of each disease. A representative example is shown in Figure 7b where only 2D-MCP identifies an association that implicates SLC30A1. This gene is a zinc transporter related to SLC30A8, which has been implicated in type 2 diabetes, and zinc transport plays a role in insulin secretion by pancreatic -cells [57], [58]. Each GWAS discovery that has a well supported association produces valuable information for understanding the etiology of the disease phenotype and such discoveries are regularly used as the foundation for studies that use the locus as a starting point [59], [60]. Given that GWAS involving a thousand to several thousands of individuals seldom return more than a few to a dozen well-supported associations (depending on the disease) the monetary, time, and resource investment in these studies often translates to a considerable expenditure per discovery. This is true even when considering additional discoveries that may occur as individual GWAS are combined together into large meta-analysis studies [4]–[10]. We have demonstrated that our PUMA framework has the potential to produce added investment return for GWAS studies by discovering additional well-supported disease loci associations that are invisible to the standard single marker analysis methods responsible for almost all reported GWAS [1], [53]. For example, our re-analysis of type 1 diabetes, Crohn's disease and rheumatoid arthritis from the original Wellcome Trust Case Control Consortium (WTCCC) [2] demonstrates that PUMA methods can identify associations that are not detectable by single marker analysis approaches but which replicate associations known from independent studies, which did not include WTCCC data, as well as novel loci with strong links to known disease etiology. These included 10 novel associations identifying genes that are linked to primary pathways of these autoimmune diseases, specifically 6 genes involved in pancreatic function, insulin pathways and immune-cell function in type 1 diabetes; 4 genes (in 3 association regions) functioning in pro- and anti-inflammatory pathways in Crohn's disease; and 1 gene involved in apoptosis pathways in rheumatoid arthritis. Applying our PUMA framework therefore has the potential to add a significant number of discoveries for a given GWAS. A critical property of our PUMA framework is it does not return the same ordering of significant markers produced by a standard single marker analysis. By simultaneously accounting for the associations of multiple loci and better reflecting the underlying polygenic architecture of complex phenotypes, PUMA can find strong statistical support for associations deemed non-significant by a single marker analysis and places these among the top list of associations. A prime example is marker rs613232 which had a p-value of 6.51 by a single marker analysis in the type 1 diabetes dataset so it would not be considered for a follow-up study. However, by taking into account the polygenic architecture of the trait, 2D-MCP assigned it a p-value of 6.96 (Figure 7b, Table 5). This marker tags the zinc transporter SLC30A1 and zinc transport has an established role in type 1 diabetes, yet this gene was only identified as a susceptibility locus by 2D-MCP. This example illustrates the power of PUMA methods to reorder the p-values of markers so that a marker that is not in the top 20,000 by a single marker test can be in the top 30 by 2D-MCP. Another example is that of LPHN2, a gene identified by an independent GWAS of type 1 diabetes, yet the association was not replicated in an independent dataset in the same study [61] or, to our knowledge, any subsequent study. In our re-analysis, 2D-MCP identified a strong signal in a nearby marker and assigned it a p-value of 3.99, while the p-value by a single marker test was 3.78 (Figure 7c, Table 5). The very weak single marker p-value found in this dataset makes the previous inability to replicate this association unsurprising. The gene encodes the G-coupled protein receptor latrophilin 2 and has a weak association with rheumatoid arthritis [62] but its relation to disease etiology is unclear. These examples illustrate that our PUMA framework returns additional and complimentary information to the results of a single marker analysis of a GWAS. In general, it seems clear that applying a spectrum of appropriate GWAS analysis methods to the same data is likely to maximize discovery. The PUMA framework and software that we present here is immediately applicable to a large number of existing GWAS and we are currently exploring extensions of the framework to address additional challenges in GWAS experimental designs and GWAS analysis. For example, GWAS discoveries are regularly being produced by consortia that combine several independently executed GWAS experiments. Such combined data introduce a number of complexities including complex batch effects, population structure, relatedness, and latent environmental variables. While meta-analysis techniques for combining p-values across studies are a good approach to normalizing for many of these issues [4]–[10], a PMR analysis directly on the genotypes can include correction for study heterogeneity, population structure and cryptic relatedness using a linear mixed model [63]–[65], and we are currently working on such extensions. Given that the increase in performance for our PMR methods compared to single marker analysis increases with increasing sample sizes, solving these problems has great potential to detect additional weak associations. There is also going to be a near-term shift towards GWAS that add millions of additional genetic markers genotyped by next-generation sequencing, which can add increased density of markers and different allele types. Our approach can already handle these large number of markers directly to take advantage of the better tagging, and in some cases genotyping, of causal disease polymorphisms. The trend of increased sample sizes and genome marker coverage in GWAS also opens the opportunity to identify genetic interactions that are currently difficult to detect, including epistasis and geneenvironment interactions, which could be identified by incorporating group penalty approaches [30], [66], [67] within our framework. Overall, our framework represents a platform for integrating richer statistical models and techniques for addressing the future needs of GWAS. Our framework is a combination of algorithms and heuristic approaches designed for robust and efficient analysis of GWAS datasets when the desired output is a ranked list of genetic markers that individually tag disease loci. The value of the framework is that tag genetic markers, which are too weak to be reliably identified by a single marker analysis, can be identified while preserving a conservative FDR genome-wide. To solve issues that have limited the value of existing PMR GWAS software for this purpose, we designed our framework to have the following properties: 1) the versatility to handle a diversity of penalties for simultaneous analysis of thousands to millions of genetic markers while incorporating unpenalized covariates, 2) the efficiency to analyze up to millions of markers after pre-screening on a standard desktop, 3) the sensitivity to tune the strength of penalties and to perform model selection when the fraction of variation accounted for by disease loci identifiable with tag markers is small (as is typical for GWAS), and 4) the capability to return a ranked list of p-values where each of the top markers identifies an independent disease association. We outline the components of our framework responsible for each of these properties in the next four sections, followed by a description of our software PUMA that implements our recommended practices and options for implementation. We note that in its entirety, this framework is a new GWAS analysis approach that incorporates novel components including (but not limited to): application of penalties not previously applied to GWAS, a new MM algorithm for GLMs, heuristics for penalty strength and model selection, and post hoc model fitting approaches for ranking associated markers. Our framework makes use of a generalized linear modeling (GLM) framework to construct the likelihood objective function. We can therefore model phenotypes measured on a large diversity of scales by implementing an appropriate link function [45], although here, we limit our implementations to an identity and logistic link function to model continuous phenotypes with normal error and case-control phenotypes, respectively. We also note that incorporating unpenalized covariates is straightforward, where these are modeled with regression coefficients with no penalty. While our current implementation is restricted to penalties that select a small number of well supported markers (i.e. feature selection penalties [51], [68]–[72]), the framework is versatile enough to implement a wide diversity of penalties approaches when making use of the algorithm described in the next section. For marker selection, we use the penalized maximum likelihood estimate (pMLE) of the regression coefficients:where is the vector of disease phenotype values, and is the vector of regression coefficients is the log-likelihood of a linear or logistic regression and is the penalty function on the magnitude of indexed by a vector of tuning parameters, . Since we are interested in identifying a small set of variables associated with the phenotype, the penalty function must have the sparsity property whereby most of the regression coefficients are set to exactly zero. Multiple penalties satisfy this condition while balancing computational tractability with desirable theoretical properties. We implement the penalties that have been applied for PMR GWAS analysis (i.e. Lasso, Adaptive Lasso, NEG, MCP) as well as a penalty that has not be previously applied to GWAS (i.e. LOG). We describe the properties of each penalty in the following paragraphs and the functional form of each penalty is given in Text S2. The Lasso penalty [68] (previously implemented for PMR GWAS in the software Mendel [31]) is a linear function of the magnitude of the regression coefficients and is the most widely used since it has a single tuning parameter. Moreover, the penalty is convex so that it yields a unique pMLE. Yet it is widely known to select too many variables with non-zero coefficients in high dimensional datasets [72] and does not satisfy the “oracle property” whereby parameter estimates are asymptotically equivalent to unpenalized estimates when the identity of the non-zero coefficients are known in advance [72]. The Adaptive Lasso penalty [69], (previously implemented for PMR GWAS by Yang, et al. [37]) unlike the Lasso, satisfies the oracle property. This two-step Lasso regression procedure is also convex (yields a unique pMLE) although it requires an initial estimate of the regression coefficients, which are then used to weight the strength of a Lasso penalty in the second step. There is no criterion for determining optimal weights, so in practice the Lasso penalty for each coefficient is weighted by the square root of the initial coefficient estimate. The NEG penalty [71] (previously implemented for PMR GWAS in the software HyperLasso [22]) has two tuning parameters and is non-convex such that it produces a multimodal likelihood surface where pMLE's are not unique. The penalty satisfies the oracle property, since the derivative of the penalties approach zero in the limit [72], although it has other less desirable properties since its derivative approaches zero much more slowly than the other penalties and its very complex functional form makes it numerically unstable for large coefficient values. In our framework, we re-implement NEG using a faster algorithm than Hoggart, et al. [22] and includes a two dimensional search of the tuning parameter space, while Hoggart, et al. [22] use asymptotic theory to set the tuning parameters. The MCP penalty [51] (previously implemented in the R package grpreg [73]) has two tuning parameters. Like NEG, this penalty is non-convex and satisfies the oracle property. However, the derivative of MCP reaches zero for finite coefficient values so that it avoids over penalizing large coefficient values. Moreover, MCP is designed to reduce the multimodality of the objective function [51]. The tuning parameters determine the slope of the penalty near the origin (i.e. ) and coefficient value at which the derivative of the penalty is set to zero (i.e. or , depending on notation). When applying this method to GWAS data, Ayers and Cordell [32] fixed the value of at 30, and identify the value of using a permutation approach. We term this approach perm-MCP. In addition, we consider a one dimensional search over the value of and a two dimensional search over both parameters, termed 1D-MCP and 2D-MCP, respectively. The latter has the most potential since it explores the value of the tuning parameter, , that determines the coefficient value at which the derivative of the penalty is set to zero, and learns the value of the parameter based on the data. We also implement the LOG penalty and apply it to GWAS for the first time. The LOG penalty [70] has two tuning parameters and is non-convex, such that it produces a multimodal likelihood surface where pMLE's are not unique, but is satisfies the oracle property, since the derivatives of the penalties approach zero in the limit [72]. This penalty is also designed to identify fewer non-zero regression coefficients. Our framework implements a highly efficient algorithm and optimized coding practices to allow fast simultaneous analysis of genetic markers in the range of hundreds of thousands to millions. We implement a new minorize-maximization (MM) algorithm for finding pMLE by using a coordinate-wise ascent approach with an upper bound on the second derivative of the likelihood function [44]. By using bounded univariate updates, the algorithm is extremely fast and is guaranteed to converge to a mode of the likelihood surface. While Newton-Raphson algorithms must evaluate the log-likelihood after each update to check if it has decreased [74], the use of an MM algorithm for logistic regression guarantees a monotonic increase and eliminates the expensive function evaluation. Derivations are given in Text S2. In addition to the algorithm, we also implement a number of optimized coding practices to accelerates these PMR methods. These include data structures to minimize access time to each marker, use of optimized linear algebra libraries and searching multiple modes of the likelihood surface for non-convex penalties in parallel. Critical to the performance of our framework is preserving a conservative control of the FDR for identified markers. To accomplish this, we employ a strategy that allows our PMR methods to automatically adapt, not only to the sample size of the dataset, but also to the number and magnitude of the non-zero regression coefficients for relevant markers associated with the phenotype. Our approach includes an adaptive tuning of penalty strength in combination with model selection and assessment of model fit. Statistical theory considering linear regression has shown that for a sample size of , the number of variables detectable as having nonzero coefficients is on the order of [75]. This is consistent with other theoretical work [76], [77] and satisfies our intuition that the number of detectable associations is directly related to the sample size of the dataset. For adaptive tuning of the Lasso and Adaptive Lasso, where the likelihood is convex and there is a single tuning parameter, the search is simple and we start with a severe penalty, which is gradually decreased to select one additional non-zero coefficient at a time, until genetic markers are selected. For the non-convex penalties [22], [32], [70], a grid search over a two-dimensional space of tuning parameters is used starting from equally spaced Lasso models (a special case of all non-convex penalties) where the non-convexity of the penalty is gradually increased until markers are selected. This approach for searching the space has been shown to avoid some suboptimal modes of the likelihood surface [70]. We note that we have previously published the approximate Bayesian methods, VBAY and VBAYNET, which incorporate a probabilistic bound, where we applied the same bound [33], [78]. In order to mitigate the problem of suboptimal modes at least to some degree, we explore multiple modes of the likelihood surface for non-convex penalties by permuting the order in which the regression coefficients are updated. For both the simulations and WTCCC analyses of this study, we found 100 reorderings was sufficient to obtain robust results. Once sets of markers with nonzero coefficients are identified for each value of the tuning parameters for a given penalty, the optimal set is determined. We assessed the overall appropriateness of the fit of a selected model based on a QQ plot by fitting an unpenalized model with selected markers, and calculating p-values for each marker in the dataset by regressing it against the residuals from the first step (Figure S12) [79]. A slight inflation of p-values at the tail of the QQ plot indicates that the PMR method has not over fit the data. We tried many model selection strategies to determine which consistently produced optimal residual-QQ plots, including Bayesian information criterion [80], cross-validation [40], asymptotic justifications [22] and permutation-based approaches [32]. We eventually selected AIC [81] with an additional restriction as our default because it tended to produce the best overall performance for simulated GWAS data, when assessed by both QQ plot and when applying a strong control of the FDR. While AIC [81] and BIC [80] have been studied extensively in the context of sparse model selection theory, these criteria do not incorporate an upper bound on the number of variables it is feasible to select for a given sample size. Therefore, while good performance may be guaranteed for an infinite sample size, these criteria can select too many variables than is plausible for a finite sample size [82] such that our upper bound on the number of markers selected seems reasonable, where we compare all models satisfying this bound using AIC. We also note that while AIC is not asymptotically model selection consistent [83], our use of AIC should not suffer from the inclusion of irrelevant genetic markers since we apply it in the context of models constrained to have fewer than genetic markers and we apply a significance test (described in the next section) before considering a marker to have an association. The most valuable final output of a PMR GWAS analysis is a ranked list of markers in decreasing order of highest confidence. Standard methods for producing such ranked lists in a PMR framework assess significance by conducting variable selection on a subset of the data and assessing significance on another subset [84], or subsetting the data many times and identifying variables selected in many of the subsets [85]. Such methods can be very computationally demanding for large GWAS datasets and have been shown to underperform a standard single marker test of association [86]. Moreover, these methods do not address the challenging problem of assessing the significance of a marker in the presence of correlated markers within the same linkage disequilibrium (LD) block. While PMR methods tend to select a single non-zero regression coefficient for an associated LD block, cases often arise where multiple markers in a LD block have non-zero coefficients. In such cases, the correlation between the markers in the block dramatically increases the sample variance of the coefficients so that the markers are not assigned significant p-values even if each would be significant if the other were dropped from the model. We deal with the problem of producing an informative ranked list of selected markers in our framework by initially fitting an unpenalized regression model with all markers selected by a given PMR and including relevant covariates, and calculating p-values for each marker using a standard likelihood ratio test by comparing the full model to null models where each marker is omitted in turn. The correlation between all pairs of selected markers is then evaluated, the pair of markers with the largest correlation is identified, the marker with the smallest absolute regression coefficient of the two is dropped, and p-values are then recalculated for the remaining markers. This process is repeated until no pairwise correlation between remaining markers exceeds 0.1 and each marker is finally assigned the smallest p-value produced for it during this processes. This heuristic procedure means that the values cannot be interpreted strictly as asymptotic p-values [31], but they can be considered as scores indicative of the significance of the association that ranks the values in terms of confidence while ensuring at least one marker in the LD block is assigned a p-value rank that can appropriately reflect a true association. PUMA implements both linear and logistic models for PMR methods, as well as single marker analysis, a conditional regression analysis, and the variational Bayes multiple regression method VBAY [33]. The software reads genotype files in TPED format and phenotypes in TFAM as used by Plink [87]. For all multiple marker methods, we employ a heuristic to remove markers with low marginal correlations with the phenotype as they are extremely unlikely to be selected to have nonzero regression coefficients [31], [46]. This is approach is not novel [31], [46], but accelerates computation and allows flexibility when analyzing extremely large GWAS datasets. The set of markers identified at each mode of the likelihood surface is stored and is saved in a text file readable by R. The software is available at http://mezeylab.cb.bscb.cornell.edu/Software.aspx. For the single marker analysis, p-values are calculated using a standard F-test or likelihood ratio test [88] for linear and logistic models, respectively. The conditional regression performs the standard single marker test of association and includes the significant markers as covariates in a second set of single marker tests. The minimum of the two p-values from the first and second stage analysis is then reported for each marker. In this analysis we used a first stage p-value cutoff of 1 and selected the single most strongly associated marker within 100 Kb to include as covariates in the second stage. We also re-implemented the approximate Bayesian method VBAY, previously developed by our group [33]. While Bayesian regularized regression methods using a number of prior distributions as been applied to association mapping using Markov chain Monte Carlo (MCMC) methods, these cannot simultaneously analyze more than a few hundred to a few thousand genetic markers at a time [34], [35]. VBAY [33] uses a variational Bayes approximation to the posterior surface [89] and applies a hierarchical mixture prior on the regression coefficients so that the large majority of coefficients have a high posterior probability of being exactly zero (see Logsdon et al. [33] for a more detailed discussion of this method). Within PUMA, we re-implemented VBAY and added the capability to analyze case-control phenotypes, where we increased the power to detect weak associations for case-control phenotypes by approximating a logistic regression by modeling the error distribution with a Student t-distribution with 7.3 degrees of freedom. This parameterization has the smallest squared error loss of any t-distribution with respect to the logistic error function [90], [91]. Moreover, we address the multimodality of the posterior surface by exploring many posterior modes and applying Bayesian model averaging [92] in order to weight the contribution of each mode to the posterior probability of association for each marker. The VBAY algorithm was run with 1000 restarts to explore the non-convex posterior surface. Due to its Bayesian formulation, VBAY reports the posterior probability, between 0 and 1, that each marker is associated with the phenotype. The default settings of PUMA are our recommended settings for a GWAS analysis, which are the same procedures we followed for our analysis of the WTCCC data here: Our approach was to simulate different sized GWAS experiments where we used the real genetic markers for unrelated European individuals from the Multi-Ethnic Study of Atherosclerosis (MESA) [95] genotyped on the Affymetrix 6 platform. Larger sample sizes were generated by drawing haplotypes from existing individuals in order to avoid the confounding effect of population structure [96]. For each simulated GWAS dataset, we considered different sample sizes () with equal numbers of case and control phenotypes simulated under an additive threshold model with a disease prevalence of 50%, using the GCTA program [97], and that different numbers of susceptibility loci () contributed to phenotype heritability, where the total contribution of these loci to heritability was varied (). Coefficients were drawn from a , independent of allele frequency, so that most effect-sizes were very small as determined by the marginal heritability calculated by GCTA [97]. We considered 20 replicates per simulation condition to give 960 simulated GWAS datasets. Causal variants affecting the phenotype were selected uniformly from the set of genetic markers with minor allele frequency (MAF) . We followed typical array-based GWAS by omitting the causal variants from the analysis so a susceptibility locus must be identified by markers in linkage disequilibrium with the causal variant. Following the performance evaluation of previous studies [22], [32], a marker was considered a true positive hit if it had with a causal marker, otherwise it was considered a false positive hit. Since a causal variant may be tagged by multiple true positive markers, the true positive count is defined as the total number of causal variants tagged when all true positive hits are considered together. Alternatively, since false positive hits will often fall in clusters in the same linkage disequilibrium block, we assign each to a 100 kb cluster centered at the most significant hit in that cluster. The false positive count is then defined as the number of such false positive clusters. We note this is a strategy for assessing the performance properties that will be of greatest interest to GWAS practitioners since it focuses on correct identification of tag markers that are in high linkage disequilibrium (LD) and in close physical proximity to the location of the true causal alleles, while considering a strict control of the FDR. To run the data analysis we used the same quality control filters as in the Wellcome Trust Case Control Consortium, first by excluding 809 individuals because of poor sample quality, non-Caucasian ancestry, or a high degree of relatedness [2]. An additional individual was removed for being an outlier by principal components analysis [96]. Marker locations and genetic map are based on reference assembly GRCh37/hg19 and dbSNP v131. Next, the same study-wide missing data rate and deviation from Hardy-Weinberg equilibrium cutoffs were used for each set of cases as in the original study [2], with an additional filter to only include markers in the analysis with a minor allele frequency greater than 0.05 in each combined case-control population, leaving approximately 360,000 markers for each combined case-control data set. We used the CHIAMO calling scores to set data to missing, where any call with a score of less than 0.90 was set to missing [2]. To impute this sporadic missing data we used Beagle [93], with the default settings and allocating a maximum of 3000 MB of memory, where the sporadic missing data for each cohort was imputed separately. The same set of controls (1958 Birth Cohort (58C) and UK Blood Service sample (NBS)) were used for each set of cases as in the original study [2]. Finally, the PMR and single marker analyses included sex as a covariate along with the first two principal components of the genotype matrix. In order to explore the biological function, relevant pathways and possible disease implications of each gene near a significantly associated marker, we mined public databases including GenBank [98], Pfam [99], KEGG [100], OMIM [101], GeneCards [102] as well as the HuGE database [53] of known GWAS hits and known gene-phenotype links. We also conducted an extensive literature search with each gene name and relevant phenotypes.

10.1371/journal.ppat.1005378

Functional Interplay between Type I and II Interferons Is Essential to Limit Influenza A Virus-Induced Tissue Inflammation

Host control of influenza A virus (IAV) is associated with exuberant pulmonary inflammation characterized by the influx of myeloid cells and production of proinflammatory cytokines including interferons (IFNs). It is unclear, however, how the immune system clears the virus without causing lethal immunopathology. Here, we demonstrate that in addition to its known anti-viral activity, STAT1 signaling coordinates host inflammation during IAV infection in mice. This regulatory mechanism is dependent on both type I IFN and IFN-γ receptor signaling and, importantly, requires the functional interplay between the two pathways. The protective function of type I IFNs is associated with not only the recruitment of classical inflammatory Ly6Chi monocytes into IAV-infected lungs, but also the prevention of excessive monocyte activation by IFN-γ. Unexpectedly, type I IFNs preferentially regulate IFN-γ signaling in Ly6Clo rather than inflammatory Ly6Chi mononuclear cell populations. In the absence of type I IFN signaling, Ly6Clo monocytes/macrophages, become phenotypically and functionally more proinflammatory than Ly6Chi cells, revealing an unanticipated function of the Ly6Clo mononuclear cell subset in tissue inflammation. In addition, we show that type I IFNs employ distinct mechanisms to regulate monocyte and neutrophil trafficking. Type I IFN signaling is necessary, but not sufficient, for preventing neutrophil recruitment into the lungs of IAV-infected mice. Instead, the cooperation of type I IFNs and lymphocyte-produced IFN-γ is required to regulate the tissue neutrophilic response to IAV. Our study demonstrates that IFN interplay links innate and adaptive anti-viral immunity to orchestrate tissue inflammation and reveals an additional level of complexity for IFN-dependent regulatory mechanisms that function to prevent excessive immunopathology while preserving anti-microbial functions.

Influenza A virus (IAV) is a leading cause of respiratory infection and induces a strong acute inflammation manifested by the recruitment of monocytes and neutrophils as well as the production of proinflammatory cytokines in infected lungs. The interferons (IFNs) are strongly induced by IAV and are known to mediate host resistance to the infection. However, in contrast to their well-studied inhibitory effect on viral replication, the effects of IFNs on host inflammatory responses are less well understood. In this manuscript, we demonstrate that anti-viral IFN signaling is also required for the orchestration of a tissue response associated with the protection against IAV infection in mice. Importantly, we identify that type I IFNs cross-regulate and cooperate with IFN-γ to inhibit monocyte activation and neutrophil infiltration, respectively. This study also demonstrates that Ly6Clo monocytes/macrophages can potentially mediate influenza virus-induced inflammation, suggesting that IFNs dictate the homeostasis versus inflammatory function of mononuclear phagocytes in viral infection. Our study reveals a novel IFN-dependent regulatory mechanism designed to prevent the excessive immunopathology while preserving its anti-microbial functions. Moreover, these observations have particular relevance for understanding the mechanisms underlying the strong inflammatory response associated with lethal IAV strains and have implications for the development of new immunotherapies to treat influenza.

Influenza A virus (IAV) is a leading cause of respiratory infection and an ongoing threat to global health. Host clearance of IAV, which infects primarily airway epithelial cells, requires the development of both innate and adaptive immune responses [1,2]. Interestingly, recent studies have suggested that the host immune response rather than the cytopathic effect of viral infection plays the key role in driving tissue pathology and host mortality [3–5]. IAV triggers an acute pulmonary inflammation associated with the recruitment of inflammatory monocytes and neutrophils in infected lungs (reviewed in [6]). While it is clear that elevated neutrophil accumulation into infected lungs is associated with increased mortality following IAV infection [7,8], monocyte recruitment can be host protective or detrimental [9,10], suggesting that monocytes may play a multifactorial role in the infection. The current understanding of monocytes suggests that there are at least two major subsets: classical Ly6Chi and nonclassical Ly6Clo monocytes [11]. The classical Ly6Chi monocytes are known to mediate various inflammatory conditions [12] and accumulate in large numbers in IAV-infected lungs [13]. In contrast, nonclassical Ly6Clo cells have been shown to “patrol” the vasculature to clear damaged endothelial cells and contribute to tissue remodeling during the resolution phase of inflammation [14]. Interestingly, the patrolling Ly6Clo monocytes have been shown to differentiate into alternatively activated macrophages during Listeria monocytogenes infection [14]. However, the phenotype and function of Ly6Clo monocytes / macrophages (Mo/Mϕ) in IAV infection are currently unknown. In addition to recruiting myeloid cells, IAV induces the production of proinflammatory cytokines, including type I and II interferons (IFNs), in infected animals. The type I IFN family consists of ~20 different members believed to be important in anti-viral and cancer immunity [15], whereas the sole member of type II IFN, IFN-γ, plays a major role in activating Mo/Mϕ and protection against intracellular bacterial and parasitic infections (Reviewed in [16]). In contrast to its extensively studied function in initiating a cell-autonomous anti-viral state, the mechanisms by which IFN signaling regulates host tissue responses to IAV infection are poorly understood. Several recent studies have shown that viral-induced type I IFNs promote the accumulation of classical Ly6Chi monocytes into the airway and lungs of IAV-infected mice [17,18]. However, it is unclear whether these cytokines also regulate the function of pulmonary monocytes for the resistance to IAV infection. Moreover, the contribution of IFN-γ to IAV-induced pulmonary tissue inflammation is not clearly defined. We report in this study that type I IFN and IFN-γ signaling each play a pleiotropic role in the pulmonary inflammatory response to IAV. Importantly, IFN cross-regulation and cooperation are essential for the suppression of monocyte- and neutrophil-driven tissue inflammation. While antagonizing IFN-γ signaling to inhibit Mo/Mϕ activation, type I IFNs synergize with IFN-γ to inhibit neutrophil infiltration. Moreover, we demonstrate that in the absence of type I IFN signaling, Ly6Clo Mo/Mϕ become more proinflammatory than their Ly6Chi counterparts. Since the Ly6Clo Mo/Mϕ population is traditionally associated with tissue remodelling rather than inflammation, our findings also reveal an unrecognized pro-inflammatory potential for Ly6Clo Mo/Mϕ and suggest that IFNs dictate the homeostasis versus inflammatory function of mononuclear cells in viral infection. Sub-lethal, intranasal (i.n.) infection with influenza A/Puerto Rico/8 (PR8) virus in wild-type (WT) mice is characterized by a progressive weight loss that peaks at day 10, after which mice recover and clear the infection (Fig 1A and [19]). To investigate the function of IFNs in shaping the host response to IAV infection, we first examined IFN gene expression and observed that while in WT mice type I IFNs Ifna and Ifnb were both rapidly up-regulated at day 3 post-infection (p.i.), the sole type II IFN, Ifng, was not significantly induced until day 7 p.i. (Fig 1B). The expression of both type I and II IFNs declined to baseline levels by day 10 p.i.. In WT mice, the cellular immune response to IAV infection was characterized by the strong influx of monocytes that peaked at day 7 and then declined by day 10, at which point T cells became the dominant immune cell subset in lungs (Fig 1C). In contrast to WT controls, Stat1-deficient mice, lacking both type I IFN and IFN-γ signaling, displayed a neutrophil-enriched cellular response (Fig 1D), highlighting the critical importance of IFN signaling in orchestrating a protective tissue response to IAV infection. Next, we quantified the viral loads at days 3 and 7 p.i. using both qRT-PCR and classical plaque forming (PFU) assays. While mRNA copy numbers of viral nucleoprotein (NP) determined by qRT-PCR were comparable in WT and Stat1—/—mice, the latter animals showed a marginal increase (0.4 log) in PFUs at day 7 (Fig 1E). The extent of defect in viral control in Stat1—/—mice observed is consistent with that reported previously [20,21], suggesting that the dysregulated tissue response in Stat1—/—mice cannot be explained fully by increased viral loads. To determine whether the altered pulmonary inflammation in infected Stat1—/—mice stems from dysregulated myelopoiesis in bone marrow (BM) or cell trafficking in the periphery, or both, we analyzed hematopoietic cells in the BM of infected WT and Stat1—/—mice by flow cytometry. The stem/progenitor cell populations are commonly defined as lineage-negative Sca1+cKit+ (CD117+) (LSK) cells [22,23]. However, since Sca1 expression is regulated by IFNs (S1 Fig and [22]), we omitted Sca1 staining in the analysis. We did not observe a significant difference in the percentage of lineage-negative CD117+ progenitor cells (Fig 1F), or mature BM-residing CD11b+Ly6G+Ly6Clo neutrophils and CD11b+Ly6G—Ly6Chi monocytes (Fig 1G) in infected WT and Stat1—/—mice, suggesting that IFN signaling plays a minimal role in regulating central myelopoiesis in the BM during IAV infection. To determine the relative contribution of type I and II IFNs to monocyte recruitment in IAV infection, we analyzed monocyte populations in the lungs of infected WT, Ifnar1—/—, Ifngr1—/—and Stat1—/—mice using flow cytometry. When compared to WT mice, Ifnar1—/— and Stat1—/—mice displayed a significant reduction in CD11b+Ly6Chi monocytes in the lungs at day 7 p.i. (Fig 2A, 2B and 2C). The difference is unlikely due to kinetic variations of the pulmonary response to IAV in IFN signaling-sufficient and deficient mice, as the significant influx of leukocytes into the infected lungs was not observed until day 7 p.i. in all groups (Fig 2B and 2C). In contrast to Ifnar1—/—animals, monocyte recruitment in Ifngr1—/—mice was largely unaffected, suggesting that type I IFN signaling alone is sufficient to trigger monocyte recruitment into infected lungs. We next determined whether type I IFNs act directly or indirectly on monocyte populations to regulate their trafficking to lungs. Lethally irradiated CD45.1+ WT recipient mice were reconstituted with equal numbers of CD45.1+ WT and CD45.2+ Ifnar1—/—BM cells. After full reconstitution (8–10 wk), chimera mice were infected with IAV and the cellular response analyzed at day 7 p.i. (Fig 2D). We found that the defect in the recruitment of Ly6Chi inflammatory monocytes observed in Ifnar1—/—mice was not restored in CD45.2+ Ifnar1—/—cells in mixed Ifnar1—/—and Ifnar1+/+ BM chimeric mice (Fig 2E). As such, the ratio of Ly6Chi / Ly6Clo monocytes in the CD45.2+ Ifnar1—/—compartment was significantly lower than that in the CD45.1+ Ifnar1+/+ compartment, indicating that direct type I IFN signaling in Ly6Chi monocytes is required for their recruitment into infected lungs (Fig 2F). We next examined whether type I IFN signaling regulates the phenotype and function of major populations of CD11b+ monocytes in IAV-infected lungs. Consistent with current knowledge of mononuclear phagocyte subsets [11], we observed that Ly6Chi monocytes in IAV-infected lungs displayed higher CCR2 expression than their Ly6Clo counterparts in WT and Ifnar1—/—mice (Fig 3A). Furthermore, expression of the integrin LFA-1 (CD11a), a key molecule responsible for the function of monocytes [24], was unchanged on either Ly6Chi or Ly6Clo cells irrespective of type I IFN signaling. Strikingly, however, loss of type I IFN signaling on Ly6Clo but not Ly6Chi populations resulted in significant increases in the expression of MHC-II (I-A), CD11c, CD16/32 and CD64, suggesting that in Ifnar1—/—mice, Ly6Clo Mo/Mϕ display a phenotype that is characteristic of activated proinflammatory monocytes. Consistent with this hypothesis, we found that Nos2, a known proinflammatory molecule produced predominantly by monocytes, was more highly expressed in the lungs of Ifnar1—/—mice than WT animals (Fig 3B), despite the reduction in Ly6Chi monocytes in Ifnar1—/—mice (Fig 2B). Consistent with Nos2 gene analysis, flow cytometric analysis revealed that there was a significant increase in NOS2-expressing cells in the lungs of infected Ifnar1—/—mice compared to WT animals (Fig 3C). Further analysis revealed that multiple mononuclear cells in IAV-infected lungs produced NOS2 (Fig 3D). However, while Ly6Chi monocytes were the major NOS2-producing cells in WT animals, both Ly6Chi and Ly6Clo subsets contributed to the NOS2 production in the lungs of infected Ifnar1—/—mice (Fig 3D and 3E). Paired analysis of the two monocyte subsets among individual mice demonstrated that in Ifnar1—/—animals, both the percentage of cells expressing NOS2 as well as the total quantity of NOS2 produced per cell was higher in Ly6Clo than Ly6Chi monocytes, suggesting that the former subset is more susceptible to type I IFN-dependent suppression (Fig 3F). Therefore, type I IFNs are able to regulate not only trafficking, but also the phenotype and effector function of mononuclear cells in the lung during IAV infection. Since the surface markers and NOS2 examined above are known to be highly sensitive to induction by IFN-γ [25–27], we suspected that IFN-γ might play a role in the up-regulation of these molecules in the infected Ifnar1—/—mice. Indeed, at day 3 p.i., a time point at which minimal levels of IFN-γ are produced (Fig 1B), there were no differences in the expression of I-A, CD11c or CD64 on Ly6Chi and Ly6Clo cells (S2 Fig). To demonstrate directly that IFN-γ is responsible for the up-regulation of the molecules, we compared the expression of I-A and NOS2 in Ly6Clo Mo/Mϕ of WT, Ifnar1—/—, Ifngr1—/—and Stat1—/—mice at day 7 following IAV infection. We found that the enhanced expression of I-A and NOS2 observed in lungs of Ifnar1—/—animals was completely abolished in Ifngr1—/—mice as well as Stat1—/—mice that lack both type I IFN and IFN-γ signaling pathways (Fig 4A and 4B), suggesting that up-regulation of I-A and NOS2 in the absence of type I IFN signaling is driven by IFN-γ. To determine whether type I IFNs act directly on monocytes to exert their suppressive effect, we analyzed NOS2 production in IAV-infected, mixed Ifnar1+/+ and Ifnar1—/—BM chimera mice and found that type I IFNs signal directly to suppress NOS2 production in both Ly6Clo (Fig 4C) and Ly6Chi (S3A Fig) mononuclear cells. Moreover, as we observed previously in Ifnar1—/—mice (Fig 3F), Ly6Clo Mo/Mϕ in BM chimera mice expressed higher levels of NOS2 than their Ly6Chi counterparts (S3B Fig), confirming that enhanced susceptibility to type I IFN inhibition is intrinsic to this Mo/Mϕ subset. We next determined whether the enhanced IFN-γ-inducible response in IAV-infected Ifnar1—/—mice was a consequence of increased IFN-γ production or signaling by examining the expression of Ifng and Ifngr1 in infected lungs. We found that although Ifng expression was comparable in the lungs of infected WT and Ifnar1—/—mice, Ifngr1 levels were significantly higher in Ifnar1—/—animals (Fig 4D), suggesting that enhanced IFN-γ signaling rather than IFN-γ production is responsible for the increased IFN-γ-inducible response in Ifnar1—/—mice. To test this hypothesis, we analyzed the cell surface expression of IFN-γ receptor 1 (CD119) by flow cytometry and found CD119 expression was significantly increased on Ly6Clo Mo/Mϕ of Ifnar1—/—mice compared to WT mice (Fig 4E and 4F). Therefore, the intrinsic regulation of IFN-γ receptor levels by type I IFNs appears to be a mechanism by which Mo/Mϕ activation is suppressed in the presence of type I IFNs. Our findings above suggest that type I IFN signaling is dominant over IFN-γ in the recruitment of Ly6Chi monocytes to the lungs of IAV-infected mice (Fig 2). To investigate the relative contribution of type I and type II IFNs on the STAT1-dependent suppression of neutrophil migration (Fig 1D), WT, Ifnar1—/—, Ifngr1—/—and Stat1—/—mice were infected with IAV and the relative abundance of neutrophils in each mouse strain determined by flow cytometry. Interestingly, we observed that both percentage and numbers of CD11b+Ly6G+ neutrophils were significantly elevated in the lungs of Ifnar1—/—as well as Ifngr1—/—mice compared to WT animals at day 7, but not day 3 p.i. (Fig 5A, 5B and 5C). Importantly, this defect was further exaggerated in Stat1—/—mice, suggesting a synergistic function of type I and II IFNs in suppressing neutrophil recruitment during influenza infection. Moreover, in contrast to the cell-intrinsic function of type I IFNs in monocyte migration described above (Fig 2D), we found that the percentage of CD11b+Ly6CloLy6G+ neutrophils was comparable among CD45.1+ WT and CD45.2+ Ifnar1—/—compartments before (S4 Fig) and following (Fig 5D) infection, suggesting that type I IFNs regulate neutrophil migration through a cell-extrinsic manner. Therefore, type I IFNs employ distinct mechanisms to regulate monocyte and neutrophil trafficking in IAV infection and act in concert with IFN-γ to prevent accumulation of tissue-damaging neutrophils at the site of infection. To investigate potential cell-extrinsic mechanisms responsible for the increased accumulation of neutrophils, we measured gene expression of the known neutrophil-attracting chemokine Cxcl1 and the cytokine Il1b in WT and IFN-signaling deficient animals. Stat1—/—mice displayed significantly higher expression of Cxcl1 and Il1b than in either Ifnar1—/—or Ifngr1—/—mice (Fig 5E), consistent with the notion that IFNs synergize to suppress neutrophil chemotactic chemokine/cytokine production in IAV infection. STAT1 is one of many transcription factors that can transduce both type I IFN and IFN-γ receptor signaling, but it can also mediate IFN-independent functions [28–30]. Therefore, it is possible that the increased neutrophilic influx in IAV-infected Stat1—/—mice, as compared to Ifnar1—/—mice, is independent of IFN-γ signaling. Moreover, it is unclear whether IFN-γ-dependent neutrophil suppression requires intact type I IFN signaling. Type I IFNs are known to regulate neutrophil recruitment by signaling directly in inflammatory monocytes to suppress their production of neutrophil chemoattracting chemokine Cxcl2 [31]. To address these questions, we generated Ifnar1—/—BM reconstituted WT chimera mice, which are deficient in type I IFN signaling in hematopoietic cells only. Similar to Ifnar1—/—mice, these chimeric mice also displayed defective pulmonary accumulation of Ly6Chi monocytes following IAV infection (S5 Fig). Ifnar1—/—BM chimera mice treated with anti-IFN-γ antibody at day 2 and 6 p.i. exhibited a marked reduction in I-A expression on monocytes when analyzed at day 7 (Fig 6A). This is consistent with our previous observation that IFN-γ augments MHC class II expression (Fig 4) and confirms that the mAb administration successfully blocked IFN-γ activity. Importantly, consistent with the results presented in Fig 5, we found both the proportion and numbers of neutrophils to be significantly elevated in the anti-IFN-γ antibody-treated mice compared to untreated animals (Fig 6B). We also enumerated neutrophils in the bronchoalveolar lavage (BAL) and observed a similar increase in neutrophils in the airway of anti-IFN-γ antibody treated mice when compared to untreated controls (Fig 6C), indicating that IFN-γ regulates neutrophil migration into both the lung parenchyma and bronchoaveolar air spaces independent of type I IFN signaling in inflammatory monocytes. These results suggest that IFN-γ produced by the adaptive immune response to IAV infection regulates the tissue inflammatory response in addition to its anti-viral activity [1]. To test this hypothesis, we infected WT and Rag2—/—mice, which lack B and T-cells, and analyzed the IFN-γ and neutrophil response at d7 p.i.. As expected, Rag2—/—mice displayed significantly reduced IFN-γ expression (Fig 6D) and increased Ly6G+ neutrophil accumulation in the lungs when compared to WT animals (Fig 6E). Together, our findings reveal a mechanism by which IFNs link innate and adaptive immune systems to orchestrate pulmonary inflammation for the resistance to IAV infection. While it is established that IFN-mediated resistance to viral infection in vitro is dependent on the inhibition of viral replication [32], the mechanisms by which IFNs protect against infection in vivo are less well understood. Nevertheless, recent studies have suggested that type I IFN signaling is important in regulating myeloid cell migration during viral infections [17,31,33], arguing that IFNs can play a broader role in anti-viral immunity beyond their well-established cell-intrinsic anti-viral activity. In this study, we report that type I IFN signaling is necessary, but not sufficient, to control the full scale of pulmonary innate responses to IAV. To this end, the functional interplay between type I IFN and IFN-γ signaling pathways is required for the regulation of both monocyte- and neutrophil-driven pulmonary inflammation (Fig 6F). The discovery that, in the absence of type I IFN signaling, IFN-γ promotes inflammatory functions of Ly6Clo Mo/Mϕ suggests that the interplay between innate and adaptive IFNs dictates the outcome of tissue inflammation for the resistance to IAV infection. The impaired Ly6Chi monocyte migration observed in infected Ifnar1—/—mice bears similarities with Ccr2—/—mice [10,18,34,35]. Indeed, the CCR2 ligands CCL2, CCL7 and CCL12 are all type I IFN-inducible and influence monocyte recruitment in a model of chronic inflammation [36]. A key difference between Ccr2—/—and Ifnar1—/—mice, however, is that CCR2 deficiency is favourable for infection outcome [10,18] whereas Ifnar1—/—deficiency leads to enhanced mortality compared to WT animals [17]. Therefore, type I IFNs must mediate other protective mechanisms beside the recruitment of Ly6ChiCCR2+ inflammatory monocytes in IAV infection. Unexpectedly, we discovered that type I IFN signaling plays a major role in suppressing mononuclear cell activation following IAV infection. Interestingly, this inhibition is particularly effective at suppressing the pro-inflammatory potential of Ly6Clo Mo/Mϕ. Our observation that Ifnar1—/—Ly6Clo Mo/Mϕ express higher levels of CD11c, MHC class II and NOS2 than their Ly6Chi (WT or Ifnar1—/—) counterparts is unexpected. These Ly6Clo cells resemble both phenotypically and functionally the inflammatory monocytes or TNF/iNOS-producing dendritic cells (Tip-DC) described in other inflammatory conditions (reviewed in [12]). Although Ly6Chi and Ly6Clo monocytes are considered to be phenotypically distinct lineages, and their developmental maturation is still under contention [37,38], our data demonstrate that both subsets can be activated by IFN-γ to mediate pro-inflammatory actions. Given these findings, it remains to be established whether CD11b+Ly6Clo Mo/Mϕ in lungs of IAV-infected mice represent a separate lineage or have developed from Ly6Chi monocytes as a result of the down-regulation of Ly6C expression upon entry into inflamed lung tissues as described in other models [33,36]. However, the latter scenario is unlikely to be the major mechanism accounting for Ly6Clo Mo/Mϕ accumulation in the absence of type I IFN signaling in our study, because of the significant difference in the numbers of total CD11b+ Mo/Mϕ in the lungs of infected WT and Ifnar1—/—mice. Our finding that type I IFNs differentially regulate neutrophil and monocyte trafficking in IAV infection is consistent with a previous report [17]. However, while the previous study analyzed exclusively the cells in the BAL, our investigation examined myeloid populations in both the BALF and lung tissues. Furthermore, in contrast to Seo et al, we did not observe any major changes in progenitor or mature myeloid cells in the BM and blood of IAV-infected Ifnar1—/—or mixed BM chimeric mice, suggesting that the IFNs play a major role in coordinating regional immunity rather than central myelopoiesis as proposed by Seo and colleagues [17]. This discrepancy may be explained by the fact that the previous study analyzed myeloid progenitor populations following infection of BM with IAV in vitro, an event not typically known to occur during natural infection [39]. Importantly, in addition to defining the function of type I IFNs, we have uncovered two novel functions of IFN-γ in the pulmonary response to IAV infection; the inhibition of neutrophil migration and the induction of monocyte activation. Interestingly, although IFN-γ is known to elicit direct anti-viral activity in infected cells [16] and is produced in high quantities following influenza infection, its function in IAV infection has been elusive. Numerous studies investigating lymphocyte function, viral clearance or survival of mice deficient in IFN-γ or IFN-γR1 collectively reported no appreciable differences compared to WT animals [20,40–42]. Therefore the current study reveals a functional role for IFN-γ in influenza infection and suggests that some IFN-γ functions are masked by type I IFN-dependent regulatory mechanisms. The role of type III IFNs in pulmonary inflammation was not examined in this report, it is possible that the cytokines also play a role in controlling innate cell trafficking and activation during IAV infection. Indeed, a recent study revealed that type III IFNs can regulate neutrophil migration and function in an experimental arthritis model [43]. Interestingly, while type III IFNs are shown to mediate immunity to viral infections, there exists a large degree of redundancy with type I IFNs [44–46]. Future studies investigating the involvement of type III IFNs in this process may provide further insight into the regulatory functions of the IFN system. Type I IFN production has previously been reported to suppress IFN-γ driven immune responses and resistance to intracellular bacteria [25,26], but it is unknown whether a similar mechanism is activated in viral infection. We demonstrate in this study that the IFN regulatory circuit also plays a pivotal role in the host response to IAV infection, particularly in preventing Ly6Clo Mo/Mϕ from IFN-γ-induced activation by regulating IFN-γR1 expression. The tightly controlled IFN-γ signaling in Ly6Clo Mo/Mϕ may explain why under some circumstance this monocyte population differentiates into alternatively activated macrophages [24], a process known to be susceptible to IFN-γ suppression [47]. Interestingly, the type I IFN cross-regulatory mechanism described here appears to function more actively in some components of the immune response to IAV infection than others, as the inhibitory effect of IFN-γ on neutrophil migration is not suppressed during IAV infection. Nevertheless, this and previous studies collectively suggest that inhibition of IFN-γ function by type I IFN signaling is an important regulatory mechanism operating under some infection and inflammatory settings, where both type I and II IFNs are produced [48]. We propose that in contrast to intracellular bacterial infection, where activation of infected monocytes by IFN-γ is essential for pathogen clearance, inhibition of IFN-γ by type I IFNs during influenza infection serves a host-protective role. Indeed, IFN-γ-inducible nitric oxide (NO) has been shown to play a major role in mediating pulmonary pathology in IAV infection [13,49–51] and, as such, must be tightly controlled to limit immune-mediated tissue damage. IFNs are potently induced by viral pathogens and mediate host immunity to infections. In this study, we demonstrate the cooperation and cross-regulation between type I and II IFN signaling pathways coordinate a multifaceted pulmonary inflammatory response to IAV infection. Interestingly, it is known that viruses have evolved mechanisms to counteract the host IFN system [52]. Therefore, our findings suggest that in addition to impairing cell-intrinsic anti-viral effector functions as proposed previously [53], blockade of type I IFN production or signaling by viral products may lead to dysregulated inflammation, thereby contributing to impaired disease resistance and possibly increased viral transmission. This hypothesis may explain why some highly virulent strains of IAV are associated with hyper-inflammatory responses [54,55] and suggests that targeted manipulation of IFN signaling pathways could lead to new therapeutic opportunities. C57Bl/6 (CD45.2+) and CD45.1+ (B6.SJL-Ptprca) mice were obtained from the Animal Resources Centre (ARC, Perth). Rag2—/—, Ifnar1—/—, Ifngr1—/—and Stat1—/—mice (all on C57Bl/6 background) were bred and maintained at the University of Sydney Bosch Rodent Facility. All mouse work was performed according to ethical guidelines as set out by the University of Sydney Animal Ethics Committee. All experiments within this manuscript were approved under protocol numbers 2013/5847 and 2013/5848. University of Sydney Animal Ethics Committee guidelines adhere to the Australian Code for the Care and Use of Animals for Scientific Purposes (2013) as set out by the National Health and Medical Research Council of Australia. Mice were anaesthetized by intraperitoneal (i.p.) injection with 2% 2-2-tribromoethanol (Avertin) and inoculated intranasally (i.n.) with 20 plaque forming units (PFU) of influenza A virus strain PR8 (A/Puerto Rico/8/1934 H1N1) in a volume of 50 μl. PR8 virus was a kind gift from A/Professor John Stambas (Deakin University, VIC., Australia). IAV was quantified by plaque assays with MDCK cells using standard methods. In brief, 0.9 x 106 MDCK cells were seeded in each well of a 6-well culture plate. Lungs were homogenized in RPMI and clarified by centrifugation for 5 minutes at 2,000 g. Homogenates were serially diluted in RPMI and added to MDCK cell monolayers. After incubation for 45 minutes at 37°C, cells were overlaid with 1% w/v Avicel (FMC Biopolymer) in L15 media (Sigma Aldrich) containing 2 μg/ml TPCK-treated trypsin (Worthington Biochemicals) and incubated at 37°C, 5% CO2 for 3 days. Cells were subsequently washed, methanol fixed and stained with crystal violet before plaques were counted. Euthanized animals were perfused with 10 ml PBS and the lungs removed into 1 ml cold 2% FCS/RPMI. Single cell suspensions were made by dissociating the lungs with a scalpel blade and then incubated in 2% FCS/RPMI supplemented with 2 mg/ml of DNaseI (Sigma Aldrich) and Collagenase IV (Sigma Aldrich) for 30 minutes at 37°C. The digested lungs were then dissociated through a 70 μm cell strainer (Falcon) and red blood cells lysed with ACK lysis buffer (Life Technologies). Cells were counted using trypan blue exclusion. For bone marrow cells, femurs and tibias were removed and cleaned of flesh before the bone marrow was flushed out with 2% FCS/RPMI. Cells were pelleted at 300 g for 5 minutes and resuspended in ACK lysis buffer (Life Technologies) for 1 minute to lyse red blood cells. Cells were washed and resuspended in 2% FCS/RPMI before being counted by trypan blue exclusion. For BAL collection, a catheter attached to a 3-way stop cock and 5 ml syringe was inserted into the trachea and the lungs flushed with 5 x 1 ml of cold PBS supplemented with 2 mM EDTA. Cells were pelleted at 300 g for 5 minutes and the cells resuspended in 250 ul of 2% FCS/RPMI and stored at 4°C until used for flow cytometric analysis. BM chimeras were generated by lethally irradiating CD45.1+ recipient mice with 10 Gy, followed by intravenous transfer of 2 x 106 BM cells from Ifnar1—/—mice into the tail vain. For mixed BM chimeras, irradiated CD45.1+ recipient WT mice were reconstituted with a total of 2 x 106 BM cells from WT (CD45.1+) and Ifnar1—/—(CD45.2+) mice at a ratio of 1:1. Mice received antibiotic-supplemented (Trimethoprim sulpha) drinking water for 3 weeks after irradiation. Mice were used at least 8 weeks after bone marrow reconstitution. For in vivo IFN-γ neutralization, mice received 500 μg of anti-IFN-γ monoclonal antibody XMG1.2 intravenously on days 2 and 6 post IAV infection. Anti-IFN-γ clone XMG1.2 hybridoma was expanded in 10% FCS/RPMI supplemented with Pen/Strep (Gibco) and affinity purified using protein G beads (GE Healthcare). Purified antibody was dialyzed into PBS and filter sterilized prior to injection into mice. Lung (1 x 106), blood (200 ul whole blood) or BM (4 x 106) cells were stained according to standard procedures. Briefly, cells were incubated for 30 minutes with UV Live/Dead stain according to the manufacturer’s instructions (Life Technologies). Cells were stained with the following antibodies in FACS wash (2% FCS/PBS): CD4 (clone GK1.5), CD8 (53–6.7), B220 (RA3-6B2), I-A/I-E (M5-114.15.2), Ly6G (1A8), Ly6C (HK1.4), CD11b (M1/70), CD45.1 (A20), CD45.2 (104), NK1.1 (PK136), Siglec-F (E50-2440), CCR2 (475301), LFA-1 (M17/4), CD119 (2E2), CD11c (N418), CD16/32 (93), CD64 (X54-5/7.1), CD45 (30-F11), Lineage cocktail, CD117 (2B8), Sca1 (D7), CD48 (HM48-1), CD150 (Q38-480). The gating strategy used for identifying cell populations is shown in S6 Fig. Monocyte and macrophage populations were identified as CD4—CD8—B220—NK1.1—SiglecF—Ly6G—CD11b+ and then categorized into Ly6Chi and Ly6Clo populations based on Ly6C expression. Neutrophils were identified as CD4—CD8—B220—NK1.1—SiglecF—CD11b+Ly6G+. For intracellular NOS2 staining, single cell suspensions were incubated at 37°C for 3 hours prior to staining with surface marker antibodies. Intracellular staining was carried out using the BD Cytofix/Cytoperm kit according to the manufacturer’s instructions (BD). Intracellular staining was performed using NOS2 antibody (clone CXNFT). All flow cytometry data acquisition was performed on a LSRII using FACSDiva software (BD Biosciences) and all analysis was performed using FlowJo X v0.7 (TreeStar). Following perfusion, mouse lung tissue was collected and submerged in RNAlater (Ambion) for 24 hours prior to long term storage at -80°C. RNA was prepared from mouse lungs using Trisure (Bioline) according to the manufacturer’s instructions (Bioline). Total RNA (2 μg) was reverse transcribed using the Tetro cDNA synthesis Kit with random primers according to the manufacturer’s instructions (Bioline). Data are expressed as fold increases over uninfected controls and were calculated by the ΔΔCT method using 18S as the reference gene. For absolute viral nucleoprotein quantification, RNA was extracted from 1 x 107 PFU PR8 using the ISOLATEII RNA kit according to the manufacturer’s instructions (Bioline) and 100 ng reverse-transcribed with the Tetro cDNA synthesis kit using IAV nucleoprotein specific primers [39]. Following amplification, the 216 bp cDNA product was gel purified using a Gel Extraction kit (Sigma Aldrich) and total copy number determined based on size and yield of product. A standard curve was generated to determine absolute viral nucleoprotein mRNA copy number among sample mRNA. All quantitative reverse-transcriptase PCR (qRT-PCR) was performed using SYBR NoROX master mix (Bioline) on a Roche LightCycler480. Forward and reverse qRT-PCR primers are listed in Table 1. All statistical analyses were performed in Prism 6 (GraphPad Software). Significance was determined using Student’s t-test when comparing two experimental groups or one-way ANOVA followed by Tukey’s post-test correction for more than two groups. Results with p<0.05 were deemed statistically significant. *<0.05, **<0.01, ***<0.001, ****<0.0001.

10.1371/journal.pcbi.1001019

Modeling Mechanisms of In Vivo Variability in Methotrexate Accumulation and Folate Pathway Inhibition in Acute Lymphoblastic Leukemia Cells

Methotrexate (MTX) is widely used for the treatment of childhood acute lymphoblastic leukemia (ALL). The accumulation of MTX and its active metabolites, methotrexate polyglutamates (MTXPG), in ALL cells is an important determinant of its antileukemic effects. We studied 194 of 356 patients enrolled on St. Jude Total XV protocol for newly diagnosed ALL with the goal of characterizing the intracellular pharmacokinetics of MTXPG in leukemia cells; relating these pharmacokinetics to ALL lineage, ploidy and molecular subtype; and using a folate pathway model to simulate optimal treatment strategies. Serial MTX concentrations were measured in plasma and intracellular MTXPG concentrations were measured in circulating leukemia cells. A pharmacokinetic model was developed which accounted for the plasma disposition of MTX along with the transport and metabolism of MTXPG. In addition, a folate pathway model was adapted to simulate the effects of treatment strategies on the inhibition of de novo purine synthesis (DNPS). The intracellular MTXPG pharmacokinetic model parameters differed significantly by lineage, ploidy, and molecular subtypes of ALL. Folylpolyglutamate synthetase (FPGS) activity was higher in B vs T lineage ALL (p<0.005), MTX influx and FPGS activity were higher in hyperdiploid vs non-hyperdiploid ALL (p<0.03), MTX influx and FPGS activity were lower in the t(12;21) (ETV6-RUNX1) subtype (p<0.05), and the ratio of FPGS to γ-glutamyl hydrolase (GGH) activity was lower in the t(1;19) (TCF3-PBX1) subtype (p<0.03) than other genetic subtypes. In addition, the folate pathway model showed differential inhibition of DNPS relative to MTXPG accumulation, MTX dose, and schedule. This study has provided new insights into the intracellular disposition of MTX in leukemia cells and how it affects treatment efficacy.

One of the primary agents used in the treatment of childhood acute lymphoblastic leukemia (ALL) is methotrexate (MTX). By better understanding its intracellular disposition, we are able to better design treatments that circumvent drug resistance and thus help improve ALL cure rates. In this study, we develop a system of mathematical models that describe the intracellular disposition of MTX along with its inhibition of important biosynthetic pathways necessary for cell division. First, we used the models to describe the disposition of intracellular MTX in a cohort of 194 patients enrolled on St. Jude Total XV protocol for newly diagnosed ALL. The results of this modeling allowed us to determine mechanisms of in vivo variability in MTX accumulation. These mechanisms related to both the influx and efflux of the drug along with the enzymes related to its metabolism. Next, we used model simulations to show the effects of changes in MTX dose and schedule on its efficacy. The results of these simulations show that longer infusions yield better efficacy and that higher MTX doses can circumvent resistance observed in ALL subtypes with lower intracellular MTX accumulate. The results from this study provide new insights into the design of more effective therapy for pediatric ALL.

Methotrexate (MTX) is one of the primary anticancer agents used for the treatment of acute lymphoblastic leukemia (ALL) [1]–[3]. The ability of cells to accumulate intracellular polyglutamate metabolites of MTX (MTXPG) is an important factor in its antileukemic effects [4]. Specifically, MTXPG inhibits the folate pathway by competitively inhibiting several important enzymes including: dihydrofolate reductase (DHFR), thymidylate synthase (TS), glycinamide ribonucleotide transformylase (GART), and aminoimidazole carboxamide ribonucleotide transformylase (AICART). This inhibition leads to reduced or blocked TS and de novo purine synthesis (DNPS), which are needed for DNA synthesis. There is large variability in MTXPG accumulation and a variety of studies have related differences in its accumulation to ALL lineage, ploidy, molecular subtype, and folate pathway gene expression [5]–[8]. Thus, developing a better understanding of the underlying mechanisms responsible for these differences in cellular disposition of MTX is important for understanding the basis for inter-patient differences in MTX's antileukemic effects and to identify strategies to circumvent mechanisms of resistance. Pharmacokinetic and pharmacodynamic modeling is a useful approach to quantify the intracellular kinetics of MTX and to aid in understanding the underlying mechanisms related to differences in MTXPG accumulation [9]. For example, modeling can be helpful in addressing whether higher accumulation of intracellular MTXPG is related to higher formation of polyglutamates via higher folylpolyglutamate synthetase (FPGS) activity, lower degradation of polyglutamates via γ-glutamyl hydrolase (GGH), or differences in MTX influx or efflux from leukemic cells, In addition, there are numerous models describing MTX inhibition of target enzymes in the folate pathway [10]–[18], which can be exploited to advance our understanding of how folate inhibitors such as MTX alter folate homeostasis leading to its antileukemic effects. In an effort to better understand the underlying dynamics of the observed differences in MTXPG accumulation along with their differential effects on folate kinetics, we used a pharmacokinetic model to characterize the disposition of plasma MTX and intracellular MTXPG [9] along with a pharmacodynamic model to describe the dynamics of perturbations in the folate pathway [12]. These models allowed us to relate differential disposition of intracellular MTXPG to changes in transport of MTX into and out of leukemic blasts along with metabolism of intracellular MTXPG. In addition, the folate pathway model allowed us to investigate how this differential disposition of intracellular MTXPG alters folate homeostasis and its downstream consequences. Therefore, the objectives of this study were to determine the intracellular pharmacokinetics of MTXPG in circulating leukemic blasts, and to assess the relationship between these pharmacokinetic parameters and covariates including ALL lineage, ploidy, molecular subtype, and gene expression of and polymorphisms in or flanking genes related to MTX transport and metabolism. In addition, we analyzed the effects of intracellular MTXPG disposition, MTX dose, and MTX infusion schedule on the folate pathway. This study included 194 patients with newly diagnosed ALL who were enrolled on the St. Jude Total XV protocol and had sufficient circulating ALL cells to permit serial measurement of MTXPG in their leukemia cells. There were no differences in demographics, lineage, ploidy, or molecular subtype between the 194 patients and all other patients on the Total XV protocol (n = 162). Not surprisingly, diagnostic WBC counts were higher in the 194 patients than those in all other patients due to the need for sufficient circulating ALL cells to perform the MTXPG assay (Table S1). A summary of the demographic, lineage, chromosomal ploidy, molecular subtype, and randomized window therapy arm for the patients included in this study are shown in Table 1. A total of 791 plasma samples in 194 patients were assayed to determine the plasma MTX disposition. Figure 1A shows the concentration vs time plot of these data along with the population average model fit of the data sub-grouped by infusion length. The median clearance of MTX was higher in the 24 hr infusion group compared to the 4 hr infusion group (122.6 ml/min/m2 vs 108.6 ml/min/m2; p<0.001). A total of 732 peripheral blood leukemia cell samples in 194 patients were assayed for intracellular MTXPG disposition. Fixing the plasma MTX pharmacokinetic parameters to each individual's estimates, the intracellular population pharmacokinetic parameters for MTXPG were determined and the descriptive statistics of the individual estimates (conditional means) are shown in Table 2. In addition, the population estimates, relative standard error estimates of the population estimates, inter-individual variability estimates, and sensitivity analysis of the individual estimates are summarized in Table S2. Figure 1 shows the concentration vs time plot of intracellular MTX (or MTXPG1) (Figure 1B) and total intracellular MTXPG2-7 (Figure 1C) along with the population average model fit (for non-hyperdiploid B-lineage ALL) of the data. In addition, several representative plots of individual model fits to the data are shown in Figure S1. It has been previously reported that there are significant differences in intracellular MTXPG accumulation by ALL lineage, ploidy, and molecular subtype [4], [5], [19]. Using the pharmacokinetic model of the intracellular disposition in peripheral blasts of MTXPG, we quantified how differences in MTXPG disposition related to the model estimated parameters describing MTXPG influx, efflux, FPGS, and GGH activity. ALL chromosomal ploidy exhibited differences in influx and efflux parameters for MTX. Specifically, NET-influx was 2 times higher (p<0.0009) in hyperdiploid ALL compared to non-hyperdiploid ALL (Figure 2). In addition, we observed higher efflux (1.8 times higher; p<0.003) and lower NET-influx (1.5 times lower; p<0.02) in patients randomized to the 24 hr infusion compared to the 4 hr infusion. The model parameters describing FPGS activity differed significantly by ALL lineage and molecular subtype. Specifically, the maximum FPGS activity was 2.1 times higher (p<0.0002) in B-lineage ALL compared to T-lineage ALL. In addition, there was a significant difference (p<0.0002) in the maximum FPGS activity among the different molecular subtypes of ALL with the highest activity in B-lineage hyperdiploid ALL followed in decreasing order by B-lineage non-hyperdiploid, t(12;21) [ETV6-RUNX1], t(1;19) [TCF3-PBX1], and T-lineage ALL (Figure 3A). These differences translated to differential net accumulation (p<0.003) of MTXPG (NET-PG) with highest accumulation in B-lineage hyperdiploid and B-lineage non-hyperdiploid ALL, followed by t(12;21) [ETV6-RUNX1], T-lineage, and t(1;19) [TCF3-PBX1] ALL (Figure 3B). We also investigated how the MTXPG model parameters related to gene expression (mRNA) in ALL cells and germline polymorphisms in or flanking genes related to MTX transport and metabolism. These data were available for 168 and 190 of the patients, respectively. First, we assessed how MTX transporter gene expression and polymorphisms related to the model estimated parameters for MTX influx and efflux. Specifically, MTX influx (Vmax-in/Km-in) increased as the expression of SLC19A1 (probe set ID: 209775_x_at) increased (p<0.0005) and NET-influx increased as the expression of SLC19A1 (probe set ID: 211576_s_at) increased (p<0.005) (Figure 4A–B). None of the polymorphisms in or flanking transporter genes that we evaluated were significantly related to the MTX influx or efflux parameters. Next, we studied how FPGS and GGH gene expression and polymorphisms related to the model estimated parameters for FPGS and GGH activity. Specifically, net accumulation of MTXPG (NET-PG) increased as the expression of FPGS (probe set ID: 202945_at) increased (p<0.005) (Figure 4C). In addition, two SNPs upstream of FPGS (DB SNP ID: rs1544105, 2782 base pairs (bp) upstream and DB SNP ID: rs7033913, 4440 bp upstream) showed a significant relation to maximum FPGS activity (CC 2.6 times higher activity compared to TT: p<0.005; CC 2.4 times higher activity compared to TT: p<0.01, respectively) (Figure 5A–B). We simulated the effects of differential MTXPG accumulation on the MTX targets in the folate pathway to assess the effects of varying dose and schedule on these targets. We used the previously described enzyme kinetic parameters [12], the MTX and MTXPG inhibition parameters [11], along with the MTX plasma and MTXPG intracellular PK parameters defined in this study. Figure 6 depicts an individual simulation of the dynamics of the various folate components after infusion of 1 g/m2 of MTX over 24 hours. This predicted a two-fold increase in DHF, a one-fold decrease in 5mTHF, and only small changes in the remaining folate components relative to the untreated steady-state levels. Because MTXPG accumulation was significantly lower in T-lineage ALL compared to B-lineage hyperdiploid ALL, we investigated how this differential accumulation affected the inhibition of DNPS by comparing the simulated baseline DNPS activity to its activity over a 72 hr post MTX treatment interval. The simulations showed that there was both greater and longer inhibition of DNPS in the B-lineage hyperdiploid group (Figure 7A). Next we used simulations to compare the 44 hr post MTX treatment inhibition of DNPS between different doses (100 mg/m2 to 5 g/m2) and schedules (4 vs 24 hour infusions). As expected, we observed that as we increased dose the simulations predict greater inhibition of DNPS. We also observed greater DNPS inhibition for doses infused over 24 hours compared to 4 hours. Specifically, while a 1 g/m2 dose infused over 24 hours was predicted to inhibit about three quarters of the patients' DNPS more than 90%, it was predicted to take approximately a 2.5 g/m2 dose infused over 4 hours to produce the same antifolate effects (Figure 7B). MTX is one of the primary anticancer agents used to treat children with ALL and its intracellular accumulation has been shown to relate to its antileukemic effects [4]. The current study allowed us to better understand the basis of differential MTXPG accumulation and how it relates to ALL lineage, ploidy, molecular subtype, gene expression, and genetic polymorphisms. We accomplished this by developing innovative mechanistic pharmacokinetic and pharmacodynamic models of intracellular MTXPG and its interaction with the folate pathway. This gave us a new approach to describing the intracellular disposition of MTXPG (e.g. influx, efflux, FPGS, and GGH activity) along with the effects of MTXPG on the folate pathway. In addition, the model allowed us to easily test hypotheses about which factors have the strongest effects on MTXPG accumulation along with which MTX doses and schedules more effectively inhibits the folate pathway. Specifically, using the pharmacokinetic and pharmacodynamic models, we were able to evaluate a) the mechanisms of intracellular MTXPG accumulation, b) the causes of differential accumulation by lineage, ploidy, and molecular subtype, c) the difference between 4 vs 24 hour MTX infusion (validating simulations in the previous study [9] which showed that longer infusions of MTX at equivalent doses related to higher accumulation of MTXPG), d) the relations between the pharmacokinetic and pharmacodynamic model parameters and mRNA expression of and polymorphisms in and flanking related genes, and e) how MTXPG accumulation affected target enzymes in the folate pathway. We observed that net influx of MTX was highest in B-lineage hyperdiploid ALL cells which also corresponded to higher RFC expression (SLC19A1). This relation has also been observed in our previous modeling [9] and experimental studies [20]. In addition, we observed differences in the influx and efflux parameters relative to the infusion length of MTX. These differences are most likely attributed to significantly different intracellular disposition of MTXPG1 in the 4 hr infusion group compared to the 24 hr infusion group. Specifically, while the population average intracellular concentration of MTXPG1 is higher during the first 6 to 8 hours after the start of infusion in the 4 hour group compared to the 24 hr group, the concentrations fall below that of the 24 hour group for the remaining time (Figure 1B). These differences could cause an overall increase in the efflux activity for the individuals with higher intracellular concentrations over much of the treatment interval of those in the 24 hr infusion group. Next, we observed differential FPGS activity and net MTXPG accumulation with respect to ALL lineage, ploidy, and molecular subtype. This is concordant with observed differences in FPGS mRNA expression and SNPs in the gene encoding FPGS in both the current study and others [21]–[23]. The folate pathway simulations allowed us to assess the effects of differential MTXPG accumulation on the inhibition of important biosynthetic pathways that are known targets of MTX. One advantage of the modeling and simulation approach was that we could efficiently evaluate multiple situations that would otherwise be difficult, time consuming, and in many cases not practical from a clinical trials perspective. In fact, this is the first time a system of models combining the intracellular disposition of MTXPG and its inhibition of the folate pathway have been used to aid in the understanding of effective MTX therapy. There are two important issues to consider when performing modeling and simulations: the availability of and the sensitivity to the model parameters. Due to the available studies of the folate cycle, there are numerous published estimates of all the primary enzyme kinetic parameters involved (see [12] for a summary). In addition, previous studies have addressed the parameter sensitivity of these folate cycle models by showing that in most cases the effects of changes in the model parameters were local [12]. For example, it was shown that changes in the enzyme kinetic parameters for DHFR had a proportional effect on THF and a much smaller effect on other folates. Figure S2 shows plots of the effects of changes in VMAX DHFR and VMAX ACAIRT on their respective activities. For these two parameters, only VmaxACAIRT had a proportional effect on ACAIRT activity and the remaining effects were all minimal. Thus, this effect is considered a local effect. Therefore, the folate model is not sensitive to the parameter choice for VMAXDHFR and only locally sensitive to the parameter choice for VmaxACAIRT. These simulations helped increase our understanding of how MTXPG accumulation, MTX dose, and MTX schedule affect antileukemic effects. In addition, our simulation results compared qualitatively to previously published studies on MTX inhibition of target enzymes, further validating them. Specifically, the simulation which compared differential accumulation of intracellular MTXPG by ALL lineage showed that in the T-lineage group only about half of the individuals had DNPS inhibition greater than 90% at 44 hrs compared to more than three quarters of the individuals in the B-lineage hyperdiploid group. Also, about half the B-lineage hyperdiploid individuals' DNPS was still inhibited greater than 90% by 72 hours post treatment. These two results are in line with our previously published results [24] which showed that individuals with higher MTXPG accumulation were more likely to achieve full inhibition of DNPS (defined as inhibition greater than 90%). In addition, the simulations describing the effects of MTX dose and schedule showed that there was increased inhibition of DNPS with larger doses and longer infusion schedules. These results are in line with our current clinically measured changes in DNPS in the subset of our patients in which DNPS was directly measured (unpublished data). These results suggest that higher doses of MTX are needed to obtain similar inhibition patterns with shorter (4H) compared to longer (24H) infusions. A recent COG study randomized patients with ALL to receive either a 2 g/m2 dose infused for 4 hours or a 1 g/m2 dose infused for 24 hours [25]. The results of this study have yet to be reported, but they will provide treatment outcome data that will complement the current study. In summary, our pharmacokinetic and pharmacodynamic model of plasma MTX, intracellular MTXPG, and the folate cycle provides an important new tool for elucidating mechanisms underlying inter-individual differences in MTXPG intracellular disposition and inhibition of target enzymes. Furthermore, this model permits assessment of how the dosage or schedule of MTX administration alters the delivery of active drug to leukemia cells of different lineage and molecular subtypes. This will facilitate the design of more effective therapy for pediatric ALL. A total of 356 patients were enrolled on St. Jude Total XV protocol for newly diagnosed ALL between 2000 and 2007 which stratified and randomized patients to receive MTX during the first day of therapy [26]. This study included the 194 patients who had adequate circulating leukemia cells for intracellular MTXPG quantification at 3 to 4 serial time points during the initial 42 hours of therapy. The institutional review board approved the study, and informed consent was obtained from parents/guardians or patients. This study was compliant with the regulations of the Health Insurance Portability and Accountability Act of 1996 (HIPAA). Patients were randomized during the first day of therapy to receive either 1 g/m2 MTX infused intravenously over 24 or 4 hours. Serial plasma samples were obtained at 1, 4, 24, and 42 hours after the start of the MTX infusion and MTX concentrations were assayed by the Abbottbase TDx-FPIA II assay (Abbott Diagnostics, Irving, TX). In addition, circulating leukemia cells were obtained at 1, 4, 24, and 42 hours after the start of the MTX infusion. Intracellular concentrations of MTXPG were assayed by HPLC as previously described [6], [19]. The pharmacokinetic model used to describe the plasma MTX was a first-order two-compartment model (see first two equations in (1)). The pharmacokinetic model to characterize the intracellular disposition of MTXPG was previously described [9]. Briefly, it involves two compartments, one for the intracellular concentration of MTXPG1, or intracellular MTX (the third equation in (1)), and the second for the intracellular concentration of MTXPG2-7, the sum of MTXPG2 through MTXPG7 (the fourth equation in (1)), where the subscripts denote the number of glutamates attached to each MTX molecule. A diagram of the model is shown in (Figure 8A) and the model is described by the following system of ordinary differential equations:(1)The parameters are defined as follows: ke, k12, and k21 (1/hr) are the first-order parameters describing the elimination of plasma MTX and the transition between the central (MTX) and peripheral (MTXp) compartments respectively; V (L/m2) is the systemic volume; Vmax-in (pmol/109 cells/hr) and Km-in (µM) are the Michaelis-Menten parameters describing the active influx of MTX into the leukemic blasts via the reduced folate carrier and various ABC transporters; kp (1/hr) is the first-order passive influx parameter; keff (1/hr) is the first-order efflux parameter; Vmax-FPGS (pmol/109 cells/hr) and Km-FPGS (pmol/109 cells) are the Michaelis-Menten parameters describing the FPGS activity; and kGGH (1/hr) is the first-order parameter describing the GGH activity. In addition, we defined several secondary parameters which were combinations of the above parameters. These included: Influx (Vmax-in/Km-in); NET-influx (Vmax-in/keff), the ratio of maximum influx activity to efflux—the net influx of MTX; FPGS (Vmax-FPGS/Km-FPGS); and NET-PG (Vmax-FPGS/kGGH), the ratio of maximum FPGS to GGH activity—the net accumulation of MTXPG. We assumed that the amount of drug in the plasma significantly exceeded the intracellular amount. Thus, we did not account for the intracellular drug efflux into the plasma. This allowed us to uncouple the system of four differential equations to two independent systems—one for the plasma pharmacokinetics and the other for the intracellular pharmacokinetics. First we estimated the plasma MTX pharmacokinetics using the maximum a posteriori probability (MAP) parameter estimation method implemented in ADAPT 5 [27] along with the prior parameter distribution obtained from previous St. Jude Total protocols [4]. Then, fixing, per individual, these plasma pharmacokinetic parameters, the intracellular MTXPG model parameters (both population estimates and individual conditional means) were determined using the Monte Carlo Parameter Expectation Maximization (MCPEM) [28] with importance sampling population estimation algorithm in ADAPT 5 [27]. This approach was used since, unlike the plasma pharmacokinetics where we had abundant prior parameter information from previous studies, minimal prior information on the distribution of the intracellular MTXPG model parameters was available. Due to the lack of identifiability of the passive influx parameter kp we fixed it to 0.4 (1/hr) —its previously reported value [29]. Finally, due to the known significant differences in the intracellular disposition between B and T-lineage ALL, we fit each lineage group separately in the population model. The individual conditional means were used for comparison to covariates and for the below described folate pathway simulations. The percent relative standard error of the population estimated parameters, as determined in ADAPT 5, was used to assess their sensitivity. In addition, the individual conditional means were estimated ten times using randomly chosen initial parameter values for each run. From these runs the sensitivity of the individual conditional means to changes in initial parameter values was determined by calculating their average relative absolute error. The model used to characterize the folate pathway was taken from Nijhout et al. [12] and modified to include the inhibitory effects of MTXPG on target enzymes (Figure 8B; equations in Figure S3). Specifically, MTXPG was modeled to stoichiometrically inhibit DHFR, TS, and AICART/GART via competitive binding. We simulated the effects of MTXPG on the folate pathway in each patient in the current study by using their respective MTX plasma and intracellular MTXPG model parameters along with published folate pathway enzyme kinetic parameters [11], [12]. We considered simulations over the dose range from 100 mg/m2 to 5 g/m2 and with a 4 or 24 hr infusion. Gene expression in ALL cells at diagnosis and germline SNPs in or flanking (within 10,000 bp of the gene) folate transporter (SLCO1B1, SLC19A1, ABCC1, ABCG2) and polyglutamation (FPGS, GGH) genes were determined by Affymetrix HgU133A Human GeneChip arrays and by Affymetrix 500K mapping array genotyping as previously described [5], [30], [31]. Differences in the individual pharmacokinetic model parameters (e.g. the conditional means determined by the above described methods) due to lineage, ploidy, molecular subtype, gene expression, and SNPs were determined by either the Kruskal-Wallis ANOVA or the Mann-Whitney U-test.

10.1371/journal.pntd.0000927

Factors Associated with the Prevalence of Circulating Antigens to Porcine Cysticercosis in Three Villages of Burkina Faso

Little is known about porcine cysticercosis in Burkina Faso. We conducted a pilot study to estimate the prevalence of antigens of Taenia solium cysticercosis and to identify associated factors in pigs of three villages in Burkina Faso, selected to represent different pig management practices: one village where pigs are allowed to roam freely (Batondo), one village where pigs are penned part of the time (Pabré) and one village with limited pig farming (Nyonyogo). A clustered random sampling design was used. Data on socio-demographic characteristics (source of drinking water, presence of latrines in the household, type and number of breeding animals) and pig management practices were collected using a standardized questionnaire. Blood samples were collected from one pig per household to determine the presence of antigens of the larval stages of T. solium by the B158/B60 Ag-ELISA. The associations between seropositivity and socio-demographic and pig management practices were estimated using logistic regression. Proportions of 32.5% (95% CI 25.4–40.3), 39.6% (31.9–47.8), and 0% of pigs, were found positive for the presence of circulating antigens of T. solium in Batondo, Pabré, and Nyonyogo, respectively. The results of the logistic regression analyses suggested that people acquire knowledge on porcine cysticercosis following the contamination of their animals. The presence of antigens in the pigs' sera was not associated with the absence of latrines in the household, the source of drinking water or the status of infection in humans but was associated with pig rearing practices during the rainy season. The results suggest that education of pig farmers is urgently needed to reduce the prevalence of this infection.

Taenia solium cysticercosis is a neglected tropical infection transmitted between humans and pigs. This infection is particularly common in areas where sanitation, hygiene and pig management practices are poor, and can sometimes lead to epilepsy in humans. There is very little information about the importance of this infection in Burkina Faso, even though pork meat is widely consumed in many villages. We conducted a pilot study in three villages: two villages where pig rearing and pork consumption are common (Batondo and Pabré) but with different pig management practices, and one village with limited pig farming and pork consumption (Nyonyogo). Blood tests were done on pigs and information on pig raising was collected from farmers. Our study demonstrated that at least one third of pigs are infected with cysticercosis in villages where they are raised, and, particularly when pigs are left to roam some or all of the time. It also demonstrated that farmers may not be aware of this disease until one of their animals is found to be infected. Thus, the study concluded that there is an urgent need for improving education in order to control this tropical disease.

Burkina Faso is one of the poorest countries in the world ranking 177th out of 182 according to the Human Development Index (HDI) [1]. Its economy relies predominantly on agriculture (40% of GDP), with cotton production being the most important, followed by livestock production, accounting for 12% of the GDP [2]. The main livestock species are cattle, small ruminants, poultry and pigs. The relative importance of each species varies across the thirteen regions of the country. The country's pig population is estimated at 2 million, and is more concentrated in the regions of Centre West, South West, and Boucle du Mouhoun (with 45% of the pig population) [3]. Pigs are kept mainly in a traditional way, with most being either tethered or allowed to roam freely for some time during the year (Figure 1). Porcine cysticercosis is a parasitic zoonosis caused by the larval stage of Taenia solium. While pigs are the intermediate host, man is the only natural definitive host. Human tapeworm carriers shed thousands of eggs daily through their feces. Pigs usually get infected by eating infected human feces or by consuming feed or water contaminated with human feces. Humans can also become accidental intermediate hosts upon ingestion of T. solium eggs. In both humans and pigs, the larval stage of T. solium can establish in the muscles and/or in the brain, the latter resulting in neurocysticercosis (NCC). Human NCC may lead to acute seizures, epilepsy and other neurological manifestations [4]. Ingestion of larvae (cysticerci) present in raw or under-cooked pork may result in human tapeworm infection. Porcine cysticercosis is common in developing countries where pigs are raised. Porcine cysticercosis is associated with poor sanitation and hygiene (absence of latrines, defecation in pigpens, poor handwashing practices), poor methods of pig husbandry (free-roaming), lack of meat inspection, and poor knowledge of the disease [5]–[11], all of which are associated with poverty. In order to avoid economic losses due to condemnation of infected pig carcasses in places where meat inspection is performed [10], farmers may sell contaminated pigs, either alive or for clandestine slaughtering. Information on porcine cysticercosis in Burkina Faso is very limited. Coulibaly and Yameogo (2000) [12] reported a prevalence of cysticercosis of 0.6% among pigs inspected at slaughter. Based on data from the main abattoir in Ouagadougou for 2005 to 2007, of 10,505, 12,651, and 11,887 pigs slaughtered during each of these three years, respectively, only 8 (0.08%), 13 (0.10%), and 10 (0.08%) were condemned because of cysticercosis; whereas 472, 186, and 133 carcasses, respectively, were reportedly condemned because of illegal slaughtering. These data clearly show that inspection of pork is very poor in Burkina, and reliable prevalence of porcine cysticercosis is lacking in the country. The aims of this study were therefore: 1) to estimate the prevalence of T. solium cysticercosis among pigs in three villages with distinct pigs raising techniques, and 2) to measure the association between potential risk factors and the prevalence of infection in pigs. We conducted a cross-sectional study in three villages including serological detection of active cysticercosis in pigs and a questionnaire survey for the identification of potential risk factors. In two villages, Batondo and Pabré, located 140 km west and 20 km north of Ouagadougou (the capital of Burkina Faso), respectively, pig breeding and pork consumption are very common. In a third village, Nyonyogo, located 30 km north of Ouagadougou, both pig breeding and pork consumption are rare. The sampling of pigs took place between 26 May and 29 October 2007, which corresponds to the rainy season and the very start of the dry season. The sampling was implemented by a research team composed of one physician, one veterinarian, two interviewers, and one translator (for Batondo only). The first step in the sampling process was to determine the sampling frame. This was done by using information from the national population census conducted in 2006 during which all villages were divided into enumeration areas (EA), a geographical unit which is intended to include approximately 1,000 persons. In each of the three villages, the sampling process started by identifying the geographic limits of each EA. In each EA, each concession (a grouping of several households, usually members of the same family) was numbered. In Batondo (4 EAs) and Nyonyogo (3 EAs), all the concessions were included because of the small number of concessions present. In Pabré (5 EAs), 50% of all concessions were randomly selected. In each concession, all the households were invited to participate. The head of each household was asked to list all the members of the household and the mother or the oldest woman (in the case of polygamy) was asked questions regarding selected characteristics of the household, such as the source of drinking water, the presence of latrines, breeding livestock, and cooking pork. The person in charge of pig farming was asked about how pigs were managed and slaughtered. Knowledge of porcine cysticercosis was assessed through questions about seeing lesions in the meat on dead animals or cysts under the tongue of live animals. Before the data collection started, the questionnaires were validated during a pre-test conducted in Tenado, a village located eight km from Batondo. One pig was randomly sampled per household for blood sample collection. Blood samples were also collected in humans (one person randomly selected per household) (see [13] for more details on the human component of the study). All questionnaires are available on request. Blood samples were left to decant and the sera were transported to the Institut de Recherche en Sciences de la Santé (IRSS) in Bobo-Dioulasso where they were kept at −20°C until analysis. The serum samples were tested in duplicate with the enzyme-linked immunosorbent assay (ELISA) for the detection of circulating antigens of the metacestode of T. solium (Ag-ELISA) [14]. Sera showing a coefficient of variation of more than 50% between the duplicate results were considered as missing values (n = 10). The Ag-ELISA has been reported to have a sensitivity of 76.3% (95%CI: 60.9%–88.6%), 86.7% (95%CI: 62.0%–98.0%) and 85.8% (95%CI: 71.9%–99.7%) and a specificity of 84.1% (95%CI: 74.4%–93.3%), 94.7% (95%CI: 90.0%–99.7%) and 98.9% (95%CI: 97.3%–100%) in pigs in South Africa, Zambia and West Cameroon, respectively [6], [14], [15]. First, the relative frequencies of categorical variables and the mean and extreme values of quantitative variables were calculated. Wealth quintiles for households were derived from a score calculated based on the asset ownership of each household (e.g., source of drinking water, presence of toilet, radio, bicycle, presence, type and number of livestock, types of roof and walls) using factor analysis as described by Gwatkin et al. (2000) [16]. Secondly, bivariate analyses were conducted to assess the associations between pigs' serological status and location (village), human socio-demographic and health factors (source of drinking water, presence of latrine, presence of antigens to the larval stages of T. solium, wealth quintile), and pig management factors. Finally, a multiple logistic regression was used to estimate the adjusted associations between the factors mentioned above and the presence of antigens to T. solium in pigs. Data from Nyonyogo were excluded from the logistic regression because of the small number of pigs in the village (only 6 pigs were sampled there). All the analyses were performed using STATA9 software. The prevalence of infection and 95% Bayesian credible intervals (BCI), adjusted for measurement error in serology, were also estimated in Batondo and Pabré using a Bayesian approach suggested by Joseph et al. (1995) [17]. We used two sets of priors for the sensitivity and specificity of the AgELISA test based on two studies conducted in infected pigs, one in Zambia and one in South Africa [6], [14]. The animal component of this protocol was approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Oklahoma Health Sciences Center. No IACUC exists in Burkina Faso. The human component of this protocol was reviewed and approved by the ethical committee of the Center MURAZ (Ref. 02-2006/CE-CM) and by the Institutional Review Board of the University of Oklahoma Health Sciences Center (IRB# 12694) for both human and porcine participants. Informed consents for the interviews of participants and the provision of blood samples were obtained separately. The consent process was done orally because a very large proportion of the population had never been to school (62.5%). Oral consent was documented on the individual consent forms by the research staff (fingerprints were collected in place of signatures). Both IRBs approved the use of oral consents. Socio-demographic characteristics of the three villages are described in Table 1. The majority of people in Batondo and Pabré drank water from traditional or open wells, whereas in Nyonyogo, most used water from bore wells. There were a few households in Batondo (1.2%) and Pabré (0.3%) in which people drank water from the river, marsh or spring. More than one-third of households in Pabré had latrines in contrast to Batonto and Nyonyogo where latrines were very rare. Wealth quintiles classified approximately one-half (50.9%) of households in Batondo among the poorest compared to around 1% in Pabré and Nyonyogo. In contrast, nearly 30% of households in Pabré and Nyonyogo were classified among the wealthiest compared to only 4.4% in Batondo. Characteristics related to pigs and pig management are reported in Table 2. Pigs were raised in two-thirds (66.4%) of the surveyed households in Batondo and in more than half (52.8%) of the surveyed households in Pabré. In Nyonyogo, most surveyed households owned ruminants and only 3% owned pigs. In Batondo, pigs were almost exclusively cared for by women (98.8%) whereas in Pabré, pigs were taken care of by men in nearly two-thirds of cases (63.1%). Similarly, the mother or eldest woman of the family was in charge of pigs in 71.7% of surveyed households in Batondo, but in only 25.0% of surveyed households in Pabré. Children were more often in charge of pigs in Pabré (10.5%) than in Batondo (1.8%). More than one-half of the surveyed households owning pigs in Pabré kept them penned during the rainy season. This was in contrast to Batondo where only 10.4% were kept penned, although over 90% were tethered. During the dry season, pigs were left to roam in all three villages to scavenge around cropped fields. A total of 336 pigs were sampled: 173 in Batondo, 157 in Pabré, and 6 in Nyonyogo. Four sera in Batondo, 3 in Pabré, and 1 in Nyonyogo were not tested because of insufficient volume. Using prior information on the sensitivity and specificity of the AgELISA from South Africa, the adjusted estimates of the seroprevalence of infection in Batondo and Pabré were 32.7% (95% BCI: 12.1%–54.3%) and 48.3% (95% BCI: 27.9%–74.8%), respectively. Using the priors from Zambia, the adjusted estimates of the seroprevalence of infection in Batondo and Pabré were 32.7% (95% BCI: 25.4%–68.3%) and 48.2% (95% BCI: 35.4%–82.6%), respectively. A total of 1.9% and 3.4% were found to be borderline positive in Batondo and Pabré, respectively. Neither serologically positive nor borderline animals were found in Nyonyogo. The results of the multivariable logistic regression are reported in Table 3. Allowing pigs to roam during the rainy season was associated with an increased prevalence odds of infection when compared to keeping pigs penned (POR = 6.48, 95%CI: 1.23–34.17). Being aware of pig cysticercosis was also associated with an increased prevalence odds of infection (POR = 4.70, 95%CI: 1.76–12.52). The village, while not statistically significant, was retained in the model because it was a confounding factor for most of the other variables explored, including those retained in the final model. Once the model is adjusted for pig management during the rainy season and awareness of the infection, the prevalence odds of infection tended to be higher in Pabré than in Batondo (OR = 1.63, 95%CI: 0.94–2.83). This is the first community-based study conducted in Burkina Faso to report seroprevalence of porcine cysticercosis. Our results show that cysticercosis is highly prevalent in the two villages where pigs are commonly raised. These estimates are nearly 100 times higher than the previous estimate of 0.57% based on reports from meat inspection services [12]. This finding supports the opinion of Zoli et al. (2003) [10] that data from meat inspection are not representative of the real situation, mainly because the method is insensitive [14] and depends on the technical skills and the motivation of the examiner. Pigs infected with cysticercosis may also never be sold in the official market. One-third (32.5%) of the pigs in Batondo and more than one-third (39.6%) in Pabré were infected, but none of the five sampled pigs in Nyonyogo was found to be positive. The latter is not surprising since pig breeding and pork consumption are very rare in this village. Similarly, Carabin et al. (2009) [13] reported very few cysticercosis seropositive inhabitants in Nyonyogo, and the few positives had all borderline reactions to the Ag ELISA. Using priors from either South Africa or Zambia to adjust for measurement error resulted in the same median prevalence, but 95% BCI were shifted towards higher values when using the Zambia priors. This is because both the prior values for sensitivity and specificity from the Zambian study were higher than the South Africa values. The B158/B60 AgELISA test used here was recently shown to have the highest sensitivity and specificity for detection of cysticercosis-infected pigs when compared to tongue examination, EITB, and HP10 Ag-ELISA [6], [14]. The seroprevalences in Batondo and Pabré were comparable to that reported in pigs in Mozambique [5] but were higher than that reported in several other African countries [9], [15], [18], [19] and lower than the seroprevalence reported in pigs slaughtered on the clandestine market in Lusaka, Zambia [14] and in the Eastern Cape province in South Africa [6]. Pigs allowed to roam some of the time during the rainy season were more likely to be seropositive than pigs penned during all of the rainy season. This association was statistically significant even though the proportion of pigs allowed to roam during the rainy season was small because the association was so strong (POR = 6.48, 95% CI 1.23–34.17). Given the imprecision of this estimate, however, the association will need to be further explored. This observation agrees with the findings of Pondja et al. (2010) [5] who identified the free-range pig husbandry system as an important risk factor for porcine cysticercosis. Conducting this pilot study lead to an appreciation that the pig management systems in Batondo and Pabré were different from what had been reported to us when the study was being planned. Indeed, in both of the two villages with pigs, animals were left roaming during dry seasons but restrained to one degree or another during crop production period (rainy season). The main difference in restraints was that pigs were confined in pens in Pabré and tethered to pillars in Batondo. The difference in pig management, mainly during the rainy season, between Batondo and Pabré might reflect the difference in who owns and is responsible for raising the pigs. In Batondo, almost all pigs were owned by women whereas they were kept by men in two-third of the cases in Pabré. This result suggests that the target for intervention when planning future cysticercosis control strategies depends on who is responsible for pig raising in any given village. Pigs owned by people who had heard of porcine cysticercosis were more likely to be seropositive than owners who were unaware of this infection. If pigs raised by the same farmers tend to become re-infected, this would suggest that people become aware of porcine cysticercosis when their pig is found to be infected. Despite the high frequency of households with latrines in Pabré and the low prevalence of active infections in humans [13], the seroprevalence in pigs in this village tended to be higher than that in Batondo. This result is counter-intuitive but could be explained by the following. First, the presence of latrines does not certify their use because of foul odors, flies, flooding, etc. Second, latrines may be inappropriately built (closing door, septic tank) allowing access of pigs to human feces [20]. More research is needed to explain this observation. It was anticipated that there would be an association of pig seropositivity and variables such as presence of a toilet in the household, household source of drinking water, and human serological status, but this was not confirmed by the statistical analysis. Indeed, the structure of community life in these villages means that any risk associated with lifestyle in a family is shared by other families nearby. For example, if a given family does not have access to latrines and thus defecates in nature, this will also affect the members of families equipped with latrines through environmental contamination. People with taeniasis may also infect a large number of pigs if they do not use latrines. Lescano et al. (2007) [21] demonstrated that a T. solium carrier can infect pigs within an area of 500 meters or more. In conclusion, a high prevalence of porcine cysticercosis was observed in pig populations in Burkina Faso in villages with traditional pig-rearing practices. The co-occurrence of knowledge about cysticercosis and owning seropositive pigs suggests that information about this disease may not be available to livestock managers until the disease is already present. If so, this emphasizes the clear need for improving education in order to control this zoonosis.

10.1371/journal.pcbi.1000992

Finding the Way with a Noisy Brain

Successful navigation is fundamental to the survival of nearly every animal on earth, and achieved by nervous systems of vastly different sizes and characteristics. Yet surprisingly little is known of the detailed neural circuitry from any species which can accurately represent space for navigation. Path integration is one of the oldest and most ubiquitous navigation strategies in the animal kingdom. Despite a plethora of computational models, from equational to neural network form, there is currently no consensus, even in principle, of how this important phenomenon occurs neurally. Recently, all path integration models were examined according to a novel, unifying classification system. Here we combine this theoretical framework with recent insights from directed walk theory, and develop an intuitive yet mathematically rigorous proof that only one class of neural representation of space can tolerate noise during path integration. This result suggests many existing models of path integration are not biologically plausible due to their intolerance to noise. This surprising result imposes significant computational limitations on the neurobiological spatial representation of all successfully navigating animals, irrespective of species. Indeed, noise-tolerance may be an important functional constraint on the evolution of neuroarchitectural plans in the animal kingdom.

The ability to navigate allows animals to vastly increase the action space for finding resources, mates, and to avoid predators. The benefits are many and it is commonly believed that modern brain functions have emerged from ancestral forms evolved for effective navigation. Since the time of Charles Darwin, it has been recognized that path integration is a navigation strategy innate to many species. Path integration involves adding the stepwise displacements during a circuitous journey to compute a net homeward direction. Over the past century, this phenomenon has been described for birds to mammals to arthropods, and a long list of mathematical, algorithmic, and neural network models have been proposed to explain the necessary computations. This work shows how the different types of models behave in the presence of noise. It turns out that only one class of models can function properly in the presence of noise. Since noise appears to be present at all levels of brain physiology, we arrive at the surprising conclusion that the general computational principles for path integration must be the same across all species. Two subtypes of path integration models share the same critical computational principles, and are compared to known neuroanatomy and physiology.

In nature, successful navigation is vital for survival. It follows that neural circuitry capable of carrying out navigation must be ubiquitous in the animal kingdom. The study of animal navigation, therefore, is not only important in its own right, but may offer general insights into the architecture, computational algorithms and evolutionary history of modern nervous systems. For convenience and consistency with previous work, we use the term ‘navigation’ in a general sense to encompass all forms of non-random locomotion, including biological path integration (PI) or ‘dead reckoning’, a process described by Charles Darwin in 1873 [1]. Darwin realized that documented feats of navigation amongst the local inhabitants of Northern Siberia were likely to have been achieved by mentally keeping track of the changes in heading and distances travelled. This observation was significant as it distilled navigation into a concise computational problem which could be tested experimentally and formalized mathematically. PI is arguably the simplest navigation strategy which requires a neural representation of space. In contrast, strategies such as chemotaxis or view-based homing, although biologically significant, do not necessarily allow us to probe at the neural representation of space. Since Darwin's time, much knowledge has accrued about the neuroethology of navigation, including PI and landmark-based navigation. With advanced in vivo recording and measurement techniques, a number of likely neuronal correlates of navigation have been identified [2]–[6]. Despite a plethora of data, it is still unclear even in principle how animals represent space, especially across the phylogenetic expanse. In fact, it is completely unknown whether there is any underlying reason for different species to obey the same rules. PI seems to be an ideal process for investigating the neural representation of space since it maintains a continual record of position in space. Systematic probing of this record could theoretically define the complete mapping between real and representational space. Furthermore, PI-related behaviour has already been documented in a wide variety of animal species [see 7], and it seems plausible that some sort of PI system may exist in most nervous systems capable of navigation. Finally, a consistent representation of space may simplify the computations necessary for combining different navigation strategies to generate a single coherent output. This supports the hypothesis that the entire neural representation of space is likely to be the same as that used for PI, based on the principle of reusing existing circuitry as well as computational parsimony. Such arguments have specific biological and modelling implications in light of the theoretical results of this work, and will be discussed further below. Tolman's cognitive map may be the first serious theoretical formulation of the spatial foundation for navigation in any animal [8]. More than two decades elapsed before the discovery of hippocampal place cells [2], which have widely been considered to be the neurophysiological correlates of the ‘cognitive map’. The more recently discovered medial entorhinal grid cells [5], [6] have already gained a remarkable level of agreement to be the neural substrate of mammalian PI [6], [9]–[13]. In contrast, neural correlates of arthropod navigation have been difficult to find, in part due to technical limitations. Nonetheless, lesion experiments suggest the mushroom bodies of cockroaches may serve similar navigational functions as the mammalian hippocampus [14]. Furthermore, the central complex of the locust has a topographical architecture with directional tuning [15], functionally reminiscent of the rodent head direction system [16]. In the arthropod literature, a vast body of behavioural evidence exists for the use of PI as a fundamental strategy of navigation, but concurrent neurophysiological data are lacking. In mammals, there is an abundance of place cell and grid cell data, showing firing fields which strongly correlate with spatial locations. These are, prima facie, the best neuronal correlates of spatial representation known. However, most of these data are obtained from animals navigating in artificial, relatively simple and spatially restricted arenas. Furthermore, the relevance of these cells has been called into question in a variety of navigation paradigms [17]–[19]. It is unclear whether strong conclusions can be drawn from either arthropod or mammalian data with respect to the true nature of the neural representation of space. Moreover, experimental and behavioural data on what exists in nature does not necessarily answer why. Our work addresses this important question by gaining an in-depth theoretical understanding of whether PI places any constraints on a neural representation of space. It turns out that the single assumption that all nervous systems (including biological PI systems) are susceptible to noise, is sufficient to differentiate existing PI models on a functional/behavioural basis. It has been shown that under ideal, noise-free conditions, a range of mathematical and neural models of PI are quantitatively equivalent for updating trajectories through metric space, and that the equivalence could be extended to descriptions of steering, searching behaviour and even account for observed systematic errors [7]. This is unambiguous theoretical validation of the wide range of models which have emerged as candidates for arthropod PI. However, since the alternative models behave equivalently, they have equal explanatory value. Is there any principled way of differentiating between the models? To properly answer questions about animal navigation, we need to build an understanding of species-independent truths about neurobiological spatial representations. Here, we approach this problem from a theoretical perspective. From first principles, we show how different PI systems will behave in the presence of noise. Using a general classification scheme for spatial representations during PI [7], we show how imperfections or noise in different neural representation of locomotion results in distinct outcomes, corresponding to two distinct types of directed walks [20], [21]. Only one type of directed walk, and hence the corresponding neural representation of space, can faithfully capture the real trajectory using PI. The other representations yield irrecoverably large errors, rendering the PI system useless beyond a few steps. Finally, we apply our understanding of directed walks to discuss the implications of the results on the neurobiology of PI and navigation. The results presented in this work can be understood as a mapping of the results of directed walk (DW) theory from a walk carried out in physical space to a walk or sequence of representational states taking place within the nervous system of an animal navigating by PI. The results section introduces the details of the mapping process, and establishes a strict equivalence between the physical and representational walks. The conclusions then follow automatically from the previously demonstrated results of DW theory. Of the two types of DW, only one can tolerate noise without quickly degenerating to the point where the animal is lost. Of the four classes of spatial representation, only one is equivalent to the robust type of DW and is therefore the only class tolerant of noise during PI. The theoretical insights arise from the combination of 1) the generally accepted assumptions of sensorimotor noise and process noise within the nervous system, 2) theory of directed walks, and 3) a recently developed classification scheme for PI systems. The contributions of each component to the final results are explained and justified next. Some technical details have been omitted for clarity but can be found in the supplementary material and listed references. Noise from the environment, the internals of a navigating agent, and the interface between the two, can all contribute to positional uncertainty during a navigation task. It has been demonstrated using simulations [22] and theoretical proofs [20], [21] that the type of directional cue used for PI is critical i.e., an external compass is a necessity for successful PI. Here we focus on the PI system per se rather than purely sensory or motor noise (which of course are inevitable). The motivations and mathematical implementations are described below. Biological noise can arise from a variety of sources within a nervous system [23]. From neurotransmitter diffusion, to ion channel kinetics, to action potential timing, stochastic behaviour appears to be pervasive throughout biological neural networks. Nevertheless, there is also growing evidence that neural systems have evolved near optimal systems-level solutions to common problems, even where optimal solutions may seem implausibly complex in explicit mathematical terms [24]. Therefore we ask whether certain types of neural representations of space may be superior in some way, in the presence of noise. In this work, we consider two major sources of noise, namely sensor noise (δ) which leads to imperfect inputs, and representational noise (ε) which manifests itself during the updating step of PI. We do not assume specific characteristics about the noise, but simply that it exists. In essence, we algebraically corrupt the input and updating processes with noise and consider the effects. While noise tolerance is obviously a key issue for any biological navigation system, the relevance of our results to non-biological systems is not so clear. Robotic systems commonly use representations allowing a very high degree of precision (typically 16 significant places or more) which are generally operated on by algorithms which do not introduce any random errors during processing. These are the kinds of conditions under which the fundamental mathematical equivalence between all spatial representations might lead to identical performance regardless of the spatial representation used [7]. Nonetheless, sensorimotor noise or rounding errors may still differentially affect the performance of non-biological spatial representations, especially for long journeys or where extreme precision is required, but their properties are subtly different to their biological counterparts (rounding errors are not truly random, sensory values can be sampled once and stored for future access without subsequent degradation) which are beyond the scope of the present account. Here we focus on navigation in the context of biological nervous systems where machine-level precision is implausible and where, even in the absence of noisy sensory data, the addition of noise by the representational system leads to positional uncertainty. Mathematically, we model two types of independent, random errors which are assumed to accrue during PI. It is of course true that noise may arise anywhere in a neural network subserving PI. Indeed, every computational and/or network variant will accrue noise in subtly different ways, leading to quantitative variations in the magnitude of PI errors. For instance, an update algorithm which requires multiple feedback steps, particularly if signal integration is involved, may well result in greater susceptibility to noise. This idea was the basis of an argument that polar representations are computationally inferior to Cartesian ones [25]. However, this type of reasoning contains at least three weaknesses: firstly, an explicit PI update algorithm needs to be assumed for each spatial representation in order to quantify the number of feedback loops, and therefore susceptibility to noise. For instance, what if the neural implementation of a Cartesian PI system actually required many more computational steps than the explicit mathematical version? Secondly, even if all mathematical operations had simple counterparts in neurobiology, it is still unclear whether a polar representation could consistently outperform a Cartesian one simply by having smaller error magnitudes. Thirdly, do these computational arguments apply to variants of PI models which are neither strictly Cartesian nor polar? We avoid these problems in two ways: firstly, we use an extended classification system for spatial representations; secondly, we do not explicitly assume any particular computational algorithm except what is directly implied by the classification system. Instead, we assume that there is a minimum amount of noise which corrupts the PI update process, at each step of the journey. The details of the noise are described below. Every allocentric heading, , and rotation measurement Δ is associated with an error term δ. Even if the ideal Δ is zero (no rotation) a finitely small amount of noise δ corrupts the signal. Since it is impossible for the PI system to “know” that there was no rotation, the best it can do is assume the input . Generally, perfect compass or rotational inputs are impossible. Furthermore, the state of the path integrator is assumed to be updated following every step and that each state parameter which is updated is associated with an error ε. Thus the updating process is assumed to be imperfect. Quantitatively rigorous results have been reported to describe the way in which cumulative noise affects navigation using idiothetic or allothetic directional cues [20], [21], as outlined in the next section. The resulting behaviours are vastly different depending on the directional cue, suggesting a mathematical framework for distinguishing between different classes of movement trajectories. Neural representations of real trajectories may be considered in the same way. Geometric constructions will be developed in this work to map trajectories in real space to representational space, via the process of PI. The logic of this mapping is critical for understanding the theoretical findings of the current work. At its simplest, a directed walk (DW) consists of a sequence of discrete movements or steps, , all intended to be in the same direction (which for convenience and consistency with previous notation, is designated as the X-axis, and is oriented vertically in Fig. 1). Ideally, without noise, at step n, the animal moves by . However, due to unavoidable noise, the animal makes a rotational error and a distance error (Eq 1). DW theory shows how errors accumulate during such a journey [20], [21]. In principle, the basic unit of locomotion of DWs should reflect the anatomy and physiology of the locomoting animal. The mathematical description can therefore range from a simple elementary step [20], to a general biased elementary step [21]. In the main text, we use simple elementary steps for clarity, whose formal description consists of just an unbiased turn error (denoted as a standalone , distinct from the prefix meaning change) and a step length L whose linear error is independent of . Note the physical turn error in real space is the unavoidable final output error and is conceptually distinct from a measured rotational error δ which represents an unavoidable input error (explained earlier). Note that the conclusions still apply even when these simplifying assumptions are relaxed to a general biased elementary step (which accounts for any locomotory pattern, error distribution, and statistical dependence between step components - see supplement, and ref [21]). This is necessary to generalize the proof to allow for any realization of each class of spatial representation. It should be noted that DWs account for errors in the step length as well, but previous results showed these are only of secondary importance compared to the angular error except for very short journeys [20], [21]. Although not discussed in the following analysis, it is understood that L incorporates some random error. Of fundamental importance to an understanding of the effect of cumulative noise in real space is the distinction between two types of directional information during navigation. This distinction turns out to be the critical determinant of the type of DW which results when an animal attempts to move in a straight line [20], [21]. One type of directional cue, like a geomagnetic compass, is an absolute external (allothetic) directional reference which is available continuously. Using a compass or other allothetic directional cue to move from point A to point B results in an allothetic directed walk (ADW, Fig. 1A - subtleties about what constitutes a compass, how a compass should be used, or fine distinction between allothetic and idiothetic cues are discussed elsewhere [21]). During an ADW, the displacement at step n, expressed in a Cartesian reference frame, can be written as(1)The alternative type of directional cue, where direction is estimated by internally accumulating measured rotations, is termed idiothetic. Using idiothetic directional information in moving from point A to point B results in an idiothetic directed walk (IDW, Fig 1B). The displacement of step n is thus(2)Note that the total angular displacement error for each step n is the sum of all angular errors up to and including that of step n (in contrast to an ADW, Eq 1). It is now well understood that the properties of ADWs and IDWs are qualitatively and quantitatively different in the presence of any noise, or error. Two important differences between ADWs and IDWs are briefly outlined. Firstly, the expected (mean) position of an IDW has a finite limit, irrespective of how many steps are taken [20], [21]. Recently, this limit was shown in walking humans to be approximately 100 m [26]. In contrast, there is no such limitation for an ADW. Secondly, the positional uncertainty (variance) of an IDW is generally greater than that of a pure random walk, which in turn is greater than following an ADW [20], [21]. In other words, an IDW leads to nonlinear systematic errors coupled with large random errors in position, whereas an ADW results in linear systematic errors coupled with small random errors in position. Assuming a biologically plausible level of noise, the positional uncertainty following an IDW becomes so large that the navigating agent is lost, even after very few steps. Under the same noisy conditions, following an ADW, the navigating agent is positioned close to the ideal location with relatively very little uncertainty. Importantly, the average trajectory of an ADW has the same shape as the ideal trajectory, only smaller by a constant factor. If we assume that accurate PI is biologically advantageous, then we would predict that natural selective pressures would favour the development of a spatial representation with minimal error. For instance, given a simple straight trajectory in physical space, a trajectory which resembled an ADW in representational space would have smaller errors and be a more faithful representation than one which resembled an IDW. During locomotion, from one step to the next, what distinguishes an ADW from an IDW? It turns out that the essence of this problem can be distilled down to one fundamental question – do angular displacement errors accumulate? If angular displacement errors accumulate, then the behaviour resembles an IDW (illustrated in Fig. 1B), otherwise an ADW (illustrated in Fig. 1A; summarized in Table S1 for general locomotion). In African desert ants, PI may be the dominant navigation strategy [27]. In honeybees, vectorial information is transmitted through their remarkable dance language [28], [29], and used for relocating a goal [30]. In the arthropod navigation literature, PI is considered to be of such importance that there exists at least one canonical and one neural network model which uses each of the four standard classes of spatial coordinates i.e., egocentric Cartesian (EC), egocentric polar (EP), allocentric Cartesian (AC) and allocentric polar (AP) representations (Fig. 2). Yet numerous other models were published which did not seem to fit into any of these categories, motivating the development of a more general classification scheme [7]. To adequately cover all the existing equational and neural models of PI, the new classification scheme differentiated models on the basis of two independent properties of the spatial reference frame. The first property was whether the representation was centred on the animal (egocentric) or outside the animal (allocentric), consistent with existing literature. Note that for simplicity, it was previously assumed that most allocentric reference frames are also earth-centred (geocentric) which adequately described most arthropod experiments on PI [7]. However, in many rodent experiments, the allocentric reference frame is purposefully disengaged from the geocentric one. To be strictly correct, here we assume the “geocentric” descriptor refers to some, but not necessarily all, members of the superset of “allocentric” reference frames. The second property used by the classification scheme was whether there was a need to update directional components during PI. If a position is represented along one or more predefined directions, like a Cartesian coordinate (Fig. 2C, 2D), then the representation was considered to be built from “static vectors” since the axes (vectors) have static directions. In contrast, if directions were variable and thus a change in position generally required a change in the direction component (such as polar coordinates – Fig. 2A, 2B), the representation was considered to be built from “dynamic vectors” since the axes (vectors) have dynamic directions. It has been noted previously [7] that the number of directions used in each model was not important. This means that a neural network consisting of three basis vectors behaves in essentially the same way as one consisting of hundreds of basis vectors. The results derived below do not require specification of the number of basis vectors. Hence the results are general for any model which fits under this classification scheme. The four extended classes, namely an egocentric static vectorial representation (ESVR), egocentric dynamic vectorial representation (EDVR), allocentric static vectorial representation (ASVR), and allocentric dynamic vectorial representation (ADVR) are summarized in Table 1, and discussed more extensively below (see also [7] for a detailed treatment). Theoretically, the process of PI continually adds measured displacements and supplies an animal with metric information about the overall distance and direction from home or some other location [7]. Such vectorial information may in principle be used in a variety of ways, including map construction, binding to place information, association with motor outputs, and to perform even more sophisticated tasks such as path planning. Indeed, an animal may store vectorial information about multiple important places in its world. However, in this work, we only consider the neural record of a single journey using PI under noisy conditions. We argue that task sophistication will generally lead to a greater sensitivity to PI error, not less. Therefore, if the simplest PI task is fundamentally impossible, then so are all ethologically relevant generalizations thereof. Discrete time models are used for ease of description of error terms, and for consistency with published theory on directed walks [20], [21]. In rodents, PI is currently thought to be an important mechanism which maintains the spatial consistency of place cell and grid cell firing fields [6], [9]–[13]. For example, the observation that such fields remain stable in darkness is generally interpreted as evidence that PI is used to maintain their spatial specificity, but non-visual localizing cues may contribute [31]. The corollary of this argument is that during PI, the neural network state can be decoded to calculate the animal's perceived current location [32]. In essence, our analysis is a theoretical comparison of the most accurate and precise neural record which could be obtained during noisy PI, with the actual path traversed by the hypothetical animal. In order to apply the results of DW theory to PI, we will consider the exact opposite situation of the original DW model. Rather than intending to walk in a straight line but suffering from random physical perturbations during locomotion, we will model an animal which is walking in a perfectly straight line, but which updates the internal representation of its location using a noisy PI process fed by noisy sensory inputs, resulting in an internal DW occurring in representational space. The displacement of the represented location associated with each step n of the internal DW will be denoted by Cartesian coordinates. This is the step displacement, not the final internal representation of position, which would be the sum total of all displacements 1 to n (e.g., see Text S1). Allocentric representations (whether static or dynamic vectorial) will be expressed as allocentric Cartesian coordinates, using the symbols (Un, Vn), whilst egocentric representations (static or dynamic vectorial) will be expressed as egocentric Cartesian coordinates, using the symbols (U′n, V′n), thus making the results of PI using the four classes of spatial representation immediately comparable with the original DW theory. This procedure in no way alters the expected outcome, which depends purely on the actual representational class being used by the animal, nor does it imply the conclusions are limited to representations based on a single pair of Cartesian basis vectors. Analogous to physical DWs, we model PI by assuming that at each step the animal intends to update its positional representation by a distance corresponding to the true step length (the scaling between physical and representational space is unimportant, and can be treated as unity) and in the direction corresponding to the true axis of physical locomotion (Figs. 3, Text S1). The actual representational step taken is of length , which deviates from the correct length by a random amount, corresponding to a failure of the animal to sense the true step length exactly. In all cases we consider the situation where allocentric directional information (e.g. a compass) is available every step, providing the measured allocentric heading, , but with associated error term, δn. This assumption was made since it has already been shown that in the absence of a compass, an IDW results irrespective of the navigation strategy [20], [21], and successful navigation is impossible (beyond a few steps). For completeness, we take the spatial representation most tolerant to noise, and examine the effect of using purely idiothetic directional information i.e., only rotation measurement, , is available each step, and also associated with an equivalent δn error term. Representational noise, εn, is also assumed to corrupt the updating of the representation by the PI process each step. In the main text, only the update errors which determine ADW-like or IDW-like behaviour are discussed (but see Text S1). Note that although it is convenient to use specific simulation examples for concept illustration, the theoretical results developed in this work are applicable to all animals which carry out PI, irrespective of the size of their nervous system, their evolutionary lineage or the known neurobiology. We now show the equivalence between each of these four classes of spatial representation for PI and its corresponding directed walk, in the presence of neural noise. We show that any egocentric (ESVR or EDVR) or dynamic vectorial (ADVR or EDVR) representation of space accumulates noise in a way analogous to an IDW during PI. Therefore, PI using any such neural representation of space is inevitably associated with large random and systematic errors. In contrast, we show that an allocentric static vectorial representation (ASVR) accumulates noise in a way analogous to an ADW and therefore suffers from relatively small random errors and no systematic error during PI. Formal proofs and stepwise geometric constructions showing the type and temporal order of error accrual are included in the supplement (Eqns S1.5–1.8 in Text S1, and Fig. S1). Here, we focus on key results which are necessary and sufficient to differentiate the performance of the four extended classes of spatial representations during PI. We use a theoretical construct, termed here a ‘neural record’, to illustrate the noise-induced divergence of the trajectory through representational space as indicated by the PI system, from the actual path of an animal. This hypothetical record is deduced from the changing internal states of the PI system during the journey, but calculated following the completion of a journey of n steps. For example, the neural record of step m is obtained by rearranging the vector equation following step n, giving current PI state in order to collect all the terms which should be associated with step m. Thus, for an egocentric and/or dynamic vectorial representation, the neural record is in fact the original step m corrupted by all subsequent errors up to and including those of step n (see below for analytical details). The neural record may therefore differ from the positions indicated by the set of initial PI states which resulted when each step was taken, but is a simple and intuitive way to track and visualize the errors arising during navigation. Firstly we show that any egocentric spatial representation incorporates input error δ during PI in the same manner as an IDW, but in reverse temporal order. An egocentric representation is one where places in the world are defined relative to the navigating agent. By definition, a right turn of a navigating agent implies the home direction has turned left by an equal magnitude. We know that using an egocentric representation, PI requires rotation (equivalent to a change in heading) as one input [7]. Due to the presence of biological noise, every step is associated with an angular error, δn, in estimated heading rotation, irrespective of whether rotational signals are available directly e.g., , or if it is estimated from true compass bearings e.g., . An input rotational error of δn results in a home direction error of −δn. The result is that an error in current heading measurement is effectively added to all future steps in egocentric space (Figs. 3B, 3D, Fig. S1B, Fig. S1C). Consider a true trajectory which is perfectly straight in physical space such that true home is always directly behind the navigating animal. Following step one, any egocentric representation must incorporate the rotation error −δ1. We can trace the first step in representational space, which in egocentric Cartesian coordinates is and where U′ represents the rostral-caudal (forward-backward) axis, and V′ represents the lateral (left-right) axis (Fig 3D). For convenience, negative values of U′ and V′ mean backward and rightward respectively. The step length is denoted Λ in representational space. Analogous to the step length L in real space, Λ is assumed to incorporate some random error in representing the magnitude of forward displacement. In Fig. 3B (δ and ε) and Fig. 3D (δ only), we are considering the situation of an animal physically heading away from home, so by definition, the position of home moves in backward (−U′) direction in egocentric space whereas the position of the animal moves in a positive direction with respect to the home i.e., if viewed in an allocentric reference frame. During step two, an input rotation error of δ2 results in a rotation of the entire current representation of home by −δ2. Hence following step two, both steps one and two have effectively incorporated the rotation −δ2, and so on (Fig. 3D shows pure δ accumulation). After n steps, the neural record of the mth step is given by(3)It is important to note that the neural record of the mth step is dependent on the total number of steps, n, which has been taken. This is due to the fact that a rotational error resulting from each new step affects the entire PI record, which was built from all previous steps. Effectively, the neural record of step m is affected by but not . Thus, the angular error in representational space from step m−1 to step m is as illustrated in Fig. 3D. In an IDW, each new angular displacement only affects step m and onwards, not steps 1 to m−1, which have already occurred [20]. However, in an egocentric spatial representation, an error at step m affects steps 1 to m in representational space i.e., those steps which have already been recorded. Careful analysis shows that the two representations can be considered as being equivalent. Without loss of generality we can renumber the steps in representational space in reverse order so that step n becomes step 1, step n−1 becomes step 2 and so on. Thus(4)which is mathematically equivalent to an idiothetic directed walk (IDW), but occurring in representational space (Eq. 2; Figs. 1B, 3B, 3D; Table S1; [22], [23]). Put simply, a straight trajectory in egocentric representational space accumulates angular errors in the reverse temporal order to an IDW in real space. Thus during an IDW in real space, recent rotational errors add to past ones, so earlier rotational errors contribute to all subsequent heading directions (Fig. 1B). In an egocentric representation, the most recent rotational input error, , rotates the entire current neural representation of home (Fig. 3B, 3D), which consists of steps 1 to n−1, with their associated errors. Note that the above arguments were developed independently of the type of egocentric representation. Therefore, the compass/rotation error δ is sufficient to cause a degradation of any egocentric representation of space so that a straight line in real space maps to an IDW in representational space (but see special case explained in Text S2). Thus an egocentric static vectorial representation (ESVR) or egocentric dynamic vectorial representation (EDVR; see Table 1) are both susceptible to the same type of path degenerescence in representational space. Of course, it is possible for other types of random errors to further degrade the egocentric representation (e.g. EDVR - see below). Next we show that a dynamic vectorial spatial representation incorporates update error ε during PI in the same manner as an IDW. A dynamic vectorial representation, typified by the polar representation, consists of vectors containing variable angular components. The path integrator's measure of direction accrues an error ε during the updating process, irrespective of the reference frame so where is the true current direction from home in an allocentric spatial representation. Unfortunately, the true net direction is not available, but only the approximation resulting from previous steps. The critical concept here is that the update error adds to the current net direction which was estimated from accumulating all previous steps i.e., effectively rotates the representation of all previous steps. Using the same analysis conventions as the previous section, we examine the mapping of a straight trajectory into representation space using a dynamic vectorial representation. Following step one, an update error distorts the true value of θ or θ′ by ε1 in representational space. In allocentric Cartesian coordinates (Fig. 3A), the straight trajectory in real space is aligned with the positive U axis so that and . Following step two, an allocentric dynamic vectorial representation accrues the update error ε2, which is effectively added to steps one and two, in a manner similar to rotation errors. After n steps, the neural record of the mth step is given by(5)Thus, the angular error in representational space from step m−1 to step m is as illustrated in Fig. 3A, Fig. S1A. In egocentric coordinates (Figs. 3B, Fig. S1C), the result of update errors is similar to the effect of rotation errors, except that the sign of the update error is preserved. Perhaps more importantly, as explained already, an egocentric representation is also affected by rotation errors. Thus following step one, , and . After n steps, the neural record of the mth step is given by(6)Clearly, both types (allocentric or egocentric) of dynamic vectorial neural records are of the same mathematical form as the egocentric neural record. Therefore, the update error ε is sufficient to cause a degradation of any dynamic vectorial representation of space (ADVR or EDVR) so that a straight line in real space maps to an IDW in representational space. Now we show why an ASVR incorporates input and update errors (δ and ε respectively) during PI in the manner of an ADW. A static vectorial representation, typified by the Cartesian representation, consists of vectors containing fixed angular components. We know that using an allocentric representation, PI requires absolute heading as one input [7]. Since biological compasses are imperfect, there is an angular error δ associated with each step, much like the rotation error of egocentric representations. Thus, . Again, we analyze the mapping of a straight trajectory in real space into representational space (Fig. 3C). Following step one, and . The neural record of the mth step is given by(7)which is mathematically equivalent to an allothetic directed walk (ADW), occurring in representational space (Fig. 1A, Fig. 3C; Table S1; [20], [21]). Update errors alone lead to a representational trajectory described by and , which is also equivalent to an ADW. Combining input and update errors,(8)which is also mathematically equivalent to an ADW. Therefore, in an allocentric static vectorial representation (ASVR; see Table 1), a straight line in real space maps to an ADW in representational space. Note that in all cases considered above, it was assumed that an allothetic directional cue was used as input. It is straightforward to show that using an idiothetic directional cue as input degrades performance further. Indeed, even for an ASVR, a straight line in real space maps to an IDW in representational space if only idiothetic directional cues are used i.e.,(9)The performances of the four classes of spatial representations were compared via computer simulation. An equational model from each class ([7], Table S2) was used to carry out PI using the same set of random trajectories (Fig. 4). For consistency all examples have directional/rotational input errors and update errors of equal magnitude. The neural record from one random example is shown (Fig. 4, A–D) overlaid on the true trajectory, scaled so that step length L = Λ. The average positional estimation error (Fig. 4E) clearly demonstrates the superiority of the example from the ASVR class, consistent with theory. Variants of this class and limitations of errors are considered further below and in Text S3. In the preceding analyses, the classification of spatial representations did not consider the modulus or length of the (static or dynamic) vectors used as the basis for a representation. The classification scheme used so far sufficed to give us the necessary insights into which types of spatial representations can map space faithfully via the process of PI. However, what does that tell us about the neural circuitry of navigation? Firstly, each class can be further subdivided to differentiate between fixed and variable length vectors [7]. By variable length static vectors we mean that the representation is based on vectors defining fixed directions (such as X and Y axes), but not fixed distances in these directions i.e., basis vectors. The representation records the (variable) distance moved along these fixed directions. Figure panels 5A and 5B show two graphical examples of a variable length SVR system used to represent the same net allocentric displacement (red disc). In Fig. 5A, there are three static basis vectors and a simple, mathematically exact, decomposition of the red disc's components is shown. In Fig. 5B, an ASVR with many basis vectors is shown, along with the components of the red disc, but also an example of a response function (analogous to distributing vector components) of the same displacement. A range of PI models have been published, particularly in the arthropod literature, which fall into this subclass [7]. The neural model implementations of an ASVR (dynamic moduli) are often drawn as a ring-like array of neurons, with each neuron representing a fixed allocentric direction, with receptive fields of various widths and shapes. Of course, the ring configuration could be an artefact of the compass input needed for an allocentric representation. Nonetheless, the computational requirement of a direct correspondence between allocentric angular space and neuron index is suggestive of a structured organization. In contrast, a set of static vectors with a range of fixed lengths, spread out over the 2-D Euclidean space is reminiscent of a grid or map. Effectively, each vector represents a point, or small region, in space. Then, the representation of a position no longer requires spatial measurements like length or angle. At its simplest, a binary output suffices, which denotes a location is either at a static vector or not (Fig. 5C). More information may be represented by a distribution of output values corresponding to probability, which could account for positional uncertainty (e.g. Fig. 5D). The intuitive division of allocentric static vectors into fixed and variable lengths naturally produces models which resemble place cell maps and neural rings, respectively. Interestingly, published arthropod neural models of navigation typically adopt a ring like structure even those which are now known to be noise-intolerant, with one notable exception [33]. In contrast, mammalian neural models have typically been based on map-like networks [9]–[11], [34]. Whether this is a coincidence or reflects fundamental biology remains unclear. Some important neural architectural and computational issues of the two ASVR subclasses are considered below. Four general classes of spatial representations were studied. The motivations for using this scheme were twofold. Firstly, standard classification systems have been insufficient to account for a number of neural network models. Secondly, the general classification scheme is consistent with mathematical results from DW theory which proved that the critical determinant of trajectory behaviour is whether angular errors accumulate. In particular, there is no dependence on the number of axes used to represent a position or whether linear errors accumulate. For instance, ring-like and map-like neural structures may be considered alongside simpler counterparts, even equational models. From first principles, we showed how real space is mapped into different classes of representational space via PI in the presence of noise. It was found that real navigation journeys represented in allocentric dynamic vectorial representations (ADVRs), egocentric dynamic vectorial representations (EDVRs), or egocentric static vectorial representations (ESVRs) are corrupted by noise in similar ways. Examples include all egocentric (e.g. EC or EP) and all polar (e.g. AP or EP) representations. A straight trajectory in real space maps to an IDW in representational space, resulting in nonlinear systematic errors and irrecoverably large random errors. Consequently, the error of spatial representation is expected to increase rapidly, rendering the animal hopelessly lost. Egocentric representations suffer particularly from input noise, δ, while dynamic vectorial representations are particularly affected by update noise ε. In this work, the magnitude of noise was not considered – only that noise exists. Although the properties of biological sensor noise may be well characterized in certain cases [35], neural processing noise is typically much more difficult to quantify. It is possible for instance that input noise is much larger in magnitude than update noise. Our results then predict ESVRs and EDVRs to show greater nonlinear systematic errors than ADVRs. However, due to the nature of IDWs, the random errors of ADVRs would eventually exceed those of ESVRs or EDVRs, thereby causing even greater PI inaccuracies, albeit delayed. In contrast, allocentric static vectorial representations (ASVRs), typified by the allocentric Cartesian (AC) representation, faithfully capture the geometric and metric properties of real trajectories. A straight trajectory in real space maps to an ADW in representational space. In principle, animals which have evolved ASVRs would have far superior navigational outcomes, particularly for long journeys. For theoretical completeness, we note that ASVRs are not entirely immune to neural noise. For instance, large systematic angular errors (e.g. >90°) can still cause failure of homing via PI using a compass plus an ASVR (Text S3). However, we believe that such extreme errors are unlikely to occur in nature, and in any case would cause even more severe problems for alternative neural representations of space. The strengths and weaknesses of existing arguments for or against using different representation systems to model arthropod PI have been reviewed [7], [36], [37]. Can the modelling literature say anything about the current results? In fact, under noisy conditions, and using evolutionary algorithms to optimize performance, the evolved PI neural networks were found to be subtypes of ASVRs [38], [39]. This is entirely consistent with the current theoretical results. Admittedly, the published models made a priori assumptions which might have unfairly favoured an ASVR. As noted previously, most models neglected the effects of noise. Even in the absence of noise, a number of computational properties discourage the use of non-ASVRs. These include large rates of change in angle needed for ADVRs and EDVRs near home, large rates of change in position needed for ESVRs far from home, feedback of current path integrator state into the update process for all non-ASVRs, among others [7]. Individually, the arguments made assuming noise-free conditions could be countered. Nevertheless, the weight of evidence seemed to favour ASVRs. In combination with the clear consequences of neural noise, the case for ASVRs is difficult to dispute. In light of this, new interpretations of previous experimental results [e.g. 40] may be required, and assumptions of non-ASVR systems for biological PI [e.g. 41] should be re-examined. At least two subclasses of ASVR exist, which for convenience can be approximately described as ring-like (Fig. 5A, 5B) and map-like representations of space (Fig. 5C, 5D). While there does not appear to be a significant difference in noise-tolerance between the two ASVR subclasses, there may be distinctions based on phylogeny or behavioural requirements. In the literature, there appears to be a lack of map-like models of arthropod PI, and a lack of ring-like models of mammalian PI. Interestingly, arthropod PI models, including the noise-tolerant ring-like varieties, were inspired by behavioural results. This suggests ring-like models are well suited to account for a variety of navigation behaviours related to PI [7]. In contrast, models of mammalian PI were inspired by in vivo recordings. In other words, neurons are known to exist which possess the necessary properties to represent space in a noise-tolerant way. It is tempting to hypothesize that this apparent dichotomy in published models reflects a fundamental difference between the nervous systems of arthropods and mammals. Unfortunately, evidence is lacking. Nonetheless, our results provide strong theoretical justification for the evolution of some sort of ASVR, in any species which needs to navigate or represent space. Differentiating between the two subclasses of ASVR, however, may present significant theoretical and experimental challenges and is the subject of ongoing research. Some important neural architectural and computational issues are briefly outlined below. It is relatively simple to envisage a direct correspondence between allocentric angular space and a neuronal array, particularly when a compass is available. This is analogous to the ring-like subclass of ASVR. Unfortunately, there is a lack of electrophysiological evidence for any particular type of PI system in arthropods. From the tenuous data, a possible candidate for a ring-like PI system may be the central complex [15]. For the map-like ASVR subclass, it is positional space which corresponds to a neuronal array - but this is not trivial to achieve. For instance, one might assume that using allocentric landmarks allows for relatively precise spatial localization, in the same way a compass allows for angular localization. It might be further assumed that a map-like spatial representation can therefore be generated using allocentrically-stable cues. Yet the association of landmarks to a particular spatial location requires visiting that location – but how could that location be encoded in the first place? What determines the spatial relationship to other positions? One possibility is that the spatial representation is generated dynamically, during the first visit to any physical location, with the recruitment of neural units en route. However, this seems unlikely since it results in a circular argument i.e., PI using a map-like ASVR requires a pre-existing “map”, which is not present until the path is recorded via PI. Although landmarks are excellent for localization of individual places in the world, they are generally poor for relating those places in a metrically consistent way (unless the spatial layout of the landmarks are already known). Hence it is likely that a pre-existing spatial representation is used during PI rather than being formed dynamically. It is worth noting that many SLAM (Simultaneous Localization and Mapping) algorithms have been implemented successfully in engineering and robotic applications [42], and superficially appear to contradict this assertion. However, careful examination of the algorithms reveals that in fact there is always a predefined representation of space, often metric and Cartesian-based, but empty to begin with (and without necessarily pre-allocating much memory resource). The SLAM algorithms serve to bind those spatial representations with objects and experiences during navigation. Even here, the spatial representation cannot be formed ab nihilo on encountering a landmark, but must be generated en route to maintain a consistent spatial relationship or spatial metric. If a map-like ASVR cannot dynamically bind physical space (and the corresponding allocentric sensory information) to neural substrates, then how can PI be achieved? An alternative model might involve pre-existing, metric relationships between all the neurons in an array – literally a “place map”. In this way, PI can be achieved by translation of an activity bump, for instance, along a network of neurons in a spatially consistent way [e.g. 9], [10]. Most mammalian PI models have been developed to explain in vivo data parsimoniously and are typically of the map-like ASVR subclass. However, from a computational standpoint further analysis is required to determine whether modified ring-like ASVR models may also explain in vivo data. Of note is the fact that the electrophysiological properties of place cells and grid cells are dynamically affected by allocentric cues, unlike simple place map models. Multiple ring-like ASVR systems may allow remapping to occur readily, yet maintain a consistent metric relationship between real places in representational space. Behavioural data may also offer clues for differentiating between variable and fixed length ASVRs. Ring-like models have been used with some success to model systematic errors of PI. Can models using map-like representations do the same? Searching behaviour is often associated with PI, particularly following displacement experiments. Interestingly, search behaviour seems to reflect both the accrued uncertainty of the outbound journey, as well as the dynamically changing prior and posterior probability distributions during searching [43]–[45]. Can ring-like models maintain sufficient information to account for such complexities? Rigorous theoretical analyses of these issues may yield further insights about the neural representation of space and is currently under way. PI is an ancient and ubiquitous navigation strategy. Even highly complex animals such as rodents and humans possess a PI system. Due to the ubiquitous presence of biological noise, a variant of an ASVR is most likely to be used. If other navigation strategies, and indeed other neural functions, evolved from a PI ancestry, there are likely to be residual ASVR signatures in modern nervous systems. Their properties remain an open topic for future investigation. The current work advances our understanding of PI, animal navigation, evolved neural systems, and further demonstrates the usefulness of DW theory. These analytical foundations will hopefully steer future experimentation, and focus modelling work, towards a deeper understanding of biological navigation and animal nervous systems. For example, a biological implementation of a ring-like ASVR model might entail neurons which are tuned to specific allocentric directions, and which behave like odometers in their preferred directions. If such an odometer is linear with respect to distance, it might be expected that PI fails catastrophically beyond a certain radial distance from home, once some ceiling value is reached. Alternatively, if distance is encoded in a saturating manner to avoid a ceiling range, the inherent uncertainty in the representation of space may increase nonlinearly with distance, which may manifest in the size of searching distributions following PI. In map-like ASVR representations, there is a need for translating the current position during PI. Is that achieved via interneurons perhaps like an attractor network? Do all interneurons receive the same allocentric heading signal? Does the map have an edge or does the map wrap around seamlessly like a torus? How is positional uncertainty represented? We believe the results presented in this work represent the strongest theoretical foundation to date for determining the type of spatial representation likely to be used by biological nervous systems for navigation.

10.1371/journal.ppat.1004937

TNF-α Induced by Hepatitis C Virus via TLR7 and TLR8 in Hepatocytes Supports Interferon Signaling via an Autocrine Mechanism

Invasion by infectious pathogens can elicit a range of cytokine responses from host cells. These cytokines provide the initial host defense mechanism. In this report, we demonstrate that TNF-α, a pro-inflammatory cytokine, can be induced by hepatitis C virus (HCV) in its host cells in a biphasic manner. The initial induction of TNF-α by HCV was prompt and could be blocked by the antibody directed against the HCV E2 envelope protein and by chemicals that inhibit endocytosis, indicating the specificity of endocytic uptake of HCV in this induction. Further studies indicated that the induction of TNF-α was dependent on toll-like receptors 7 and 8 (TLR7/8) but not on other intracellular pattern recognition receptors. Consistently, siRNA-mediated gene silencing of the downstream effectors in the TLR7/8 signaling pathway including MyD88, IRAK1, TRAF6, TAK1 and p65 NF-κB suppressed the expression of TNF-α. The role of p65 NF-κB in the induction of TNF-α via transcriptional up-regulation was further confirmed by the chromatin immunoprecipitation assay. TNF-α induced by HCV could activate its own receptor TNFR1 on hepatocytes to suppress HCV replication. This suppressive effect of TNF-α on HCV was due to its role in supporting interferon signaling, as the suppression of its expression led to the loss of IFNAR2 and impaired interferon signaling and the induction of interferon-stimulated genes. In conclusion, our results indicate that hepatocytes can sense HCV infection via TLR7/8 to induce the expression of TNF-α, which inhibits HCV replication via an autocrine mechanism to support interferon signaling.

Hepatitis C virus (HCV) patients have increased levels of circulating tumor necrosis factor-α (TNF-α). In this report, we demonstrate that HCV can directly induce the expression of TNF-α in hepatocytes in a biphasic manner via NF-κB. The induction of TNF-α by HCV in the first phase is prompt, requires no HCV gene expression and is dependent on TLR7 and TLR8 and their downstream effectors. TNF-α induced by HCV supports interferon signaling via an autocrine mechanism and suppresses HCV replication, as abolishing the expression of TNF-α or its receptor TNFR1 results in the loss of IFNAR2, a subunit of the type I interferon receptor, and an increase of HCV replication. Our studies thus reveal an interesting interplay between HCV and hepatocytes, with the virus attempting to blunt the IFN response by depleting IFNAR2 and the host cell overcoming this blunting effect of HCV by using TNF-α to restore the expression of IFNAR2.

Hepatitis C virus (HCV) is an enveloped virus with a single-stranded RNA genome of 9.6-Kb [1]. After binding to its receptors on hepatocytes, HCV is internalized by receptor-mediated endocytosis, and its genomic RNA is subsequently released into the cytosol to direct the synthesis of viral proteins using the internal ribosome entry site (IRES) located near its 5’-end. This leads to the production of a polyprotein with a length of approximately 3000 amino acids. The HCV polyprotein is proteolytically cleaved by host and viral proteases to give rise to individual viral proteins including the core protein, E1 and E2 envelope proteins, the p7 viroporin, and nonstructural proteins NS2, NS3, NS4A, NS4B, NS5A, and NS5B [2]. Pattern recognition receptors (PRRs) including toll-like receptors (TLRs) and RIG-I-like receptors are important components of the innate immune response. Upon the activation by the pathogen-associated molecular patterns (PAMPs), these PRRs induce the expression of various cytokines via the downstream signaling pathways. Some TLRs are located on the cellular surface and sense extracellular PAMPs and some TLRs are located in the endosomes to detect internalized pathogens [3]. The TLR signaling is mediated by the TIR domain-containing cytosolic adaptors MyD88, TIRAP/Mal and TRIF. The initial association of MyD88 to the receptor leads to the sequential recruitment and activation of IRAK4 and IRAK1. The activated IRAK1 then binds to TRAF6, after which the complex dissociates from the receptor for further signaling events including the activation of TAK1. TAK1 can activate NF-κB and AP1 to stimulate the production of pro-inflammatory cytokines. IRAK1 and TRAF6 can also activate IRF7 to induce the expression of type I interferons (IFNs) [4, 5]. Tumor necrosis factor-α (TNF-α) is a pro-inflammatory cytokine produced in response to infectious pathogens. The soluble TNF-α is produced as a result of cleavage from its precursor transmembrane TNF-α by the TNF-α-converting enzyme (TACE). The secreted TNFα binds to its receptors, namely TNFR1 and TNFR2, to exert its biological effects [6]. Multiple studies indicate that the blood level of TNF-α is increased in HCV patients and its level is positively correlated with HCV pathogenesis and the severity of liver diseases [7–9]. The major source of TNF-α in response to HCV infection is unclear and thought to be immune cells such as T lymphocytes and macrophages [10, 11]. In this report, we provide evidence to demonstrate that hepatocytes can also produce TNF-α in response to HCV infection. This TNF-α induction is prompt and mediated by TLR7 and TLR8. Furthermore, we also demonstrate that TNF-α, through an autocrine mechanism, prevents the depletion of IFNAR2 by HCV and is required to support interferon signaling in HCV-infected cells. To determine whether HCV infection can directly induce the expression of TNF-α in its host cells, we infected Huh7 hepatoma cells with a cell culture-adapted HCV JFH1 variant using a multiplicity of infection (MOI) of 0.25 and collected the incubation media at different time points after infection for quantification of TNF-α using ELISA. The soluble TNF-α was initially detectable at 48 hours post-infection and its level further increased at 72 hours (Fig 1A). When the quantitative RT-PCR (qRT-PCR) was used to analyze the expression of TNF-α RNA in cells, a similar induction profile was observed (Fig 1A), although a ~10-fold induction of TNF-α was also observed at 24 hours (see below). The induction of TNF-α in Huh7 cells could be detected by immunoblot at 24 hours post-infection when the MOI used was 1 or higher (Fig 1B). To ascertain that the TNF-α induction by HCV was not specific to Huh7 cells, we also infected primary human hepatocytes (PHH) with HCV. As shown in Fig 1C, HCV infection of PHH also induced the expression of TNF-α after 24 hours when a semi-quantitative RT-PCR was used for the analysis. To determine how early after infection HCV could induce the expression of TNF-α, we analyzed the induction of TNF-α by HCV within the first 24 hours of infection using an MOI of 2. As shown in Fig 2A, HCV could induce the expression of TNF-α as early as one hour post-infection, when the semi-quantitative RT-PCR was used for the analysis. This induction was increased at 2 hours, reduced at 4 and 8 hours and increased again at 24 hours (Fig 2A). We also used qRT-PCR to analyze the effect of MOI on the induction of TNF-α and found that the induction of TNF-α by HCV was slight at 2 hours but significant (~10-fold) at 24 hours post-infection when the MOI used was 0.25 (Fig 2B). However, when the MOI of 1 was used, the fold induction of TNF-α at 2 hours and 24 hours post-infection was similar at about 15, indicating a dose-effect of HCV on the induction of TNF-α at the early time points of infection. To rule out the possibility that the early induction of TNF-α was due to nonspecific factors in the HCV inoculum, we treated the HCV inoculum (MOI = 1) with an anti-E2 antibody, which neutralized the infectivity of HCV [12, 13]. As shown in Fig 2C, this neutralization antibody reduced the TNF-α RNA in Huh7 cells to almost the basal level, confirming the specificity of TNF-α induction by HCV. In addition, the early induction of TNF-α was inhibited, if cells were treated with actinomycin-D, an inhibitor of RNA synthesis, prior to HCV infection (Fig 2D), indicating a transcriptional up-regulation of TNF-α. The effect of actinomycin-D on TNF-α was unlikely due to its effect on HCV entry, as this treatment slightly increased the HCV RNA level in cells (S1 Fig). Considering that the induction of TNF-α occurred almost immediately after HCV infection, it did not appear likely that the translation or the replication of HCV genome RNA was involved. To test this possibility, we used UV-irradiation to inactivate HCV prior to infection. This UV inactivation did not inhibit the induction of TNF-α at two hours post infection (Fig 2E), indicating that the integrity of the HCV genome was not essential for the induction of TNF-α at this early time point. However, the UV-inactivation of HCV reduced the second-phase induction of TNF-α at 24 hours, indicating that the HCV gene expression and/or replication was required for the efficient induction of TNF-α at this later time point. Besides TNF-α, the induction of other cytokines including IL-6 and IL-1β was also observed in both Huh7 cells and PHH at 2 hours post-infection (S2 Fig). The finding that the induction of TNF-α was detected almost immediately after infection without the apparent need of an intact HCV genome suggested an early event of HCV infection in the induction of TNF-α, possibly during the viral entry. After binding to its co-receptors, HCV enters the cell via the clathrin-mediated endocytosis [1, 14]. To test whether this endocytic uptake is required for HCV to induce TNF-α, we treated Huh7 cells with Dynasore, a cell-permeable inhibitor of dynamin GTPase, which mediates the scission of clathrin-coated vesicles from plasma membranes. As shown in Fig 3A, the inhibition of dynamin with Dynasore significantly inhibited the induction of TNF-α by HCV. This result suggested an important role of endocytic uptake of HCV in the induction of TNF-α. After the dissociation of clathrin, endocytic vesicles fuse with endosomes. The acidic content of endosomes then triggers the fusion of HCV envelope with endosomal membranes for the release of the viral genome into the cytosol. To examine the possible importance of endosomal acidification in the induction of TNF-α, we treated cells with chloroquine, an inhibitor of endosomal acidification. As shown in Fig 3B, the treatment of Huh7 cells with chloroquine abolished the TNF-α induction by HCV. Altogether, these results suggested that both the scission of endocytic vesicles from plasma membranes and the endosomal acidification were important for the induction of TNF-α by HCV. RIG-I and MDA5 are two cytosolic PRRs that recognize double-stranded RNAs (dsRNAs), and the former has been shown to recognize the 3’-end UC-rich sequence of the HCV RNA [15]. To test the possible roles of these two PRRs in the induction of TNF-α by HCV, we performed the siRNA knockdown experiment to suppress the expression of these two proteins prior to HCV infection. As shown in S3A Fig, the suppression of RIG-I and MDA5 expression in Huh7 cells had only a marginal effect, if any, on the induction of TNF-α. The lack of effect of RIG-I on the induction of TNF-α by HCV was also confirmed by the infection of Huh7.5 cells with HCV. Huh7.5 cells were derived from Huh7 cells and expressed a defective RIG-I [16]. As shown in S3B Fig, HCV could induce TNF-α in Huh7.5 cells, confirming that RIG-I was not essential for the expression of TNF-α. Due to the lack of significant effect of RIG-I and MDA5 on the induction of TNF-α by HCV, we turned our attention to TLR3, TLR7, TLR8 and TLR9, which are PRRs that reside in endosomes. Among them, TLR3 is activated by dsRNA; TLR7 and TLR8, which share a high degree of structural similarity and are functionally active in Huh7 cells (S4 Fig), are activated by single-stranded RNA (ssRNA); and TLR9 is activated by the unmethylated CpG motif of DNA [3]. To test the possible roles of TLR7 and TLR8 in the induction of TNF-α by HCV, we also performed the siRNA-knockdown experiments. HCV infection induced the expression of TNF-α at two hours post-infection and the simultaneous knockdown of both TLR7 and TLR8 (TLR7/8) impaired this induction (Fig 4A). It is not likely that the knockdown of TLR7/8 impaired HCV entry, as their simultaneous knockdown increased HCV RNA levels in cells at 24 hours post-infection (S5A Fig). The single knockdown of TLR7 or TLR8 had only a marginal effect on the induction of TNF-α (S5B Fig). This lack of significant effect of single knockdown of TLR7 or TLR8 on TNF-α was likely due to the compensatory increase of the expression of the other when the expression of either one of these two TLRs was suppressed (S5C Fig). In contrast to TLR7/8, the knockdown of TLR3 and TLR9 as well as TLR4, which recognizes lipopolysaccharides, had no apparent effect on the induction of TNF-α by HCV (S5D Fig). Note that HCV infection increased TLR7 and TLR8 RNA levels (Fig 4A and S5B Fig) and appeared to also slightly increase the TLR3 and TLR9 levels (S5D Fig). The reason for this is unclear, but the induction of TLR7 and TLR8 was apparently mediated by NF-κB, as the knockdown of p65, a subunit of NF-κB, largely abolished their induction by HCV (S6 Fig). If HCV indeed induced the expression of TNF-α via TLR7 and TLR8, then the suppression of expression of their downstream adaptor molecules should also inhibit the induction of TNF-α by HCV. As shown in Fig 4B, the suppression of expression of MyD88, IRAK1, TRAF6 or TAK1, which mediates TLR7/8 signaling, all led to the reduction of TNF-α induction by HCV. The knockdown efficiency of these adaptor molecules was shown in S7 Fig. TAK1 activates the transcription factor NF-κB by phosphorylating IκB kinase β (IKKβ). Its role in the induction of TNF-α by HCV was further confirmed by the observation that Celastrol, an inhibitor of the TAK1 kinase, impaired the induction of TNF-α by HCV (Fig 4C). These adaptor molecules are not known to affect HCV entry and thus it is unlikely that their knockdown led to the reduction of TNF-α via the inhibition of HCV entry. Indeed, as shown in S8 Fig, the knockdown of TRAF6 did not decrease but rather increased the HCV RNA level. Taken together, our results demonstrated that HCV induced the expression of TNF-α via TLR7/8 and its downstream signaling molecules including MyD88, IRAK1, TRAF6, and TAK1. TAK1 activates IKKβ, which phosphorylates and destabilizes the NF-κB inhibitor IκB to result in the nuclear translocation of NF-κB. To determine whether NF-κB was indeed activated in the early time point of HCV infection, we performed the subcellular fractionation experiment. As shown in Fig 5A, HCV infection indeed induced the nuclear translocation of p65, a subunit of NF-κB, at 2 hours post-infection. To further test the role of NF-κB in the activation of the TNF-α gene, we performed the chromatin immunoprecipitation (ChIP) assay to determine the binding activity of p65 NF-κB to the TNF-α promoter, which contains the NF-κB binding site [17]. As shown in Fig 5B, an enhanced binding of p65 NF-κB to the TNF-α promoter was observed in HCV-infected cells at 2 hours post-infection. The binding of p65 NF-κB to the TNF-α promoter was also detected at 24 hours, albeit to a lesser degree. To further verify the role of p65 in TNF-α induction, we performed the p65 NF-κB knockdown experiment. As shown in Fig 5C, p65 NF-κB knockdown reduced the ability of HCV to induce the expression of TNF-α at both 2 hours and 24 hours post-infection. Consistently, Bay-11-7085, a chemical that inhibits the phosphorylation of IκBα and the activation of NF-κB, reduced the TNF-α expression (Fig 5D). Taken together, our results clearly demonstrated a role of NF-κB in the induction of TNF-α by HCV. Note that previous studies indicated that HCV could induce oxidative stress, which could activate NF-κB [18, 19]. However, it does not appear likely that oxidative stress was involved in the induction of TNF-α during the early time points of HCV infection, as the treatment of cells with the antioxidant N-acetylcysteine (NAC) had little effect on the induction of TNF-α by HCV at 2 hours and 48 hours post-infection (S9 Fig). To determine whether TNF-α induced by HCV could directly affect HCV replication, we knocked down the expression of TNF-α using its specific siRNA prior to HCV infection. As shown in Fig 6A, the inhibition of TNF-α expression increased the levels of intracellular HCV RNA as well as the level of the HCV core protein, comparing with cells treated with the control siRNA. It also increased the HCV yield, as evidenced by the significant increase of HCV-positive cells when the progeny virus in the incubation media was harvested and used to infect naive cells (Fig 6B). To confirm that TNF-α could indeed suppress HCV replication, we also treated Huh7 cells with recombinant TNF-α, which also reduced the HCV RNA level (S10 Fig). Soluble TNF-α exerts its effect through its receptors TNFR1 or TNFR2. TNFR1 is expressed in most cell types and believed to be responsible for most of the biological effects of TNF-α, while the expression of TNFR2 is primarily limited to endothelial cells and cells of hematopoietic lineages [20]. To understand how TNF-α exerted its inhibitory effect on HCV, we analyzed the expression of TNFR1 in HCV infected cells. As shown in Fig 6C, HCV caused the loss of TNFR1 at 24 hours post-infection. This loss of TNFR1 was a post-transcriptional event, as the TNFR1 RNA level was not affected by HCV (S11 Fig). As TNFR1, upon binding to TNF-α, is internalized by receptor-mediated endocytosis and degraded in lysosomes [21, 22], this result suggested an autocrine activation of TNFR1. To test this possibility, we first treated naive Huh7 cells with TNF-α. This treatment indeed caused the loss of TNFR1, which could be restored if cells were treated with Bafilomycin-A1 (BafA1) (Fig 6D), which inhibits the vacuolar ATPase activity and endocytic protein degradation in lysosomes [23]. We then treated HCV-infected cells at 24 hours post-infection with BafA1. As shown in the same figure, BafA1 also restored the TNFR1 protein level in HCV-infected cells. These results were consistent with the degradation of TNFR1 in lysosomes. Similar to the TNF-α depletion, silencing TNFR1 with its siRNA also increased HCV RNA and core protein levels (Fig 6E). These results strongly supported the argument that TNF-α suppressed HCV replication via an autocrine mechanism. TNF-α unlikely suppressed HCV replication via the induction of apoptosis, as the knockdown of its expression using the siRNA enhanced, rather than suppressed, apoptosis of HCV-infected cells (S13 Fig). HCV infection can induce a modest level of type I interferons, which stimulate the expression of interferon stimulated genes (ISGs) [24–26]. To understand how TNF-α suppressed HCV replication, we analyzed the possible effect of TNF-α on IFN signaling in HCV-infected cells using a firefly luciferase reporter linked to the interferon-stimulated response element (ISRE). As shown in Fig 7A, in agreement with the previous reports [24–26], HCV infection slightly increased the ISRE activity and this increase was reduced to below the background level when TNF-α was depleted with the siRNA. The treatment of HCV-infected cells with IFN-α significantly increased the ISRE activity. Similarly, this increase was impaired, if TNF-α was depleted. The effect of TNF-α on ISRE was confirmed by analyzing the expression of OAS1, ISG56 and MxA, three IFN-stimulated genes (ISGs). As shown in Fig 7B, HCV could also increase the expression of these ISGs in Huh7 cells. This result was consistent with the previous reports that HCV could induce a low level of interferon response [24–26]. However, this induction was largely abolished, if the expression of TNF-α was inhibited with the siRNA. These results indicated that TNF-α, produced by the cells in response to HCV infection, was required to support interferon signaling and the expression of ISGs in HCV-infected cells. This is likely how TNF-α suppressed HCV replication. Type I IFNs bind to the IFN-α receptor (IFNAR), which is composed of two subunits IFNAR1 and IFNAR2. This binding activates Janus kinase 1 (JAK1) and tyrosine kinase 2 (TYK2), which then phosphorylate and activate STAT1 and STAT2 to result in the activation of ISRE in the promoters of ISGs. To understand why TNF-α was required to support the expression of IFNs and ISGs, we analyzed the effect of TNF-α on the activation of STAT1 and STAT2. As shown in S14B Fig, HCV did not apparently affect the phosphorylation of STAT1 and STAT2 at 24 hours post-infection, but it slightly increased the STAT1 phosphorylation at 48 hours post-infection, in agreement with its modest effect on ISRE. The phosphorylation of both STAT1 and STAT2 was clearly visible after the treatment with IFN-α. Interestingly, when TNF-α was depleted with its siRNA, the phosphorylation of STAT1 induced by IFN-α was significantly inhibited in HCV-infected cells (Fig 7C). TNF-α could also enhance the phosphorylation of STAT1 and the expression of ISGs induced by IFN-α in naive Huh7 cells (S14C Fig). To further investigate why the activation of STAT1 by IFN-α was inhibited when TNF-α was depleted, we analyzed the expression levels of IFNAR1 and IFNAR2 in HCV-infected cells. As shown in Fig 7D, the depletion of either TNF-α or TNFR1 led to the loss of IFNAR2, but not IFNAR1, in HCV-infected cells. This loss of IFNAR2 was a post-transcriptional event, as the level of IFNAR2 mRNA was not apparently affected by the depletion of TNF-α or TNFR1 (S14D Fig), and most likely mediated by proteasomes, as its loss could be inhibited by the proteasome inhibitor MG132 but not by Bafilomycin A1 (S15 Fig). Taken together, the results shown in Fig 7 indicated that TNF-α induced by HCV was required to maintain the IFNAR2 expression level to support IFN signaling. In the absence of TNF-α, IFNAR2 was depleted by HCV and IFN signaling was impaired. HCV patients have an elevated serum level of TNF-α, and this level is positively correlated with the severity of liver diseases [7–9]. The source of TNF-α is unclear, but it is generally assumed that it is produced by immune cells such as macrophages [27]. In this report, we demonstrated that TNF-α could also be induced in HCV-infected cells. Although the amount of TNF-α produced by HCV-infected hepatocytes might be lower than that produced by professional immune cells such as macrophages [28], it was sufficient to trigger an inhibitory response on HCV replication. Our finding is consistent with a previous report, which described an increased level of TNF-α in the hepatocytes of HCV patients [9]. The induction of TNF-α by HCV was specific, as it could be blocked by the antibody that neutralized the infectivity of HCV (Fig 2C). This induction was biphasic, with the first phase of induction peaked at 2 hours post-infection (Fig 2A). The induction of TNF-α in the first phase was dependent on TLR7/8 (Fig 4A) and required no HCV gene expression or replication (Fig 2E). As TLR7 and TLR8 are activated by ssRNA, HCV either has to release the viral genomic RNA into endosomes during endocytosis to activate TLR7/8 or the HCV genomic RNA released into the cytosol after uncoating must be delivered immediately back into the endosomes. We favor the first scenario, as, if HCV RNA is released first into the cytosol, then it will likely also activate RIG-I and/or MDA5, which are cytosolic PRRs. However, we found that these two PRRs did not play a significant role in the induction of TNF-α (S3A Fig). If HCV indeed activates TLR7/8 during endocytosis, then the HCV virion must be disintegrated during this process for the genomic RNA to be released. This may be triggered by the acidic content of endosomes/lysosomes, which may destabilize HCV virion to release the viral RNA. The activation of the TLR7/8 signaling pathway by HCV led to the activation of NF-κB (Fig 5A). This pathway was required for the induction of TNF-α by HCV in the first phase. The induction of TNF-α in the second phase also required NF-κB, as the depletion of p65 NF-κB also suppressed the second-phase induction of TNF-α by HCV (Fig 5C). It does not appear likely that this second-phase induction of TNF-α was due to the second-round of infection by progeny virus particles, as this second-phase induction of TNF-α was long-lasting (Fig 1A). A number of factors had been shown in the past to activate NF-κB in HCV-infected cells. These factors include TLR3 and protein kinase R (PKR), which could both be activated by the double-stranded HCV RNA replicative intermediates [29, 30]. These factors are likely the reasons why HCV was also able to induce TNF-α in the later phase of infection. TNF-α induced by HCV suppressed HCV replication (Fig 6A). Our results indicated that this was likely due to its role in interferon signaling and the induction of ISGs (Fig 7A and 7B). We found that both TNF-α and TNFR1 participated in IFN signaling by maintaining the stability of IFNAR2, as in the absence of either one of them, IFNAR2 was lost in HCV-infected cells (Fig 7D), apparently due to degradation by proteasomes (S15 Fig). How HCV induced the degradation of IFNAR2 and how TNF-α antagonized this effect of HCV are interesting questions that remain to be determined. It is noteworthy that the degradation of IFNAR1 and IFNAR2 had previously been shown to be regulated by different mechanisms [31], and thus the selective degradation of IFNAR2 by HCV without affecting IFNAR1 was not unexpected. Nevertheless, our results reveal an interesting interplay between the virus and the host cell, with the virus attempting to blunt the IFN response by depleting IFNAR2 and the host cell overcoming this blunting effect of HCV by using TNF-α to restore the expression of IFNAR2. Although our results indicated that TNF-α could support IFN signaling to suppress HCV replication in cell cultures, the role of TNF-α in HCV replication and pathogenesis in vivo may be more complicated. This is due in part to our observation that TNF-α suppressed apoptosis of HCV-infected cells (S13 Fig), which would favor HCV persistence, in part to the pro-inflammatory activities of this cytokine, and in part to a recent report that TNF-α could depolarize liver cells to enhance HCV entry [28]. Thus, it is tempting to speculate that TNF-α induced in the first phase may enhance HCV entry whereas it induced in the second phase may suppress HCV replication. This may explain why in the clinical trial of a limited number of HCV patients, the TNF-α inhibitor Etanercept was found to improve the therapeutic effect of IFN-α and ribavirin on HCV rather than to suppress it [32]. Huh7 cells were maintained in Dulbecco’s modified eagle medium (DMEM) supplemented with 10% fetal bovine serum (FBS). Huh7.5 cells were maintained in DMEM supplemented with 10% FBS and 1% nonessential amino acids. Primary human hepatocytes (PHHs) were obtained from the Cell Culture Core Facility of the USC Research Center for Liver Diseases. They were maintained in DMEM medium supplemented with what10% FBS. The JFH1 (HCV genotype 2a) variant, which produced a high level of infectious virus particles [33], was propagated in Huh7.5 and used in our infection studies. Huh7 cells were infected with HCV with an MOI of 0.25, and the incubation medium was replaced with the fresh medium after 3 hours of infection. The incubation medium was collected at the time points indicated and TNF-α was assayed using the human TNF-α Instant ELISA kit (eBioscience) following the instructions of the manufacturer. Cells were lysed in M-PER Mammalian Protein Extraction Reagent (Thermo scientific) containing the protease inhibitor cocktail, 1mM PMSF, 1mM sodium orthovanadate, and 1mM sodium fluoride for 10 minutes on ice followed by a brief sonication. Cell lysates were cleared by centrifugation at 14,000 x g for 2 minutes. The supernatant was collected, boiled in Laemmli buffer for 5 minutes, and used for immunoblot analysis or stored at −80°C for future use. The rabbit anti-HCV core antibody was prepared in our laboratory [34]. TNFR1, p65 NF-κB, PARP, Caspase-8, and IRAK1 antibodies were from Cell signaling, and actin and IFNAR2 antibodies were from Sigma. IFNAR1 antibody was from Abcam, and GAPDH, TLR7, TLR8, and TRAF6 antibodies were from Santa Cruz. Horseradish peroxidase (HRP)-conjugated goat anti-rabbit and rabbit anti-mouse secondary antibodies were also purchased from Abcam. siRNAs targeting TNF-α, TNFR1, TLR3, TLR4, TLR7, p65 NF-κB, MyD88, IRAK1 and TAK1 were from Sigma. siRNAs targeting TLR7, TLR8, MDA5 and RIG-I were from Qiagen, and the siRNA targeting TRAF6 was from Santa Cruz. The transfection of siRNA into Huh7 cells was performed using Lipofectamine RNAiMax (Invitrogen). Briefly, Huh7 cells seeded in 100 mm dishes were transfected with 600 pmole siRNA for 6 hours. The transfected cells were further incubated in fresh media for 48 hours prior to infection with HCV. For generation of stable cell line, lentiviral particles were first produced in 293T cells through the coexpression of pLKO.TRC plasmid with shRNA insertion that targets TLR7 or TLR8 and packaging vectors. Lentiviral particles were harvested 60 hours post transfection and filtered. Huh7 cells were then infected with these lentiviral particles and selected with puromycin (2 μg/mL). Protein lysates were extracted 10 days post selection and immunoblotted with antibodies against TLR7 and TLR8. Total cellular RNA was isolated using TRIzol (Invitrogen) and reverse-transcribed to cDNA using SuperScript II First-Strand Synthesis System for RT-PCR (Invitrogen). The cDNA was then subjected to PCR amplification with the primer sets listed in the table in the supplemental information (S1 Table). For the quantification of HCV RNA, total cellular RNA was subjected to qRT-PCR using the TaqMan EZ RT-PCR Kit (Applied Biosystems, Foster City, CA) following the manufacturer’s instructions. HCV JFH1 primers 5′-TCTGCGGAACCGGTGAGTA-3′ (forward) and 5′-TCAGGCAGTACCACAAGGC-3′ (reverse) and the probe 5′-CACTCTATGCCCGGC CATTTGG-3′ were used for the qRT-PCR. The control GAPDH primer set with the probe was purchased from Applied Biosystems. For detection of other gene expressions, 100 ng total RNA was analyzed using the Power SYBR Green RNA-to-CT 1-Step Kit (Applied Biosystems, Foster City, CA). The primers used are shown in S1 Table. Huh7 cells with or without HCV infection were rinsed once with phosphate-buffered saline (PBS) and then trypsinized. Cells were washed once more with PBS before the subcellular fractionation using NE-PER Nuclear and Cytoplasmic Extraction Reagents (Thermo Scientific) following the manufacturer’s instructions. Binding of p65 NF-κB to the TNF-α promoter was analyzed using the Abcam ChIP kit following the manufacturer’s instructions. Briefly, Huh7 cells were infected with HCV for the indicated time period. After which, 1 x 106 cells were fixed with 4% formaldehyde for 10 minutes at the room temperature. Cells were then lysed and the chromatin was sheared by sonication for 10 minutes. The chromatin was then immunoprecipitated using the anti-p65 antibody or a control IgG. The immunoprecipitate DNA samples were analyzed for the TNF-α promoter by PCR.

10.1371/journal.pntd.0000436

Cholera Epidemics, War and Disasters around Goma and Lake Kivu: An Eight-Year Survey

During the last eight years, North and South Kivu, located in a lake area in Eastern Democratic Republic of Congo, have been the site of a major volcano eruption and of numerous complex emergencies with population displacements. These conditions have been suspected to favour emergence and spread of cholera epidemics. In order to assess the influence of these conditions on outbreaks, reports of cholera cases were collected weekly from each health district of North Kivu (4,667,699 inhabitants) and South Kivu (4,670,121 inhabitants) from 2000 through 2007. A geographic information system was established, and in each health district, the relationships between environmental variables and the number of cholera cases were assessed using regression techniques and time series analysis. We further checked for a link between complex emergencies and cholera outbreaks. Finally, we analysed data collected during an epidemiological survey that was implemented in Goma after Nyiragongo eruption. A total of 73,605 cases and 1,612 deaths of cholera were reported. Time series decomposition showed a greater number of cases during the rainy season in South Kivu but not in North Kivu. Spatial distribution of cholera cases exhibited a higher number of cases in health districts bordering lakes (Odds Ratio 7.0, Confidence Interval range 3.8–12.9). Four epidemic reactivations were observed in the 12-week periods following war events, but simulations indicate that the number of reactivations was not larger than that expected during any random selection of period with no war. Nyiragongo volcanic eruption was followed by a marked decrease of cholera incidence. Our study points out the crucial role of some towns located in lakeside areas in the persistence of cholera in Kivu. Even if complex emergencies were not systematically followed by cholera epidemics, some of them enabled cholera spreading.

With the number of cholera cases up to 73,000 during the last eight years and successive wars that have persisted for fifteen years, the North and South Kivu provinces of the Democratic Republic of Congo are currently heavily hit by both cholera outbreaks and war-related population displacements. Prior to this study, no research had been done to identify the sources of epidemics and the pathways used by cholera to spread throughout the Kivu provinces. Here we show that a few cities located on the lakeshore of Lake Kivu and Lake Tanganyika act as the main sources of cholera epidemics and that the number of cholera cases tends to increase during the rainy season. We also found that only a minority of population displacements were followed by cholera outbreaks. Finally, we think that the low number of cholera cases recorded after the Nyiragongo eruption is one more argument to implement programs aiming at restoring, and if possible improving, drinking water access following natural disasters

Numerous factors have been postulated to increase the risk of cholera outbreaks in a given area where cholera is already circulating among the population. The main environmental risk factors identified include heavy rainfall, blooms of plankton, and an increase in sea surface temperatures [1]. However, most studies have been performed in coastal areas and very little is known about environmental factors involved in the recurrence of cholera epidemics in inland areas. In this context, it has been recently shown that the lake areas have been the source of iterative cholera outbreaks in the inland areas of Katanga, a province located south-east of the Democratic Republic of Congo (DRC) [2]. Deadly cholera outbreaks have also been reported during complex emergencies (CEs) that are defined as “a humanitarian crisis in a country, region or society where there is a total or considerable breakdown of authority resulting from internal or external conflict, and which requires an international response that goes beyond the mandate or capacity of any single and/or ongoing United Nations (UN) country programme” [3]. Lastly, probably by analogy with CEs, the risk of cholera epidemics is also an often-repeated assertion in the aftermath of large-scale natural disasters [4]. The provinces of North and South Kivu, bordering Lake Kivu in the east of the DRC, present an exceptional accumulation of these risk factors. They have been the site of numerous dramatic events, including invasion and occupation by foreign forces, civil war, population displacements, a major volcano eruption and earthquakes. According to a recently published mortality survey, it is estimated that the conflicts and humanitarian crises, which ravaged the eastern part of the DRC, have taken the lives of 5.4 million people since 1998, and continue to leave as many as 45,000 dead every month [5]. Despite the tragic events encountered in the Kivu provinces, an epidemiological surveillance system was set up at the end of the 1990's, and it is still recording data on cholera and other communicable diseases. Here, we present a study designed to describe epidemiological patterns of cholera outbreaks in the Kivu provinces and to elucidate the influence of specific environmental and geographical factors, CEs and disasters on outbreaks of cholera. This study and the previous work that described the patterns of cholera outbreaks in Katanga and Eastern Kasaï [2] constitute the first step of a plan aiming to fight cholera in the DRC. From January 2000 through December 2007, reports of cholera cases and deaths were collected weekly from each health district (HD) of North and South Kivu provinces. Case-patients of cholera were defined as recommended by the World Health Organization (WHO): “any person 5 years of age or older in whom severe dehydration develops or who dies from acute watery diarrhoea”, with an age limit lowered to 2 years for cases associated with confirmed cholera outbreaks [6]. Also as recommended by the WHO, each new important outbreak was confirmed by culture and identification of Vibrio cholera O1 from stool samples [6]. North Kivu (53,855 km2, 4,667,699 inhabitants, 19 HDs) and South Kivu (65,000 km2, 4,670,121 inhabitants, 14 HDs) are located in the Great Rift Valley, and border Lake Edward, Lake Kivu, and the north edge of Lake Tanganyika. In Kivu, the climate is characterized by a rainy season from October to the end of May and a dry season the rest of the year. However, the rainy season is partially interrupted by a short dryer period in January and February. The relief of the Kivu provinces is dominated by several volcanic chains. Nyiragongo, the most active volcano, is located approximately 20 kilometres north of the city of Goma (400,000 inhabitants) near Lake Kivu. Nyiragongo last major eruption occurred on January 17, 2002, when lava flow destroyed one third of the city of Goma [7], claiming 147 victims (Didier Bompangue, personal data). In response to this disaster, the international community brought quick and massive help to the population by providing safe drinking water. Moreover, during the 12-week period following the disaster, access to health care facilities was improved, due to the humanitarian response. In particular, a program of drug supply was implemented to support the primary health centres in Goma and health facilities were made free for a six-week period, followed by another six-week period with reduced prices (0.2 $ instead of 1 $ for ambulatory care services, including drugs). Reports on CEs which occurred in the Kivu provinces from 2000 to 2007 were collected from the Reliefweb Website, which compiles information from a variety of sources, including UN Agencies and non-governmental organizations (NGOs) [7]. Among them, we further selected the CEs which were subjected to a medical assessment by humanitarian organizations and which involved more than 1000 internally displaced persons (IDPs). A geographic information system was established, based on the data collected from the 33 HDs of the two provinces. Following a previously described procedure [2], we statistically examined the relationship between the number of cholera cases in each HD and geographic and environmental variables (area, population, and presence/absence of cities with >100,000 inhabitants, of at least one commercial port, of major tracks or roads, and of lakeside location for each HD). Population and area were log transformed and log(population) included as an offset term in the model. Due to the overdispersion of cholera incidence, several kinds of generalized linear models were compared using quasi-Poisson, and type I and type II negative binomial distributions and they were checked for spatial structure. Stepwise selection of variables was performed in each case and the best models of each family were compared using the Akaïke index criterion, according to Venables and Ripley [8] and Rigby et al. [9] The relationship between the number of cholera cases in health districts and geographical variables was finally modelled using the type II negative binomial family (log link function for both the mean and the distribution parameter). The residuals were checked for spatial structure by plotting an empirical variogram where the distances were computed depending on the geographical coordinates of the centroid of each health district. A variogram envelope was then computed by performing 1000 permutations of the residual values on the spatial locations and the envelope limits were then compared to the variogram. In the present study, we failed to detect residual spatial autocorrelation. The rate of the cholera cases was mapped for the 33 HDs using ESRI shapefiles. Cross-correlations between time-series of HDs were computed [10]. Time series, which were synchronous (i.e. with no time lag) in a geographical area, were merged. This led to define 5 zones (zone 1: Mutwanga; zone 2: Goma, Kirotshe; zone 3: Bukavu, Katana; zone 4: Uvira, Nundu, Fizi; zone 5: Pinga, Walikale). The time-series obtained were decomposed into a trend, a seasonal component and a remainder using a seasonal-trend decomposition procedure based on loess regression (STL) following Cleveland et al. (1990) [11]. In time series analysis, non-parametric STL methods have the advantage of robustness and simplicity but do not allow predictions and detailed quantification of time-series parameters, not necessary for our purpose in the present study. The remainder was examined and, each week, zones with an above-average number of cases were marked as an epidemic reactivation. If epidemic reactivation was typically fostered by war events in a non-epidemic period, one would expect more epidemic reactivations within the 12 weeks following a war event than within the 12 weeks following any randomly selected no-war week. This hypothesis was tested on a basis of 1000 simulations, checking for the occurrence of at least one (or the absence of any) epidemic reactivations during the 12 weeks following each week randomly selected during a non-epidemic period. The number of randomly selected weeks considered was proportional to the number of war events actually observed in each zone. The impact of the Nyiragongo 2002 eruption was also analysed in relation to the dynamics of cholera epidemics. In addition, during the 12-week period following the eruption of Nyiragongo, an epidemiological survey was implemented in primary care patients of Goma. During this period, all cases of acute diarrhoea, upper and lower respiratory tract infection and fever were recorded from five local health centres located within the western area of Goma. These data were compared to the records of the centres corresponding to the three weeks before the disasters. Computing and graphical displays were done using ArcGIS 8.3 and R 7.2 [12], and the following additional packages: MASS version 44 [8], maptools [13], sp [14], GAMLSS [15], and geoR [16]. From January 2000 through December 2007 (416 weeks), a total of 73,605 cases and 1,612 deaths (lethality: 2.2%) from cholera were reported in North and South Kivu. Vibrio cholerae O1 El Tor Ogawa was isolated in 8 samples collected from North Kivu (among 38 samples) and 3 from South Kivu (among 29 samples). Vibrio cholerae O1 El Tor Inaba was found only in South Kivu (6 positive samples among 29). In Bukavu (South Kivu), Ogawa and Inaba serotypes were isolated during the same outbreak of cholera in 2005. All isolates were found to be sensitive to ciprofloxacin, erythromycin and nalidixic acid, and resistant to tetracycline, ampicillin and cotrimoxazole. During the eight-year study, both provinces experienced at least one outbreak of cholera per year, with peaks ranging from 130 to more than 700 cases a week (Figure 1). In South Kivu, cholera cases were reported in every week except for two short periods in 2001 and 2002 (Figure 1). Time-series decomposition showed a marked seasonal influence, with a greater number of cases during the rainy season. In North Kivu, no period without cholera could be identified. Seasonal patterns were notably different in North Kivu from those in South Kivu, with epidemics occurring during both dry and rainy seasons (Figure 1). Occasionally, periods of partial remission occurred, during which the weekly incidence of the disease was below 1/50,000 inhabitants (in 2001, weeks 29, 30, 31, 32, 33, 35 and 40, in 2004, weeks 23 and 28). Each time, cholera epidemics again appeared, stemming from residual cases located in Goma (North Kivu) and Uvira (South Kivu). The spatial distribution of cholera cases was heterogeneous, with a higher number of cases in HDs bordering lakes whereas two remote HDs, Kaziba and Shabunda reported no cases of cholera (Figure 2). Table 1 shows that the number of cholera cases was significantly higher in the presence of a lake (odds ratio [OR] 7, 95% confidence interval range [CI] 3.8–12.9). According to UN and NGOs reports, a total of 18 large-scale population displacements related to CEs has been recorded during the period. In six cases, these population displacements occurred during already ongoing cholera epidemics. Among the 12 remaining war events with displacements of population, four were followed by a cholera outbreak within a period of 12 weeks. Two of these cholera epidemics occurred in IDP camps, starting six and eight weeks after the arrival of the first IDPs in the settlements. However, simulations indicate that the number of reactivations was not larger than expected after any random selection of a week with no war event in a non-epidemic period (Figure 3). In 2002, the Nyiragongo volcanic eruption was not followed by any exacerbation of cholera incidence. A survey performed in five health centres located in the western area of Goma showed that, during this period, diarrhoeas accounted for only 6% of patients who were seen in the health centres. In the entire city of Goma, only 140 cholera cases (8 cases per week), without any deaths, were reported from January to April 2002. This low number of cases contrasts with an average of 29 cases per week usually encountered in Goma at this period of the year. Since the beginning of the 1990's, Kivu provinces have been identified as one of the most active foci of cholera in the world. In 1994, the refugee camps located around Goma and Bukavu experienced the deadliest cholera epidemics recorded during the last hundred years. This explosive outbreak of cholera, which affected Rwandan refugees, resulted in about 70,000 cases and 12,000 deaths [17]. Since that period, cholera epidemics have been common in the Kivu provinces, and, in a review of reported cholera outbreaks worldwide, from 1995 to 2005, Griffith et al. notes that the eastern DRC provinces are among the most affected zones in Africa [18]. Notwithstanding the inherent limitations associated with epidemiological data collected in developing countries, especially in a context of civil war, our study is the first that provides spatial data collected weekly during an eight-year period for a population of ten million living in an area severely affected by cholera. Even though the number of cases reported in this study is impressive (more than 73,000 cases), it is likely to be underestimated due to the situation of insecurity which prevailed in this region during the studied period, and which led to poor access to health facilities. By contrast, one cannot exclude that some patients suffering from other diarrhoeal diseases may have been included in the study: indeed, only a few clinical cases have been confirmed by culture, even if that may be due to inappropriate sampling procedures (samples collected from buckets containing chlorine) as well as an excessive delay to handle the samples to the laboratories. Moreover, given the lack of local laboratory facilities, other causes of acute watery diarrhoea could unfortunately not be investigated. For instance, in Bangladesh a number of epidemics of watery diarrhoea are actually caused by ETEC (Enterotoxigenic Escherichia coli), which would have a similar clinical presentation to that of the cholera cases here [19]. Therefore, our results obtained using a clinical definition of cases, represent only an estimate of the real cholera burden in the Kivu provinces. Nevertheless, we believe our results to highlight some significant aspects of cholera epidemiology that need further discussion. First, our findings, obtained from a study performed in a lakeside region, are consistent with what was recently found in the province of Katanga, located on the south side of Kivu in the eastern part of the DRC [2]. This previous study showed that 60 percent of cases that were observed in Katanga and Eastern Kasaï between 2002 and 2005 actually occurred in a few lake areas. In these two provinces, the number of cholera cases was significantly higher in the presence of a lake, (OR 7.5, 95% CI: 3.9–14.2). The present study confirms that the same trends can be observed in the Kivu Provinces. Similar to the role played by the towns of Kalemie and Bukama in Katanga, the cities of Goma, Bukavu and Uvira seem to act as the main sources of cholera epidemics in the Kivu provinces. From an operational point of view, this finding implies that more attention should be paid to cholera in these towns, especially in periods when outbreaks are starting with a small, but rising number of cases. Here, the influence of seasons, and the effect that some lakeside areas have on the persistence of the disease, could make the epidemiological pattern of cholera in Kivu comparable to patterns studied more comprehensively in Asian coastal areas [1],[20],[21]. Two studies have stressed the link between having cholera and living on the shores of a lake or a river, which includes drinking the water and bathing in it. One of these studies was carried out in Rumonge, a city in Burundi bordering Lake Tanganyika, the same lake where the port of Uvira is located [22]. Another one, carried out in nearby Lake Victoria in Kenya, suggested the possible existence of at least a transient environmental reservoir for cholera in the lake and evoked the possible role of water hyacinths in maintaining environmental sources of toxigenic cholera strains during inter-epidemic periods.[23]. However, lake water differ from estuarine brackish water that is known to be the natural reservoir of V. cholerae [24]–[26]. Even though lake water can sometimes be rich in plankton [27], the role of lakes as reservoirs for V. cholerae is not formally established because no study has yet demonstrated a long term persistence of toxigenic V. cholerae in the East African Rift Valley lakes. Our results show that there is a need for further studies to explore the role of lake environments in the persistence of cholera in inland Africa. Indeed, some endemic V. cholerae strains have long been isolated from fresh water [28] and Kirschner et al. recently demonstrated the permanent existence of non-toxigenic V. cholerae strains, which can rapidly grow in a free-living state in one natural lake water in Austria [29]. The eruption of Nyiragongo, the largest natural disaster reported during this period, was not followed by a cholera outbreak, or by any other disease outbreak. On the contrary, during the months that followed the disaster, the rates of cholera infection were among the lowest that have been recorded in the city of Goma during the 8 years of the study. Several hypotheses can be advanced to explain this low number of cholera cases. This can be due to the impact of the emergency program, but other possible explanations cannot be ruled out, including the fact that the volcanic eruption could have decreased the likelihood of a cholera outbreak secondary to the alterations in water sources and usage patterns. Our finding is in agreement with a recent study showing that earthquakes, tsunamis and volcano eruption were not usually followed by epidemics [30]. In particular, for 20 years, not a single cholera outbreak has been recorded in the aftermath of geophysical disaster, even after the cataclysmic tsunami in Asia in 2004. Here, we show that even in a place where cholera outbreaks are common and during a period known to be favourable for epidemics (the rainy season), a disaster that destroyed approximately 12,000 homes and partially destroyed the water supply pipelines of a town of 400,000 inhabitants, the occurrence of cholera epidemics was not unavoidable. The search for the impact of the CEs indicates that they do not systematically represent a triggering factor for cholera outbreaks. However, in our study, we also saw that in four cases, the occurrence of a CE led to the exacerbation of cholera, including, in two cases a cholera outbreak, which started in a displaced settlement a few weeks after the arrival of IDPs. Actually, several conditions are necessary for a CE to father a cholera outbreak. Among these conditions, there is the fact that some of the IDPs fleeing from the conflict area should have a previously acquired cholera infection (symptomatic or in incubation), and/or the fact that the IDPs should move into areas where cases of cholera are already present. It would also be necessary for there to be insufficient or no assistance to the IDPs (i.e. provisions for safe drinking water and a system of free health care). These circumstances were met in 1994 when one million refugees coming from Rwanda settled in makeshift camps around Goma, overwhelming the capacities of humanitarian staff already present in the town. More recently, due to the insecurity that prevailed around Goma in the summer and fall of 2008, a cascade of cholera outbreaks that began in Rutshuru in the North of Goma have been recorded in North Kivu: Rutshuru (beginning on week 37), Goma and Karisimbi (beginning on week 40), Walikale and Birambizo (beginning on week 44). In each of these HDs, the outbreak was introduced by people escaping from battle-hit areas located North of Goma (D. Bompangue, personal data). Simultaneously, due to excessive danger for fieldworkers, numerous NGOs fled from Goma and neighbouring areas as cholera outbreaks started, leading to a disorganization of the programs aimed to limit the spread of the epidemic. In conclusion, the epidemiology of cholera in both Kivu provinces confirms our previous findings in Katanga and eastern Kasaï, and highlights the role of some towns located in lakeside areas as sources of cholera outbreaks. The results of this study show that, even if each CE with numerous IDPs is not systematically followed by a cholera outbreak, CEs may facilitate spreading of already existing outbreaks due to the fleeing of infected IDPs to new areas where NGOs cannot reach them due to an excessive danger for fieldworkers. By contrast, even in a context of CE and natural disaster, the occurrence of epidemics is not unavoidable. For example, the number of cholera cases was lower than expected after the partial destruction of the town by the Nyiragongo eruption followed by the implementation of an emergency program. We think that this low number of cholera cases is one more argument to implement programs aiming to restore, and if possible to improve, drinking water access following natural disasters.

10.1371/journal.pbio.1001537

Transsynaptic Coordination of Synaptic Growth, Function, and Stability by the L1-Type CAM Neuroglian

The precise control of synaptic connectivity is essential for the development and function of neuronal circuits. While there have been significant advances in our understanding how cell adhesion molecules mediate axon guidance and synapse formation, the mechanisms controlling synapse maintenance or plasticity in vivo remain largely uncharacterized. In an unbiased RNAi screen we identified the Drosophila L1-type CAM Neuroglian (Nrg) as a central coordinator of synapse growth, function, and stability. We demonstrate that the extracellular Ig-domains and the intracellular Ankyrin-interaction motif are essential for synapse development and stability. Nrg binds to Ankyrin2 in vivo and mutations reducing the binding affinities to Ankyrin2 cause an increase in Nrg mobility in motoneurons. We then demonstrate that the Nrg–Ank2 interaction controls the balance of synapse growth and stability at the neuromuscular junction. In contrast, at a central synapse, transsynaptic interactions of pre- and postsynaptic Nrg require a dynamic, temporal and spatial, regulation of the intracellular Ankyrin-binding motif to coordinate pre- and postsynaptic development. Our study at two complementary model synapses identifies the regulation of the interaction between the L1-type CAM and Ankyrin as an important novel module enabling local control of synaptic connectivity and function while maintaining general neuronal circuit architecture.

The function of neuronal circuits relies on precise connectivity, and processes like learning and memory involve refining this connectivity through the selective formation and elimination of synapses. Cell adhesion molecules (CAMs) that directly mediate cell–cell interactions at synaptic contacts are thought to mediate this structural synaptic plasticity. In this study, we used an unbiased genetic screen to identify the Drosophila L1-type CAM Neuroglian as a central regulator of synapse formation and maintenance. We show that the intracellular Ankyrin interaction motif, which links Neuroglian to the cytoskeleton, is an essential regulatory site for Neuroglian mobility, adhesion, and synaptic function. In motoneurons, the strength of Ankyrin binding directly controls the balance between synapse formation and maintenance. At a central synapse, however, a dynamic regulation of the Neuroglian–Ankyrin interaction is required to coordinate transsynaptic development. Our study identifies the interaction of the L1-type CAM with Ankyrin as a novel regulatory module enabling local and precise control of synaptic connectivity without altering general neuronal circuit architecture. This interaction is relevant for normal nervous system development and disease as mutations in L1-type CAMs cause mental retardation and psychiatric diseases in humans.

Transsynaptic interactions mediated by cell adhesion molecules (CAMs) control the formation, function, and stability of synaptic connections within neuronal circuits. While a large number of synaptogenic CAMs controlling the initial steps of synapse formation have been identified [1],[2], we have only limited knowledge regarding the identity or regulation of CAMs selectively controlling synapse maintenance or plasticity. Information processing within neuronal circuits is adjusted by the selective addition or elimination of individual synapses both during development and in response to activity [3],[4]. These changes in connectivity can occur in very close proximity to stable synapses [5],[6] indicating the existence of mechanisms capable of local alterations of transsynaptic adhesion. Potential mechanisms to alter binding affinities of CAMs include direct alterations of extracellular domains through binding of ligands like metal ions (e.g., Ca2+) or indirect mechanisms through the selective association of CAMs with the intracellular cytoskeleton via adaptor proteins [7]. Modulation of intracellular interactions via posttranslational modifications can alter mobility, clustering, and adhesive force of CAMs [8]. For example, it has been demonstrated for the Cadherin–Catenin complex that changes in biophysical properties can induce changes in synapse morphology, strength, or stability and modulate transsynaptic signaling [9],[10]. To identify cell adhesion molecules potentially controlling synapse maintenance and plasticity we performed an unbiased in vivo RNA interference (RNAi) screen at the larval neuromuscular junction (NMJ) of 287 transmembrane proteins that are predicted to function as synaptic CAMs based on their domain structure [11]. These included Ig-domain containing proteins, Leucine-rich repeat proteins, Cadherins, Integrins, Semaphorins, and others (Table S1). In this high-resolution screen we identified the Drosophila L1-type CAM Neuroglian (Nrg) as a key regulator for synapse stability. Nrg encodes the Drosophila ortholog of the L1-type protein family [12] that is composed of four closely related members in vertebrates: L1, CHL1 (close homolog of L1), NrCAM (neuronal CAM), and Neurofascin [13],[14]. L1-type IgCAMs usually consist of 6 Ig-domains, 3–5 fibronectin type III domains, a single transmembrane domain, and an intracellular tail. The extracellular domain of L1 family proteins can mediate cell–cell adhesion via homophilic interactions and can also engage in a variety of heterophilic interactions with other Ig-domain proteins (e.g., NCAM, TAG-1, Contactin, and others), extracellular matrix proteins, or integrins [13]–[15]. The intracellular tail contains distinct protein–protein interaction domains potentially controlling the localization and function of L1 proteins [16]–[18]. Most prominent is a central Ankyrin interacting motif that is highly conserved among all vertebrate L1 family proteins and Drosophila Neuroglian [16],[19]. Phosphorylation of the tyrosine within this FIGQY motif abolishes binding to Ankyrins [20]–[23]. The Ankyrin-binding domain is essential for mediating neuronal function in vivo in C. elegans, however it is dispensable for L1-mediated homophilic adhesion in transfected cells in culture [24],[25]. Importantly, we previously identified Drosophila ankyrin2 (ank2) as an essential gene for synapse stability at the larval neuromuscular junction [26],[27]. Ank2 together with the presynaptic spectrin cytoskeleton and the actin capping protein Hts/Adducin controls NMJ formation and maintenance and provides a scaffold to link the actin and microtubule cytoskeleton to synaptic cell adhesion molecules [28],[29]. Based on this potential biochemical interaction Nrg might encode the CAM upstream of this Ank2/Spectrin scaffold to control synapse development. Human mutations in L1CAM cause a broad spectrum of neurological disorders (L1 or CRASH syndrome) including MASA syndrome (mental retardation, aphasia, shuffling gait, adducted thumbs), agenesis of the corpus callosum, and spastic paraplegia. In addition, hypomorphic mutations in L1CAM and NrCAM have been linked to psychiatric diseases [14],[15],[30]. In correspondence with the human disease, animal models implicated L1-type proteins in nervous system development [13],[14]. At the cellular level L1-type proteins are involved in the control of neurite outgrowth, axon pathfinding, and fasciculation and synapse development [14],[16],[31]. The subcellular localization of L1-type proteins contributes to the establishment and maintenance of specialized neuronal membrane compartments including the axon initial segment (AIS) and nodes of Ranvier [32]–[35]. While these studies highlight essential functions of L1-type proteins, potential redundant or antagonistic functions between different L1-type proteins may mask the full extent of their importance for nervous system development. Indeed, evidence for redundant functions between L1-type proteins was provided by a double mutant analysis of L1CAM and NrCAM [36]. Together with the requirement of L1-type proteins for early nervous system development, this confounds our current understanding of the contribution of L1-type CAM to synapse development and plasticity. In addition, mechanistic insights into the in vivo control of L1-type CAM function at synapses are lacking to date. Nrg encodes the sole homolog of L1-type CAMs in Drosophila with equal homology to all four vertebrate proteins. This provides a unique opportunity to unravel the contributions and mechanisms of regulation of L1-type CAMs in synapse development and maintenance. Here, we generate a series of Pacman-based mutations that allowed us to identify the specific contributions of extra- and intracellular domains of Nrg for synapse stability. We then provide evidence that binding of Nrg to Ank2 is critical for the control of mobility of Nrg in vivo. We demonstrate that modulation of the Nrg–Ank2 interaction allows precise control over the balance between synapse formation and stability. Finally, we demonstrate that dynamic regulation of the Ankyrin-binding domain of Nrg is essential for the coordination of pre- and postsynaptic development via transsynaptic signaling mechanisms at central synapses. To identify cell adhesion molecules necessary for the maintenance of synaptic connections we performed a transgenic RNAi-based screen [37] of 287 transmembrane proteins encoding potential cell adhesion molecules based on their domain structure and previously described functions in axon guidance or synapse development (Table S1). We knocked down candidate genes simultaneously in presynaptic neurons and postsynaptic muscles and analyzed third instar NMJs for defects in synapse stability using selective pre- and postsynaptic markers (Figure 1A–F). In wild-type animals the presynaptic active zone marker Bruchpilot (Brp) is found in close opposition to postsynaptic glutamate receptor cluster at all individual synapses within the presynaptic nerve terminal demarcated by the membrane marker Hrp. In contrast, NMJs displaying postsynaptic glutamate receptor clusters without opposing presynaptic active zone markers and a fragmentation of the presynaptic membrane indicate synapse retractions [27],[28]. We identified Drosophila Nrg as the major hit in our screen resulting in synaptic retractions at more than 50% of all NMJs on muscle 4. We first tested whether specific knockdown of Nrg either in the motoneuron or the muscle also impairs synapse stability. Presynaptic knockdown of Nrg was sufficient to cause synaptic retractions equivalent to the simultaneous pre- and postsynaptic knockdown (Figure 1B,G). In contrast, muscle-specific nrg RNAi did not lead to a significant increase in synaptic retractions (Figure 1C,G). We obtained similar results when we expressed an independent nrg RNAi line (Figure 1E–G) and were able to enhance the phenotype by combining different motoneuron Gal4 drivers or by co-expressing UAS-dcr2 to enhance RNAi efficacy (Figure 1G). In addition, we observed similar rates and severities of synapse retractions when using independent pre- and postsynaptic markers and when analyzing different subsets of muscles (Figure 1D–F; Figure S1; Table S2). To monitor the efficiency of our RNAi mediated knockdown, we directly analyzed Nrg protein levels. Drosophila nrg encodes two specific isoforms, the ubiquitous isoform Nrg167 and the neuronal specific isoform Nrg180 (Figure 2A) [38]. Nrg180 was present throughout motoneuron axons and within the presynaptic nerve terminal (Figure S2A), and in addition, Nrg167 was present in muscles and glial cells (Figure S2E). Western blots of larval brain extracts and analysis at the larval NMJ demonstrated that all combinations of presynaptic nrg RNAi efficiently knocked down Nrg180 (Figure 1H; Figure S2B). Similarly, muscle-specific knockdown of Nrg caused a loss of Nrg3c1 staining in the muscle that can be attributed to a loss of Nrg167 (Figure S2F). The loss of postsynaptic Nrg resulted in a significant change in presynaptic Nrg levels and distribution at the NMJ (Figure S2C, F; reduction to 62.9±3.7% of wild-type protein within the presynaptic terminal, p<0.001), indicating a requirement of postsynaptic Nrg for presynaptic Nrg localization. However, this reduction in presynaptic Nrg180 levels was not sufficient to impair synapse stability (Figure 1C,F,G), implicating the potential existence of alternative postsynaptic Nrg interaction partners essential for NMJ maintenance [39]. Likewise, presynaptic knockdown of Nrg reduced Nrg staining in the axon and at the synaptic terminal but did not significantly alter Nrg distribution within the postsynaptic subsynaptic retriculum (SSR) (Figure S2G,H). To validate the specificity of our RNAi-mediated knockdown we aimed to rescue the synaptic phenotypes by co-expressing wild-type Nrg180. Simultaneous expression of UAS–nrg180 significantly rescued synapse stability by restoring Nrg180 protein levels both in larval brains and at the NMJ (Figure 1G,H and Figure S2D). In contrast, co-expressing UAS–mCD8–GFP or UAS–fasciclinII (NCAM homolog) failed to rescue the synaptic retraction phenotype (Figure 1G). Thus, the specific loss of pre- but not postsynaptic Nrg caused an impairment of synapse stability. To gain insights into the molecular processes inducing synaptic retractions in animals lacking presynaptic Nrg, we analyzed the distribution of two presynaptic components, Ank2 and Futsch, at early stages of synaptic retractions in comparison to active zone and vesicle markers. The presynaptic adaptor protein Ank2 is an essential molecule for synapse stability and Ankyrins can directly bind to Nrg [26],[27],[40]. The microtubule-associated protein Futsch serves as a marker for presynaptic microtubules as loss of microtubules represents an early step in synaptic retractions at the Drosophila NMJ [27],[29],[41]. At wild-type NMJs, Ank2 and Futsch were present in all terminal boutons together with Brp and DvGlut. In contrast, after presynaptic knockdown of Nrg, we observed NMJs lacking Ank2 or Futsch in terminal boutons that still contained presynaptic Brp or DvGlut (Figure 1I–K). Thus loss of Ank2 and of the associated microtubule cytoskeleton may represent early steps during synapse retractions caused by the loss of Nrg. Two domains of Nrg that may be essential for synapse maintenance are the extracellular Ig domains mediating association with postsynaptic CAMs and the intracellular Ankyrin binding domain that provides a link to the presynaptic cytoskeleton. To directly test for a potential role of these domains we generated genomic rescue constructs that allow expression of wild-type and mutated nrg at endogenous levels using a site-directed Pacman-based approach [42],[43]. We first generated a transgenic construct encompassing the entire nrg locus including 25 kb upstream and 10 kb downstream regulatory sequences (P[nrg_wt]; Figure 2A). This construct fully rescued the embryonic lethal nrg null mutations nrg14 and nrg17. We then used galK-mediated recombineering [44] to generate a deletion of the extracellular Ig domains 3 and 4 (P[nrgΔIg3–4]), thereby completely disrupting hetero- and homophilic binding capacities of Nrg [45]. In addition, we generated specific deletions of the Ankyrin-binding domains of Nrg167 and Nrg180 that are encoded by unique exons (P[nrg167ΔFIGQY] and P[nrg180ΔFIGQY]; Figure 2A). All constructs were inserted into the genomic insertion site attP40 to ensure identical expression levels. While P[nrg167ΔFIGQY] and P[nrg180ΔFIGQY] rescued the embryonic lethality of nrg14 mutants similar to the wild-type construct, P[nrgΔIg3–4] failed to rescue lethality. In order to analyze the larval NMJ of nrg14; P[nrgΔIg3–4] mutant animals, we combined the Pacman rescue approach with the MARCM technique [46]. This allows the generation of mCD8–GFP-marked motoneurons expressing only the mutated form of Nrg. First we analyzed motoneurons completely lacking nrg using the nrg14 null mutation and observed two striking phenotypes. We found synapse retractions indicated by NMJs displaying remnants of the presynaptic MARCM membrane marker opposite postsynaptic glutamate receptors but lacking the presynaptic marker Brp (Figure 2C). While synapse retractions were only observed at low frequency, about 50% of all MARCM motoneuron axons ended in “bulb-like” structures within nerve bundles and were not connected to a postsynaptic muscle (Figure 2F and Figure S3E). The wild-type nrg Pacman construct fully rescued the axonal and NMJ phenotypes (Figure 2B,F and Figure S3A,D), however, P[nrgΔIg3–4] failed to rescue these defects and the presence of P[nrg180ΔFIGQY] resulted only in a partial rescue. In both genotypes we observed synaptic retractions as well as axons ending in bulbs distant from a potential target muscle (Figure 2D–F; Figure S3C,G; Table S1). In contrast, no defects were observed in the presence of P[nrg167ΔFIGQY] indicating that only the Ankyrin binding motif of Nrg180 is essential within motoneurons (Figure 2F; Figure S3B,F; Table S2). Prior studies showed a delay of axonal outgrowth in nrg mutant embryos [47]–[49]. Our axonal phenotypes would be equally consistent with a stalling of axons or with a retraction of axons after initial innervation of target muscles. Importantly, we observed mutant MARCM motoneurons where we could link synapse eliminations to axons ending in bulb-like structures. Figure 2E shows an example of a complete elimination of an NMJ indicated by the loss of presynaptic vesicles, while fragments of the mCD8–GFP-marked motoneuron membrane were still present opposite postsynaptic Dlg. We traced fragmented membrane remnants over a distance of more than 150 µm to the retraction bulb-like structure (Figure 2E; see Figure S3C for another example). Additionally, we observed large axonal swellings in the same axon further proximal toward the cell body of the motoneuron (Figure 2E). Rates of retraction bulbs and axonal swellings were identical in MARCM clones of nrg14 and nrg14; P[nrg180ΔIg3–4] animals (Figure 2F). Finally, analysis of the innervation pattern of motoneuron axons forming stable NMJs at this larval stage demonstrated that loss of Nrg did not result in obvious axon guidance defects. We observed similar rates of innervations for all four major classes of motoneurons in all genotypes (Figure 2G and Figure S2H). In summary, while we cannot exclude a role for Nrg in axonal outgrowth, we provide clear evidence that Nrg is required for the maintenance of the NMJ and that this function requires both the extracellular domain and the intracellular Ankyrin-binding domain of Nrg180 but not of Nrg167. Based on these results we aimed to unravel the molecular mechanisms controlling the synaptic function of Nrg through the interaction with the Ankyrin-associated cytoskeleton. Prior studies in vertebrates demonstrated that phosphorylation of the conserved tyrosine residue within the FIGQY motif of L1-type proteins has the potential to abolish the interaction with Ankyrins [20]–[23]. Similarly, Yeast-2-Hybrid assays showed that Nrg can bind to Drosophila Ankyrin1 and Ankyrin2 and that replacing the tyrosine with a phenylalanine (Y-F) reduces binding capacities [50]. Therefore, we first tested whether the neuronal isoform Nrg180 can directly bind to the large isoform of Ank2 (Ank2-L) that is present within the presynaptic nerve terminal (Figure 1I,J). Using the Nrg180-specific antibody Nrg180BP104 we were able to co-immunoprecipitate Ank2-L from larval brain extracts demonstrating that Nrg180 and Ank2-L interact in vivo (Figure 3A). Next we used IP-assays to further characterize the interaction between Nrg180 and Ank2. We generated tagged Nrg180 and Ank2 UAS constructs (Ank2-S, short isoform of Ank2 containing all potential Nrg interacting domains) and co-expressed the constructs in Drosophila S2 cells. We were able to efficiently pull-down Nrg180 using Ank2-S and vice versa (Figure 3B and unpublished data). To alter the binding properties and to potentially mimic a gradual increase in phosphorylation levels of Nrg180 in vivo, we generated a series of mutations replacing the tyrosine with a phenylalanine (Y-F), aspartate (Y-D), or alanine (Y-A) or by deleting the entire FIGQY motif (ΔFIGQY). Compared to wild-type we observed a 30% decrease in binding capacities for the Y-F, a 70% decrease for the Y-D, and a 90% reduction for the Y-A mutation. The deletion of the FIGQY motif essentially abolished the Nrg–Ank2 interaction (Figure 3B,C). Thus, we have identified a series of mutations that allows us to characterize the function and regulation of Nrg in vivo. These mutations potentially allow the differentiation between processes depending on Nrg bound to Ankyrins and processes depending on a differential regulation of the FIGQY motif. Studies in Drosophila and vertebrates demonstrated that impairing the interaction of CAM with the cytoskeleton results in an increase in lateral mobility and a simultaneous reduction of adhesive properties [20],[51],[52]. To test whether Nrg is regulated in a similar manner we analyzed the impact of the FIGQY-mutations on the biophysical behavior of Nrg in vivo. Therefore, we generated GFP-tagged UAS-constructs of all nrg180-FIGQY mutations and used site-specific integration to generate transgenic flies that will express equal protein levels after Gal4 activation. Analysis of larval brain extracts after expression of the constructs in motoneurons (ok371–Gal4) in a wild-type background demonstrated equal protein levels comparable to endogenous nontagged Nrg180 (Figure 3E upper and middle panels, analysis with an anti-GFP antibody and with a newly generated anti-Nrg180cyto antibody that recognizes the cytoplasmic tail of Nrg180 C-terminal to the FIGQY motif). The Nrg180-specific antibody Nrg180BP104 specifically recognizes the Nrg180 FIGQY motif as the antibody detects wild-type GFP-tagged Nrg180 but not any mutant proteins (Figure 3E, lower panel). We used fluorescence recovery after photobleaching (FRAP) to test whether the FIGQY mutations affect the mobility of Nrg180-GFP within motoneurons in vivo. For wild-type Nrg180 we observed a mobile fraction of about 40% of total protein (Figure 3D,F,G). The deletion of the Ank2 binding domain increased the mobile fraction of Nrg180 by a factor of two to about 80%. For the tyrosine-specific point mutations, we observed a significant increase in the mobile fraction compared to wild type but to a lesser extent than for Nrg180ΔFIGQY (Figure 3D,F,G). Thus, selectively impairing the interaction between Nrg180 and Ank2 significantly changes the mobility of Nrg180 in motoneurons in vivo. To address the relevance of this Nrg180–Ank2 interaction in vivo, we introduced all FIGQY-specific mutations into the wild-type nrg Pacman construct. In addition, we generated a deletion of the Nrg180 specific C-terminus including the FIGQY motif, a complete deletion of the FIGQY motif of Nrg167 as well as a specific deletion of the last three amino acids of Nrg180 as this potential PDZ-protein interacting domain has been implicated in axon outgrowth of mushroom body neurons (Figure 2A and Table S2) [53]. All constructs were inserted into the attP40 genomic landing site and crossed into the background of the nrg14 null mutation to create flies that express only mutant Neuroglian protein under endogenous control, thereby mimicking the effect of knock-in mutations. All modifications of the intracellular cytoplasmic domains of Nrg167 and Nrg180 rescued the embryonic lethality associated with nrg null mutations (nrg14 and nrg17), allowing an analysis at the third instar larval stage. To confirm the specificity of our mutations and the expression levels and localization of the mutant proteins, we analyzed the animals with specific antibodies recognizing either both isoforms of Nrg (Nrg3c1) or only Nrg180 (Nrg180BP104) [38]. As observed for the UAS-constructs, Nrg180BP104 recognizes only wild-type Nrg180 but none of our FIGQY mutations (Figure S4A,B; Figure 3E; unpublished data), thereby demonstrating that our Pacman-rescued flies indeed only express the mutant version of the protein and no wild-type Nrg180 protein. We verified this using the Nrg180cyto antibody as well as a second new antibody, NrgFIGQY, that recognizes the wild-type FIGQY motif of both Nrg167 and Nrg180 but not any mutant versions (Figure S4A,B). We were able to unambiguously identify all mutated proteins to demonstrate that all constructs are expressed at wild-type levels within larval brains (Figure S4B,D). In addition, all constructs enabled Nrg180 localization to the presynaptic nerve terminal and Nrg167 expression within glial cells and muscles (Figure S4A,C). Our data demonstrate that the FIGQY domain is not essential for presynaptic localization of Nrg180. Next, we systematically determined the requirement of the different domains for synapse development in third instar larvae. Our analysis of synapse stability revealed a significant increase in the frequency and severity of synaptic retractions in mutants with severely disrupted Nrg180–Ank2 interactions (nrg180Y-D, nrg180Y-A, nrg180ΔFIGQY) but not in animals expressing Nrg180 with only slightly impaired Ank2 binding capacities (nrg180Y-F) (Figure 4A–E and Table S2). The FIGQY motif of Nrg167 and the PDZ protein-binding domain of Nrg180 are not essential for normal NMJ development (Figure 4D,E; Figure S5B; Table S1). These phenotypes are consistent with our results for the nrg14; P[nrg180ΔFIGQY] and nrg14; P[nrg167ΔFIGQY] MARCM clones (Figure 2F and Table S2). The observation that impairing the Nrg–Ank2 interaction only in motoneurons weakened synapse stability provided an alternative way to test for a potential contribution of postsynaptic (muscle) Nrg for synapse maintenance. Therefore, we selectively knocked down Nrg in muscles of nrg14; P[nrg_wt] or nrg14; P[nrg180ΔFIGQY] animals. Indeed, we observed a significant increase in the frequency and severity of synaptic retraction when knocking down postsynaptic Nrg in the sensitized animals lacking the Ank2 interaction domain but not in animals expressing the wild-type Pacman construct (Figure S6A–D). We next asked whether the Nrg180 FIGQY motif is required for the synaptic localization of Ank2-L to mediate NMJ stability. Interestingly, we did not observe obvious alterations of presynaptic Ank2-L localization or protein levels in P[nrg_wt, 180Y-F, or 180ΔFIGQY] mutant animals at stable synapses when compared to control animals (Figure S7A; p>0.05 for comparison of protein levels; unpublished data). Thus, Nrg and Ank2 do not depend on each other for initial synaptic localization but display a high sensitivity toward normal levels of their interaction partner as they are among the first proteins to be lost at ank2 or nrg mutant semi-stable NMJs (Figure 1I,J and Figure S7B–D). In addition to the synapse stability defects, we observed a second striking defect in these animals. With an increasing reduction in Ank2 binding capacities of Nrg180 we observed an increase in growth of the NMJ as reflected by an increase in the span of the presynaptic nerve terminal and an increase in the number of synaptic boutons (Figure 5A–E). At the same time we observed a corresponding decrease in synaptic bouton area (Figure 5F). Interestingly, only subtle alterations were observed for the Nrg180 Y-F mutation that still binds Ank2 efficiently (Figure 5B,E,F). The identical phenotypes of the FIGQY deletion and of the C-terminal deletion indicate that control of NMJ growth critically depends on the Nrg180-FIGQY motif (Figure 5E,F). Similar to the analysis of synapse stability we did not observe any phenotypes in nrg14; P[nrg167ΔFIGQY] or nrg14; P[nrg180ΔPDZ] mutant animals (Figure 5E,F and Figure S5D). Finally, we tested whether we could mimic these growth defects by ectopically expressing mutant Nrg180 (UAS-nrg180ΔFIGQY-GFP) in wild-type animals. High levels of expression of Nrg180ΔFIGQY-GFP but not of Nrg180-GFP resulted in an almost 2-fold increase in bouton number and in the span of the presynaptic nerve terminal (Figure S8A–D). Together, our data demonstrate that the loss of Ank2 binding capacities of Nrg180 correlates with both a loss of NMJ growth control and an impairment of synapse stability, suggesting that these two parameters are tightly coupled. To address synaptic functions of the Nrg180–Ank2 interaction in the central nervous system (CNS), we extended our analysis to the adult Giant Fiber (GF) circuitry. We used the GF to TTMn (Tergo-trochanteral motoneuron) connection as a model neuro-neuronal synapse as it provides precise genetic control of pre- and postsynaptic neurons [54]. Previous analysis of nrg mutations affecting either homophilic cell adhesion properties (nrg849) or protein levels (nrg305) identified both axon guidance and synaptic defects at the GF terminal [55],[56]. Our Pacman-based mutants enabled us to directly determine potential requirements of the intracellular regulation of the Nrg–Ankyrin interaction for GF circuit formation and function. We analyzed the function of the GF to TTM (Tergo-Trochanteral Muscle) pathway in all viable nrg14; P[nrg-mutant] animals by intracellular recordings from the TTM using either brain or thoracic stimulation to differentiate between potential GF-TTMn synapse or TTMn NMJ defects. Importantly, presence of the wild-type nrg construct in nrg null mutants (nrg14; P[nrg_wt]) established normal function of the GF-TTMn circuit. We observed no significant differences in average response latencies or following frequencies after a train of stimulations at 100 Hz when compared to wild-type control animals (Figure 6). In contrast, all mutations affecting the Nrg180 FIGQY motif caused equally severe impairments of GF circuit function. The average response latency, a measure for synaptic strength, was significantly increased (Figure 6), and mutant animals were not able to follow trains of high-frequency stimulations; in some animals we observed a complete absence of responses (Figure 6B,D). In contrast, when we bypassed the GF and stimulated the motoneurons directly using thoracic stimulation, both response latency and ability to follow high-frequency stimulation were normal in all tested animals (unpublished data). This indicates that the observed defects were specific to the GF-TTMn synaptic connection. Similarly, we observed a disruption of synapse function when expressing UAS-Nrg180ΔFIGQY-GFP simultaneously pre- and postsynaptically at the GF synapse in wild-type animals (average latency increased to 1.01±0.057 ms; following frequency reduced to 52.34±7.17%), further demonstrating the importance of the Nrg180 FIGQY motif for normal GF synapse development. In contrast, neither the Nrg167 FIGQY nor the C-terminal Nrg180 PDZ protein-binding domains were essential for GF circuit function (Figure 6C,D). In order to identify potential morphological phenotypes and to distinguish between axon guidance and synaptic defects, we co-injected large (Rhodamin-dextran) and small (Biotin) fluorescent dyes into the GF. In wild-type animals the large dye is confined to the GF and reveals the morphology of the synaptic terminal. In wild-type animals the GF-TTMn synapse grows to a large presynaptic terminal with mixed electrical and chemical synapses [54]. Biotin can pass through gap-junctions and thereby dye-couple pre- and postsynaptic neurons in animals with a synaptic connection. While we observed no obvious morphological alterations of GF terminals in nrg14; P[nrg_wt], nrg14; P[nrg167ΔFIGQY], or nrg14; P[nrg180ΔPDZ] mutant flies, all mutations affecting the Nrg180 FIGQY motif resulted in severely disrupted GF terminals (Figure 7A,B). The GFs were present within the synaptic target area, however large areas of the synaptic terminals were either thinner or swollen and often contained large vacuole-like structures (Figure 7A, insets). Similar to the electrophysiological phenotypes, we observed no obvious qualitative or quantitative differences between different Nrg180 FIGQY mutations. Next, we directly tested for the presence of a synaptic connection between the GF and the postsynaptic TTMn using the dye-coupling assay. We found a residual synaptic connection in more than 90% of animals carrying mutations in the Nrg180 FIGQY motif (Figure 7A,C). However, dye-coupling was often weaker or required longer injection times in mutant animals when compared to animals rescued with the wild-type construct. These results indicate that at most GF terminals of Nrg180FIGQY mutants, synaptic connections with at least a small number of gap junctions were established. When we correlated the ability to dye-couple with the electrophysiological properties of these synapses, we observed that approximately 40% of Nrg180FIGQY mutant animals that were positive in the dye-coupling assay did not show any functional response (Figure 7C). This suggests that the synaptic strength was below the threshold to trigger an action potential in the postsynaptic TTMn. In contrast, neither the deletion of the FIGQY motif of Nrg167 nor of the PDZ protein-binding domain of Nrg180 affected GF morphology or function (Figure 7). Thus, we conclude that a wild-type Ankyrin binding motif of Nrg180 but not of Nrg167 is essential for normal GF-TTMn synapse maturation and function, but it is not required for GF axon guidance or synapse targeting. The similar phenotypes of nrg180 mutations affecting Ank2 binding either weakly (Y-F) or strongly (ΔFIGQY) indicate that normal GF synapse development requires a dynamic regulation of this interaction by phosphorylation, a feature disrupted by all mutations. Our Pacman-based mutants enabled us to determine temporal and spatial requirements of a wild-type, modifiable, Nrg180-FIGQY motif as we can express wild-type Nrg180 in the background of our mutants using the Gal4/UAS system. We used previously characterized Gal4-driver lines that allow expression of Nrg180 either simultaneously in pre- and postsynaptic neurons of the GF-TTMn synapse, only in one of the two partner neurons, or only during late stages of synaptic development in the GF (Figure 8A) [57]. Simultaneous expression of wild-type Nrg180 in pre- and postsynaptic neurons throughout GF circuit development was able to rescue all electrophysiological and morphological defects associated with the Nrg180-FIGQY mutations (using Y-F, Y-A, and Nrg180ΔFIGQY as representative examples) (Figure 8B–D and unpublished data). Thus, this assay is suitable to determine specific pre- or postsynaptic requirements of the Nrg–Ank2 interaction. To our surprise, we were able to rescue the anatomical and physiological phenotypes to a similar extent by expressing wild-type nrg180 either in the pre- or the postsynaptic neuron in the background of the Pacman-based mutants (Figure 8). We did not observe any nonresponding animals, the average response latency was significantly restored, and only subtle and rare defects in the ability to follow multiple stimuli at 100 Hz were evident (Figure 8B–D). Furthermore, pre- or postsynaptic expression was also sufficient to rescue the morphological phenotypes of the GF synapse terminal of nrg14; P[nrg180ΔFIGQY] mutant animals (Figure 8E). Finally, we tested whether there might be different temporal requirements of Nrg180 wild-type expression during GF synapse development. For this, we used a Gal4 line that drives presynaptic expression only after the initial connection between the GF and the TTMn has been established (Figure 8A) [54],[57]. Strikingly, this late expression of wild-type Nrg180 was not sufficient to rescue nrg14; P[nrg180ΔFIGQY] or nrg14; P[nrg180Y-A] mutant animals but efficiently restored electrophysiological properties in nrg14; P[nrg180Y-F] mutants (Figure 8B–D). Precise control of synaptic connectivity is essential for the formation, function, and maintenance of neuronal circuits. Here we identified the L1-type CAM Neuroglian as a key regulator of synapse stability in vivo. By combining biochemical, biophysical, and genetic assays at two complementary model synapses, we demonstrate that regulation of the Nrg180–Ankyrin2 interaction plays a critical role in controlling synapse growth, maturation, and maintenance. Several important findings arise from our work: (1) Control of synapse stability requires Nrg-mediated cell adhesion, which can be controlled by direct coupling to the presynaptic Ankyrin-associated cytoskeleton. (2) Synapse elimination and axonal retraction display striking phenotypic similarities to developmentally controlled synapse elimination at the vertebrate NMJ, suggesting common cellular mechanisms between developmental and disease processes. (3) Local regulation of the Nrg180–Ankyrin2 interaction provides a mechanism to gradually control the delicate balance between synapse growth and stability. (4) Transsynaptic Nrg signaling mediates the coordination of pre- and postsynaptic development in the CNS and requires a dynamic regulation of the Ankyrin-motif both temporally and spatially. (5) The FIGQY motif enables synapse-specific regulation of L1-type proteins to differentially control connectivity and function in distinct neuronal circuits. A large number of cell adhesion molecules have been implicated as important mediators of synapse development, but the regulatory mechanisms controlling structural synapse plasticity and maintenance remain largely unknown. In an unbiased RNAi screen, we identified the Drosophila L1-type CAM Neuroglian as essential for synapse stability at the neuromuscular junction. We demonstrate that knockdown of presynaptic Nrg induces synapse disassembly that shares all cellular hallmarks of synapse retractions observed in ank2, spec, or hts mutant animals [26]–[29]. By analyzing individual motoneurons lacking any Nrg expression, we verified this presynaptic requirement of Nrg for synapse maintenance. In addition, this allowed us to unravel the cellular events occurring in response to loss of cell adhesion at the presynaptic nerve terminal. In nrg mutant motoneurons, we observed both synaptic retractions and motoneuron axons ending in “retraction bulb”–like structures. Excitingly, we directly observed eliminated NMJs that were still connected via traces of clonally marked presynaptic membrane remnants to retraction bulb–like structures. This demonstrates that loss of synapse stability can induce a cellular program resulting in the retraction of the motoneuron axon accompanied by shedding of presynaptic membrane. This phenotype shares striking similarities with developmental synapse elimination at the vertebrate NMJ [58] and points to similar cellular programs underlying synapse loss in development and disease. It will be of particular interest to analyze the contribution of glial cells in this process as they are part of a pro-degenerative signaling system at the NMJ and actively clear membrane remnants of degenerating or pruning axons in both Drosophila and vertebrates [58]–[61]. It is important to note that some of the axonal phenotypes would also be consistent with a stalling of the axonal growth cone before reaching the appropriate target. Indeed, prior studies in both Drosophila and vertebrates demonstrated a function of Nrg and L1CAM during neurite outgrowth [14],[16],[47]–[49], indicating that both defects in axon growth and loss of synapse stability may have contributed to the observed phenotypes. Although Nrg is critical for NMJ maintenance, our observation that 50% of larval NMJs were still stable in nrg null mutants indicates that redundant mechanisms control synapse stability at the level of synaptic cell adhesion molecules. A candidate to provide such redundancy would be the Drosophila NCAM homolog FasII, which has been previously implicated in NMJ maintenance [62] and can substitute for Nrg during axonal outgrowth of ocellar neurons [48]. However, we demonstrate that FasII cannot compensate for the loss of presynaptic Nrg at the larval NMJ (Figure 1G and Figure S1D). The identification of the entire combinatorial code of CAMs contributing to synapse stability will be of high interest in the future. The dynamic nature of many neuronal circuits requires controlled changes in synapse assembly and disassembly without a disruption of neuronal circuit function. While interactions of synaptic cell adhesion molecules are essential to maintain synaptic connectivity, mechanistic insights regarding the regulation of these interactions to alter transsynaptic adhesion are limited to date. The process is probably best understood for Cadherins where adhesive properties are modulated either via binding of extracellular calcium or by altering their association with intracellular Catenins via posttranslational phosphorylation [7],[8]. These changes alter localization, clustering, and transsynaptic signaling of Cadherins leading to modulations of synaptic connectivity and function [9],[10]. Here we identify the interaction between the L1-type CAM Nrg and the adaptor protein Ank2 as a similar control module. First, we demonstrate that Nrg180 directly interacts with Ank2 in vivo. Second, a series of specific mutations in the Ankyrin binding motif allowed us to differentially modulate the Ankyrin-binding capacity of Nrg180. We demonstrate that decreasing Ank2-binding capacities correlate with an up to 2-fold increase in lateral mobility of Nrg180 in motoneurons. This is consistent with studies in vertebrates demonstrating that phosphorylation of the conserved tyrosine of the FIGQY motif reduces or abolishes binding to Ankyrins and increases mobility of L1-type CAMs [20]–[23],[52]. Finally, Pacman-based nrg mutants with altered Ankyrin-binding capacity caused two striking phenotypes. There was a significant increase in synapse retractions in mutants with severely impaired Ank2 binding but not in mutants with partial binding (Nrg180Y-F). In addition, we observed increased NMJ growth that correlated in a similar manner with the decrease in Ank2 binding capacities. The reduction in Ank2 binding potentially decreases the static population and adhesive force mediated by Nrg and thereby impairs synapse stability. At the same time this reduction in transsynaptic adhesion might allow for increased NMJ growth. We previously identified similar switch-like alterations of synapse growth and stability in animals lacking the spectrin-binding and actin-capping protein Hts/Adducin [28]. Importantly, studies of adducin2 mutant mice demonstrated that Adducin2 provides a similar function in vertebrates and is essential to mediate changes in synaptic connectivity relevant for learning and memory [63],[64]. Interestingly, we did not observe significant alterations in presynaptic Nrg180 or Ank2 levels in these animals similar to previous observations for the axonal localization of these proteins [50],[65]. However, we found a clear dependence on the respective partner protein at semi-stable nrg and ank2 mutant synapses, indicating that the Nrg–Ank2 interaction is required to maintain their synaptic localization. A similar late loss of AnkyrinG has been observed in neurofascin mutant Purkinje cells, demonstrating a function of the L1-CAM paralog for maintenance but not for initial localization of AnkG to the AIS [33],[35] and likewise AnkG is required for the maintenance of Neurofascin [32]. Together these data indicate that modulation of the Nrg–Ank2 interaction balances synapse growth and stability. Changing the interaction via posttranslational phosphorylation could thus locally decrease synapse stability, thereby allowing the formation of new synapses without impairing general neuronal circuit architecture. Despite our detailed knowledge regarding the expression of synaptic cell adhesion molecules, mechanistic insights into the transsynaptic control of synapse maturation or function are only recently emerging [1],[2],[9],[10]. Here we provide evidence that transsynaptic coordination of synapse development can be controlled via a dynamic regulation of the L1-type CAM Nrg. In contrast to the larval NMJ, the lack of significant differences in phenotypic strength between the mutations in Nrg180-FIGQY motif demonstrates that normal GF synapse development requires a dynamic regulation of the Nrg–Ank2 interaction via phosphorylation. To address the importance of this regulation for transsynaptic development, we selectively reintroduced wild-type Nrg180 either pre- or postsynaptically in the background of the different FIGQY motif mutations. Surprisingly, pre- or postsynaptic expression of wild-type Nrg in the presence of mutant Nrg on both sides of the synapse was sufficient to restore synaptic function in all mutants, but late presynaptic expression could only rescue the Nrg180Y-F mutation. This highlights two important novel aspects of Nrg function at central synapses. First, while Nrg180 is required on both sides of the synapse, regulation of the FIGQY motif is sufficient on either side of the synapse, demonstrating that Nrg can control synapse development in a transsynaptic manner. Second, constitutive binding to Ank2 is sufficient for early stages of synapse development, but GF synapse maturation requires dynamic regulation of the Nrg–Ank2 interaction. A potential function of the phosphorylation of Nrg could be an increase of lateral mobility of Nrg to allow precise spatial alignment with postsynaptic CAMs. Alternatively, phosphorylation may enable an interaction with proteins that bind only phosphorylated Nrg. One candidate would be the microtubule binding protein Doublecortin that binds only phosphorylated Neurofascin [66], but physiological relevance for this interaction in nervous system development is lacking to date. While we observe distinct functions and modes of regulation of Nrg at peripheral versus central synapses in both cases, the Nrg180–Ank2 interaction did not influence axon outgrowth or guidance. In addition, we did not observe any requirements for the Nrg167 FIGQY motif or for the PDZ-binding motif of Nrg180, which has recently been implicated in controlling axonal outgrowth in Drosophila mushroom bodies [53]. A surprising observation from studies of vertebrate L1 family proteins was that mutations within the intracellular domain that are linked to human L1/CRASH syndrome and neuropathological diseases [30] resulted in significantly weaker phenotypes in mice compared to the complete L1 knockout [14],[31],[67],[68]. While extracellular interactions are essential for early nervous system development including neurite outgrowth and axon targeting [16], here we provide evidence that reversible phosphorylation of the intracellular Ankyrin binding motif might provide a regulatory module to fine tune synaptic connectivity without impairing overall circuit stability. The expansion of the L1-type CAM family to four independent proteins in vertebrates may provide the means to cope with the diversity and complexity of synaptic connectivity in the vertebrate CNS. Indeed, while mutations in the L1CAM Ankyrin motif did not affect the general organization of the nervous system, they resulted in specific impairments of particular neuronal circuits and at subsets of synapses [31],[67],[68]. The functions of the different L1-type proteins may be distinct, partly opposing or redundant as evident by an analysis of cerebellar granule cell development in L1CAM and NrCAM double mutants [36]. Our data suggest that the coordinated phosphorylation of a subpopulation of synaptic L1 family proteins may allow differential modulation of biophysical properties of L1 complexes to precisely control distinct aspects of synapse development. Elucidating the synaptic L1-family protein code at specific synapses and identifying their phosphorylation status during synapse development and in response to activity might uncover new mechanisms controlling synaptic plasticity in development and during learning and memory. Flies were maintained at 25°C on standard food. Crosses and most experiments were performed at 25°C, while RNAi assays were performed at 27°C. The following fly strains have been used in this study: w1118 (wild-type), nrg14 (nrg1), nrg17 (nrg2), UAS–mCD8–GFP, UAS–fasII, elavC155–Gal4, ok371–Gal4, sca–Gal4, mef2–Gal4, BG57–Gal4, UAS–dcr2, ok307–Gal4 (A307–Gal4), P(hsFLP)86E, P(hsFLP)1, P(neoFRT)19A (all Bloomington stock center), c17-Gal4, c42.2–Gal4, and shakB–Gal4 [57]. RNAi lines were obtained from the Vienna Drosophila RNAi Center [37]: Nrg RNAi line1 (stock ID6688) and Nrg RNAi line2 (stock ID107991). The full-length Nrg180 ORF was amplified from the plasmid pMT–Neuroglian and the Nrg167 ORF from cDNA GH03573 (both obtained from the Drosophila Genomic Research Center, Indiana, USA). Full-length ORFs were cloned into pENTR vector via TOPO cloning (Invitrogen). To obtain pUASTattB-10xUAS destination vectors suited for gateway cloning, a gateway cassette with a C-terminal 3xHA or EGFP tag was introduced into the pWALIUM10-moe plasmid (TRiP collection, Harvard Medical School). Final expression constructs were generated via gateway cloning using standard procedures (Invitrogen). Deletions and point mutations were introduced into pENTR clones using the QuickChange II site-directed mutagenesis kit following the manufacturer's instructions (Agilent Technologies). All constructs were verified by sequencing (FMI sequencing facility). The P[acman] clone CH321-4H20 was obtained from BACPAC Resources Center (BPRC, Oakland, California) and modified using galK-mediated recombineering [44] according to [42] (NCI Frederick National Laboratory). Site-specific integration via the phi-C31 system [43] was used to generate insertions at the attP40-landing site for both pUAST and Pacman constructs. Primers used in this study are listed in Table S3. Wandering third instar larvae were dissected in standard dissecting saline and fixed with Bouin's fixative for 2–3 min (Sigma-Aldrich). Primary antibodies were incubated at 4°C overnight. Primary antibodies were used at the following dilutions: anti-Nrg180 (BP104) 1∶250 [38], anti-Bruchpilot (nc82) 1∶250, anti-Futsch (22c10) 1∶500, anti-Synapsin (3c11) 1∶100 (all obtained from Developmental Studies Hybridoma Bank, IA), rabbit anti-Dlg 1∶30,000, rabbit anti-DGluRIII [28] 1∶2,500, rabbit anti-DvGlut, rat anti-CD8 (Caltag Laboratories) 1∶1,000, anti-Nrg (3c1, gift from M. Hortsch, Ann Arbor, MI, USA [38]) 1∶500, rabbit anti-NrgFIGQY (raised against the peptide: TEDGSFIGQYVPGKLQP) 1∶100, and rabbit anti-Nrg180cyto (raised against the peptide: NNSAAAHQAAPTAGGGSGAA) 1∶500. Monoclonal rat anti-Ank2-L 1∶40 was generated against a protein fragment containing aa 3134–3728 (according to the 4,083 aa isoform of Ank2-L). Rabbit anti-Ank1-4 antibody was generated against the Ankyrin domains 1–4. This antibody recognizes both Ankyrin1 and Ankyrin2. Antibodies were generated at David's Biotechnology (Regensburg, Germany). Alexa conjugated secondary antibodies (Invitrogen) were used at 1∶1,000 for 2 h at RT. Directly conjugated anti-Hrp (Alexa or Cy-dyes) were used at 1∶100–1,000 (Jackson Immunoresearch Laboratories). Larval preparations were mounted in Prolong Gold (Invitrogen). Images were captured at room temperature using a Leica SPE confocal microscope. To process, analyze images, and quantify phenotypes Adobe Photoshop, Imaris (Bitplane), Image Access (Imagic), and the open source tool FIJI/ImageJ were used. Synaptic retractions were quantified using presynaptic Brp and postsynaptic DGluRIII staining and counting the number of unopposed postsynaptic footprints. Complete loss of the presynaptic marker Brp was considered an elimination of the presynaptic nerve terminal. Synapse retraction frequencies are presented as values per animal. NMJs on the indicated muscles in segments A2–A5 (10 NMJs/animal) were scored. n indicates the number of independent animals per quantification. Bouton area, number, and NMJ length were quantified using Synapsin, Dlg, and Hrp staining. Bouton area and NMJ length were quantified using the Image access software (Imagic). Hrp staining was used to visualize the bouton area and 10 A3 muscle 4 NMJs were quantified per genotype. To measure NMJ length 20 muscle 4 NMJs (segments A3 and A4, 10 each) were analyzed. Bouton number was quantified on muscle 4 in segments A2–A6 using Synapsin/Dlg staining. Larval brains were dissected and transferred into 2× sample buffer (Invitrogen). Five brains per lane were analyzed on NuPage gels (Invitrogen) according to standard procedures. Primary antibodies were incubated overnight at 4°C. Secondary Hrp-conjugated goat anti-mouse and goat anti-rabbit antibodies were used at 1∶10,000 (Jackson Immunoresearch) for 2 h at RT. PVDF-membranes were incubated with ECL substrate (SuperSignal West Pico Kit, Thermo Scientific) and developed on film (Fujifilm). For immunoprecipitations (IPs) 100 larval brains were collected, grinded in NP40-based lysis buffer, and incubated on ice for 30 min. The supernatant was split equally between control IPs using empty protein-G beads (Dynabeads, Life Science) and protein-G beads pre-incubated with Nrg180BP104 antibody. IPs were analyzed with anti-Nrg180BP104 (1∶200) and rat anti-Ank2-L (1∶20). For IPs of mutated Nrg180 proteins, S2 cells were co-transfected with act5C–Gal4, UAS–Ank2-S–EGFP [27], and UAS–Nrg–3xHA plasmids using Fugene (Roche) following the manufacturer's instructions. IPs were analyzed using mouse anti-HA (12CA5) 1∶200 and rabbit anti-GFP (Molecular Probes) 1∶500 antibodies. Rabbit anti-Ank2 (anti–Ank1–4) 1∶1,000 was used for visualization of the input. Quantification of Ank2 binding between Nrg180 mutants was performed using four independent IP experiments and Odyssey2.1 software (LI-COR). Wandering third instar larvae expressing nrg–GFP using ok371–Gal4 were dissected in HL3 saline and prepared for live imaging using a magnetic pinholder device. 1-Naphthylacetyl-spermin-trihydrochloride (NSH) (100 mM; Sigma, St Louis, MO) was added to the HL3 saline to block postsynaptic glutamate receptor activation and muscle contractions. Six to nine motoneuron axons from three to four independent animals were analyzed. Motoneuron axons were photo-bleached using a Zeiss LSM700 by scanning the targeted region for 30 iterations at 100% laser-power using the 488 nm line. Ten images were acquired before the bleach and 40 after the bleach with a time interval of 5 s at low laser power. Images of the FRAP series were corrected for animal movement using the FIJI registration plugin (StackReg option). Images were corrected by substracting background fluorescence from regions outside the axons and corrected for bleaching using a control area within the same axon. The recovery curves were fit to a double exponential curve as follows: The maximum was calculated from the fitting curve (maxfitting). To calculate the real max value, the following formula has been used: The mobile fraction was calculated using the following formula: The nrg null mutation nrg14 was recombined with the P(neoFRT)19A chromosome. The indicated Pacman constructs were crossed into this background to create stable stocks. These lines were crossed to P(hsFLP)1, P(neoFRT)19A, tubGal80; ok371-Gal4, UAS–CD8–GFP; MKRS, P(hsFLP)86E flies. Embryos were collected for 2 h, aged for 3 h, and heat shocked for 1 h at 37°C. All statistical analyses were performed using Microsoft Office Excel and an online source for unpaired Student's t test (http://www.physics.csbsju.edu/stats/t-test.html). p≤0.05 was accepted as statistically significant (*p≤0.05, **p≤0.01, ***p≤0.001). Adult Drosophila nervous system was dissected, dye filled, and fixed as previously described [55]. Young 2- to 5-d-old flies were used for all the experiments. To visualize the morphology of GF–TTMn connection either a 10 mM Alexa Fluor 568 Hydrazide (Molecular Probes) in 200 mM KCl or a dye solution of 10% w/v Neurobiotin (Vector labs) and tetramethyl rhodamine-labeled dextran (Invitrogen) in 2 M potassium acetate was injected into the GF axons by passing hyperpolarizing or depolarizing current, respectively. Preparation of GF samples for confocal microscopy has been described previously [55]. Samples were analyzed using a Nikon C1si Fast Spectral Confocal system. Images were processed using Nikon Elements Advance Research 4.0 software. Electrophysiological recordings from the giant fiber circuit were obtained as described in detail in [69]. The flies were given 10 single pulses at 30–60 mV for 0.03 ms with a 5-s interval between the stimuli and the shortest response latency of each fly was averaged. To determine the reliability of the circuit, the ability to follow frequencies at 100 Hz was determined. For this 10 trains of 10 stimuli were given at 100 Hz with an interval of 2 s between the trains and percent of the total responses was calculated. All traces were recorded, stored, and analyzed using pClamp 10 (Molecular Devices) software. Mann–Whitney Rank sum test was used to determine significant differences between different genotypes in average response latencies and following frequencies (Sigma Plot 11 software).

10.1371/journal.pgen.1003531

PARP-1 Regulates Metastatic Melanoma through Modulation of Vimentin-induced Malignant Transformation

PARP inhibition can induce anti-neoplastic effects when used as monotherapy or in combination with chemo- or radiotherapy in various tumor settings; however, the basis for the anti-metastasic activities resulting from PARP inhibition remains unknown. PARP inhibitors may also act as modulators of tumor angiogenesis. Proteomic analysis of endothelial cells revealed that vimentin, an intermediary filament involved in angiogenesis and a specific hallmark of EndoMT (endothelial to mesenchymal transition) transformation, was down-regulated following loss of PARP-1 function in endothelial cells. VE-cadherin, an endothelial marker of vascular normalization, was up-regulated in HUVEC treated with PARP inhibitors or following PARP-1 silencing; vimentin over-expression was sufficient to drive to an EndoMT phenotype. In melanoma cells, PARP inhibition reduced pro-metastatic markers, including vasculogenic mimicry. We also demonstrated that vimentin expression was sufficient to induce increased mesenchymal/pro-metastasic phenotypic changes in melanoma cells, including ILK/GSK3-β-dependent E-cadherin down-regulation, Snail1 activation and increased cell motility and migration. In a murine model of metastatic melanoma, PARP inhibition counteracted the ability of melanoma cells to metastasize to the lung. These results suggest that inhibition of PARP interferes with key metastasis-promoting processes, leading to suppression of invasion and colonization of distal organs by aggressive metastatic cells.

Metastasis is the spread of malignant tumor cells from their original site to other parts of the body and is responsible for the vast majority of solid cancer-related mortality. PARP inhibitors are emerging as promising anticancer therapeutics and are currently undergoing clinical trials. It is therefore important to elucidate the mechanisms underlying the anti-tumor actions of these drugs. In our current study, we elucidated novel anti-neoplastic properties of PARP inhibitors that are responsible for the anti-metastatic effect of these drugs in the context of malignant melanoma. These effects appear to be the result of PARP-1's ability to regulate the expression of key factors, such as vimentin and VE-cadherin, involved in vascular cell dynamics and to limit pro-malignant processes such as vasculogenic mimicry and EMT.

Metastatic melanoma is a fatal malignancy that is remarkably resistant to treatment; however, the mechanisms regulating the transition from the primary local tumor growth to distant metastasis remain poorly understood. Metastasis, defined as the spread of malignant tumor cells from the primary tumor mass to distant sites, involves a complex series of interconnected events. Understanding the biochemical, molecular, and cellular processes that regulate tumor metastasis is of vital importance. The metastatic cascade is thought to be initiated by a series of genetic alterations, leading to changes in cell-cell interactions that allow the dissociation of cells from the primary tumor mass. These events are followed by local invasion and migration through proteolitically modified extracellular matrix (ECM). To establish secondary metastatic deposits, the malignant cells evade host immune surveillance, arrest in the microvasculature, and extravasate from the circulation. Finally, tumor cells can invade the local ECM, proliferate, recruit new blood vessels by induction of angiogenesis, and then expand to form secondary metastatic foci [1]. Several key steps in metastatic progression involve tumor-associated endothelial cells (EC) [2]. Both angioinvasion and angiogenesis require disruption of endothelial integrity for tumor cell transmigration across the endothelium, EC migration and EC access for mitogenic stimulation. An essential step in angioinvasion and angiogenesis is the disruption of the adherent junctions between EC. Vascular endothelial cadherin (VE-cadherin; also known as cadherin 5) is the most important adhesive component of endothelial adherent junctions [3]; while ectopic expression of VE-cadherin in malignant melanoma cells confers this tumor the capability to form vessel-like structures that contributes to the lack of efficient therapeutic strategies and increases the risk of metastatic disease [4]. Epithelial-mesenchymal transition (EMT) is a trans-differentiation characterized by decreased epithelial markers such as E-cadherin[5]. EMT is a dynamic process resulting in the acquisition of cell motility with decreased adhesive ability for body organization that includes embryonic development and wound healing. Currently, EMT is thought to be a key step in the process of cancer metastasis [6]. Molecular markers of EMT include E-cadherin down-regulation, responsible for the loss of cell-cell adhesion, up-regulation of matrix-degrading proteases and mesenchymal-related proteins such as vimentin and N-cadherin, actin cytoskeleton reorganization, and up-regulation and/or nuclear translocation of transcription factors underlying the specific gene program of EMT, such as β-catenin and members of the Snail1 family [6]. The nuclear protein PARP-1, known to function as a DNA damage sensor and to play a role in various DNA repair pathways, has recently been implicated in a broad variety of cellular functions, including transcriptional regulation [7]. PARP inhibitors exhibit antitumor activity in part due to their ability to induce synthetic cell lethality in cells deficient for homologous recombination repair [8], [9], [10], [11]. PARP inhibitors also possess anti-angiogenic properties [12], [13], [14], [15], and recently, our group reported that PARP inhibition results in the down-regulation of Snail1 by accelerating the degradation of this protein [16]. In the present study we aimed to address the potential of PARP inhibition as modulators of metastasis [16]. The results presented here indicates that PARP inhibition, through down-regulation of the intermediary filament vimentin in both endothelial and melanoma cells, led to a reversion of mesenchymal phenotype in both cell types and prevented malignant melanoma cells from developing vasculogenic mimicry. As monotherapy, PARP inhibition displayed an anti-metastatic effect in a model of murine melanoma. Moreover, we identified vimentin as an upstream modulator of EMT: forced expression of vimentin was sufficient to induce tumor cell transformation through the ILK/GSK-3β signaling axis. The ability of PARP inhibition to modulate vimentin levels (and hence EMT), the interference with vasculogenic mimicry, and the modulation of endothelial plasticity allowed PARP inhibitors to exert a multifaceted antimetastatic effect to counteract the progression of malignant melanoma. A number of reports from various laboratories, including ours, have identified a novel and unexpected effect of PARP inhibitors on angiogenesis, raising the possibility that PARP inhibitors may be useful as anti-angiogenic agents [13], [17]. In our present study, we disrupted PARP activation in HUVECs in an attempt to elucidate the mechanism by which PARP-1 influences endothelial cell dynamics. We have previously shown that PARP inhibitors reduced angiogenesis both in vitro and in vivo ([13] and Figure S1). To further characterize this effect of PARP inhibition on endothelial cell plasticity, we performed a proteomic analysis using primary HUVEC in the presence or absence of the PARP inhibitor DPQ (Figure 1A, Figure 2 and Figure S2). The expression levels of a number of proteins were altered following PARP inhibition, as detected by 2D DIGE electrophoresis (Figure S2) and mass spectrometry analysis (Figure 1A, Figure 2). A statistically significant down-regulation of vimentin (a class III intermediary filament), tropomyosin alpha-4 chain (involved in stabilizing actin filaments), endoplasmin (a molecular chaperone involved in processing and transport of secreted proteins), mitochondrial ATP synthase ATPB5, protein disulfide isomerase PDIA6, heat-shock 70 kD protein-5 (glucose-regulated protein, 78 kD), heat shock protein 90 kDa alpha (cytosolic), class B member 1, and HSP90AB1 occurred following PARP inhibition. An increase in the expression of the mitochondrial heat shock protein HSPD1 was also observed after PARP inhibition. Due to its important role in the biology of endothelial cells, we focused our study on vimentin, the main structural protein of intermediary filaments. It has been reported that vimentin can be targeted for tumor inhibition due to its specific up-regulation in tumor vasculatures [18], [19]. To confirm the results of our proteomic analysis, we performed western blot in HUVEC either treated with DPQ (right) or left untreated. In Figures 1B and 1C, western blot and indirect immunofluorescence analysis indicated that vimentin expression was down-regulated in HUVEC cells treated with DPQ. Figure 1B and 1D show that PARP inhibition affected not only vimentin levels but also Snail1 and VE-cadherin protein and mRNA levels. Endothelial to mesenchymal transition (EndoMT) is a process by which endothelial cells disaggregate, change shape, and migrate into the surrounding tissue. The process of endoMT is characterized by the loss of endothelial cell markers, such as vascular endothelial VE-cadherin, and the expression of mesenchymal cell markers, such as vimentin and Snail1 [20]. Endothelial cell migration was strongly inhibited by PARP inhibition (Figure 1E). These results suggest that PARP inhibition prevented the acquisition of a mesenchymal phenotype by endothelial cells. Vimentin is a well-known marker of EMT, which is a hallmark of primary tumor progression to a metastatic phenotype. We tested the impact of vimentin down-regulation (induced by PARP inhibition or vimentin silencing) on EMT differentiation in various melanoma cell lines and in endothelial cells. One major event induced by PARP inhibition, in the process of EMT is the up-regulation of E-cadherin expression through the inactivation of the transcription factor Snail1. Snail1 and vimentin levels were both down-regulated following PARP inhibition, indicating a disruption EMT in the absence of PARP activation (Figure 3A in G361 cells and Figure S3B in B16-F10 cells). Down-regulation of PARP activity was confirmed in G361 following H2O2 treatment as a positive control of PARP-1 activation and poly(ADP-ribose) (PAR) synthesis (Figure S4). Vimentin and Snail1 mRNA levels were decreased after PARP inhibition (Figure 3C and Figure S3C). In Figures 2B and Figure S3A, indirect immunofluorescence showed that vimentin expression was down-regulated in melanoma cells treated with DPQ or KU0058948 (G361 cells, Figure 3B) or PJ-34 (B16-F10 cells, Figure S3A). Using two different luciferase reporter plasmids under the control of a Snail1 responsive sequence and the E-cadherin promoter, we found that PARP inhibition affected negatively the activation of Snail1 and activated the expression of the E-cadherin promoter (Figure 3D and Figure S3D). Wound healing experiments also revealed decreased wound closing following treatment with a PARP inhibitor, PJ-34 (Figure 3E). We have also evaluated the effect of both PARP-1 and vimentin silencing on the expression of Axl, a key determinant of cell migration and EMT promotion [21]. Following PARP-1 silencing in HUVEC or G361 cells, the EMT marker Snail1 decreased while E or VE-cadherin were upregulated (Figure 4A and 4B respectively). Interestingly, Axl expression was also down-regulated in parallel with decreased levels of vimentin. Vimentin knockdown also caused a global alteration in the expression of EMT markers. Under these conditions, Axl levels were decreased (Figure 4A and 4B), suggesting that vimentin down-regulation was sufficient to drive tumor cells toward an epithelial state. We next sought to determine if alterations in vimentin levels were sufficient to alter or reverse EMT progression. Vimentin is known to positively influence tumor cell migration. To test the impact of vimentin expression on cell migration and invasion we performed either silencing or over-expression in endothelial and melanoma cells. Following vimentin knockdown, wound healing closure in HUVEC cells was significantly diminished (Figure 4C) while its over-expression increased wound healing efficiency (Figure 4D). The same approach was used in B16F10 melanoma cells where over-expression of vimentin increased significantly cell migration (Figure 4E). Nonetheless, inhibition of PARP had a less impact on cell migration after vimentin over-expression, suggesting that the levels of vimentin were implicated in the effect of PARP inhibition on cell motility (Figure 4E), although a multifactorial mechanism for downstream effect of PARP inhibition could not be excluded. To further confirm the role of vimentin in PARP-inhibitor-induced impaired cell migration we decided to analyze the effect of vimentin over-expression and PARP inhibition in a well-established model of epithelial cells, MDCK, that undergo EMT after hepatocyte growth factor (HGF) treatment, including fast movement and circularity (scattering) [22]. The trajectories of cell migration were determined under video-microscopy and analyzed using MetaMorph image analysis software. Global trajectories after expression of GFP-vimentin in the presence or absence of PARP inhibitor and HGF were determined. Treatment with the PARP inhibitor PJ-34 or olaparib resulted in decreased cell motility in cells transfected with empty GFP vector (Figure 4F). Vimentin expression increased cell motility (Figure 4F, right), and PARP inhibition was unable to prevent this increase, suggesting that vimentin down-regulation is needed for the effect of PARP inhibition in reversing the EMT phenotype. To characterize more in-depth the implications of vimentin expression in the context of EMT, we expressed GFP-vimentin in both a human melanoma cell line (Figure 5A) and a human breast tumor cell line with an epithelial phenotype (MCF7) (Figure 5B and 5C); MCF7 cells were chosen due to the lack endogenous vimentin expression compared with melanoma G361 cells (Figure 5A, G361 cells and Figure 5B and 5C, MCF7 cells). GFP-vimentin over-expression alone induced a mesenchymal phenotype characterized by Snail1 up-regulation, loss of E-cadherin, increased pGSK-3β (inactive form) and β-catenin expression (Figure 5A, 5B and 5C). The most remarkable effect of PARP or vimentin silencing observed in our model was the down-regulation of ILK and GSK-3β (Figure 4A and 4B). In order to get mechanistic information on the interaction between vimentin over-expression and the activation of EMT signaling pathway, we focused in the axis ILK/GSK-3β, which plays a central role in EMT commitment, upstream of Snail1. Inhibition of GSK-3β was achieved by LiCl treatment while its activation was driven by silencing the kinase, ILK, which is the upstream inhibitory kinase for GSK-3β (Figure 5, central panel). Specifically, inhibition of GSk-3β (which was confirmed by an increase in the level of inhibitory phosphorylation of GSK-3β at Ser9) with LiCl, activated EMT and resulted in E-cadherin down-regulation, Snail1 accumulation and increased levels of β-catenin (Figure 5B); concomitantly, E-cadherin was down-regulated following GSk-3β inhibition by LiCl (Figure 5B) or exogenous expression of vimentin (Figure 5). GSk-3β activation is achieved through the silencing of its upstream inhibitor integrin-linked kinase (ILK). ILK knockdown resulted in Snail1 down-regulation and increased E-cadherin expression (Figure 5C). Interestingly, exogenous vimentin expression completely prevented siILK-induced E-cadherin up-regulation and partially prevented the reduction of Snail1 expression. These results suggested that vimentin, when over-expressed, is sufficient to drive the phenotypic changes associated with a mesenchymal cell status, depending on the activation of GSk-3β, whose inhibition accentuated vimentin-induced changes, while its activation (following ILK-silencing), abolished vimentin-induced E-cadherin decrease and Snail1 accumulation (Figure 5C). The formation of patterned networks of matrix-rich tubular structures in three-dimensional culture is a defining characteristic of highly aggressive melanoma cells. It has been demonstrated that aggressive melanoma cells in which VE-cadherin was repressed, could not form vasculogenic-like networks [23], suggesting that tumor-associated misexpression of VE-cadherin (observed in melanoma cells) is instrumental in allowing endothelial cells to form vasculogenic networks. We measured VE-cadherin protein levels in B16-F10 cells after treatment with the PARP inhibitor PJ-34 or KU0058948. VE-cadherin expression was strongly down-regulated following PARP inhibition. We tried to confirm this result by indirect immunofluoresce of VE-cadherin, however the protein was barely detected, as was the case for the protein in western blot (Figure 6A). Phosphorylation of VE-cadherin has been shown to correlate with loss of function of VE-cadherin and increased vascular permeability [24], as is the case for pseudo vessels during VM. PARP inhibition was able to impact negatively on the levels of both total and phosphorylated VE-cadherin, which, indeed, had a membrane and cytoplasmic distribution (Figure 6A). The consequences for the down-regulation of both total and phosphorylated VE-cadherin by PARP inhibitors during VM are now being investigated in our laboratory. VM was measured in vitro using B16F10 cells cultured in matrigel coated plates (Figure 6B). All markers of VM structure formation (covered area, tube length, branching points and loops) were significantly decreased after inhibition of PARP with PJ-34 (Figure 6C). We next aimed to examine the effect of PARP inhibition on melanoma tumor growth of cells subcutaneously implanted in C57BL/6 mice. Mice were treated every two days with 15 mg/kg (i.p.) of the PARP inhibitor DPQ or vehicle. A significant difference in tumor growth was found after 14 days of tumor implantation in the DPQ-treated group compared to the control (Figure 7 and Figure S5A). To evaluate the direct effects of the PARP inhibitor DPQ on tumor metastasis, we used a well-characterized model of experimental lung metastasis [25]. Experimental metastasis model provide several advantages for investigation. The time course for model maturity is generally rapid, the biology of metastasis is reproducible and consistent, and we control de number and type of cells that are introduced to the circulation [26]. B16-F10 cells were tail vein injected into mice, and the mice were then treated with 15 mg/kg of the PARP inhibitor DPQ or vehicle three times per week over a three-week period. Tail vein injection results primarily in pulmonary metastases. Photon emission was acquired every two days. Seven days after B16-F10 cell injection, a photon signal was already detected in the lungs (Figure 7B), and DPQ treatment significantly suppressed lung metastasis compared to the control throughout the duration of the experiment (21 days). Similar results were obtained using the clinically relevant PARP inhibitor olaparib (Figure S6). Metastatic foci were also detected in other organs upon mice autopsy. These organs included the liver, kidney, spleen, gut, stomach and heart (Figure S5B). In all cases, the incidence of metastatic foci was reduced compared to lung metastasis. DPQ-treated mice exhibited a decreased incidence of extra-pulmonary metastasis compared to the control. Pathologic analysis of the lungs showed a decrease in size and number of metastatic foci (more than 80%) after DPQ treatment (Figure 7C) that was accompanied by a reduced number of tumor vessels in both primary subcutaneous tumors and lung metastasis (Figure 7D), suggesting that the anti-angiogenic effect of PARP inhibition may be involved in the observed reduction in metastatic progression. Apoptotic and mitotic rate were not significantly different in tumors derived from DPQ-treated or untreated mice (Figure S7). To investigate in vivo the effect of PARP inhibition on the expression of Snail1 and E-cadherin, we performed immunohistochemistry for these EMT markers in metastatic lung tumors (Figure 7E). We observed that Snail1 was highly expressed in the vessels of tumors derived from the untreated group. This expression exhibited both nuclear and cytoplasmic distribution as previously reported [27]. Metastatic lung tumors derived from DPQ-treated mice displayed reduced expression of Snail1 as well as an increase in E-cadherin expression, similar to the results obtained in cultured melanoma cells. These data indicate that the in vivo expression of EMT markers within tumors is also reduced following treatment with PARP inhibitor. We also performed a Kaplan Meyer curve to compare the mortality of both groups of mice, and we observed a statistically significant difference in the survival rate from <4 weeks in the untreated group to >8 weeks in the DPQ-treated mice (Figure 7F). Survival of mice injected with B16-F10 cells stably expressing shRNA targeting PARP-1 (Figure 7G), was also significantly increase. To determine the correlation between PARP-1 expression and disease progression in human melanoma, we used IHC to analyze the levels of vimentin, PARP-1, Snail1, E-cadherin and MITF in nodular and metastatic melanoma frozen biopsies. Vimentin was expressed in all biopsies derived from both nodular and metastatic melanoma; however, the level of expression was elevated in nodular melanoma, which is the initial stage of the disease. PARP-1 expression was positively correlated with vimentin expression, suggesting an association between the in vivo expression of both proteins (Figure S8, Table S1). Expression of the Snail1 and microphthalmia-associated transcription factor (MITF), which is a melanocyte marker, is also increased in metastatic melanoma. Interestingly, nodular melanoma did not express Snail1 while 40% of metastatic melanoma samples displayed Snail1 expression. Loss or reduction of E-cadherin and increased expression of EMT markers is frequently associated with the development of an invasive phenotype in cancer. Expression of E-cadherin in normal melanocytes is significantly reduced during the initial steps of melanoma progression [28]; however, elevated levels of E-cadherin are found at advanced stages of the disease [29]. E-cadherin expression was similar in both nodular and metastatic melanoma (Table S1), which is in agreement with previous publications. These findings suggest that in human melanoma, there is a complex interconnection between the expression levels of various disease markers and the expression of PARP-1, although we have detected a strong correlation between vimentin and PARP-1 expression (Figure S8). PARP inhibitors are a novel and important class of anticancer drugs, and there are now more than 40 clinical trials that are ongoing or in development to study the effectiveness of PARP inhibitors in the treatment of various cancers. Given the enormous interest in this target, it is important to understand the underlying mechanisms by which PARP-1 and other PARPs function in tumor cell biology. Until recently, the development of PARP-1 inhibitors has focused almost exclusively on the function of this enzyme in DNA repair. Emerging literature, however, indicates other activities of PARP-1 that may explain the in vivo potency of some PARP-1 inhibitors that cannot be entirely attributed to their apparent in vitro activity and that could provide additional targets for anti-cancer therapies. In addition to its direct role in DNA-damage recognition and repair, PARP-1 can regulate the function of several transcription factors, including p53 and NF-κB. In the context of certain cancers, PARP-1 interacts with the transcription factors HIF1 [13] and Snail1 [16]. The mechanisms underlying the effects of PARP inhibition on vascular plasticity and metastasis remain relatively unknown. Our current study identifies PARP-1 as a pivotal modulator of the molecular and functional changes characteristic of EndoMT (involved in the loss of function of tumor-associated vessels) and of the phenotypic switch that facilitates the acquisition of pro-metastatic capacities by tumor cells. Proteomic analysis of endothelial cells that have been treated with a PARP inhibitor identifies the intermediary filament protein vimentin as a target of PARP inhibition. Intermediary filaments such as vimentin and keratins are known to play non-mechanical roles in protein trafficking and signaling (reviewed in [30]), which in turn influence cellular processes such as cell adhesion and polarization. Vimentin is abundantly expressed by mesenchymal cells and plays a critical role in wound healing, angiogenesis and cancer growth. Vimentin has also been described as a tumor-specific angiogenesis marker, and targeting endothelial vimentin in a mouse tumor model significantly inhibited tumor growth and reduced microvessel density [31]. Vimentin is both an EMT and EndoMT marker and is also over-expressed in tumor samples compared to normal tissues. This protein also contributes to tumor phenotype and invasiveness [18], [19]. Our findings indicate that PARP inhibitors reduce the metastatic potential of melanoma cells, at least in part, through their ability to down-regulate vimentin expression. Vimentin expression has been shown to be transactivated by β-catenin/TCF and thus increasing the tumor cell invasive potential [19]. It has been shown that NF-κB, a key protein regulating the immune and inflammatory process, also plays an important role in regulating EMT process and its inhibition in the mesenchymal cells reversed the EMT process, suggesting the importance of NF-κB in both activation and maintenance of EMT [32]. Since vimentin is over expressed during EMT process, and NF-κB being one of the transcription factors binding to vimentin promoter, it would be tempting to speculate that this over-expression of vimentin is a result of activated NF-κB in cancer cells. Also, TGFβ1 response element was found within the activated protein complex-1 region of the vimentin promoter and was involved in regulation of vimentin expression in myoblasts and myotubes [33]. Interestingly, ADP-ribosylation of Smad proteins by PARP-1 has been shown to be a key step in controlling the strength and duration of Smad-mediated transcription [34]. Regulation of vimentin levels by PARP inhibition may also involve other transcription factor such as Snail1 and HIF-1/2. Our results also reveal that vimentin levels are not merely a hallmark of EMT. While silencing of vimentin in melanoma cells can reverse the EMT phenotype, in part by promoting down-regulation of the protein kinase Axl that is involved in cell motility, forced expression of vimentin in tumor cells lacking this protein is sufficient to trigger the switch from epithelial to mesenchymal phenotype. GSK-3β is an upstream regulator of key factors involved in EMT such as Snail1 and β–catenin. We hypothesized that vimentin may be involved in the modulation of this upstream regulator of EMT. Indeed, vimentin expression potentiated LiCl-(a GSK-3β inhibitor) induced EMT (Figure 5B) and counteracted the inhibitory action of ILK-silencing (leading to GSK-3β activation) in the context of EMT (Figure 5C). Mechanical signals can inactivate GSK-3β resulting in stabilization of β-catenin. Intermediate filaments are important in allowing individual cells, tissues and organs to cope with various types of stress, and they play a significant role in the mechanical behavior of cells [35]. It is possible that the signaling pathway that integrates PARP activation with altered vimentin expression and fluctuations in GSk-3β activity could be related to the capability of PARP inhibitors to inactivate AKT signaling [36], which would result in GSk-3β activation and the modulation of its downstream signaling, ultimately resulting in the reversal of EMT. Vasculogenic mimicry, as a de novo tumor microcirculation pattern, differs from classically described endothelium-dependent angiogenesis. This is a unique process characteristic of highly aggressive melanoma cells found to express genes previously thought to be exclusively associated with endothelial cells and is characteristic of aggressive melanoma tumor cells. HIF-1α and HIF-2α, transcription factors that are stabilized during conditions of oxygen depletion (hypoxia), are the master regulators of VE-cadherin. HIF-mediated transcriptional regulation during hypoxia is critical as this process induces genes that are essential for tumor cell adaptation to the stress of oxygen depletion. As a result, the expression of HIF target genes is associated with increased malignancy. Although the expression of VE-cadherin is not hypoxia-regulated, HIF-2α, but not HIF-1α, activates the VE-cadherin promoter by binding to the HRE during normoxic conditions [37]. HIF-2α expression is associated with developing endothelium, proper vascular development and increased tumor malignancy [38], [39], raising the possibility that it may be an important protein that functions in the induction of tumor cell plasticity. Using a mouse model of melanoma lung metastasis, we also present in vivo evidence indicating that targeting PARP strongly reduces metastatic dissemination of melanoma cells, at least in part through inducing a reduction in tumor microvessel density along with changes in the expression pattern of EMT markers (Snail1, vimentin and E-cadherin) within the tumor. Snail1 is a master regulator of EMT, and the activation of this protein can mediate tumor invasiveness through the transcriptional repression of E-cadherin expression. Regulating the activity of E-cadherin repressors represents a potentially beneficial strategy to fight cancer progression, and PARP-1 inhibitors accomplish this function by interfering with Snail1 activation. Results from human tissue arrays of melanoma suggest a complex interaction between PARP-1 expression and melanoma progression. It is difficult to verify EMT experimentally in vivo due to the reversible and dynamic nature of the process. Although melanoma cells are not epithelial in nature, the EMT for this tumor is well characterized and the relevance of the cadherin switch has been previously described using several experimental approaches, demonstrating that melanoma cell lines transfected with N-cadherin are morphologically transformed from an epithelial-like shape to a fibroblast-like shape [37]. Adenoviral re-expression of E-cadherin in melanoma cells down-regulates endogenous N-cadherin and reduces the malignant potential of these cells [37]. Globally, our study shows that PARP inhibition is perturbing metastatic transformation at least at three levels (Figure 8): i) decreasing abnormal tumor angiogenesis through its ability to counteract Endo-MT; ii) preventing from acquisition of EMT and iii) limiting vasculogenic mimicry in melanoma cells. Over the past few years, PARP has emerged as a strong and effective target for first line anticancer therapy. Due to its ability to regulate a number of cellular functions (from DNA repair to cell death and transcription), inhibition of PARP may affect multiple facets of tumor metabolism. These findings strongly indicate that several novel activities of PARP-1 may contribute to the effects of anti-cancer therapy targeting this protein by interfering with tumor physiology and the tumor microenvironment. Given these findings, it is of vital importance that we elucidate mechanisms regulating novel functions of PARP-1 and poly (ADP-ribose) in tumor biology so that PARP inhibitors can ultimately make the transition to routine clinical use. Human umbilical vein endothelial cells (HUVEC) were cultured in EGM-2 Endothelial Cell Growth Medium-2 (LONZA). Cells were subjected to experimental procedures within passages 3–6. B16-F10-luc-G5 cells stably expressing plasmids pGL3 control (SV40-luc) (Promega) and pSV40/Zeo (Invitrogen). Human (G361), murine (B16-F10) malignant melanoma cells and breast cancer (MCF7) cells were cultured in DMEM containing 10% fetal bovine serum, 0.5% gentamicin (Sigma, St. Louis, MO), and 4.5% glucose. All cells were cultured at 37°C (5% CO2). The tumor cell lines have been developed as described in detail previously [40]. Melnikova et al. [41] found that unlike human melanomas, the murine melanomas cell lines did not have activating mutations in the Braf oncogene at exon 11 or 15. All of the cell lines also expressed PTEN protein, indicating that loss of PTEN is not involved in the development of murine melanomas. This B16-F10 cell has previously been shown to be sensitive to stable depletion of PARP-1 in vivo melanoma growth [17]. Previous publication from our lab in G361 cells show similar results [16]. Cells were treated with the PARP inhibitors 3,4-dihydro-5-[4-(1-piperidinyl)butoxyl]-1(2H)-isoquinolinone (DPQ), [N-(6-Oxo-5,6-dihydro-phenanthridin-2-yl)-N,N-dimethylacetamide] (PJ-34) (Alexis Biochemicals, San Diego, CA) (as described [42], KU0058948 (as we shown in previous publications [16] or Olaparib (KU0059436, Selleckchem) for 22 hours. For capillary-like formation assays, 25 µL of Matrigel (BD Biosciences) were spread onto eight-chamber BD Falcon glass culture slides (BD Biosciences) or onto 96-well plates. Cells were seeded at 2.5×104 cells per well (high density) in eight-chamber slides and at 5×103 cells per well (low density) in 96-well plates and maintained in RPMI supplemented with 1% FBS [13]. These assays were performed according to previously published methods [13]. Primary antibodies used in these studies consisted of vimentin and VE-cadherin (mouse monoclonal), E-cadherin (rabbit polyclonal) (Santa Cruz Biotechnology), Snail1 and pVE-cadherin (rabbit polyclonal) (Abcam), ILK (rabbit monoclonal) (Millipore), Axl (rabbit polyclonal), total-GSK-3β (mouse monoclonal) and pGSK-3β (rabbit monoclonal) (Cell Signaling), β-catenin (mouse monoclonal) (BD Transduction Laboratories), PARP-1 (monoclonal) (Alexis) as well as β-actin (Sigma Aldrich). Quantitation of western blots was performed using Quantity One software analysis and all densitometries were normalized for loading control (Table S2). Luciferase activity was determined after transfecting the constructions into the B16-F10 cells. Firefly Luciferase was standardized to the value of Renilla Luciferase. Cells were co-transfected with 0.5 µg renilla as control of transfection together with 0.5 µg of the Snail or E-cadherin plasmid using jetPEI cationic polymer transfection reagent according to the manufacturer's instructions. The expression of Firefly and Renilla luciferases was analysed 48 h after transfection, according of the manufacturer's instructions. Cloning of the human Snail1 promoter (−869/+59) in pGL3 basic (Promega), was described previously [43]. E-cadherin promoter were cloned into pGL3-basic (Promega) to generate pGL3-E-cadherin (−178/+92). HUVEC or G361 cells were transiently transfected with an irrelevant siRNA [44], PARP-1 siRNA or vimentin siRNA (Thermo Scientific) for 24 h using JetPrime (Polyplus transfection) according to the manufacturer's recommendations. At 48 h post-transfection, the expression of PARP-1, vimentin, Axl, E-cadherin, Snail1, ILK, β-catenin, pGSK-3β and total-GSK-3β was measured. Cells were washed twice in phosphate-buffered saline (PBS) and scraped in Laemmli buffer (1 M Tris, 20% SDS and 10% glycerol) and sonicated. The protein concentration was determined using the Lowry assay. Levels of β-actin were monitored as a loading control. We used the GFP-vimentin expression vector supplied by Dr. Goldman (Department of Cell and Molecular Biology, Chicago, Illinois). For transfection, JetPrime was used according to the manufacturer's protocol. 24 h post-tranfection, 5 µM of LiCl (Sigma Aldrich) was added in MCF7 cells and 48 hours later, the expression of vimentin, ILK, pGSK-3β, total-GSk-3β, E-cadherin, Snail1 and β-catenin was measured. In other experiment, co-transfection of GFP-vimentin and ILK siRNA (Sigma Aldrich) was used the according of the manufacturer's protocol. GFP and an irrelevant siRNA [44] were used as a control. HUVEC and B16-F10 cells were cultured on coverslips in six-well cell culture dishes. Monolayer cultures were stained with CellTracker Green CMFDA in HUVEC cell (5-chloromethylfluorescein diacetate) (Invitrogen) according to manufacturer recommendations or with 4′,6′-diamidino-2-phenylindole dihydrochloride (DAPI) (post-fixation). A wound was induced in the confluent monolayer cultures, and the cultures were then treated with the indicated inhibitor. The cells were fixed with 3.7% buffered formaldehyde and then prepared for immunofluorescence. Images were captured using a confocal microscope (LEICA TCS SP5 Argon Laser 488 nm, HeNe Laser 543 nm) when the cells were stained with CellTracker Green CMFDA Abs [522 nm] and Em [529 nm] and Zeiss Axio Imager A1 microscopy for cells stained with DAPI. The method used to Wound Healing using a service provided by Wimasis with permits users to upload their images online at any time and form anywhere and allows their images to be analyzed and the results uploaded back to the researcher's serve. Madin-Darby canine kidney (MDCK) cells (1,5×104) were seeded in 12-well tissue culture dish. After 24 h, cells were transfected with GFP or GFP-vimentin and 1 day after, cells were incubated with HGF (hepatocyte growth factor, Sigma Aldrich) or PBS. HGF is a mitogenic growth factor that is well known to induce the dissociation of islands of cells into individual cells, termed “cell scattering” or EMT. When inhibitors were used, cells were preincubated with PARP-1 inhibitor, PJ-34 or Olaparib for 2 h before addition of HGF. After 48 h, representative photographs were taken at 10× magnification using a Leica Spectral confocal laser microscope. The results were analyzed using the MetaMorph image analysis software. The effect of PARP inhibitors on the formation of tube-like structures in Matrigel (BD Biosciences) was determined according to manufacturer instructions. Briefly, 24-well plates were coated with 100 µl of BD Matrigel™ Basement Membrane Matrix and allowed to solidify at 37°C in 5% CO2. Cells were treated with DPQ (40 µM) or PJ-34 (10 µM). After 22 h of incubation at 37°C in 5% CO2, the cells were fixed with 3.7% formaldehyde, and images were acquired using an Olympus CKX41 microscope. The formation of tube-like structures was then quantified. Each treatment was performed in triplicate, and the experiment was independently repeated at least three times. C57BL/6 mice background (8 weeks old) were subcutaneously (s.c.) flank-injected with 600 µl of matrigel (BD Biosciences) supplemented with VEGF (100 ng/ml) (Peprotech) and heparin (Sigma, 19 U). The negative controls contained heparin alone. Each group consisted of four animals. After seven days, mice were sacrificed and matrigel plugs were extracted. The angiogenic response was evaluated by macroscopic analysis of the plug at autopsy and by measurement of the hemoglobin (Hb) content within the pellet of matrigel. Hb was mechanically extracted from pellets reconstituted in water and measured using the Drabkin (Sigma-Aldrich) method by spectrophotometric analysis at 540 nm. The values were expressed as optical density (OD)/100 mg of matrigel. This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the Bioethical Committee of CSIC. The protocol was approved by the Committee on the Ethics of Animal Experiments of the CSIC. All surgery was performed under isoflurano anesthesia, and every effort was made to minimize suffering. Eight-week-old male C57BL/6 albino mice (The Jackson Laboratories, Bar Harbor, MN, USA) were injected subcutaneously with B16-F10-luc-G5 cells (1×105) and intravenously with B16-F10-luc-G5 cells (1×105 or 5×105). Three times per week mice were injected intraperitoneally with DPQ dissolved in phosphate-buffered saline/10% DMSO at a dose of 15 mg/kg body weight or olaparib at 50 mg/kg. Mice were injected intraperitoneally with D-luciferin solution dissolved in phosphate-buffered saline at a dose of 150 mg/kg body weight. After 5 to 8 minutes, the animals were anesthetized in the dark chamber using 3% isoflurane in air at 1.5 L/min and O2 at 0.2 L/min/mouse, and animals were imaged in a chamber connected to a camera (IVIS, Xenogen, Alameda, CA). Exposure time was 3 min in large binning, and the quantification of light emission was performed in photons/second using Living Image software (Xenogen). Tumor growth was monitored at 0, 2, 7, 14 and 21 days by in vivo imaging and bioluminiscence measurement. After 21 days, mice were sacrificed, and their organs were removed and stored in buffered formalin (3.7%) until histological staining. Immunostaining for vimentin, VE-cadherin, pVE-cadherin, Snail1 and E-Cadherin was performed on cells plated onto coverslips and grown for 22 h prior to experimental treatments. The culture medium was removed, and the cells were fixed (Paraformaldehyde 3%, Sucrose 2% in PBS) for 10 minutes at room temperature. Permeabilization was performed using 0.2% Triton X-100 in PBS. The coverslips were rinsed three times in PBS prior to incubation with primary antibody for 1 h at RT and then rinsed three times in PBS before incubation with the secondary antibody. Secondary antibodies were FITC-conjugated anti-mouse IgG or anti-rabbit (Sigma, St. Louis, MO). Antibodies were diluted in PBS containing 2% bovine serum albumin. Nuclear counterstaining with 4′,6′-diamidino-2-phenylindole dihydrochloride (DAPI) was performed after removal of excess secondary antibody. Slides were prepared using Vectashield mounting medium (Vector Lab., Inc., Burlingame, CA 94010), cover slipped and stored in the dark at 4°C. Immunofluorescence images were obtained in the linear range of detection to avoid signal saturation using a fluorescent microscope (Zeiss Axio Imager A1) or confocal microscopy (Leica SP5). For conventional morphology, three buffered 4% formaldehyde-fixed, paraffin-embedded skin longitudinal tissue sections were stained with periodic acid schiff (PAS) at the end of treatment. The study was done in blinded fashion on 4-µm sections with light microscopy. The mitosis and apoptosis cells were assessed by examining their number in ten high power field (hpf) at 600× magnifications. The results were expressed as number of cells per mm2. For evaluation of blood vessels density, tissue sections of different groups were dewaxed, hydrated, and heat-treated in 0.01 M citrate buffer for antigenic unmasking. The rest of the procedure was carried out using an automatic immunostainer (Autostainer480, Labvision, Fremont CA, USA). The incubation time with lectin Ulex europaeus biotin conjugated was 60 min, the dilution was 1∶200, and the streptavidin-biotin-peroxidase method (Master Diagnóstica, Granada, Spain) with diaminobezidine was used as visualization system. A millimeter scale in the eyepiece of a microscope BH2 (Olympus) with 40× objective was used to count the vessel per mm2 of tissue section. The morphological and immunohistochemistry study was done in a double-blinded fashion by two pathologists. For data shown in Figure 7 and FigureS7 we have fitted the values of the average number of tumors per mouse during carcinogenesis treatment using the Mann-Whitney u-test. Statistical analysis of other experiments used unpaired Student's t-test.

10.1371/journal.pntd.0007494

Functional illiteracy burden in soil-transmitted helminth (STH) endemic regions of the Philippines: An ecological study and geographical prediction for 2017

Soil-transmitted helminth (STH) infections remain highly endemic across the Philippines, and are believed to be important contributors to delayed cognitive development of school-aged children. Identification of communities where children are at risk of functional illiteracy is important for the attainment of Sustainable Development Goals target for literacy. We aimed to quantify the associations between the spatial variation of STH infections and functional literacy indicators adjusting for other important contributors, and identify priority areas in the Philippines in need of interventions. We used data from 11,313 school-aged children on functional literacy indicators collected in 2008. Nested fixed-effects multinomial regression models were built to determine associations between STH endemicity and geographical distribution of functional literacy, adjusting for demographics, household level variables, and the prevalence of malaria. Bayesian multinomial geostatistical models were built to geographically predict the prevalence of each level of functional literacy. The number of school-aged children belonging to each of the functional literacy indicator classes was forecast for 2017. We estimated 4.20% of functional illiteracy burden among school-aged children in Mindanao might be averted by preventing T. trichiura infections. Areas predicted with the highest prevalence of functional illiteracy were observed in localised areas of the eastern region of the Visayas, and the south-eastern portion of Mindanao. The study demonstrates significant geographical variation in burden of functional illiteracy in school-aged children associated with STH infections suggesting that targeted helminth control could potentially promote the development of cognitive function of school-aged children in the Philippines. The benefits of a spatially targeted strategy should be tested by future studies.

While previous studies in the Philippines indicated an association between STH infection and cognitive development measured by memory and school performance, the contribution of STH infections on the overall functional illiteracy burden in the Philippines is unknown. This study presents the first use of geographical risk models of functional literacy adjusted for a wide array of probable confounders to uncover the associations with STH infections. This study also explores how the application of spatial epidemiology in mapping functional illiteracy provides an evaluation-planning tool for the design and implementation of STH-associated morbidity control intervention strategies, estimating the number of school-aged children with functional illiteracy associated with STH infections, and the number of interventions needed in the Philippines.

Functional literacy is one of the targets of the Sustainable Development Goals (SDG) of the United Nations, launched in September 2015 [1]. Functional literacy is a key indicator of cognitive function, especially information processing and comprehension, and it has been used to measure cognitive function in school-aged children [2]. According to the latest United Nations Educational Scientific and Cultural Organization report on literacy, there are globally 114 million illiterate adolescents and youths [3]. Despite widespread acknowledgement of this problem, between 2000 and 2015, global literacy rates were estimated to have improved by just 4%. Of particular relevance to the current work, progress in addressing national literacy rates in the Philippines has been slow [4]. Debate has recently intensified regarding the role of soil-transmitted helminths (STH), Ascaris lumbricoides, Trichuris trichiura, and hookworms (Ancylostoma duodenale, A. ceylanicum and Necator americanus) in childhood cognitive function, and by extension functional literacy [5]. To date, relatively few studies have investigated this link and evidence remains inconclusive, partly because differences in study designs and the use of cognitive development measurement tools makes it difficult to compare results of studies [6]. STH infections are among the most common infections in school-aged populations, and are particularly common in impoverished communities where the provision of water, sanitation, and hygiene education is limited [7]. STH infections are estimated to incur 4.98 million years lived with disability, related to anaemia, chronic nutritional imbalances, stunting, and cognitive and motor developmental delay [8]. Possible mechanisms for the effect of STH on cognitive function in children may include the interaction between the host immune system and the species of STH, the direct effect of metabolites excreted during infection, anaemia, or inflammation, and through secondary effects such as malnutrition and micronutrient deficiencies [9]. A recent experimental study demonstrated that T. trichiura infection may influence cognitive function of animals, although a definitive prospective human study has not been performed [10]. STH infections remain highly endemic across the Philippines among both primary and secondary school children [11, 12]. While previous studies in the Philippines indicated an effect of STH infection on children’s cognitive development, the contribution of STH infections on the overall functional illiteracy burden in the Philippines is unknown [13–15]. Processes that result in reduced functional literacy are complex and multifactorial. Compromised nutritional and family contextual factors such as poverty, unemployment, low maternal education, and household education stimuli have been linked consistently with cognitive function and educational performance of school-aged children [16]. In developing countries such as the Philippines, cognitive dysfunction could be caused by malnutrition in complex combinations with other factors including deprivation of environmental and emotional stimulation, biological factors, and infections such as pneumonia, meningitis, STH, and malaria [17]. Predictive risk maps, together with geospatial modelling, are emerging as important decision tools for public health policy makers [18]. However, to date there have been no studies looking at geographical variation in functional literacy and the associations with its key determinants. In this study, we aimed to quantify the associations between the spatial variation of STH infections and functional literacy in school-aged children in the Philippines, adjusting for probable confounders. In doing so, we also aimed to develop the first prediction maps of each functional literacy indicator in order to quantify the number of school-aged children at risk of reduced functional literacy in the Philippines. Ethical clearance for this analytical study was provided by the University of Queensland School of Medicine Low Risk Ethical Review Committee (clearance number 2014-SOMILRE-0100). We obtained population level data on functional literacy indicators of 11,313 school-aged children (aged 10–19 years) collected in 2008 during the nationwide Functional Literacy, Education and Mass Media Survey (FLEMMS). In brief, a total of 1,506 barangays (the smallest administrative unit in the Philippines: average diameter of 11 km) were included, 851 located in Luzon, 254 in the Visayas, and 401 in Mindanao (A map of 2008 FLEMMS survey locations is shown in S1 Fig [19]). All data analysed were anonymised. Functional literacy levels were stratified into four classifications – 1) those who cannot read, write, compute and comprehend were classified as functional illiterate; 2) those who can only read and write (with understanding a simple message) were considered as low functional literate; 3) those who can read, write and compute were considered as moderate functional literate; and 4) those who can read, write, compute and comprehend (with a significantly higher level of literacy that includes not only reading and writing skills but also numerical and comprehension skills) were considered as functional literate. Further detailed information on the FLEMMS can be found in S1 Text. A wide array of plausible contributing factors were obtained from the FLEMMS individual questionnaires and FLEMMS household questionnaires, including individual level sociodemographic indicators (e.g. age, sex, education attainment level, marital status, and employment status), and household level factors such as socioeconomic status (SES), access to water, sanitation, and hygiene (WASH), and household education stimuli. Information regarding head of households such as adult functional literacy was measured for a total of 10,339 heads of households. We obtained data of spatial predicted values of Plasmodium infections from the Malaria Atlas Project (S2 Fig) [20]. We used predictive maps of STH infection prevalence generated from spatial analysis of the data collected during the National Schistosomiasis Survey in the Philippines in 2005 to 2007 (S3 and S4 Figs) [12, 21]. Further information on each indicator is detailed in S2 Text. A total of 11,313 school-aged children and 10,339 heads of household with complete information were included in the analysis. The highest prevalence of functional illiteracy in school-aged children was observed in the Visayas, followed by Mindanao and Luzon (7.5%, 6.9%, and 3%, respectively; Table 5). The observed prevalence of functional illiteracy of heads of households was also higher in the Visayas and Mindanao compared to Luzon (15%, 15%, and 6%, respectively; S2 Table). Full descriptive results of the dataset are presented in S3 and S4 Tables, and S5 Fig [19]. While the relative importance of determinants of functional literacy varied between regions, some findings were consistent across all regions (Table 6). We found that highest education attainment, low socioeconomic status (SES) and adult functional illiteracy rates are major contributors to functional illiteracy across Luzon, the Visayas and Mindanao. Population attributable fraction (PAF) due to highest education attainment: 83.30% ([ratios of relative risks (RRR)], 9.69), 85.49% ([RRR], 10.69), and 81.96% ([RRR], 8.33), respectively; PAF due to low SES: 24.36% ([RRR], 1.97), 29.45% ([RRR], 1.98), and 53.22% ([RRR], 3.28), respectively; PAF due to adult functional illiteracy: 6.61% ([RRR], 2.16), 18.57% ([RRR], 2.53), and 26.71% ([RRR], 3.43), respectively. Our results indicated that the estimated risk of functional illiteracy attributable to poor sanitation facilities in the Visayas is 13.21% ([RRR], 1.86). Our results indicated that the estimated risk of functional illiteracy attributable to P. vivax infection, T. trichiura monoinfection, and moderate/high infection intensity class for T. trichiura in Mindanao were 0.53% ([RRR], 1.26), 4.20% ([RRR], 1.40), and 3.96% ([RRR], 1.82), respectively (Table 7). In Luzon and Mindanao, being female was negatively and significantly associated with the prevalence of functional illiteracy compared to being male (Table 6). Full results of the multinomial logistic regression models are presented in S3 Text. Across all regions, we observed reduced propensity of clustering and larger cluster sizes after adjusting for the effect of covariates (Table 8 and S6 Fig). Our predictive maps indicate significant spatial heterogeneity in the prevalence of each level of functional literacy within each region of the Philippines (Fig 1 and S7 Fig). Our predictive maps also demonstrate that the prevalence of functional illiteracy ranges between 1 and 3%, with the highest rates predicted in localized areas of the eastern region of the Visayas (Eastern Samar), and the centre (Davao Del Norte province) and the southwestern tip of Mindanao (Davao Occidental province) (up to 10.5%) (Fig 1). The models showed reasonable discriminative ability for functional illiteracy (area under the curve [AUC] = 0.75 for Luzon, 0.75 for the Visayas, and 0.71 for Mindanao), low functional literacy (AUC = 0.70 for Luzon, and 0.74 for the Visayas), and moderate functional literacy (AUC = 0.74 for Luzon, 0.72 for the Visayas and 0.82 for Mindanao) (Table 9). For 2017, it was estimated that Luzon had the highest estimated number of school-aged individuals with functional illiteracy (estimated total 2,185), followed by Mindanao (estimated total 1,550) and the Visayas (estimated total 1,212) (Table 10). The highest number of school-aged individuals with functional illiteracy (Fig 2), low functional literacy (Fig 3), and moderate functional literacy (S8 Fig) was circumscribed to areas around metropolitan Manila, with some municipalities exceeding 63 people per square kilometre. In the Visayas (Fig 2) and Mindanao (Figs 2 and 4), the predicted number of school-aged individuals with functional illiteracy was widely distributed in a number of provinces. The highest numbers of school-aged individuals with moderate (S8 Fig) or low functional literacy (Fig 3) were predominantly localised in the western region of the Visayas (some municipalities with more than 24 persons per square kilometre). In Mindanao, the highest numbers of school-aged individuals with moderate (S8 and S9 Figs) or low functional literacy (Figs 3 and 5) were localised in the central and southern parts of Mindanao with some municipalities exceeding 50 people per square kilometre. Names of provinces and municipalities with the poorest functional literacy indicators based on our analyses are provided in Table 11. Overall, our findings suggest significant spatial heterogeneity in the prevalence of functional literacy indicators within each region of the Philippines, reflecting variability in the determinants and the need for location-specific interventions. Our findings are consistent with previous studies, indicating that multiple factors exert a negative impact on functional literacy in school-aged children. Evidence suggests a possible link between low socioeconomic status and cognitive function, including children’s levels of language and literacy skills [17]. This relationship is mediated by different mechanisms such as parenting behaviour and household linguistic stimulation [17]. The observed association could also be explained by the effect of malnutrition in the poorest areas of the Philippines [27]. We found that the determinants of functional literacy are region-specific. For example, in Luzon and Mindanao, our results indicate that female school-aged children are at less risk of functional illiteracy compared to males, suggesting that girls may be less exposed to factors that affect their cognitive development compared to boys. The gender difference identified in our study could also be partly explained by integrated helminth control programs, which provide more attention to female adolescents [28], and the fact that boys are often involved with agricultural activities in resource-limited areas, limiting their participation in schooling or academic learning as agriculture in the Philippines has been dominated by males [29]. In contrast, our results for the Visayas showed that higher prevalence of functional illiteracy was associated with households with poor sanitation facilities. School-aged children who live in socioeconomically deprived environments face multiple challenges such as non-utilisation of sanitary facilities, open defecation, and limited education services. Previous studies have demonstrated that open defecation, a practice highly prevalent in the Visayas [11], is associated with high prevalence of STH infections [30]. Deficient sanitation promotes not only the transmission of STH infections, but also water-borne infections and diarrhoea. It also increases the associated risks of malabsorption, malnutrition, and iron-deficiency anaemia, which could aggravate cognitive dysfunction [31]. Our results also showed that the utilisation of unprotected water sources such as wells and lakes as main sources of drinking water at home is negatively associated with the prevalence of functional illiteracy in the Visayas. Our findings may be confounded by the fact that unprotected drinking water sources are more likely to be present in agricultural communities where access to food and nutritional security are assured through local food production [32]. Further investigation is needed to examine the factors mediating the relation between access to water sources and the prevalence of functional illiteracy identified in this study. The risk of STH-associated morbidity depends on the intensity of STH infection and the species of STH [33]. This study demonstrated significant geographical variation in the burden of functional illiteracy in school-aged children, which could possibly be explained by T. trichiura infection. Indeed, our findings suggest the risk of functional illiteracy among school-aged children in Mindanao might be reduced by 4.20% by preventing T. trichiura infection. These results could be explained by the pathophysiological impact of T. trichiura infection, including chronic diarrhoea, malnutrition, and iron-deficiency anaemia, all of which are associated with impaired cognitive function [34]. Furthermore, a recent experimental study demonstrated that T. trichiura contributes to pathological changes in the hippocampus and amygdala [10]. Existing preventive chemotherapy shows low to moderate efficacy against T. trichiura in high endemic countries [35]. This finding suggests that new solutions such as alternative treatment (e.g. oxantel pamoate) are needed to eliminate STH-associated morbidity. Our results show that large areas in the Philippines still lag in meeting functional literacy targets. Our estimates indicated that for 2017, Luzon had the highest estimated number of school-aged individuals with low levels of functional literacy. However, when these estimates were adjusted by the geographical variation in population density we found that areas in Mindanao had the highest density of school-aged individuals with functional illiteracy. Geographically targeted functional literacy interventions should thus be prioritised in at-risk areas identified in each region. These interventions should consider the region-specific determinants highlighted above. For example, in Luzon, the gender difference in the prevalence of reduced functional literacy identified in our study suggests that the current integrated helminth control program, which recently expanded its target age group from 1 to 12 years old to individuals 1 to 18 years old [28], should provide more attention to male adolescents. In the Visayas, given the high attributable risk of functional illiteracy from poor sanitation facilities, health educational programs promoting appropriate hygiene and sanitation practice such as educational videos (e.g. The Magic Glasses), which have proven efficacy in China [36], and are currently being tested in the Philippines are recommended. In Mindanao, given the perceived risk of functional illiteracy that could be associated with T. trichiura infections, and the chemotherapeutic failure for this particular parasite, educational and health promotion programs to this region should be considered. This study has some limitations that need consideration. Our data are from the 2008 Functional Literacy, Education and Mass Media Survey (FLEMMS) and may not accurately reflect the current situation. That said, these data constitute the best available and most contemporaneous dataset available and represents a cross section of the school-age population in the Philippines. Additionally, the rate of functional literacy has not seen much improvement in the last three FLEMMS (83.8% in 1994, 84.1% in 2003, and 86.4% in 2008) suggesting that data from the 2008 survey were unlikely to differ notably from the current situation. The prediction figures for 2017 may have overestimated the numbers of children currently at risk of functional illiteracy because this model assumes a constant increase in population and it does not account for changes in the prevalence of risk factors of functional illiteracy and STH infection such as poverty and WASH, with the rapid urbanisation that is happening in some parts of the Philippines [37]. However, according to the Philippines Statistics Authority (PSA), while the Philippines has shown some progress achieving the SDGs for literacy, poverty, and water sanitation, the levels have fluctuated between each survey period, showing no distinct trend, however there remain some parts of the Philippines, such as Autonomous Region in Muslim Mindanao (ARMM), that constantly record above average rates of poverty and less access to safe drinking water and sanitation facilities [38, 39]. Presently, there are no data available on functional literacy in the Philippines for 2017 to validate our prediction results, therefore further investigation is required. In addition, our predictive prevalence for P. falciparum and P. vivax are likely to represent underestimates, as malaria data for the targeted age group of our study were not available. However, given the low endemicity level of both species of malaria across the Philippines, the effect of malaria was subtle. Small-scale factors such as nutrition are also known to be important determinants of cognitive dysfunction [6]. Unfortunately, data on these factors were not available. In addition, our estimated PAF could be biased in the presence of unaccounted confounding factors. In conclusion, our findings support the need for spatially targeted strategies that can lead to a reduction in the transmission of STH infections and other determinants of functional illiteracy in school-aged children in the Philippines. In the context of the current work, this is particularly relevant in order for the Philippines to achieve the SDG target for functional literacy by 2030.

10.1371/journal.ppat.1003834

An In-Depth Comparison of Latent HIV-1 Reactivation in Multiple Cell Model Systems and Resting CD4+ T Cells from Aviremic Patients

The possibility of HIV-1 eradication has been limited by the existence of latently infected cellular reservoirs. Studies to examine control of HIV latency and potential reactivation have been hindered by the small numbers of latently infected cells found in vivo. Major conceptual leaps have been facilitated by the use of latently infected T cell lines and primary cells. However, notable differences exist among cell model systems. Furthermore, screening efforts in specific cell models have identified drug candidates for “anti-latency” therapy, which often fail to reactivate HIV uniformly across different models. Therefore, the activity of a given drug candidate, demonstrated in a particular cellular model, cannot reliably predict its activity in other cell model systems or in infected patient cells, tested ex vivo. This situation represents a critical knowledge gap that adversely affects our ability to identify promising treatment compounds and hinders the advancement of drug testing into relevant animal models and clinical trials. To begin to understand the biological characteristics that are inherent to each HIV-1 latency model, we compared the response properties of five primary T cell models, four J-Lat cell models and those obtained with a viral outgrowth assay using patient-derived infected cells. A panel of thirteen stimuli that are known to reactivate HIV by defined mechanisms of action was selected and tested in parallel in all models. Our results indicate that no single in vitro cell model alone is able to capture accurately the ex vivo response characteristics of latently infected T cells from patients. Most cell models demonstrated that sensitivity to HIV reactivation was skewed toward or against specific drug classes. Protein kinase C agonists and PHA reactivated latent HIV uniformly across models, although drugs in most other classes did not.

HIV establishes a state of latency in vivo and this latent reservoir, although small, is difficult to eradicate. To be able to better understand this state of latency, and to develop strategies to eliminate it, many groups have developed in vitro models of HIV latency. However, notable differences exist among cell model systems because compounds that reactivate latent HIV in a particular system often fail to do so uniformly across different models. To begin to understand the biological characteristics that are inherent to each HIV model of latency, we compared the response properties of five primary T cell, four J-Lat cell models and those obtained with patient-derived infected cells. A panel of thirteen stimuli that are known to reactivate HIV by defined mechanisms of action was selected and tested in parallel in all models.

The possibility to achieve HIV eradication has been limited, at least in part, by the existence of latently infected cellular reservoirs [1]–[3]. The major known cellular reservoir is established in quiescent memory CD4+ T cells, providing an extremely long-lived set of cells in which the virus can remain transcriptionally silent [1]–[3]. Reactivation of latent viruses followed by killing of the infected cells has been proposed as a possible strategy (“shock and kill”) to purge the latent reservoir [4]. Studies to examine the control of HIV latency and potential reactivation have been hindered, however, by the small numbers of latently infected cells in vivo and the absence of known phenotypic markers that can distinguish them from uninfected cells. In this setting, cell-line models of latency have been very useful due to their genetic and experimental tractability. Major conceptual leaps have been facilitated by the use of latently infected T cell lines [5]–[10], including the ability to conduct genetic screens [11]. On the other hand, latently infected cell lines are limited by their cycling nature and inherent mutation in growth controls, and the clonal nature of the virus integration sites. Such transformed cell lines lack the ability to differentiate and naturally oscillate between phases of quiescence and active proliferation in response to biological signals. Because of these limitations, a number of laboratories have recently developed primary cellular models of HIV-1 latency that capitalize on specific aspects of the T cell reservoir, found in vivo (reviewed in references [12]–[14]). These newer models afford investigators the ability to easily and rapidly study proposed mechanisms governing latency and to evaluate novel small molecule compounds for induction of viral reactivation. One significant complication, associated with the present variety of available latency models, is that notable differences exist among the cell model systems. Disparities relate to: the T-cell subsets being represented; the cellular signaling pathways that are capable of driving viral reactivation; and the genetic composition of the viruses employed, ranging from wild-type to functional deletion of multiple genes. Additional differences reside in the experimental approaches taken to establish latent infection in these primary cell models, which involve either infection of activated cycling cells that are later allowed to return to a resting state [15]–[19], or direct infection of quiescent cells [20], [21]. Because of such system variables, screening efforts in specific cell models with identified drug candidates for “anti-latency” therapy often fail to reactivate HIV uniformly across the different models. Therefore, the activity of a given drug candidate, demonstrated in a particular cellular model, cannot predict reliably the activity that will be seen in other cell model systems or in infected patient cells, tested ex vivo. The current situation in this research field represents a critical knowledge gap that is adversely affecting our ability to identify promising treatment compounds and their associated molecular mechanisms and is hindering the advancement of drug testing into relevant animal models and ultimately, human clinical trials. The present work represents a broad collaborative effort to compare and contrast induction of HIV reactivation across a battery of well-characterized cell models of viral latency, employing a highly coordinated and standardized testing approach. This work is based on the premise that it is unlikely that a single in vitro cell model can completely recapitulate the biological properties of the latent reservoir in vivo, let alone reflect accurately the response characteristics of infected patient cells ex vivo. Therefore, it is important to define both the common and unique properties among the available cell models of HIV latency in order to design a rational approach to employ such models in the identification of valid candidate drugs to induce HIV reactivation. Examples of how such an approach also can inform the underlying mechanistic actions of experimental compounds are available in the field. For instance, in the latency model developed by Bosque et al. [15], the derived central memory CD4+ T cells (TCM) are highly responsive to stimuli that activate the nuclear factor of activated T-cells (NFAT); on the other hand, virus reactivation from J-Lat clones [8] tends to be highly responsive to stimuli that activate the nuclear factor kappa of B cells (NFκB), such as protein kinase C (PKC) activators and tumor necrosis factor-alpha (TNF-α). Although the use of these two model systems would predictably yield different types of hits during a compound library screen, it is important to note that known compounds, which signal through either of these activation pathways, are capable of reactivating HIV replication in latently infected CD4+ T cells from patients ex vivo, and by inference, perhaps in vivo. To begin to understand the biological characteristics that are inherent to each model of HIV-1 latency, we compared the properties of six models (Table 1), to those obtained with a standard viral outgrowth assay using patient-derived infected cells [1], [22]. As no specific denominations have been assigned to these models, we have for simplicity referred to them by the name of the senior investigator in whose laboratory they were developed. They included the following (details are provided within the Methods section): The Greene laboratory model [23] is a modification of the original O'Doherty model of latency [20] and establishes HIV infection directly in quiescent primary CD4+ cells, using spinoculation delivery of virus. Replication-competent NL4-3 reporter virus is used, which contains Luciferase in the nef reading frame (Δnef/luciferase). After a short 3 day-culture, induction of provirus activation from latency is performed in the presence of integrase inhibitor to prevent viral spread and the contribution of any unintegrated viral species. Quantification of HIV replication by Luciferase expression is population-based. While only approximately 5–10% of the culture contains latently infected cells, this assay permits the generation and analysis of test compounds within 6 days. The model developed by Lewin and colleagues uses exposure of primary resting CD4+ T cells to chemokines that bind to receptors CCR7, CXCR3 or CCR6 to effectively establish infection with wild-type NL4-3 virus [21], [24]. Incubation with the chemokines does not cause significant cellular activation, but induces changes in the cellular actin cytoskeleton, which allows for efficient virus nuclear localization, integration, and establishment of latent infection [24]. Treatments to reactivate virus are followed by co-culture with amplifying feeder cells. Productive HIV replication is determined on a total population basis by quantification of soluble reverse transcriptase (RT) activity released into culture. The Planelles model [14], [15] establishes viral latency in cultured primary CD4+ T cells that have been differentiated by TCR stimulation in the presence of TGF-β, and αIL-4 and αIL-12 monoclonal antibodies into a non-polarized subset, representative of central memory cells (TCM) [14], [25]. Spinoculation with a packaged env defective NL4-3 clone establishes a single round of infection in the majority of the cells that transition into latency. Induced reactivation of HIV is monitored on a per-cell basis, using staining and flow cytometry detection for intracellular Gag (p24) expression. The Siliciano model [17] uses a two-step derivation of latency in cultured primary CD4+T cells, isolated from peripheral blood. In the first step, cells are TCR stimulated, transduced with the EB-FLV lentiviral vector, for constitutive expression of Bcl-2, expanded in culture with IL-2 and allowed to return to a resting state. In the second step, the cells are reactivated and infected with a trans-packaged, replication defective NL4-3 GFP-reporter virus clone (NL4-3-Δ6-drEGFP). After 3–4 weeks of culture, the GFP-negative cell subset, expressing a quiescent effector memory cell (TEM) phenotype, is isolated by flow cytometry sorting. Approximately, 2–6% of the recovered cells carry latent HIV infection. Reactivation of virus replication is tracked by GFP expression, on an individual cell basis. The Spina model (unpublished results; manuscript submitted) is based on early work demonstrating that HIV-1 can establish infection directly in resting primary CD4+ T lymphocytes in vitro [26], [27], and on recent work showing that during acute HIV infection in a heterogeneous population of primary CD4+ T cells, undergoing varying degrees of cell activation, viral latency is established early and preferentially in non-dividing and minimally activated cells. This model uses the experimental approach of deriving latent NL4-3 infection (wild-type) in non-dividing “bystander” cells during brief co-culture with autologous productively infected, proliferating cells. When the quiescent bystander cell population is isolated from co-culture, the latently infected subset ranges from 1 to 12% cells containing integrated HIV DNA, and 0.5–5% cells with inducible provirus, as measured by expression of intracellular Gag. Latent infection is found in all of the major phenotypic subsets of T cells: naïve, central memory and effector memory. After incubations with experimental compounds, reactivation of virus replication is measured on a population basis, through quantification of tat mRNA by RT-qPCR. Verdin and colleagues have generated a number of Jurkat cell line-derived clones, bearing latent HIV-1 in single integration sites, that were engineered to express GFP in lieu of nef [8] (J-Lat). J-Lat cells have been used in numerous studies that have contributed a wealth of knowledge in the area of viral latency. In contrast to several other models of HIV latency in cell lines, where mutations are present in the HIV tat gene or the TAR element, the J-Lat cell model contains wild-type tat and TAR. Three J-Lat clones established in the Verdin laboratory, 6.3, 8.4, 11.1 and one clone generated by the Greene laboratory, 5A8, have been included in this comparison. J-Lat 5A8 was derived by specifically selecting for cells that would be more responsive to αCD3/αCD28 co-stimulation than the parental J-Lat line [28]. Under untreated basal conditions, little or no GFP expression is detected. However, reactivation of latent provirus is readily monitored by flow cytometry analysis of GFP expression. Results obtained with the above cell models were compared to results obtained in quantitative viral outgrowth assays (QVOA; patient cell assay) performed in the Margolis laboratory, with resting CD4+ T cells obtained from the leukopheresed peripheral blood of aviremic, ART-treated HIV-infected patients. This assay, as first described by three laboratories [1]–[3], was later modified to its present design [22]. Following negative selection, resting CD4+ T cells are incubated with integrase and reverse transcriptase inhibitors to ensure the decay of any HIV genomes in the state of pre-integration latency [29]. The cells are exposed briefly to test compounds, and then plated in replicate microwells in a terminal-dilution assay and cultured with PHA-stimulated, allogeneic irradiated peripheral blood mononuclear cells (PBMC) from a sero-negative donor, and rIL-2. After 19 days, the microcultures are scored for virus replication by soluble p24 production, and the number of cells containing replication-competent HIV is expressed as infectious units per million CD4+ T cells (IUPM). Induction of viral reactivation across all cell models was assessed using a selected common panel of stimuli that are known to function by distinct and defined mechanisms of action. The panel included 13 treatments (Table 2) that modulate T cell processes such as T-cell receptor engagement, protein kinase C (PKC) activation, calcium influx, cytokine signaling, histone deacetylation, and release of P-TEFb from the HEXIM/7SK RNP complex. This study was designed to answer the following questions: 1) are certain models of latency biased towards or against particular cell signaling pathways; 2) can stimuli be identified that work uniformly in multiple models; 3) can a central uniting theme or a single signaling pathway be responsible for control of viral latency; and 4) can a model or limited group of models predict experimental drug activity in authentic latently infected cells from patients? Thirteen stimuli shown in Table 2 were chosen on the basis of their known or proposed activities in reactivating latent HIV-1 in various systems. For primary cell assays, experiments were performed with cells from three different donors, and replicate samples (duplicate or triplicate) were used for each treatment variable tested (refer to Methods Section for details). For J-Lat clones, experiments were performed in triplicate for each clone. Figure 1, panels A and B depict the average responses (mean +/− SEM) obtained with each of the cell models and the patient cell outgrowth assay. In all cases, the stimulus providing maximal reactivation response was used as a reference and assigned a 100% value, and the results from all other stimuli were normalized as a percentage of the maximal response (Figure 1). Within each individual experiment (e.g., donor cells), the untreated baseline value was first subtracted from each treatment response value, prior to the normalization step. The average relative response was then calculated across all experiments (donors) for each stimulus tested. While transformed and primary cell models could be tested at three concentrations of each stimulus, assays with patient cells/QVOA were only performed at a single drug concentration due to limiting cell numbers. Therefore, two comparisons were performed: one which included all concentration points for each drug, and did not include patient cell assay data; and a second one in which a single concentration point was considered, to provide an analysis that could include patient cell results. The maximal response for all primary models, except for the Lewin model and the patient-cell outgrowth assay, was obtained with αCD3+αCD28 antibody stimulation. In the Lewin model and the QVOA, PHA was the stimulus yielding a maximal response. The maximal response in the four J-Lat clones was obtained with PMA+Ionomycin. In all the J-Lat clones, except 5A8, CD3 surface expression is normally downregulated in culture (E.V., W.C.G., unpublished data). CD3 downregulation makes these cells unresponsive to αCD3/αCD28 antibody stimulation, although they remain responsive to PHA (most likely through engagement of the CD2 receptor). An additional representation of the data is shown in Figure S1, where, for each treatment, only the concentration of compound that was most active is represented. T-cell receptor engagement is effectively mimicked by the binding and cross-linking of antibodies against CD3ε, one of the signal transduction subunits in the CD3 complex [30] and the co-stimulatory molecule, CD28 [31]. Phytohemagglutinin (PHA-M) is a lectin that binds to carbohydrate moieties on surface glycoproteins. PHA is a polyclonal mitogen for T cells. Both PHA and αCD3/αCD28 antibody treatments stimulate signaling cascades that encompass TCR/LCK/p38 activation leading to calcineurin and NFAT activation, as well as PKC stimulation leading to NFκB activation. Incubation with αCD3/αCD28 antibody-coated beads produced strong responses in all primary cell models, with the exception of the Lewin model (Figure 1A). In contrast, all J-Lat clones, except 5A8 were completely unresponsive to αCD3+αCD28 incubation (Figure 1B). The response of J-Lat 5A8 cells after stimulation by αCD3+αCD28 coated beads, although detectable, was lower than that displayed by most primary cell models. However, the levels of stimulation can be improved using plate-bound αCD3 and free αCD28 antibodies, if so desired (D.R. and W.C.G., data not shown). Moreover, these cells are highly responsive to PHA, which indicates that these cells contain an intact signaling pathway downstream of TCR engagement. PHA reactivated latent viruses in all primary cell models and in the J-Lat clones, although with variable efficiency (Figure 1, Panels A and B). Therefore, the lack of responsiveness of J-Lat clones and of cells in the Lewin model to αCD3+αCD28 antibody treatment cannot be attributed to a lack of signaling mediators, since these cells respond to PHA through a highly similar signaling pathway. PKC is a family of ten kinases that are activated by phorbol esters [32]. In general, phorbol esters promote activation and differentiation of monocytes and monocytoid cells, as well as potent T-cell activation. Three PKC agonists were tested, namely PMA, prostratin (both phorbol esters); and bryostatin-1 (a cyclic polyketide). PKC agonists activate the DAG-PKC-NFκB signaling pathway. PMA has long been used as a T-cell mitogen. PMA was tested at 2 nM in primary cell models and 16 nM in J-Lat clones. At these concentrations, PMA elicited maximal or near-maximal responses in J-Lat cells, except in clone 8.4. Responses to PMA were near maximal in the Planelles and Siliciano models; the rest of the primary cell models and the QVOA also showed viral reactivation in response to PMA, although at more modest levels. Prostratin is a unique phorbol ester in that it induces potent T cell activation signals but, unlike PMA, is not tumorigenic. The ability of prostratin to induce T-cell activation through PKC, without tumor promoting ability, has made prostratin the subject of studies for its possible use as an inductive adjuvant therapy in the context of anti-retroviral therapy (ART) [33]. Another unique property of prostratin is that, despite being able to reactivate latent HIV-1, it exerts an inhibitory effect on active HIV-1 replication through downregulation of CD4 [34], [35]. The relative reactivation efficiencies observed in response to prostratin were similar to those obtained with PMA treatment. Thus, the models with the highest responses to PMA (J-Lat 6.3 and 11.1 clones, and Siliciano and Planelles models) showed the highest responses to prostratin as well. Conversely, poor to intermediate responses to PMA, observed in the Greene, Lewin and Spina models, and the quantitative patient-cell outgrowth assay (QVOA) were paralleled by similar responses to prostratin (Figure 1, Panels A and B). In the specific case of the Greene model, it has been observed that only a minority of cells, about 5%, respond to PMA, although the reasons for this observation are unknown. Bryostatins are a family of natural products found in several species of bryozoans. Bacterial symbionts of the bryozoan species are thought to be responsible for bryostatin synthesis (reviewed in [36]). Bryostatins bind to the diacylglycerol-binding region within the C-1 regulatory domain of PKC. Bryostatin-1 was recently shown to reactivate latent HIV-1 in vitro in monocytoid and lymphoid cell line models of latency [37] and was approximately 1,000-fold more potent than prostratin. More recently, DeChristopher and colleagues achieved the chemical synthesis of several analogs of bryostatin-1, which demonstrated potent activity in J-Lat cells [38]. Bryostatin-1 was very potent in J-Lat clone 11.1, but had only modest activity in the other J-Lat clones (Figure 1B). In primary cell models, bryostatin-1 induced maximal response in the Siliciano model, and about half-maximal responses in the Lewin and Planelles models. However, the Greene and Spina models, and the patient cell outgrowth assay, showed very low to non-detectable responses to bryostatin-1 (Figure 1A). A commonly utilized T-cell activation regimen in the laboratory, which mimics the signaling pathway used in TCR engagement, is the combination of PKC activation via PMA along with the calcium ionophore, Ionomycin, which bypasses the requirement for both CD3/TCR and CD28 receptor engagements. Signaling downstream of TCR engagement involves the formation of inositol triphosphate, which triggers an increase in the intracellular Ca2+ concentrations, which in turn activate the phosphatase, calcineurin. Calcineurin then dephosphorylates cytoplasmic NFAT transcription factor, which translocates to the nucleus. A combination of PMA and Ionomycin induced vigorous viral reactivation in most cell models tested, but not in the Spina model. Viral reactivation in response to PMA+Ionomycin was generally increased when compared to that of PMA alone, with the exception of the Lewin and Spina models (Figure 1, panels A and B). Unexpectedly, PMA+Ionomycin stimulation of primary T-cells in the Spina model caused inhibition of Tat mRNA transcription, the readout in this assay, to below initial basal levels (Figure 1A). It has been reported previously that PMA induction of HIV replication can be Tat-independent [39]; and in this case, the combination with Ionomycin appeared to actually suppress Tat transcription at 24 hrs. following stimulation. In the patient cell outgrowth assay/QVOA, PMA+Ionomycin produced a strong reactivation response that was higher than that observed with each compound alone. Previous reports showed that incubation with IL-7, alone [40] or in combination with IL-2 [41] can reactivate latent HIV-1 in resting CD4+ T cells isolated from infected individuals. IL-7 also reactivated latent HIV-1 in thymocytes in a SCID-hu mouse model of HIV latency [42] and in cultured TCM in the Planelles model [43]. In the Planelles model, IL-2+IL-7 stimulation of latently infected cells was previously shown to be inefficient (10–20% of the reactivation obtained with αCD3/αCD28) and to promote division of infected cells in the absence of viral reactivation [43]. Responsiveness to IL-7, or IL-2+IL-7 stimulation is physiologically relevant as these cytokines, along with IL-15, are known to drive the homeostatic proliferation of memory T cells in vivo [44]. A recent study found that IL-7, when administered to HIV-1 infected patients undergoing ART, promotes viral persistence by enhancing residual levels of viral production and inducing proliferation of latently infected cells without reactivation [45]. Robust responsiveness to IL-2+IL-7 was observed in the Siliciano and Spina primary cell models, and minimal activity was observed in the Greene model. Cells in the Lewin and Planelles models and the patient cell outgrowth assay responded poorly or not at all (<5% of maximal); whereas, cells in the Greene model exhibited a weak response. It is interesting to note that IL-7 used alone at 25 ng/ml induced robust reactivation in the Lewin model [46]. J-Lat cells failed to reactivate virus in response to IL-2+IL-7 stimulation. Jurkat cells, the parental tumor cell line from which J-Lat clones were derived, are IL-2-independent for their growth and survival, do not express the high-affinity IL-2 receptor, CD25 [47], [48], and express low levels of the IL-7 receptor alpha [49]. TNF-α is a potent inducer of viral gene expression in certain tumor cell lines harboring integrated, latent HIV-1, through the activation of NFκB [5], [8], [50], [51]. As previously reported, TNF-α treatment activated virus expression in J-Lat cells, especially in clones 6.3 and 11.1 (Figure 1B). However, among the primary cell models, TNF-α failed to induce any detectable viral reactivation in the Greene and Planelles models and showed only minimal activity in the Lewin and Siliciano models. In contrast, the patient cell outgrowth assay responded robustly to TNF-α, and cells in the Spina model showed an intermediate response. In order to better understand the responsiveness, or lack thereof, to TNF-α, we analyzed the levels of TNF-R in primary and Jurkat cells. We isolated bulk PBMC from two donors, selected memory CD4+ cells using CD45RO expression, and then stained the cells for CCR7, CD27 and the TNF-α receptor. These experiments showed that none of the freshly selected memory subsets tested (specifically, TCM, TEM and transitional memory T cells, TTM) expressed detectable levels of the TNF-α receptor (Figure S2). TNF-R expression was extremely low in cultured TCM from the Planelles model (Figure S2). In contrast, J-Lat 10.6 cells expressed high levels of TNF-R (Figure S2). HIV reactivation in response to TNF-α in vitro and in vivo is likely linked to whether cells under the specific culture or physiological conditions upregulate the expression of the TNF-α receptor. Hexamethylene bisacetamide (HMBA) is a hybrid bipolar compound that induces differentiation and apoptosis in transformed cell lines in culture [52], [53]. HMBA was shown to activate HIV transcription in vitro [7], [54], to reactivate latent HIV in vitro [19], [55] and to reactivate HIV in primary cells from aviremic, infected patients [56]. The activity of HMBA on HIV transcription is a result of its ability to induce dissociation of P-TEFb from the inhibitory 7SK ribonucleoprotein complex [19], [55]. HMBA treatment had significant reactivation activity in the QVOA and the Lewin model, but demonstrated little to no activity in the rest of the primary cell models and J-Lat clones tested (Figure 1, panels A and B). The “histone code” model states that a variety of covalent, post-translational modifications (PTM) on histone tail residues regulate the interaction of transcriptional regulators with chromatin to determine gene expression levels. The nature and localization of such post-translational modifications is broad, and their ability to act in a combinatorial manner provides an attractive model for how a finely tuned regulation can be effected. Histone code modifications include acetylation, phosphorylation, methylation, ubiquitination and sumoylation, among others [57], [58]. Acetylation of lysine residues in histone tails can have two important effects on chromatin organization (reviewed in [58]). First, this PTM results in neutralization of a basic charge on the lysine residue, which results in disruption of histone contacts with other histones and with DNA, diminishing the degree of compaction of the local chromatin. Second, proteins containing a specialized domain known as bromodomain specifically recognize acetylated lysine residues and then trigger downstream regulatory effects. Acetylation of histones is regulated by the concerted action of HATs and HDACs. Acetylated histones have long been associated with actively transcribed genes [59] and, therefore, inhibitors of HDAC (HDACi) are considered as general activators of transcription. Two main categories of HDACs have been described: Class I (HDAC 1, 2, 3 and 8), and Class II (HDAC 4, 5, 6, 7, 9, 10 and 11). Inhibition of Class I, but not Class II, HDACs has been shown to induce reactivation of latent HIV [60], [61]. Suberoylanilide hydroxamic acid (SAHA; also known as vorinostat) is a pan-HDAC inhibitor that targets both Class I and Class II HDACs, and can induce reactivation of HIV in models of HIV latency [11], [62]–[65], and in resting cells from ART-treated, aviremic HIV-infected patients [65]–[67], although it failed to induce reactivation in patient cells in another study [68]. Recently, a single administration of SAHA to ART-treated, aviremic patients was shown to induce global cellular acetylation and increases in viral RNA in resting CD4+ cells from these patients [69]. To test the ability of HDAC inhibitors to reactivate latent HIV in the various models of latency, we utilized three such inhibitors, provided by Merck Research Laboratories. SAHA potently blocks the Class I HDACs (i.e., 1, 2, 3, and 8) and has modest activity against Class II HDACs (i.e., 6, 10 and 11). MRK-1 is a selective inhibitor of the Class I HDACs (i.e., 1, 2 and 3) and HDAC6 (Class II); whereas, MRK-11 selectively blocks Class II HDACs (i.e., 4, 5, 6 and 7) and HDAC8 (Class I) [60]. SAHA was moderately potent in the Lewin and Spina cell models and QVOA, but was marginally active or inactive in the rest of the primary cell models and the J-Lat clones. The activity profile of MRK-1 was similar to that of SAHA in the primary models, showing the best activity in the Lewin cell model and the patient cell outgrowth assay. All the J-Lat clones had modest responses to MRK-1, which contrasted with the poor activity seen with SAHA in these cells (Figure 1B). The differences between SAHA and MRK-1 responses could, potentially be explained by the slightly different specificities of these HDAC inhibitors. In general, MRK-11 was inactive or minimally active (<20% response) in the QVOA and all J-Lat and primary cell models, except in the Lewin model, where it exhibited close to 50% activity. Cells in the Lewin model are unique in this study, in that they are very sensitive to viral reactivation by both Class I and Class II HDAC inhibitors. In contrast, other models tested are either insensitive to HDACi or show sensitivity to Class I inhibitors but not to Class II. The relationship between models based on the ability of compounds to activate latent HIV within each model was investigated by hierarchical clustering and heatmap visualization (Figures 2A and 2B). Two comparisons were performed. First, all the cell models for which data was available for all compounds and at all concentrations were compared (Figure 2A). This comparison excluded the patient cell outgrowth assay for which data for only certain concentrations of activators were available. In the second comparison, all models were included but only those concentrations that were universally tested were included (Figure 2B). In both comparisons, reactivation values obtained with PHA at 10 µg/ml were used as a reference, to which all other reactivation values were normalized to. Both comparisons yielded strikingly similar results. Three significant clusters of models were identified, with one robust outlier, the Spina model. The Lewin and J-Lat 5A8 clustered very close in both comparisons (Figures 2A and 2B), with the patient cell assay/QVOA being the next closest to those two (Figure 2B). Therefore, the first subcluster is defined by the Lewin, J-Lat-5A8 and QVOA models. The second subcluster is defined by the Planelles and Siliciano models, closest to each other, and the Greene model. The first two subclusters have a close association with each other, that separates them from the three remaining J-Lat clones (8.4, 6.3 and 11.1), which form the third and more distant subcluster. This clustering conforms to what would be expected biologically with the majority of primary cell models clustering together and the majority of cell line models clustering separately, with the exception of J-Lat 5A8, which clusters among the primary models. In addition, this clustering pattern was largely maintained when the QVOA data was included and a reduced compound set analyzed (Figure 2B). Since all primary cell models clustered together, this suggests that the resting phenotype of these models compared with the proliferating phenotype of J-Lat cells may influence the responsiveness to different agents. The QVOA model appears to cluster robustly with the Lewin model and the J-Lat 5A8, suggesting that these two models may represent the best proxy currently available for the activation capabilities of compounds when analyzing cells from HIV-infected subjects. However, this interpretation should be treated with caution as the clustering in Figure 2B, when the QVOA data was included, was performed with a reduced compound set and may not be as robust as the analysis that included all compounds at all concentrations (Figure 2A). The relationship of compounds to each other, based on their ability to activate HIV across the different models, was also investigated by hierarchical clustering and heatmap visualization (Figures 2A and 2B). The first analysis (Figure 2A) revealed that PMA+Ionomycin and, separately, αCD3+αCD28 antibody stimulation represented treatments that were strong outliers. The rest of the compounds then fell into one of two significant major clusters. The first cluster contained the majority of the HDACi, but also IL-7+IL-2 treatment, Ionomycin, and HMBA. The second cluster contained all concentrations of the PKC activators (i.e., prostratin, PMA and bryostatin) as well as PHA, TNF-α and the 6 µM concentration of MRK-1. This pattern of compound clustering was supported when data from the QVOA was included and a reduced compound set analyzed (Figure 2B). It is noteworthy that HMBA clustered interspersed with the HDAC inhibitors, which suggests potential similarities in the mechanism of action. The recent finding that the HDAC inhibitor, SAHA, can release P-TEFb from the inhibitory 7SK snRNP complex [70] provides a potential explanation for the close clustering of HMBA and HDAC inhibitors. In fact, a provocative finding in that study was that the viral reactivating ability of SAHA did not correlate with histone H3 or tubulin acetylation but, rather, with release of P-TEFb [70]. As shown in Figures 2A and 2B, the NFκB agonists PMA, prostratin, bryostatin, PHA and TNF-α cluster together. This result indicates that NFκB agonists consistently work as latency-reversing drugs across the different models, and that NFκB may play a central role in viral reactivation from latency, independent of the model used. In agreement with that, PHA and PMA were active in all the models tested. In summary, the clustering of compounds based on their activation of HIV across models conforms to what would be expected biologically and validates the analytical approach utilized in the current study. This study represents the first experimental comparison among several broadly used HIV latency systems, including primary cell models, transformed cell lines and patient-derived cells. To establish these comparisons in an unbiased manner, we chose a panel of known stimuli that were tested in parallel in the selected cell models. The methodology was designed to circumvent variations due to batch, formulation or concentration differences in the compounds tested. To the extent possible, the duration of exposure to each stimulus, the inclusion of appropriate controls and the maximal-response stimulus were standardized as well. PHA was the only stimulus that uniformly reactivated latent viruses in all systems tested. Most T cells, whether transformed or primary, express CD3ε or CD2, both of which are triggered by PHA. Unfortunately, the therapeutic potential of agonists of the CD3/CD28/CD2 signaling pathway is uncertain, given the plethora of undesirable side effects, including transient lymphopenia, previously observed in patients treated with OKT3 antibodies [71], [72]. PMA also reactivated viruses across models. Responsiveness to PMA was roughly, although not exactly, paralleled by responsiveness to the other PKC agonists tested, prostratin and bryostatin. For example, patient cells were responsive to PMA and prostratin, but not to bryostatin. Differences may be explained by the repertoire of PKC isoforms that is activated by each PKC agonist. This issue will require further exploration, as it is likely that certain PKC isoforms may be more involved than others in the reactivation of latent HIV. It is also plausible that certain PKC isoforms may be able to mediate viral activation with only minimal induction of cellular activation and/or proliferation, which, if true, would clearly be desirable in an eradication strategy. The addition of Ionomycin to PMA generally provided an enhancement of the activity observed with PMA alone, with the exception of cells in the Spina model. This is intriguing, and contrary to expectations. Ionomycin induces calcium influx, which activates the calcineurin phosphatase that, in turn, activates NFAT. A possible explanation for the loss of activity with PMA+Ionomycin in the Spina model might be the onset of apoptosis, due to a high level of stimulation. However, this was not the case; increased cell death was not observed in these cultures during testing. Virus reactivation in the Spina model was measured by levels of tat mRNA transcription after 24 hrs. following exposure to stimulus. In other studies, in which HIV reactivation was tracked by production of soluble p24, virus replication was detected readily 4–5 days after PMA+Ionomycin stimulation (C.A.S., unpublished results). Because PMA+Ionomycin stimulation delivers such strong and immediate cell activation signals, it may be possible that at early time points, limited “signaling resources” in primary T cells could be redirected away from the viral LTR and initiation of tat transcription [39]. Additional studies will be necessary to address this mechanistic point. The activities of cytokines are usually dependent on the presence or absence of their respective receptors on the target cells. TNF-α showed remarkable activity in several J-Lat clones and in patient cells, but was inactive or had low (Lewin) to moderate (Spina) activity in the primary cell models. As stated above, the TNF-R was not found in cultured or fresh TCM. Therefore, the high level of responsiveness in patient cells may underlie upregulation of the receptor under the culture conditions utilized, including perhaps the incubation with TNF-α itself. It will be informative to ascertain whether such upregulation occurs, and the specific conditions influencing it. This putative upregulation of the TNF-R is potentially exciting because, if appropriately targeted to cells in the latent reservoir, it would render cells exquisitely responsive to TNF-α or an agonist thereof. Responsiveness to TNF-α clusters among PKC agonists (Figures 2A and 2B), which likely reflects the fact that both types of stimuli culminate in NFκB activation. However, in the analysis displayed in Figure 2A, TNF-α clusters closest with MRK-1, an HDACi, at 6 µM. HDAC inhibitors are the first drug class to be utilized in clinical trials for HIV eradication and the results so far have been promising [69] because intracellular increases in HIV transcription were induced in vivo during SAHA treatment. Future development of HDAC inhibitors should be directed at ascertaining which HDAC isoforms are more involved in maintaining HIV latency, so that they can be specifically targeted. In general, the Lewin model clustered closely with the J-Lat 5A8 cells and both of these clustered with the patient cell outgrowth assay. However, one of the major differences between both models pertains to responsiveness to the HDACi, MRK11, which blocks Class II enzymes. Cells in the Lewin model displayed very high sensitivity to all tested HDACi, and were the only ones in this study to exhibit a substantial response to MRK11. In contrast, patient cells in the outgrowth assay did not respond to MRK 11. Three primary cell models, Greene, Planelles and Siliciano, had extremely low or no sensitivity to HDACi treatments. It is unclear what aspects of the biology of the cells or the latent viruses in these models renders the latent viruses so refractory to the effects of HDAC inhibition. As we suggest below, the low levels of active P-TEFb components in resting cells may constitute a major barrier to efficient transcription, which may not be overcome simply by inhibition of HDACs. Recent observations indicate that incubation of primary resting cells with stimuli that induce P-TEFb allows the cells to then become responsive to HDAC inhibition (Matija Peterlin, UCSF; personal communication). Cells in the Lewin and patient cell/QVOA models shared responsiveness to HMBA, while most other models had very low or no responsiveness to this agent. HMBA facilitates the dissociation of P-TEFb from the 7SK snRNP complex and makes P-TEFb more readily available to interact with Tat, and then to be recruited to the TAR loop on nascent viral RNAs. This is an early step in the transcriptional activation of the silent provirus and, therefore, it is viewed as a “gate keeper” step. Recent reports [18], [70], [73] have suggested that resting T cells contain very low levels of cyclin T and phosphorylated CDK9, leading to the hypothesis that the activity of the P-TEFb complex is inherently low, and not controlled by recruitment to the inactive 7SK snRNP complex. In view of these observations, Budhiraja et al. explained the lack of responsiveness of cells in the Planelles model as a result of the low levels of cyclin T and of CDK9 phosphorylation [73]. However, the previous model does not explain two of the observed responses in the present studies. First, patient cells and those in the Lewin model responded strongly to HMBA, while also being quiescent. Future studies should be undertaken to test the levels of P-TEFb in these model systems and examine the correlation between levels of P-TEFb and sensitivity to HMBA. Second, J-Lat cells seemed unresponsive to HMBA, while they would be expected to have high levels of active P-TEFb, given that they are dividing cells. We speculate that P-TEFb is not limiting in J-Lat cells, and that the rate-limiting step to active proviral transcription is either at the transcription initiation level, prior to the participation of Tat, or downstream of P-TEFb recruitment. A plausible mechanism for the lack of activity of HMBA in J-Lat cells is through transcriptional interference imposed by a proximal cellular promoter, as was shown for certain J-Lat clones, including J-Lat 6.3 and 8.4 [74]. Ionomycin was a poor inducer of reactivation in all primary cell models and patient cells, and had no detectable activity in the J-Lat cells. Calcium influx is necessary for activation of the NFAT transcription factor, but is not sufficient by itself for optimal viral reactivation. It appears that the full effect of NFAT on HIV reactivation, at least in cultured TCM, requires an additional signal provided by LCK activation [15]. PKC agonists were generally potent reactivators in most models tested here. Bryostatin is of particular interest because it stands as the only PKC agonist that is FDA approved and, consequently, data on its pharmacokinetics and toxicity in humans are available [75], [76]. Bryostatin has been tested in clinical trials for cancer and Alzheimer's disease [75], [76]. In addition, bryostatin was shown to synergize with the HDACi, valproic acid, in reactivation of latent HIV in a J-Lat model [76]. Although bryostatins are emerging as potential therapeutics for HIV eradication, they typically induce cellular activation, proliferation and secretion of pro-inflammatory cytokines. Thus, future research will need to identify analogs with diminished capacity to induce such undesirable cellular effects, while preserving the ability to reactivate latent HIV. No single experimental system of HIV latency completely recapitulated responsiveness to all types of stimuli tested here. The Lewin in vitro model displayed the broadest responsiveness. Similarities between responsiveness of patient-derived cells and the Lewin model cells were observed more frequently than with any other model. However, several notable differences separated the previous two models. These were: the high responsiveness of the Lewin model to MRK-11 and bryostatin, which contrasted with the lack of responsiveness of patient cells; and the lack of response of Lewin cells to αCD3/αCD28. The lack of responsiveness of the Lewin model cells to IL-2+IL-7 contrasted with the high responsiveness of the Spina and Siliciano models. Therefore, secondary screening of latency reversing drugs obtained through high throughput systems could be accomplished by using a combination of testing in the Lewin system plus a system that shows complementary properties, such as the Spina or the Siliciano models. The site of proviral integration can modulate the levels of viral transcription and has been proposed as a mechanism to explain latency [77], [78]. Specifically, integration in the vicinity of actively transcribed cellular genes can lead to transcriptional interference effects [74], [79], [80]. The present study did not attempt to analyze the influence of integration on proviral latency status. However, in a separate study [81], the influence of host cell gene transcription on proviral latency was analyzed and compared for five different models of latency including the Siliciano [17] and Planelles [15] models, a Jurkat model with polyclonal integration [78], infection of primary resting CD4+ T cells [82], and infection of primary activated CD4+ T cells [82]. When the influence of positioning in the chromosome (regardless of orientation) was examined, proviruses integrated in nearby positions shared the same latency status more often than predicted by chance. However, this trend was only statistically significant when comparing proviruses within each model, but not when comparing proviruses across models. This was interpreted by the authors to mean that local chromosomal features affecting latency are model-specific. Regarding proviral orientation with respect to cellular genes, the Siliciano model exhibited a modest, but statistically significant preference for latent proviruses to be in the same orientation as proximal cellular genes, confirming a previous report [80]. In contrast, the other models exhibited no statistically significant deviation from 50% of latent integrations being in the same orientation as cellular genes. Rational design of drugs to target HIV latency is not possible at the moment, because we do not have precise knowledge of all the cellular factors and activation pathways that impact viral transcription, leading to productive replication. A second obstacle to rational drug design for viral eradication lies in the notion that while the desired compound should trigger HIV reactivation, it should induce minimal or no cellular activation/proliferation. Therefore, drug screening studies should include an evaluation of the ability of candidate compounds to induce expression of cellular activation markers and proliferation. Studies involving human peripheral blood mononuclear cells were conducted at the following institutions, and approved by the respective internal boards as indicated: The test compounds, listed in Table 2, were obtained, and stocks prepared and distributed centrally to each of the participating laboratories by the CARE Pharmacology Core of the University of North Carolina. The compounds were tested in each cell model at the following final concentrations: αCD3/αCD28-conjugated beads (Dynal) at 1∶1 bead∶cell ratio; PHA-M (Sigma) at 1.1, 3.3, 10 µg/mL; PMA (Sigma) at 2 nM for primary T cells, 16 nM for J-Lat cells; Ionomycin (Sigma) at 0.5 µM; prostratin (LC Laboratories) at 0.3, 1, 3 µM; bryostatin (provided by the National Cancer Institute) at 10, 33, 100 nM; SAHA/vorinostat (Merck) at 0.11, 0.33, 1 µM; MRK-1 (class I HDACi, Merck) at 0.67, 2, 6 µM; MRK-11 (class II HDACi, Merck) at 3, 10, 30 µM; HMBA (Sigma) at 0.3, 1, 3 mM; TNF-α (Peprotech) at 10 ng/mL; IL-2 (Peprotech) at 30 IU/mL; IL-7 (Peprotech) at 50 ng/ml. IL-2, IL-7, TNF-α, and αCD3/αCD28 bead stocks were prepared in RPMI culture medium; HMBA stock was prepared in water. All the other compounds were prepared in DMSO solvent. Unless otherwise specified, each cell model tested and the HIV outgrowth assay included the controls: untreated (base culture medium), 0.1% DMSO, 0.5% DMSO (specific to 10 µg/mL PHA). The exposure time of cells to compounds was standardized across the models to 24 hrs., except for PHA (48 hrs.), αCD3/αCD28 beads (48–72 hrs.), and IL-2+IL-7 (5 days). The timing of assay read-outs for HIV reactivation was specific to each model system, dependent on unique cellular and viral characteristics. Initially any compound or any concentration of a compound that was not used universally across all models was removed. The untreated control, representing background activation, was subtracted from each compound for each donor in each model. Activation values for each compound were then averaged across donors within each model and any activation resulting from the DMSO condition was subtracted from those compounds that were dissolved in DMSO. DMSO has structural similarity to HDAC inhibitors, as some of these compounds were derived from DMSO following the observation of DMSO effects on transformed cells [85]. Average activation values for each compound were then normalized within each model by dividing by the average activation value for the highest concentration of PHA used so that models could be compared to each other. Finally, examining the distribution of average activation values across compounds revealed right-skewed data for each model and thus a log10 transformation was performed. Constants were added to the average activation value for each compound to account for negative values prior to log10 transformation and to shift activation values into a range that reflected their actual activation level. An unsupervised approach was used to determine the relationship between compounds based on their ability to activate HIV across models and also between models based on their response to compounds. Cluster 3.0 [86] was used for hierarchical clustering of compounds and models such that distances were calculated using the Euclidean based metric and then clustered using the average linkage method. The results were visualized in a heatmap using Java TreeView [87]. The statistical significance associated with clustering was determined using pvclust [88] (R package), which calculates approximately unbiased (AU) p-values that are computed using multiscale bootstrap resampling such that 95% equates to a p-value cut-off of 0.05. These normalization procedures and hierarchical clustering approaches were performed twice since not every compound was assessed at every concentration in the QVOA model. Specifically, the were performed once using a complete list of compounds but without data from the QVOA and a second time with a subset of compounds but now with the inclusion of data from the QVOA. TNFR surface expression was determined using anti-human TNFRI-APC (R&D Systems, Minneapolis, MN). Briefly, 1×105 cells were incubated with 1∶100 anti-human TNFRI-APC in 100 µl of PBS/3%FBS Buffer during 30 min at 4°C followed by flow cytometric analysis in a BD FacsCanto II flow cytometer using the FACSDiva software (Becton Dickinson, Mountain View, CA). Data was analyzed with FlowJo (TreeStar Inc., Ashland, OR).

10.1371/journal.pbio.1000620

Extinction Risk and Diversification Are Linked in a Plant Biodiversity Hotspot

It is widely recognized that we are entering an extinction event on a scale approaching the mass extinctions seen in the fossil record. Present-day rates of extinction are estimated to be several orders of magnitude greater than background rates and are projected to increase further if current trends continue. In vertebrates, species traits, such as body size, fecundity, and geographic range, are important predictors of vulnerability. Although plants are the basis for life on Earth, our knowledge of plant extinctions and vulnerabilities is lagging. Here, we disentangle the underlying drivers of extinction risk in plants, focusing on the Cape of South Africa, a global biodiversity hotspot. By comparing Red List data for the British and South African floras, we demonstrate that the taxonomic distribution of extinction risk differs significantly between regions, inconsistent with a simple, trait-based model of extinction. Using a comprehensive phylogenetic tree for the Cape, we reveal a phylogenetic signal in the distribution of plant extinction risks but show that the most threatened species cluster within short branches at the tips of the phylogeny—opposite to trends in mammals. From analyzing the distribution of threatened species across 11 exemplar clades, we suggest that mode of speciation best explains the unusual phylogenetic structure of extinction risks in plants of the Cape. Our results demonstrate that explanations for elevated extinction risk in plants of the Cape flora differ dramatically from those recognized for vertebrates. In the Cape, extinction risk is higher for young and fast-evolving plant lineages and cannot be explained by correlations with simple biological traits. Critically, we find that the most vulnerable plant species are nonetheless marching towards extinction at a more rapid pace but, surprisingly, independently from anthropogenic effects. Our results have important implications for conservation priorities and cast doubts on the utility of current Red List criteria for plants in regions such as the Cape, where speciation has been rapid, if our aim is to maximize the preservation of the tree-of-life.

The rapid loss of biodiversity is the most significant ecological challenge we face today. Over the past few years, the International Union for Conservation of Nature has published Red Lists documenting the inexorable slide towards extinction of species; recent losses include the Hawaiian crow, golden toad, Baiji dolphin, and the West African black rhino. In groups we know well, such as mammals, the risk of extinction has been related to biology, with the most vulnerable species tending to be large, slow breeding, and narrowly distributed. Although plants are the basis for life on Earth, our knowledge of the drivers of plant extinctions is poor. Here, we disentangle the causes of plant extinctions. We show that the processes linked with extinction risks in plants of the Cape, South Africa differ from those for vertebrates more generally. The most vulnerable species are found within young and fast-evolving plant lineages, opposite to patterns in vertebrates. Our results illustrate the intricate link between the processes of speciation and extinction. We also show that the most threatened species are marching towards extinction at faster rates, but surprisingly, the risk appears independent of human effects.

The rapid and accelerating loss of biodiversity is the most significant ecological challenge we face today. Current rates of extinction are already estimated to be several orders of magnitude greater than background rates [1] and are projected to increase another order of magnitude within the next few hundred years [2]. The terrestrial environment is now dominated by people—approximately one-third of land area has been transformed for human use [3] and one-fourth of global productivity diverted to human consumption [4]. The main direct human-induced drivers that impact biodiversity now are habitat loss and fragmentation, whereas climate change is likely to become a dominant future driver [5]. With each extinction event, we lose an element of biodiversity along with the associated ecosystem services it provides and the unique evolutionary history it represents. For over four decades the Red List of Threatened Species from the International Union for Conservation of Nature (IUCN; http://www.iucnredlist.org/) has provided a record of the incremental slide towards extinction of much of current biodiversity [6]–[8]. Based on detailed, peer-reviewed assessments [9], species are placed into one of the following seven categories, in order of increasing extinction risk: least concern (LC), near-threatened (NT), vulnerable (VU), endangered (EN), critically endangered (CR), extinct in the wild (EW), and finally, extinct (EX). There are currently 47,978 species on the IUCN Red List, of which 17,315 are classified as threatened with extinction; the vast majority (75%) of these records are from animals. In vertebrates, including mammals, birds, and amphibians, the proportion of species falling within the different threat categories differs significantly between higher taxa [10]–[13], indicating taxonomic selectivity in species vulnerabilities. Species traits linked with body size, generation times, and geographic range size are commonly associated with threat status [10],[13]–[16], with the most vulnerable species tending to be nested within species-poor clades [12],[13]. Identifying the key traits linked to high extinction risk is critical for predicting future declines and provides an opportunity for implementing preemptive conservation measures [17], although the particular attributes that influence vulnerability can differ among clades and geographic regions [18],[19]. Our knowledge of extinction in plants is much poorer than for vertebrates, with <5% of known plant species assessed (10,916 species) by the IUCN using current criteria. Nonetheless, current listings are still informative, especially if we focus on regions where taxonomic sampling has been more complete. Within flowering plants, over 70% of currently listed species are classified as at risk of extinction (category VU or higher; Figure 1), a much higher percentage than that reported for vertebrate groups (22% listed species). In large part, this likely reflects the relative incompleteness of the dataset for plants, and recent efforts suggest the proportion of threatened plants might be similar to that for mammals [20], although these estimates assume an even distribution of threatened species within higher taxa. Parallel to trends for vertebrates, plants also demonstrate an uneven distribution of threat across taxonomic ranks in both local floras [21]–[23] and globally, as shown here (p<0.001 and p = 0.002 from randomizations for families and orders, respectively; see Materials and Methods, Tables S1 and S2) and suggested elsewhere [24]. However, plant studies based on local floras show that life history traits (e.g., pollination syndrome, sexual system, habit, height, and dispersal mode) seem to only correlate poorly with species rarity (relative frequency of occurrence) or threat (index of species' decline or vulnerability to future decline) [25]–[27], and putative key traits differ between studies [23],[26],[28],[29]. Significant taxonomic structure in the distribution of threat in the absence of any strong correlation with heritable biological traits requires an explanation—although it remains possible that important traits have yet to be identified. Further, in contrast to findings for vertebrates, threatened plant species, as indexed by rarity, tend to be over-represented within species-rich taxa [21],. This trend suggests a potential link between the processes of speciation and extinction [30]. We have not attempted to differentiate between rarity and threat status here, because rarity does not have a standard definition or geographical scale and might refer to both local and global scales [21],[30] or relate more directly to extinction risk [21],[22]. Whilst rarity may not always match to estimates of global threat status as defined by the IUCN, it is a central criterion in Red List (e.g., Criteria A, C, and D), and regional rarity would be expected to provide a good indicator of regional Red List status in the absence of more comprehensive IUCN assessments. Further, a recent study [31] demonstrated that regional Red Lists typically correlate strongly with global listings, and for endemic species, regional status will translate directly to global status. In this study, we explore the distribution of extinction risks for two of the best-studied floras, the Cape of South Africa and the United Kingdom, for which practically complete regional Red List data have recently been published (Materials and Methods). These two regions represent very different floristic histories, the former assembled via post-glacial recolonization and range expansion, the latter a global biodiversity hotspot [32] with extraordinary high endemism [33] suggesting a rich history of in situ diversification. If the processes of speciation and extinction are coupled, the Cape flora likely provides the best opportunity for detecting the imprint of any such links. Moreover, if there are similarities in the distribution of extinction risks between the UK and Cape floras, it would be strong evidence for common trends across angiosperms. Here, we contrast the taxonomic distribution of extinction risks between the Cape and the UK floras. We then use detailed phylogenetic data for the Cape to identify the factors likely driving extinctions. Phylogenetic approaches not only allow us to correct for the non-independence of characters given the evolutionary relationships between species but also provide information on species' evolutionary histories. We reveal an unusual phylogenetic signal in extinction risks for plants and demonstrate how species' present-day vulnerabilities can be explained by their recent evolutionary past. Taxonomic structure in the distribution of threat has provided the stimulus to search for heritable traits that predispose some species towards extinction [12],[34]. If biological traits were the main determinant of extinction risk, and key traits were evolutionarily conserved so that species within clades tended to share similar vulnerabilities, we would predict the taxonomic distribution of threatened species to be broadly similar among regions. We contrasted the distribution of threat between the UK and South Africa. We revealed that taxonomic patterns differ dramatically between these two regions (Figure 1; Tables S3–S10). For example, some taxa show congruent patterns between the two floras (e.g., Brassicales – cabbages and allies), whereas others differ strikingly (e.g., Asterales – daisies and allies). This mismatch indicates that, as shown for other groups [18],[19], geography as well as biology are important in determining vulnerabilities in plants. To disentangle the factors underlying extinction risks we need, therefore, to access not only comprehensive Red List data but also detailed information on geography and phylogeny. Uniquely, the Cape of South Africa, a renowned biodiversity hotspot for plant life, provides an ideal case study. The Cape flora has been the focus of recent Red List assessments [35], and a phylogenetic tree depicting the evolutionary relationships among 735 plant genera based upon molecular data is available [36], along with fine-scaled regional distribution records [37]. Here, although we detected an uneven distribution of threatened species across higher taxa, we found no evidence for more closely related lineages to contain similar proportions of threatened species (p>0.05 from randomizations using Blomberg et al.'s K-statistic [38]), consistent with the weak correlations between life history traits and species vulnerabilities in plants. Previous studies have suggested a positive link between rarity and species richness in plants [21],[30]. Using generalized linear modeling (GLM) of threat in genera endemic to the Cape, we show that lineages with a higher proportion of threatened species are not only species rich but also young and rapidly diversifying (z = 5.86, p<0.001; z = −6.99, p<0.001; z = 5.54, p<0.001, from GLMs for richness, age, and diversification, respectively; Table 1). Moreover, in multiple regression including both age and richness as predictor variables, the relationship between threat and species richness is weaker, with taxon age the dominant predictor in the model (partial deviance explained  = 0.11 versus 0.17 for richness and age, respectively; Materials and Methods). Therefore, by incorporating information from phylogeny, we demonstrate that the link between richness and threat is in part a likely product of both factors co-varying with clade age, which correlates tightly with diversification rate—younger clades (genera) have diversified at faster rates (Spearman's rho  = −0.95). Analogous results were obtained when weighting GLMs either by number of listed species within each genus or the ratio of listed to non-listed species within each genus (Materials and Methods; Tables S11 and S12), and when controlling for phylogenetic covariance, but with diversification rate the better predictor (Table S13). Threatened species cluster in young, rapidly diversifying lineages. To evaluate further the link between clade diversification and extinction risks, we then searched for more information on radiating lineages within in the Cape. We identified 11 “Cape clades,” clades thought to have initially diversified in the Cape and with the majority of species indigenous to the Cape Floristic Region [39], for which near-complete species-level phylogenies have been produced previously or for this study (Table 2). For each clade, we used a continuous linear scale between 0 (LC) and 5 (EW) to quantify extinction risks [15]. We partitioned the variance in risk among and between clades to derive an index of the risk disparity through time (DTT) [40]. The disparity value at a given point in time is then the ratio of the average disparity of subclades for which ancestral lineages were present at that time relative to the disparity of all species within the clade [40]. We compared DTT to two alternative null models: first, assuming a Brownian motion model, in which species differences accumulate over time in a manner analogous to a random walk, and second, a punctuated model in which extinction risk was apportioned asymmetrically between daughter species at speciation (Materials and Methods). Although DDT plots are somewhat autocorrelated through time, large within-clade variance at the tips does not necessarily constrain within-clade variance to be low towards the root, or vice versa. Indeed, within the limits defined by the clade origin and the present, it is possible for clades to fall completely above expected values from our null models (e.g., compare DDT for Disa to the Brownian expectations). A Brownian motion model of trait change was strongly rejected (Table 2, Figure 2), and biologically significant phylogenetic signal in threat was only detected within Moraea (K = 0.39 versus an expectation of K = 1.0 for a Brownian motion model and K = 0 for absence of phylogenetic signal; see Materials and Methods). These results are inconsistent with a simple, heritable, trait-based model of extinction, although it remains possible that the important traits influencing extinction risks evolve in a non-Brownian fashion. However, we observed two common trends. First, most variation in extinction risk was between species at the tips of the phylogenetic trees—disparity within clades above the line derived from Brownian expectations (Figure 2) (e.g., Cypereae, Indigofera, Muraltia, Podalyrieae, Protea, and Restionaceae). Second, towards the root of the tree more variation in risk was observed between clades, disparity within clades being below the line derived from Brownian expectations (Figure 2) (e.g., Disa, Indigofera, Moraea, and Muraltia). The distribution of extinction risk therefore can be described as fitting a late-burst model of evolution, in which threat is phylogenetically conserved within deep (early diverging) clades but differs between closely related species at the tips of the tree. What process might explain this unusual phylogenetic distribution of extinction risks? In vertebrates, the least threatened taxa are found in the more diverse clades [12],[13], which might be expected from a simplistic assumption that low extinction elevates net diversification (speciation – extinction) rates [12]. We have shown that in plants, the pattern is reversed—more rapidly diversifying clades have more vulnerable species. We suggest this contrast between plants and animals reflects differences in their predominant mode of speciation. Speciation in plants is often associated with the establishment of small, reproductively isolated populations, for example, via occasional long-distance dispersal, flower-pollinator co-evolution, hybridization, and polyploidization [41]. Because small range size is a key IUCN criterion for assessing Red List status [9], rapidly diversifying lineages will then tend to have a high proportion of threatened species. Further, under a model of peripatric speciation, the geographic ranges of recently diverged taxa are also predicted to display large asymmetry in range size [42], explaining large differences in threat status among recently diverged species, and consistent with the late-burst evolutionary model suggested from the DTT plots. Our simulations assuming punctuated evolution capture well the asymmetry in trait values predicted from a peripatric model of speciation and better predict the disparity between species within clades than the Brownian null (Figure 2). To evaluate more directly the mode of speciation, we repeated the DTT plots with range size as the trait of interest. We show a strikingly similar trend for large disparity within clades towards the tips of the phylogenetic trees (Figure S1), fitting well the peripatric speciation model. However, less apparent is any trend for greater between-clade disparity towards the root of the trees (exception Moraea), consistent with evidence showing only weak phylogenetic signal in range size [43],[44]. If threat in plants is a product of the speciation process, we might predict extinction risk to be largely independent from anthropogenic drivers, such as habitat loss, which would have important implications for how we manage the conservation of threatened species. We tested this hypothesis using detailed taxonomic distribution data for the Cape [37] and an index of habitat transformation that aggregated the combined impacts of urbanization, cultivation, and alien invasion [45]. We found significant geographic clumping of threat even after correcting for taxonomic richness (Moran's I = 0.07, z = 10.11, p<0.01; Figure 2). However, “extinction hotspots” did not correlate with habitat transformation (Pearson's r = 0.038, F = 0.065, p = 0.8, adjusting degrees of freedom to correct for spatial autocorrelation; Figure 3), suggesting that the current threat status of Cape plants is independent from anthropogenic drivers, although that is not to say that they might not be important in the future. In contrast, but consistent with our findings above, we find that “extinction hotspots” reflect locations where lineages have recently diversified (Pearson's r = 0.504, corrected F = 0.534, corrected p<0.001; Figure 3). Finally, we looked at the trend in species' risk status over recent years. Because we found that threatened species often represent recently diversified taxa, we might expect over time young species to expand their geographic distributions as they become established and, as a consequence, decrease in perceived vulnerability. If this was the case, IUCN Red List classifications may be misleading, erroneously listing species with small but potentially expanding distributions. However, by comparing consecutive Red Lists of the South African flora (Materials and Methods), we found that the most threatened taxa are marching towards extinction at the fastest pace (ratio of species increasing in threat status against those remaining unchanged or decreasing in status: 43∶44 versus 28∶62 for genera within the top and bottom tertiles, respectively, ranked using 1996 Red Lists: G-test: G = 12.61, p<0.001). The species identified as most vulnerable by the IUCN Red List appear firmly committed to extinction. Our results suggest extinction hotspots may therefore represent both cradles and graveyards of diversity—linking the processes of speciation and extinction [46],[47]. Our results explain the paradox of strong phylogenetic structure in extinction risk in the absence of biological predictors at the species level. Speciation via peripheral isolates will result in asymmetry in range size—and therefore threat status—between closely related species that tend to share similar suites of biological traits. However, differences among lineages in the propensity to diversify will result in deeper phylogenetic structure of threat. Our simulations, assuming asymmetry in daughter lineages, fit well the observed disparity in species values within clades. However, we did not replicate the trend for deeper phylogenetic structure because we made the simplifying assumption that diversification rates were independent from trait values (a necessary limitation allowing us to simulate trait evolution along the branches of the empirical set of phylogenies). Finally, because diversification is also influenced by locations, traits that predispose species to diversify in one environment may fail to do so in another [48],[49], explaining regional variation in both threat and species richness. As we move towards assembling the complete tree-of-life, there has been increasing emphasis on preserving phylogenetic diversity [50],[51]. Recently, Vamosi and Wilson [24] suggested that globally we risk losing a disproportionate amount of angiosperm evolutionary history as we lose the most vulnerable species to extinctions. Using more detailed phylogenetic information, we show that plant extinctions may result in little loss of evolutionary history, at least in biodiversity hotspots where much of present-day diversity is a product of recent speciation. It is possible that the processes driving extinction in relict lineages are different from those for young diversifying lineages and that there exist two classes of globally threatened plant species. Estimating the true impact of plant extinctions on the loss of evolutionary history across the angiosperm tree-of-life will therefore require detailed knowledge of the interspecific phylogenetic relationships within higher taxa. Our results linking speciation and extinction derive from an analysis of the unique flora found within the Cape region of South Africa and might not extrapolate across less diverse biomes with different evolutionary histories. However, we note that the trend for threatened species to fall within species-poor clades has been observed within the relatively depauperate floras of the UK [22] and North America [21]. We suggest that species turnover (speciation and extinction) might be rapid generally for plants, and high extinction rates may not be unusual over evolutionary timescales. Because even rapid speciation occurs over timescales too long for practical management [52], conservation efforts must focus urgently on reducing rates at which species are being lost [53]–[55]. However, if we wish to maximize the preservation of the tree-of-life [11], we must consider whether plants and animals may be best served by different assessment criteria when deciding upon conservation priorities. For example, for plants, investing in currently less threatened but still vulnerable species in more evolutionary distinct clades might be the most sensible conservation strategy, whereas for vertebrates the IUCN Red List may provide a more straightforward index for conservation decision making. Plant species Red List data were extracted from the following sources: (1) Global data: The IUCN Red List of Threatened Species (http://www.iucnredlist.org, accessed August 2009); (2) UK flora: Joint Nature Conservation Committee's Vascular Plant Red Data List for Great Britain 2006 (http://www.jncc.gov.uk); and (3) South African flora: the Interim Global Status for plants published online by the South African National Biodiversity Institute (SANBI: http://www.sanbi.org/; accessed June 2008), and the Beta version of South African Plant Red List including all assessed taxa and their status (SANBI: http://www.sanbi.org/; accessed October 2010). An updated print version is now available from the Pretoria National Botanical Garden [35]. Plant names were synonymized to agree with the Angiosperm Phylogeny Group [56] taxonomy. We used IUCN Red List categories to classify each listed species as either threatened (vulnerable [VU], endangered [EN], critically endangered [CR], extinct in the wild [EW], and extinct [EX]) or non-threatened (least concern [LC] and near-threatened [NT]). Taxonomic structure was evaluated by recording the ratio of threatened to non-threatened species recorded within higher taxa (families, orders, and, for the local floras, APG [56] higher taxonomic class) and calculating the variance across taxa. Significance was assessed by randomizing species membership among taxa and recalculating the ratio of threatened species within each random assemblage, keeping number of species per taxon constant. Taxa with significantly more or less threatened species were identified by comparing the observed proportion of threatened species with expectations from the randomizations. The p values were determined from 1,000 random draws. Last, we synonymized 1996 [57] and 2007 Interim Global Status for the South African flora so that extinction risk categories were broadly equivalent, and we scored species included in both listing as +1 if threat increased or −1 if threat decreased. We then derived an index of change in threat status by summing species values within each genus. Because criteria used in the 1996 and 2007 listing were not directly comparable, we evaluated whether species identified as the most vulnerable to extinction in the 1996 listing were more or less threatened in the subsequent assessment by contrasting trends between the top third most threatened taxa with trends within the bottom third (least threatened taxa), ranked using the 1996 listing. We quantified phylogenetic signal in extinction risk for the Cape flora using Blomberg's K-statistic [38] on the recent comprehensive phylogenetic tree of Cape genera [36]. We use the proportion of threatened species within genera, as described above, as our index of threat. Significance was calculated by randomizing the tips of the tree and recalculating K (1,000 replicates). The K statistic compares the distribution of phylogenetically independent contrasts across nodes within the clade [58], to expectations under a Brownian motion model of trait evolution. Because our metric of extinction risk is bounded between 1 and 0 and therefore violates assumptions of normality, we report only the p values from randomizations. Within the Cape flora, we characterized the phylogenetic distribution of extinction risk across 11 exemplar clades resolved at the species level (Table 2). For each clade we then described variance in extinction risk between and within clades using Harmon and colleagues' index of disparity through time (DTT) [40]. DTT is derived from the standardized mean pair-wise distance between species and therefore does not necessitate the reconstruction of ancestral states. Values of DTT near 0 indicate that most of the variation is partitioned between clades, whereas values near 1 indicate that most variation is among species within subclades. Because at the limits DTT must be 1 at the root and 0 at the tips, we compared observed values to two alternative null models. First, we derived expectations under a null model of Brownian motion. Second, we simulated a model of punctuated evolution in which daughter lineages are assigned trait values asymmetrically at speciation. Following divergence, one daughter lineage inherits a value that differed from the parental lineage by some constant factor, and the other daughter lineage assumes the parental value. Prior to subsequent diversification, both lineages evolve trait distances with a drift factor taken from a normal distribution with mean zero and a given standard deviation plus an evolutionary trend for the smaller lineage to expand in size. This latter simulation might be considered to approximate a model of range size evolution assuming speciation via peripheral isolates, in which one daughter lineage has restricted initial geographic distribution. To evaluate the link between speciation mode and extinction risk directly, we then repeated the analysis of DTT using range size as the trait of interest. Here we estimate range size as the number of quarter degree squares with presence data for each species [37]. If speciation via peripheral isolates is common, we would expect large disparity in range size between species within clades towards the tips of the phylogeny. We did not attempt to link speciation and extinction rates with trait values because to compare DTT plots phylogenetic topology must be identical. Simulations were implemented in the Geiger R-library [59]. R-code for the punctuated model is available from the authors. For all Cape genera we compiled data on species richness (n) [60], time to most recent common ancestor (millions of years; my), and net diversification rates (log[n]/my). We then constructed a series of regression models to describe the relationships between threat and taxon age, richness, and diversification rate. Statistical models were constructed in the R statistical package (http://www.r-project.org/). First, we generated a series of single-predictor generalized linear models (GLMs) with the proportion of threatened species in each genus as the response and assuming binomial errors. Species richness and diversification rates were log transformed, and taxon ages square-root transformed prior to model fitting. Second, we constructed a two-predictor model including both species richness and age as explanatory variables. Model fits were assessed using Akaike's information criterion. The marginal contributions of each variable in the two-predictor model were estimated as the additional percent deviance explained by inclusion of that variable to the reduced model. Because missing lineages might result in overestimation of taxon ages and hence underestimation of diversification rates, we focused our analyses on genera endemic to the Cape. Cape endemics are most likely to have their sister clades also included in the phylogenetic tree, and therefore subtending branches will not be broken by the addition of missing taxa; nonetheless, models including all Cape genera also supported the significant relationship between taxon age and threat (not shown). We evaluated model sensitivity in two ways. First, we repeated the set of regression models weighting the data by (1) the logarithm of the number of Red List records for each genus, thereby assuming our estimates of threat are more reliable for genera where multiple species have been assessed, although this will also down-weight the influence of species-poor taxa, and (2) the ratio of listed species to total clade species richness, providing an indication of the impact of missing (unlisted) species. Second, we generated a phylogenetic distance matrix (√my) and used partial Mantel tests to determine the relationship between threat and diversification whilst controlling for phylogenetic relatedness. We extracted the list of genera within each quarter degree square in the Cape from the PRECIS database [37]. We then calculated the mean proportion of threatened species and mean diversification rate for genera within cells, as well as a per-cell index of human impact on the environment [45]. Spatial correlation coefficients were calculated using Pearson's correlation coefficient (r) for grid cells and controlling for spatial covariance when estimating the statistical significance by adjusting degrees of freedom [61]. First, we estimated the correlation strength between mean threat and human impact, because habitat transformation is thought to be a key driver of species extinctions. Second, we estimated the correlation strength between threat and diversification, because this was the best of our predictor variables from the regression models (above). Spatial analysis was performed in ArcMap (9.2 Environmental Systems Research Institute Inc.) and SAM (Spatial Analysis in Marcoecology v3.0) [62].

10.1371/journal.pgen.1000631

The Ataxia (axJ) Mutation Causes Abnormal GABAA Receptor Turnover in Mice

Ataxia represents a pathological coordination failure that often involves functional disturbances in cerebellar circuits. Purkinje cells (PCs) characterize the only output neurons of the cerebellar cortex and critically participate in regulating motor coordination. Although different genetic mutations are known that cause ataxia, little is known about the underlying cellular mechanisms. Here we show that a mutated axJ gene locus, encoding the ubiquitin-specific protease 14 (Usp14), negatively influences synaptic receptor turnover. AxJ mouse mutants, characterized by cerebellar ataxia, display both increased GABAA receptor (GABAAR) levels at PC surface membranes accompanied by enlarged IPSCs. Accordingly, we identify physical interaction of Usp14 and the GABAAR α1 subunit. Although other currently unknown changes might be involved, our data show that ubiquitin-dependent GABAAR turnover at cerebellar synapses contributes to axJ-mediated behavioural impairment.

Cerebellar ataxia describes a combination of motor symptoms and uncoordinated movements that originates from various hereditary and non-hereditary diseases. Although functional disturbances of cerebellar inhibitory output signals are thought to cause ataxia, the underlying molecular mechanisms are barely understood and medical treatment therefore remains difficult. We analysed a behavioural abnormality up to the molecular level in a mouse mutant (axJ) representing a model for ataxia. The axJ mutation reduces the expression level of a ubiquitin protease (Usp14) leading to an abnormal turnover of neurotransmitter receptors. Despite other yet unknown changes in axJ mutants, our data show that intracellular protein turnover contributes to a motor behavioural syndrome.

A number of heterogeneous hereditary and non-hereditary disorders lead to ataxia characterized by coordination failures [1],[2],[3]. The spontaneous axJ mutation affects the locomotory system, causing an ataxic phenotype in mice [4]. The mutated gene encodes the deubiquitinating enzyme (DUB) Usp14 [5], a member of the ubiquitin-specific protease family [6],[7]. Due to insertion of an intracisternal A-particle into intron 5, expression levels of full-length Usp14 in brains of axJ mice are reduced to about 5% [5]. Usp14 catalyzes the hydrolysis of isopeptide bonds in ubiquitin-protein conjugates [8]. Upon alternative splicing of exon 4, two isoforms of Usp14 are generated. The full-length isoform contains an addition of 33 amino acids, required for proteasome binding. Accordingly, binding of Usp14 to the proteasome is thought to be necessary for efficient hydrolyse activity of Usp14 [9],[10]. AxJ mice display an exclusive downregulation of the full-length isoform, thereby representing a specific knockdown of the proteasome binding form of Usp14. Although the proteasome is likely to be involved in the neurological dysfunctions [9], Usp14 is unable to process polyubiquitin chains [11]. Since, its physiological substrate is thought to be mono- or oligoubiquitinated [5], rather than representing a polyubiquitinated protein destined for degradation at the proteasome, Usp14 may have several functions in ubiquitin-signaling pathways. Ubiquitination is a key process in the regulation of synapse formation and function [6],[7],[12]. Following endocytosis, ubiquitinated receptors are sorted for lysosomal degradation, thereby preventing their recycling to the plasma membrane [13],[14],[15]. For instance, the surface expression of glycine receptors (GlyRs) depends on ubiquitination, suggesting an important role for this process in the regulation of synaptic receptor levels [16]. Moreover, ubiquitination of inhibitory GABAA receptors (GABAARs) has recently been shown to be activity-dependent and to regulate synaptic GABAAR accumulation [17]. GABAARs mediate the majority of fast synaptic inhibition in the mammalian brain. In the cerebellum, 75% of all GABAARs contain the α1 subunit [18], whereas PCs exclusively express α1-containing GABAARs [19],[20],[21]. PCs transform excitatory afferent signals to inhibitory efferents that target the neurons of the deep cerebellar nuclei (DCN) and vestibular nuclei (VN) [1],[22],[23]. Their inhibitory influence on DCN and VN neurons is a prerequisite for normal motor coordination, and even minor disturbances of cerebellar inhibition has been shown to cause uncoordinated movements and ataxia [1]. Hence, mouse mutants characterized by Purkinje cell degeneration, such as pcd or leaner mice suffer from ataxia [24],[25],[26],[27]. Here, we show that the downregulation of Usp14 in axJ mice is accompanied by a marked redistribution of intracellular α1-containing GABAARs to PC surface membranes, leading to enlarged IPSC amplitudes. We further demonstrate physical interaction of Usp14 and GABAAR α1, suggesting that Usp14 directly participates in the regulation of synaptic GABAAR turnover. Consistently, interference with GABAAR-Usp14 binding in a heterologous system mimics the in vivo observations. Our data demonstrate a new concept with the ubiquitin-proteasome system (UPS) representing a key player in synaptic neurotransmitter receptor regulation. Mice carrying the axJ mutation display reduced expression levels of the full-length Usp14 isoform in brain, whereas expression of the short Usp14 isoform remains unaltered ([5], Figure 1A). Phenotypically, axJ mice demonstrate severe coordination failures and ataxia (Figure 1B) [4], often linked to dysfunctions within cerebellar circuits [28]. Although increased apoptotic cell death is reported in axJ-derived cerebellar granule cell layers [29], application of nuclear staining revealed that the overall architecture of the cerebellum remains normal (Figure 1C). To visualize PC bodies (arrows) and the molecular layer, representing PC dendrites, the PC marker protein Calbindin was immunolabeled. (Figure 1D, upper panels, green). Parallel staining of GABAARs using antibodies specific for the α1-subunit (Figure S1A-S1D) that represents the only α-type subunit in PCs, demonstrated strongly increased GABAAR cluster intensities in the molecular layer of axJ compared to wt cerebellum (Figure 1D, upper panels, red; lower panels, white). This phenomenon appeared to be mosaic and is in agreement with previous reports of variable expression levels of Usp14LF throughout different tissues [5]. Notably, these effects were specific and due to decreased Usp14 levels, since neuron-specific transgenic expression of Usp14 on the background of axJ mice [9] genetically reversed this effect, thereby leading to similar GABAAR α1 signal intensities as detected in wt PCs (Figure 1D, axJ x tg, right). Together, these data indicate that Usp14 regulates either the gene expression or the subcellular distribution of α1-containing GABAARs in PCs. To investigate the underlying mechanism of increased GABAAR clusters, PCs of wt and axJ mice were analyzed at the subcellular level. Immunostaining of GABAAR α1 using either fluorophore- (Figure 2A) or biotin-labeled (Figure 2B) secondary antibodies revealed a marked increase in GABAAR α1 clusters at the surface of cell bodies and proximal dendrites (Figure 2A and 2B, arrows). At the ultrastructural level, electron microscopy confirmed that large areas of the axJ PC surface, including extrasynaptic sites, were covered by α1-containing GABAARs (Figure 2C). Notably, the cytoplasm of PCs did not show increased signal intensities between the genotypes (Figure 2A–2C). In addition, western blot analysis of cerebellar protein extracts from wt and axJ mice (Figure 2D) as well as mRNA levels upon in situ hybridization (Figure 2E) demonstrated equal signals of GABAAR α1 proteins and mRNAs in both genotypes, indicating that the total gene expression of GABAAR α1 is not increased. We therefore conclude that a major loss of Usp14 expression leads to a surface redistribution of intracellular α1-containing GABAARs rather than to a significant increase in GABAAR α1 expression levels. Consistent with the immunohistochemical data, analysis of IPSCs (n = 10,000 events) indeed revealed that 67% of axJ PCs displayed a significant increase in GABAergic current amplitudes (Figure 3A–3D, and Figure S2A and Figure S2B). In parallel and as expected for a postsynaptic receptor phenomenon, the kinetic parameters, such as rise-time (10–90%) and decay-time (τ) remained unaltered under these conditions, indicating that both genotypes display no major changes in neurotransmitter uptake or release mechanisms (Figure S2C). However, the maximal amplitudes in axJ animals (>150 pA) displayed significantly (p = 0.005) higher decay time constants (τ = 12.1±3.1 ms), as compared to the decay time constants (τ = 10.3±2.9 ms) of maximal IPSC amplitudes in wt animals (80–150 pA; Figure S2C). Such differences are consistent with increased perisynaptic receptor numbers and support the immunochemical and EM observations. We therefore conclude that increased GABAAR levels at PC plasma membranes induce increased cerebellar inhibition, that leads to reduced inhibitory output levels of PCs. Notably, also PC degeneration (pcd) mutant mice display a severe decrease of PC inhibitory output and develop ataxia [1],[27]. In addition, GAT1 deficient mice, represented through prolonged GABA actions, due to disturbed GABA reuptake, suffer from ataxia [30]. Hence, altered inhibitory input to PCs leads to similar behavioral consequences compared to the loss of cerebellar inhibitory output upon Purkinje cell degeneration. However, if the observations in the present study contributed to the ataxia phenotype of axJ mice, one should identify a molecular link between GABAAR turnover and Usp14-mediated pathways. To determine, whether Usp14 and GABAAR α1 might physically interact, we applied the lexA-based MATCHMAKER yeast two-hybrid system using Usp14LF (prey) and the large intracellular loop of GABAAR α1 (aa 334–420, bait). These experiments indeed revealed that Usp14 represents a direct GABAAR α1 binding partner (Figure 4A). Fine mapping, using systematic GABAAR α1 deletion mutants, identified the Usp14 binding region within the first 13 amino acids of the α1 loop sequence (aa 334–346) (Figure 4A). Vice versa, the GABAAR α1 binding site within the Usp14 protein was localized at its C-terminal domain (Figure 4B). In order to biochemically verify this interaction, we then applied a pull-down experiment using the GST-tagged GABAAR α1 loop (aa 334–420). Endogenous Usp14 protein derived from mouse brain lysates specifically bound to the immobilized GST-tagged GABAAR α1 loop, but not to GST alone (Figure 4C), indicating in vitro binding of the protease and the receptor polypeptide. Differential centrifugation of brain extracts revealed that both endogenous proteins cofractionate at P2 plasma-membrane (10,000×g), P3 vesicular (100,000×g), and P4 protein complex (400,000×g) fractions. However, while GABAARs are enriched at the plasma membrane (P2), Usp14 binds to the proteasome and is consequently enriched in fraction P4 (Figure 4D). This marginal overlap is consistent with a transient enzyme-substrate complex, however turned out not to be sufficient to obtain coimmunoprecipitation under standard conditions. Nevertheless, a GFP-tagged Usp14 mutant (GFP-Usp14(H434A-D450A)), harboring two point mutations within its functional catalytic domain [31], stabilized the complex, and enabled coprecipitation of both full-length binding partners derived from HEK293 cells (Figure 4E). Together, these data demonstrate physical interaction of GABAAR α1 and the ubiquitin-specific protease Usp14, and suggest that the observed GABAAR redistribution in ataxia mice (Figures 1 and 2) is directly caused by the loss of Usp14, thereby indicating that GABAAR turnover is ubiquitin-dependent. We next asked whether both proteins colocalize at the subcellular level. While GABAAR α1 subunits have been extensively characterized in both tissues and cells [32],[33], immunohistochemical analysis of Usp14 displayed a wide distribution in all layers of the cerebellum (Figure 5A) and was detected at synaptic vesicle protein 2 (SV2)-positive synaptic sites in both cultured hippocampal and cerebellar neurons (Figure 5B, and Figure S1E, turquoise). For analysis at ultrastructural resolutions, we performed immunoelectron microscopy using biotin- (Figure 5C, upper panel) or gold-labeled (Figure 5C, middle and lower panels) secondary antibodies. In this assays Usp14 was detected in close proximity to and directly at postsynaptic sites (Figure 5C, upper panel, arrows), at both the pre- and post-synapse, as well as directly at synaptic plasma membranes (Figure 5C, middle and lower panels, arrowheads). In addition and in consistence with the in vitro binding data, coimmunostaining with antibodies specific for Usp14 and the GABAAR α1 subunit revealed partial colocalization in both cultured hippocampal and cerebellar neurons (Figure 5D, and Figure S1F, yellow, white arrows). At ultrastructural levels, this could be confirmed using gold-labeled secondary antibodies of different particle sizes. In accordance to the literature [33], GABAAR α1 (black arrows) was localized opposed to unlabeled (Figure 5E, left, arrows) or synaptophysin-positive presynaptic boutons (Figure 5E, middle, arrows), while colabeling of Usp14 and GABAAR α1 was rather detected at submembrane tubular organelles (Figure 5E, right, white arrow), described in both dendritic shafts and spines [34]. In addition to the smooth endoplasmic reticulum (SER), tubular compartments are generated through merge of internalized vesicles and multivesicular body (MVB)-tubule complexes and serve as intracellular stores of material destined for recycling or degradation [34],[35],[36],[37]. Given the fact that organelles that mediate neurotransmitter receptor sorting are localized subsynaptically [34],[38] with ubiquitin serving as a signal for internalization [39],[40], both the observed in vitro binding (Figure 4) and colocalization data (Figure 5) suggest that Usp14 represents a direct regulator of GABAAR turnover. To investigate whether GABAAR α1 might be a putative substrate for Usp14, we examined whether this subunit could be ubiquitinated in cells. Thus, HEK293T cells were transfected with GFP-tagged GABAAR α1, GABAAR β3, HA-tagged ubiquitin and either Usp14 wildtype (wt) or a catalytic mutant of Usp14, respectively (Figure S4). Extracts of untransfected HEK293T cells served as controls. Upon immunoprecipitation, using anti-GFP antibodies, GABAAR α1-GFP was precipitated from extracts containing GABAAR α1-GFP (Figure S4A, lower panel). Upon the use of HA-antibodies ubiquitinated forms of GABAAR α1 could be detected in extracts from transfected but not untransfected HEK293T cells (Figure S4A, upper panel). The detection of ubiquitinated GABAAR α1 is in line with a recent publication that reported ubiquitinated GABAAR β subunits [17], suggesting that GABAARs in general are subject to ubiquitin conjugation. In particular the abundance of ubiquitinated GABAAR α1 forms between 75 and 100 kDa (Figure S4A, upper panel, asterisk) is slightly increased in the presence of the Usp14 catalytic mutant, represented by a more intensive blurred signal (see magnified image in Figure S4B). This observation suggests a stabilization of mono-/oligoubiquitinated GABAAR α1 polypeptides upon binding of a functionally inactive form of Usp14. Thus, Usp14 might represent a critical DUB to GABAAR α1. A balanced control of GABAAR ubiquitination and deubiquitination might therefore be an important determinant in regulating GABAAR surface expression in neurons. If the above interpretations were true, a minimal heterologous system should verify that Usp14 directly affects GABAAR turnover. In addition to the loss of Usp14 in mice, we aimed to proof, whether heterologous overexpression of an isolated Usp14 binding site (compare with Figure 4A) of GABAAR α1 would mimic the receptor surface distribution phenotype upon competitive interference with GABAAR α1–Usp14 binding. Thus, HEK293 cells were transfected with constructs encoding GABAAR α1-GFP, GABAAR β3 and the monomeric red fluorescent protein (mRFP)-tagged Usp14 binding site of GABAAR α1 (mRFP-GABAAR α1(334–346)) or with mRFP, respectively. Biotinylation of surface proteins, followed by immunoprecipitation, indeed revealed a 2.5-fold increase of GABAAR α1-GFP surface membrane levels in the presence of the competing peptide (Figure 6A and 6B), thereby leading to the same functional consequence, as observed in axJ mice. To verify both the expression and catalytical activity of Usp14 in HEK293 cells, we performed western blot analysis of protein extracts from kidney and cultured HEK293 cells, using a HA-tagged ubiquitin vinyl methyl ester (HAub-VME), active site probe [41],[42]. Western blot analysis using HA-specific antibodies confirmed that HEK293 cells express catalytically active Usp14 (Figure S3A). Hence, we conclude that the disruption of GABAAR α1-Usp14 binding, and consequently the gene expression knockdown of Usp14 in axJ mice, is directly causal for increased GABAAR surface membrane expression, known to result in enlarged IPSC amplitudes in vivo. Since Usp14 represents a protease, its enzymatic activity should also be critical in this respect. To test this, we transfected HEK293 cells with constructs encoding GABAAR α1-GFP, GABAAR β3 and a catalytically inactive Usp14 mutant or with mRFP, respectively. Overexpression of the loss-of-function mutant, notably resulted in 3-fold enrichment of GABAAR α1-GFP surface expression (Figure 6C and 6D), as compared to control experiments, indicating that a balanced turnover of α1-containing GABAARs indeed requires a catalytically intact Usp14 enzyme. In summary, these and other data in this study suggest that the function of Usp14 is directly involved in GABAAR turnover. Regulation of synaptic strength requires the precise control of neurotransmitter receptor numbers at synaptic sites. Our in vivo and in vitro observations in this study indicate the novel concept that DUB-dependent pathways regulate neurotransmitter receptor density and might participate in synaptic plasticity mechanisms. Since axJ mice represent mutants that are exclusively deficient for the proteasome-associated form of Usp14, with deubiquitinating enzymes playing an important role in the UPS, our data suggest a role of the proteasome in GABAAR turnover. It has been shown, that epidermal growth factor receptors (EGFRs), once activated, undergo ubiquitination and internalization from the plasma membrane [43]. Prior to their sorting into multivesicular bodies (MVBs), EGFRs require deubiquitination, a process that depends on proteasomal activity, although EGFRs, such as most transmembrane proteins, undergo degradation at lysosomes. Recent data in yeast further confirm a proteasomal contribution in similar processes, by showing that the proteasome-associated deubiquitinating enzyme Doa4 removes ubiquitin from cargo proteins prior to their entry into internal vesicles of MVBs [14],[44],[45]. Usp14, as it binds to the proteasome and mediates deubiquitination, is therefore a candidate factor to serve in similar pathways in neurons (Figure 6E). Consequently, Usp14 might represent the responsible DUB to trigger GABAAR transport from early endosomes into MVBs/lysosomes, although it might already interfere with GABAARs at the cell surface or right after internalization (Figure 6E). Interruption of this function, either by loss of Usp14 (axJ mice), or by interference with Usp14-GABAARs α1 binding (HEK293 cells), might either induce (i) backpropagation of disturbed endocytic pathways, leading to maintenance of the receptor at the cell surface or (ii) increased recycling of receptors back to the cell surface. Although an exact molecular mechanism remains to be elucidated, Usp14 represents a novel candidate to participate in the regulation of GABAAR turnover and synaptic plasticity at GABAergic synapses. Since axJ animals show further neuronal abnormalities, such as impaired synaptic transmission at neuromuscular junctions, their neurological phenotype might be due to a combination of deficits. However, we conclude that impaired GABAAR turnover in PCs due to the loss of Usp14 significantly contributes to the ataxic phenotype in axJ mice. This view is supported by previous studies, which report that mouse mutants with altered GABAAR densities or reduced GABAergic terminals in the cerebellum, also develop severe motor impairments [1],[25],[27],[46]. For instance the pcd mouse mutant, characterized by a complete loss of PCs, shows ataxia. PC degeneration in pcd mice results in a loss of inhibitory PC output and consequently a reduced inhibitory input to the vestibular nuclei, representing one of the direct PC target regions. In addition, altered Purkinje cell input leads to ataxia. Hence, mice deficient for the GABA transporter GAT1 develop severe ataxia due to disturbed GABA re-uptake. Consequently, this functional deficit leads to an increased GABAA receptor-mediated tonic conductance and prolonged IPSCs in both cerebellar granule and Purkinje cells [30]. It is therefore a possible scenario that motor impairments are closely linked to the level of inhibition within cerebellar circuits. The observed increase in GABAAR surface expression and IPSC amplitudes in axJ mice, as reported in this study, negatively affect PC functions and are likely to contribute to the ataxic symptoms. In parallel other yet unknown changes might be involved, too. Hence, in addition to GABAergic transmission, ubiquitin-mediated pathways might be putative targets for therapeutic treatment against certain forms of cerebellar ataxia. The entire cDNA sequence of GABAAR α1 was subcloned as an EcoRI/SalI fragment into the pEGFP-N vector (BD Biosciences). The cDNA of Usp14 was subcloned as a BamHI fragment into the pFLAG-CMV-2 vector (Sigma). HA-tagged Ubiquitin was subcloned as an EcoRI/XhoI fragment into the pcDNA3f vector (Invitrogen). Single mutations (Usp14: C114A; GABAAR α1 loop: introduction of stop codons to generate deletion mutants and the competitive peptide, respectively) or group mutations (Usp14: H434A-D450A) were introduced using the site-directed mutagenesis kit (Stratagene). Immunofluorescence: rabbit anti-GABAAR α1 (1∶250 Upstate); guinea pig anti-GABAAR α1 (1∶6000); rabbit anti-Usp14 [10] rabbit anti-Usp14 (138-R/PB+2, SM Wilson-lab.); mouse anti-Calbindin (1∶100, Sigma); mouse anti-synaptic vesicle (SV2, 1∶100, Hybridoma bank, University of Iowa); Secondary antibodies: CY3-, CY2- or CY5-conjugated donkey-anti rat, mouse, guinea pig or rabbit (all 1∶500, Dianova). Immunoprecipitation/Western blot analysis: rabbit anti-GABAAR α1 (1∶500 Upstate, 1∶1000 AbD serotec); guinea pig anti-GABAAR α1 (1∶6000 JM Fritschy-lab.); mouse-anti-Usp14 (IA4; 1∶1000, SM Wilson-lab.) mouse anti-pan Cadherin (1∶100, Abcam); mouse anti-N-Cadherin (1∶4000, Cell Signaling Technology); rabbit anti-actin (1∶2000, Sigma); mouse anti-HA (1∶1000, Sigma; 1∶1000, Santa Cruz Biotechnology); anti-GFP (1∶1000, Roche), mouse anti-rpt4 (1∶1000, Biomol) Secondary antibodies: HRP-conjugated goat-anti rabbit, guinea pig and mouse (all 1∶10.000, Dianova); HRP-conjugated protein A (1∶1000, KPL); biotinylated secondary antibodies (1∶1000, Vector laboratories), gold-labeled secondary antibodies. Preembedding immunocytochemistry: mice were anaesthetized and perfused with 4% PFA with 0.1% glutaraldehyde in PBS. Sagittal vibratome sections of the cerebellum were cut (60 µm). After washing in PBS, sections were treated with 0.3% H2O2 and 1% NaBH4 in PBS for 30 min. After rinsing in PBS, sections were incubated with 10% horse serum (HS) containing 0.2% BSA for 15 min and subsequently incubated over night with primary antibodies in PBS, containing 1% PS and 0.2% BSA (carrier). Sections were washed in PBS, incubated with biotinylated secondary antibody and diluted in carrier for 90 min. After rinsing, sections were incubated with ABC (Vector Labs) and diluted to a 1∶100 concentration in PBS for 90 min. Afterwards they were washed in PBS and further incubated in diaminobenzidine (DAB)-H202 solution (Sigma) for 10 min. Sections were then either mounted on glass coverslips (light microscopy) or postfixed with 1% OsO4, dehydrated in an ascending series of ethanol and embedded in Epon (Roth). Ultrathin sections were examined with a Zeiss EM 902. Immunocytochemistry of ultrathin frozen sections: mice were perfused and cerebellar sections of the cerebellum were cut (100–200 µm), as described above. Small blocks of cerebellar tissue containing all layers were immersed in 12% gelatin in PBS at 37°C for 15–30 min. Blocks were transferred into vials containing 2.3 M sucrose in PBS and incubated over night. Thereafter they were frozen on specimen holders in liquid nitrogen. Ultrathin sections were prepared at a Reichard Ultracut microtome, equipped with a cryochamber and placed on copper grids (Sciences Services). Single and double immunogold labeling was performed according to Slot and Geuze using secondary 10 nm large protein A gold to label rabbit primary antibodies and 6 nm large gold (Dianova) to label guinea pig primary antibodies [49]. Mice were anaesthetized and the cerebellar vermis prepared in ice-cold carboxygenated ACSF (NaCl 135 mM; KCl, 5 mM; CaCl2 2 mM; MgCl2 1 mM; glucose 10 mM; Na2HCO3, 30 mM; NaHPO4, 1.5 mM; pH 7.4 (bubbled with carbogen)). The tissue was cut into 200 µm sagittal sections (Microm HM 650V; Histoacryl glue, Braun), that were transferred to carboxygenated ACSF at 35°C for 20–30 min before being kept at RT (22–24°C), until further use. Slices were placed in a recording chamber (RC26GLP, Warner Instr.) under an upright microscope (BX51WI, Olympus). Individual PCs were visually identified and recorded with borosilicate capillaries of approximately 5 MO resistance (Hilgenberg) using the whole-cell patch-clamp configuration. Spontaneous synaptic events were recorded under equimolare Cl− concentrations at −60 mV and the GABAergic input isolated using the AMPA-type glutamate antagonist CNQX; the remaining IPSCs could be blocked by 20 µM bicuculline (Figure S2A). IPSC were recorded at 10 kHz for 60 s every 10 min. over approximately 1 h using the Patchmaster 2.05 software (HEKA). Data were analysed with the MiniAnalysis 6.02 program (Synaptosoft; converted with the supplied ABF Utility) using identical parameters for evaluating all IPSCs. Intracellular electrode: CsCl, 125 mM; MgCl2, 2 mM; EGTA 0.1 mM; TEA 5 mM; Na2-ATP, 4 mM; Na-GTP, 0.5 mM; HEPES, 10 mM; pH 7.3 (CsOH). HEK293 cells were harvested in 1% Triton X-100, 48 h after transfection. Antibodies were coupled to 30 µl of protein G beads (Dynal Biotech) in IP washing buffer (50 mM TrisHCl, 150 mM NaCl, 5 mM MgCl2, PH 7.1). Cell extracts were incubated with the beads over night, then washed and boiled in SDS sample buffer. For GST-pulldown experiments, HEK293 cells were harvested 48 h after transfection in 1 ml 1% Triton X-100. E. coli BL21 lysates were obtained by sonification and centrifugation at 10,000×g for 30 min. Bacterial lysates were coupled to glutathione-sepharose beads (Amersham) for 3 h. HEK293 cell lysates were applied to the beads for 10–12 h. Beads were washed and boiled prior to Western blot analysis. Proteins separated by SDS PAGE were transferred to PVDF or nitrocellulose membranes and unspecific binding sites were blocked using PBS containing 0.1% Tween and 5% skim milk powder. Primary and secondary antibody incubation was performed in blocking solution. HEK293T cells were transfected using GeneJuice Transfection Reagent (10 µg DNA/10 cm dish). 48–72 h after transfection, cells were lysed in lysis buffer (50 mM HEPES, 150 mM NaCl, 10% glycerol, 1 mM EGTA, 1 mM EDTA, 25 mM NaF, 10 µM ZnCl2 pH 7.5) supplemented with 10 mM NEM to inhibit deubiquitinating enzymes as well as protease and phosphatase inhibitors (10 µg/ml Aprotinin; 2 µg/ml Leupeptin; 1 mM PMSF; 1 mM Na-orthovanadate). Cell lysates were preincubated with 20–25 µl Protein A/G PLUS Agarose (Santa Cruz Biotechnology) to remove unspecifically bound proteins. Immunoprecipitation using anti-GFP antibodies (1.0–1.2 µg; Roche) was performed overnight at 4°C. Proteins bound to GFP-Antibodies were precipitated by adding 25 µl Protein A/G PLUS Agarose (Santa Cruz Biotechnology) followed by incubation for 45 min at 4°C. Precipitates were analysed by western blot analysis as described above considering the following differences. Unspecific binding sites were blocked using TBS (150 mM NaCl, 50 mM Tris, 0.1% Na-Azide, 0.5% v/v phenol red) containing 5% BSA. Primary antibodies were diluted in TBS/5% BSA. 48 h after transfection, HEK293 cells were incubated (20 min; 4°C) with HEPES containing 1 mM biotinamidohexanoic acid 3-sulfo-N-hydroxysuccinimid-ester sodium salt (Sigma). Remaining biotin reagent was quenched by adding 100 mM glycine (twice for 20 min at 4°C). Cells were washed with ice cold PBS and lysed in PBS containing 1% Triton X-100 and protease inhibitor cocktail (MiniComplete 1 tablet/10 ml Roche). After a 30 min incubation step on ice, followed by a brief centrifugation-step at 1,000×g (5 min, 4°C), 30 µl of the supernatants were loaded on a gel to evaluate the amount of GABAAR α1-GFP. After quantification adjusted volumes of supernatants were added to 30 µl of prewashed magnetic Streptavidin MyOne beads (Dynal) to achieve equal amounts of GABAAR α1-GFP used for precipitation. Beads were incubated at 4°C for 3 h on a rotation wheel, washed 3 times, collected and boiled in SDS sample buffer. For protein-protein interaction analysis, the Matchmaker LexA yeast Two-Hybrid system (Clontech, Heidelberg, Germany) was used. Interactions of bait (pGilda) and prey (pJG4-5) fusion proteins were examined by activation of a LEU2 and a lacZ reporter gene [47]. For detection of GABAAR α1 mRNAs, antisense oligonucleotides were synthesized encoding the large intracellular loop region between transmembrane domains M3 and M4 (aa 342-356) [51]. In situ hybridization was performed as previously described [51],[52].

10.1371/journal.pbio.1002549

Short Time-Scale Sensory Coding in S1 during Discrimination of Whisker Vibrotactile Sequences

Rodent whisker input consists of dense microvibration sequences that are often temporally integrated for perceptual discrimination. Whether primary somatosensory cortex (S1) participates in temporal integration is unknown. We trained rats to discriminate whisker impulse sequences that varied in single-impulse kinematics (5–20-ms time scale) and mean speed (150-ms time scale). Rats appeared to use the integrated feature, mean speed, to guide discrimination in this task, consistent with similar prior studies. Despite this, 52% of S1 units, including 73% of units in L4 and L2/3, encoded sequences at fast time scales (≤20 ms, mostly 5–10 ms), accurately reflecting single impulse kinematics. 17% of units, mostly in L5, showed weaker impulse responses and a slow firing rate increase during sequences. However, these units did not effectively integrate whisker impulses, but instead combined weak impulse responses with a distinct, slow signal correlated to behavioral choice. A neural decoder could identify sequences from fast unit spike trains and behavioral choice from slow units. Thus, S1 encoded fast time scale whisker input without substantial temporal integration across whisker impulses.

Sensory input is rich in temporal patterns, but how the brain processes this temporal information is not well understood. This process is important in the whisker tactile system of rodents, in which active whisking on objects generates dense streams of stick-slip and contact events. Rats can discriminate vibrotactile sequences applied to the whiskers, and prior studies show that this often involves behavioral integration over time to calculate mean whisker-speed. How the brain represents and integrates vibrotactile input is not known. We recorded neural activity in primary somatosensory cortex as rats discriminated rapid vibrotactile sequences. We found that neurons in the primary somatosensory cortex encoded whisker sensory information at very fast time scales (<20 ms), without evidence for substantial temporal integration. A subset of neurons encoded relatively little stimulus information but strongly encoded the rat’s behavioral choice on each trial. Thus, primary sensory cortex represents immediate sensory input, suggesting that temporal integration occurs in downstream brain areas.

Natural sensory input comprises dense temporal series of discrete events, which animals often temporally integrate to guide perceptual decisions. The temporal integration process has been studied in primate somatosensation and vision [1,2], but less in rodents, in which modern tools could reveal the underlying circuit mechanisms. In the whisker tactile system, active whisking generates dense streams of stick-slip events on surfaces (5–10 ms duration, ~60 ms interval) [3,4] and contact events on object edges [5,6]. These temporal series constitute the whisker vibrotactile signal. While animals can perceive individual brief whisker impulses alone or within trains [7–11], behavioral discrimination of vibrotactile sequences is often based on a time-averaged composite feature, mean whisker speed, rather than the kinematics or precise pattern of individual deflections [12,13]. This suggests that the brain generates both short time-scale (individual impulse) and temporally integrated, long time-scale (mean speed or intensity) representations of whisker input. How these time scales are represented in the cortex is unknown. We tested which time scale(s) of information are represented in S1 in awake behaving rats discriminating rapid whisker sequences. Under anesthesia, most S1 neurons spike phasically to whisker deflections, and responses adapt strongly during stimulus trains. This suggests that S1 does not temporally integrate across impulses (we use “integration” to mean temporal summation or averaging) [14–18]. Most S1 neurons also spike phasically to whisker deflection in basic detection tasks [7,9,10,19] or when rats must detect kinematically distinct impulses within ongoing stimulus trains [8]. However, these tasks do not require stimulus integration for behavioral performance [7–10]. Whether temporal integration occurs in S1 during tasks in which animals behaviorally integrate whisker information is unknown. A subset of S1 neurons exhibit sustained responses to stimulus sequences in awake mice [20], but whether these contribute to perceptual integration is unclear. We trained rats to discriminate rapid sequences of three brief whisker impulses with an ~60 ms interpulse interval. This interval matches the median interval between stick-slip events during texture palpation [21]. S1 is required for passive vibrotactile discrimination [13,22,23]. Stimuli differed in both rapid temporal structure (kinematics and order of individual impulses) and time-integrated information (mean speed of the entire sequence). Rats could use either for discrimination. Behavioral choice correlated with mean speed, suggesting that rats temporally integrated whisker impulse sequences, as shown explicitly in similar prior studies in which both rapid kinematic and slow intensity cues were available [12,13]. In tetrode recordings during behavior, most S1 units accurately encoded single-impulse kinematics on a rapid (≤20 ms) time scale with modest adaptation. A minority of units responded weakly to individual impulses but exhibited slowly increasing or decreasing spiking during the stimulus period. However, these units did not effectively integrate across impulses and instead combined transient impulse responses with a distinct, slow signal correlated to behavioral choice. Thus, S1 appears to represent only short time-scale information about whisker impulse trains during vibrotactile discrimination. This suggests that temporal integration may occur downstream of S1. We developed a novel whisker vibrotactile discrimination task in which rats initiated trials by entering a nose poke with their right whiskers resting on a wall panel coupled to a hidden piezoelectric actuator (Fig 1). The panel delivered a rapid sequence of three up-down impulses. Each impulse was 16–26 ms long and had Fast (F), Medium (M), or Slow (S) rise/fall velocity. Sequences had FFF, FMS, SMF, or SSS pulse order (34 ms interval from end of a pulse to beginning of next pulse; 120–148 ms sequence duration). Sequences were constructed so that mean speed was greatest for FFF, lowest for SSS, and equal and intermediate for FMS and SMF sequences (Fig 1; Table 1; S1 Fig). One sequence was delivered per trial, beginning 75–100 ms after nose poke entry. Rats had to maintain nose poke for 250 ms to ensure delivery of the entire sequence and then discriminate by selecting a right or left drink port for water reward. FFF and FMS sequences were rewarded right, and SMF and SSS were rewarded left. Training was conducted under infrared light, and sound cues from the piezo were masked. In a subset of trials (43 trials, 4 rats), we verified with high-speed video that whiskers remained on the panel throughout the stimulus period and that rats did not whisk while in the nose poke, as shown previously [22]. Head movement averaged 0.8 mm in right-left position and 1.0 mm in rostrocaudal position during the stimulus period. Rats initially trained on FFF versus SSS discrimination and then FMS and SMF stimuli were added (see Materials and Methods). These sequences differed in both rapid stimulus features, like identity of individual impulses, and slow features, like mean speed of the entire sequence. We designed the task so that fully correct discrimination is only possible if rats attend to fine time-scale information, like precise internal structure of the train (FFF or FMS indicates choose right, SMF or SSS indicates choose left), or identity of the first impulse (F indicates choose right, S indicates choose left). In contrast, if behavior is guided by mean speed (or duration) of the entire sequence, then rats should respond to FFF and SSS correctly but make mistakes in which they treat SMF and FMS identically and intermediate to FFF or SSS. Using a similar task design in which both rapid and slow, integrated cues were available, two prior studies found that rats choose to guide vibrotactile discrimination by the integrated variable, mean speed or intensity [12,13]. After 14.2 ± 4.4 (standard deviation [s.d.]) (range: 8–22) d of training on FFF-FMS-SMF-SSS discrimination, all eight rats successfully discriminated FFF from SSS stimuli, but failed to respond appropriately to FMS and SMF stimuli, instead treating them as equivalent and intermediate between FFF and SSS (Fig 2A and 2B). Seven out of eight rats failed to differentiate at all between FMS from SMF stimuli (proportion test, Bonferroni-adjusted p-value >0.00625). One rat (62SC) showed modest but significant discrimination, with more right-side choices to FMS than SMF stimuli (p = 0.0039). Behavior was stable, on average, across the training period (S2 Fig). Thus, seven out of eight rats showed behavior consistent with guiding decisions by time-integrated whisker information. To examine this further, we plotted the mean behavioral performance of each rat versus the mean speed of panel movement across the entire sequence (150 ms). Behavioral performance was computed as (fraction of right drink port choices for each stimulus)–(mean fraction of right drink port choices for all stimuli), to account for right-left choice bias by some rats (Fig 2B). Right drink port choice was strongly related to mean sequence speed for all rats (Fig 2C). To confirm that rats guided behavior by panel movement, we ran a “fixed panel” control in six rats, immediately after the final normal training session. The panel was fixed in place, while the piezo behind it moved normally. Panel fixation strongly impaired behavioral discrimination in all but one rat (example rat, Fig 2A; population data using d-prime analysis, Fig 2D; population data using a simpler non-parametric analysis, S2B Fig). Some residual discrimination did persist and may have been mediated by inadequately masked piezo sound cues. Further analysis showed that three rats treated the average fixed-panel stimulus similarly to SSS stimuli; one rat responded by choosing right or left randomly; and one rat stopped completing trials in the fixed-panel condition (S2C Fig). Thus, different rats had different strategies for handling the unfamiliar fixed panel trials. These results suggests that, as in prior studies [12,13], rats used slow, integrated information (mean speed or intensity) to guide discrimination, rather than rapid information (first or last impulse identity or impulse order). This may reflect either a predisposition for intensity cues, or task factors such as our use of strong intensity cues in initial training or the nose poke time requirement, which may have promoted an integration-based strategy. Rats are known to sense fast kinematic cues during ongoing sequences [7–11], and they can utilize these cues for discrimination in some cases [8]. We did not apply additional stimuli to further dissociate slow from rapid information (as was done in [12,13]), and thus we cannot independently rule out the possibility that rats guided behavior from a hidden fast cue (e.g., second impulse identity) that correlated with mean speed. To test whether failure to discriminate FMS versus SMF reflected insufficient training on these sequences or the presence of easier FFF and SSS stimuli on 50% of trials, we trained two rats on a modified task. This used a very simple task structure with only two stimuli: an FSFS sequence (rewarded at the right drink port) and an SFFS sequence (rewarded at the left drink port). F and S impulses had 216 and 120 mm/s peak velocity and 1.2 and 0.7 mm amplitude, respectively. Both trains had 34 ms interpulse interval and 188 ms total duration (Fig 3A). We constructed two sets of stimuli: a “same-intensity” version in which FSFS and SFFS trains had nearly identical mean speed (25.7 and 26.4 mm/s, calculated across the full sequence), and a “different-intensity” version in which FSFS and SFFS stimuli were scaled in amplitude so that mean speed was 27.8 and 8.7 mm/s, respectively. Two rats (58B and 60W) were initially trained to discriminate the different-intensity sequences (>65% correct over 3 d). Then, we replaced these stimuli with the same-intensity FSFS and SFFS sequences, so that discrimination could only occur by detecting differences in fine temporal structure, not mean speed. Performance dropped to chance and did not improve over 5 d of training (Fig 3B). We then alternated weekly training on different- and same-intensity sequences. Both rats consistently discriminated FSFS from SFFS when they had different mean speed (58B: 70 ± 1.5% correct, 60W: 69.2 ± 1.6%), but not when they had the same mean speed, even after >20 cumulative days of training (58B: 52 ± 0.8% correct; 60W: 53 ± 0.8% correct). This was evident in the d-prime measure of discrimination between FSFS and SFSF stimuli, which was 1.02 for different-intensity stimuli and 0.12 for same-intensity stimuli (Fig 3C). Thus, behavior correlated with the presence of a slow, integrated cue. To study S1 coding of whisker sequences during vibrotactile discrimination, we recorded S1 spiking during the FFF-FMS-SMF-SSS behavioral task using chronic multi-tetrode microdrives. Four tetrodes (~350 um lateral spacing) were driven as a group, enabling simultaneous recording of many neurons in several whisker-related columns (Fig 4A). Tetrodes were initially implanted into mid-L2/3 and advanced by ~140 μm every one to two recording sessions, sampling neurons from L3 to L6 over 12–22 d of recording. Spike sorting yielded 3.8 (range: 0–11) well-separated single units per recording session (Fig 4B). Additional units showed clear separation from noise but failed the interspike interval criterion for single units and were classified as multi-units. We obtained 306 single units and 167 multi-unit clusters (total: 473 units) across 80 recording sessions in five rats (18FB, 18Ri, 18Ro, 62BS, 107St), spanning across L3 to L6 (Fig 4C). Fast-spike (FS) and regular-spike (RS) units were well separated by spike width. Recordings were localized to C1-4, D2-4, and E3 columns based on receptive field mapping under light isoflurane anesthesia and recovery of marking lesions. These whiskers were visually confirmed to contact the panel, as in a prior study using this behavioral apparatus [22]. Mean firing rate during a 25-ms prestimulus baseline period in the nose poke was 6–10 Hz across layers for RS units, 8–32 Hz for FS units, and higher for multi-unit clusters (S1 Table). Lowest firing rates were observed in L2/3, L4, and L6. Firing rate distributions were positively skewed (S3A Fig). Firing rates for RS units were higher than in prior studies using cell-attached or whole-cell recording in rodents whisking mostly in air [6,24,25]. This likely reflects recording bias for more active units and the fact that whiskers contacted the stimulus panel through the entire nose poke duration, including the baseline period. We first identified units whose average firing rate was significantly temporally modulated with any dynamics during the nose poke period (p < 0.05, temporal modulation permutation test, see Materials and Methods). Three hundred five out of 473 units (63.5%) showed significant temporal modulation. Temporally modulated units were distributed uniformly across whisker columns and layers (Fig 4C) and had higher baseline firing rates than non-modulated units (S3B Fig). Subsequent analysis focused only on these temporally modulated (i.e., task-involved) units. Single- and multi-units showed similar response properties and were combined for analysis unless indicated. The average population response, compiled across all temporally responsive units in each layer, was dominated by a brief, phasic increase in firing rate following each panel impulse (Fig 4D). This was greatest in L2/3, L4, and L5b, and weakest in L5a and L6. The mean impulse-evoked firing rate modulation (in Hz above pre-impulse baseline) was 14.2 ± 2.3 in L2/3, 15.2 ± 1.9 in L4, 6.3 ± 1.2 in L5a, 14.4 ± 2.3 in L5b, and 7.0 ± 1.5 in L6 (n = 28–82 units per layer). Among units with significant impulse responses, peak response latency was shortest in L4, L5a, and L5b (9.8, 10.3, and 12.0 ms) and longest in L2/3 and L6 (13.8 and 16.1 ms). Superimposed on these phasic responses to individual impulses was a gradual decrease in average firing rate during the nose poke period, observed in all layers except L5a (Fig 4D). Individual units most commonly showed phasic responses to individual impulses (examples, Fig 5A and 5B). However, some units instead showed cumulatively increasing firing rate during the stimulus period (Fig 5C and 5D) or decreasing firing rate (not shown). These were intermixed in the same columns and recording sites. To quantify the time scales of stimulus representation in S1, we performed a multiple regression analysis for each temporally modulated unit (n = 305), whose goal was to identify the time window of stimulus integration that best predicted the neuron’s firing rate (Fig 6). The dependent variable was firing rate, in 5 ms bins, calculated over all trials for each stimulus sequence. The regressors were integrated speed of panel movement over a variety of temporal integration windows (5, 10, 15, … 180 ms, for a total of 36 regressions). Firing rate in each 5 ms bin was predicted from the integrated panel speed in the preceding bin. Two hundred four units showed a significant regression for at least one stimulus integration window (α = 0.05/36 = 0.0014, using Bonferroni correction for the multiple regressions). For each unit, we defined the best fit integration window as the stimulus integration window with the highest R2 value. Most units had a short best fit integration window (5–20 ms), indicating that firing rate was best predicted by stimulus speed on a short time scale (examples, Fig 6A and 6B). However, some units exhibited slowly increasing or decreasing firing that was correlated with integrated speed over long timescales, most often the whole stimulus period (example, Fig 6C). Individual cells had high R2 values for either short or long integration windows but rarely both (Fig 6D). Most units showed a positive regression slope for the best integration window, indicating that firing rate increased with integrated stimulus speed, while ~20% showed a negative slope (Fig 6D and 6E). Empirically, units with 5–20 ms best integration windows (Fast units; n = 158) had positive slopes. Units with 25–55 ms integration windows were rarer (Medium units; n = 52) and had largely negative slopes. Units with slow (55–180 ms) integration windows had either positive regression slope (Slow Positive units; n = 51) or negative regression slope (Slow Negative units; n = 44). Fast units were 73% of temporally modulated units in L2/3 and L4, 50% in L5, and 23% in L6. Overall, 52% of temporally modulated units were Fast units. Both Fast and Medium units were most prevalent in L2/3, L4, and L5b. In contrast, both Slow Positive and Slow Negative units were located primarily in L5 and L6 (Fig 6F and 6G). Overall, slow units were 13% of temporally responsive units in L2/3 and L4, 31% in L5 and 56% in L6. Fast, Medium, Slow Positive, and Slow Negative categories each contained both single- and multi-units and both RS and FS units. Fast time scale units showed temporally precise coding of individual panel impulses and sequences (Fig 7A–7C). Population peri-stimulus time histograms (PSTHs) for the fastest units (5 ms best integration window) showed responses to F impulses (16 ms duration) that lasted just ~20 ms and responses to S impulses that tracked impulse onset and offset separately. Units with 10 ms and 15–20 ms best integration windows had somewhat slower responses, as expected, but still tracked individual impulses. Adaptation within each train was quantified as mean firing rate to pulse N/pulse 1 and was modest in FFF trains (2/1: 0.80 ± 0.11, 3/1: 0.70 ± 0.14, p < 0.05 by t test, n = 61 single RS units with significant response to F impulses) and statistically absent in SSS trains (2/1: 1.09 ± 0.26, 3/1: 0.86 ± 0.35, all mean ± SEM) (Figs 7A and S4). This is less adaptation than reported for non-whisking, non-task-engaged rats [16,26] and is similar to passive whisker detection [10]. To determine whether Fast units accurately discriminate impulse velocity, we calculated the average response to all individual F, M, or S impulses (compiled across all sequences). The firing rate of Fast units (n = 158) in a brief window after each impulse was greater for F versus S impulses, and intermediate for M impulses (Fig 7C, left). Mean firing rate measured over the entire duration of a sequence (0–150 ms after sequence onset) varied closely with mean speed of the sequence, being highest for FFF, lowest for SSS, and intermediate and equal for FMS and SMF (Fig 7C, right). Thus, population average firing rate of Fast units over the entire sequence closely matched the mean behavioral performance of the animals (Fig 2C). In addition to coding pulse velocity, Fast unit coding was also influenced by pulse order because of adaptation. Fast RS single units (n = 61) showed greater adaptation during FFF than SSS sequences. Consistent with this, the middle M pulse in FMS sequences appeared weaker than in SMF sequences, though this did not achieve statistical significance (p = 0.08, paired t test, n = 61 units) (S4A Fig). Thus, Fast units represent impulse velocity, but with some history dependence due to adaptation, and no sign of positive temporal integration across impulses. In contrast, medium time scale units responded to impulses with a modest decrease in firing rate, rather than an increase, consistent with the negative regression slope for most of these cells (Figs 6E and 7D). In firing rate analysis, these cells were inhibited by F, M, and S impulses and did not distinguish either individual impulse identity or whole sequence identity (Fig 7D and 7E). Thus, medium time scale units do not represent stimulus information useful for this discrimination task. Slow positive units (n = 51) also showed a time-locked increase in firing rate after panel impulses, on average, but mostly to the second and third impulses in the sequence. Responses were small and sustained (unlike the large, transient responses by Fast units) and were evident for F and M impulses but not S impulses (Fig 8A). However, mean firing was not different for FFF, FMS, SMF, or SSS trains, suggesting that these neurons do not appreciably integrate impulse information for sequence discrimination (Fig 8A). Slow negative units did not respond to impulses at all, and firing rate steadily declined over time, not locked to panel impulses (Fig 8B). Unexpectedly, firing of Slow Positive units correlated with the animal’s behavioral choice on each trial. Fig 8C shows population PSTHs for Slow Positive units in L5a and L5b, divided into trials in which the rat chose the right- or left-side drink port. Slow Positive units fired more on trials when the rat chose right (contralateral to the S1 recording). This was true for both FFF and FMS stimuli, for which right was the correct response, and SMF and SSS stimuli, for which right was the incorrect response. We quantified right-choice bias as the firing rate difference on right versus left trials, measured 5–50 ms after the start of the final impulse. Right-choice bias was significant for Slow Positive units in L5a and L5b, but not other layers (Fig 8D). Firing rate began to diverge on right versus left choice trials after the second impulse and was consistently significant by 125 ms, which is during the third impulse (p < 0.05, sliding paired t test) (Fig 8E). This preceded the earliest withdrawals (150 ms) and mean withdrawal time (190 ms). Choice-related activity was absent in fast time scale units in L4 (Fig 8E). Thus, L5 Slow Positive units exhibited weak impulse-evoked spiking and strong choice-related spiking (Fig 8). We tested for stimulus integration in these units by comparing firing rate during each impulse of FFF, FMS, SMF, and SSS sequences on right- and left-choice trials separately, which removes choice as a factor (S5 Fig). Evoked firing was minimal for pulses 1 and 2 and was not correlated with pulse velocity. Pulse 3 firing rate was higher but was essentially identical for FFF, FMS, SMF, and SSS sequences and did not correlate with mean speed of the entire sequence or of the last two impulses. Thus, these units did not effectively summate stimulus information across impulses. We asked whether choice-related firing could reflect a feed-forward sensory reafferent signal generated by decision-related movements in the nose poke. Reafference from fast whisker deflections is unlikely, because L4 Fast units did not exhibit choice-related firing (Fig 8). However, a distinct slow reafferent signal is possible. We tested for choice-related postural movements by analyzing high-speed videos in 43 trials (22 left choice, 21 right choice) from four rats. In each trial, we tracked head position, head angle, and whisker tip position with ~100 μm precision at 8.4-ms intervals from 0 to 150 ms after stimulus onset. Head angle and whisker tip trajectories were invariant between right- and left-choice trials. Head position differed modestly between right- and left-choice trials beginning at 100 ms, with a 0.6 mm difference at 125 ms (S6 Fig). Thus, slow head movements are a potential reafferent driver of choice-related firing in L5. Fast, Medium, Slow Positive, and Slow Negative response classes all included RS, FS, and multi-unit clusters, although few FS cells were found in the Slow classes (S7 Fig). Among Fast units, all three unit types had similar sequence-related PSTHs. Among L5 Slow Positive units, both RS units and multi-unit clusters had similar choice-related firing, and no FS units existed in this category (S7 Fig). Thus, all response classes involved RS units. S1 neurons spike sparsely, with individual whisker deflections eliciting mostly zero spikes, occasionally one spike, and, very infrequently, two spikes on a single trial [21,27,28]. We also observed this highly variable, sparse single-trial spiking behavior (Fig 5). To test whether S1 accurately encodes whisker sequences on single trials, we constructed a neural population decoder that predicted stimulus identity from single-trial spike trains. In the model, each recorded neuron was represented by a separate, independent one-vs-all (OVA) classifier that predicted the probability of each sequence (FFF, FMS, SMF, or SSS) given one spike train, chosen randomly from that neuron’s recorded spike trains in vivo, and binned in discrete time bins. Each OVA classifier was trained by logistic regression from a randomly chosen subset of spike trains for that unit. The output of each classifier was the probability of each stimulus type versus all others, based on the presented spike train. To create a population prediction, stimulus probabilities were summed across units, and the sequence with highest summed probability was taken as the population stimulus prediction (Fig 9A). This model assumes independence between neurons and allows stimulus prediction by both firing rate and temporal information within spike trains. We first constructed a decoder from all Fast and Medium units, using 10 ms time bins. This model predicted sequence identity, using one single-trial spike train per model unit, with 83% overall accuracy (range: 74% for FMS to 88% for FFF spike trains). Chance performance is 25% (Fig 9B). The individual neurons with best stimulus prediction were those with 5–10 ms best integration windows (Fig 9C). Remarkably, this model identified SMF and FMS sequences with 78% accuracy, even though rats could not. A second decoder constructed of all Slow units, also using 10 ms bins, predicted sequence identity at near chance levels (32% correct, not significantly different from chance, p = 0.47) (Fig 9B). Decoding from mean firing rate in a single 150-ms bin substantially reduced Fast/Medium decoder accuracy (43% correct) and did not improve Slow decoder accuracy (Fig 9D). To test whether the Fast/Medium model recognized sequences by mean firing rate or temporal spike pattern, we rate-normalized the spike train data (preserving temporal information across the 10-ms bins) or time-scrambled spike trains within trials (preserving firing rate information). Fast/Medium decoders trained on rate-normalized data performed well (80% correct), but time-scrambling spikes abolished performance (Fig 9E). Thus, the Fast/Medium decoder primarily identified stimuli by temporal spike patterns, which varied between FFF, FMS, SMF, and SSS sequences (Fig 7). Thus, sequence identity was primarily encoded in short time-scale spiking information, carried by Fast units. We constructed a similar decoder to predict behavioral choice. This was trained on spike data from all four sequences and was tested for prediction of right versus left drink port choice separately for FFF, FMS, SMF, and SSS trials. A choice decoder based on Fast/Medium units was unable to predict drink port choice, either using 10 ms bins (not shown), mean firing rate in a single 150-ms bin, or mean firing rate in the last 100 ms prior to nose poke withdrawal (Fig 9F). A choice decoder based on Slow units successfully predicted drink port choice using a single 150-ms bin, or mean firing rate in the last 100 ms before nose poke withdrawal (65% correct for both models) (Fig 9F). Post-hoc analysis showed that units with best choice prediction were Slow Positive units located primarily in L5b (Fig 9G). Thus, spiking of Slow Positive units was sufficient to decode behavioral choice but not sequence identity. Cortical sensory systems temporally integrate sensory signals for many types of perceptual decision-making [2]. Where and how integration is performed is unclear. In fingertip vibrotactile discrimination by primates, S1 neurons spike to each rapid skin deflection, and this information is temporally integrated downstream of S1 to guide behavioral discrimination [1,29]. In the rodent whisker system, passive vibrotactile discrimination is often based on slow, time-integrated input [12,13], although rats are also capable of discrimination based on rapid kinematics [8]. Integration is also implicated in discrimination of surface texture (roughness), in which surface whisking generates temporally dense sequences of stick-slip whisker micromotions, whose mean statistics, including mean whisker speed, correlate with roughness [3,4,21,30–33]. S1 neurons spike phasically to stick/slip events and other features such as dynamic changes in whisker bend [3,21,34], and behavioral judgments of surface roughness correlate with mean firing rate and rate of synchronous spiking across S1 neurons [21,35,36]. Thus, roughness discrimination likely involves temporal integration of stick/slip events and S1 spike trains. Integration is useful because it reduces the complexity of the vibrotactile signal to a single scalar quantity of stimulus intensity. Intensity-based discrimination is common across modalities and is a defining feature of texture discrimination in people and non-human primates [37]. Integration is also evident in whisker-based object localization, in which S1 spikes are time-locked to object contact, but mice judge object location by behaviorally integrating spike counts over ~50 ms, rather than using precise timing [19]. In our task, rats were able to distinguish FFF versus SSS sequences that differed in mean speed, but not FMS versus SMF sequences that had the same mean speed, and choice behavior was strongly related to mean speed across the sequence (Fig 2). Similar performance was observed in the SFSF versus FSSF task (Fig 3). Task performance was relatively low (d-prime for FFF versus SSS: 0.5–1.5), as in a prior study [13], indicating the difficulty of these tasks. The results suggest that rats utilized slow, time-integrated information for task performance, even though simple, short time-scale cues (e.g., identity of the first impulse) would have led to more rewards. This hypothesis is consistent with two prior vibrotactile discrimination studies using a similar design, in which rapid kinematics and slow intensity cues were manipulated separately to prove that rats guided discrimination by slow, time-integrated cues [12,13]. We did not test this causally in our study, so we cannot rule out that rats may have solved our task using a hidden short time scale cue. Integration is not required for simpler detection tasks [7,9,10] or detection-of-change tasks [8], and rodents can perceive single brief whisker impulses within ongoing deflection trains [7–11,38]. This suggests that rats generate neural codes for both rapid and integrated features that guide different aspects of sensory-guided behavior. Rats may differentially use these codes depending on task demands and training strategies. In our task, initial training involved strong intensity cues, which may have promoted adoption of an integration-based strategy. An intensity-like feature of vibrotactile stimuli is encoded in primate dorsolateral prefrontal cortex during a working memory task [39], but no explicit intensity representation is known yet in the rodent whisker system. We tested for stimulus integration in S1 during task performance but found that S1 encoded whisker sequences almost exclusively at very rapid time scales. Forty-four percent and 52% of temporally responsive units showed very fast (5–10 ms) and fast (5–20 ms) stimulus integration, respectively (Fig 6E). These units spiked to individual whisker impulses, with firing rate encoding impulse velocity, and mean firing rate correlated with mean whisker speed across the sequence (Fig 7A–7C). Seventeen percent of units showed firing rate modulations on medium (25–55 ms) time scales, but these were inhibited by whisker impulses and did not discriminate different impulses or sequences (Fig 7D and 7E). Sequence identity could be decoded accurately from Fast units but not Medium units, and stimulus information was abolished by scrambling spike times across 10-ms bins. Thus, Fast units encode sequence identity by representing the velocity and timing of individual impulses. Fast units accurately distinguished FMS from SMF sequences, even though rats could not (Fig 9B). Thus, accurate short time-scale representation of vibrotactile sequences exists in S1 but does not appear to be used efficiently to guide behavior in our task. This is identical to primate S1, in which precise spike timing discriminates vibrotactile flutter more accurately than the animal [40]. Fast units had phasic whisker responses similar to classic anesthetized studies [14,41] and S1 units recorded during detection tasks [7,8,11]. Responses were weak in L5a and L6 (Fig 4), which may reflect involvement of this layer in active whisking, which was absent in our task [42]. Adaptation was minimal: ~25% for FFF trains and absent for SSS trains (Fig 7A–7C). This level of adaptation is less than occurs under anesthesia [15,18] or in quiescent, non-task engaged rats [16,26] and is similar to that during active exploration [16,26] or in a whisker detection task [10]. While adaptation generates history dependence and thus carries information about prior impulses [43,44], Fast units showed no evidence of positive integration across impulses. Seventeen percent of units, primarily in deep layers, were Slow Positive units with small, sustained responses to individual whisker impulses and progressively increasing firing rate during the stimulus period. However, these units did not accurately encode or integrate whisker impulses. Responses were generally absent to the first impulse of sequences, and firing rate did not differ between FFF, FMS, SMF, and SSS sequences or correlate with mean speed (Figs 8A and S5). Thus, Slow Positive units do not appear to carry integrated stimulus information for sequence discrimination. Slow Negative units had slowly decreasing firing rate and no stimulus-related firing modulation at all (Fig 8). Consistent with these observations, sequence identity could not be decoded from Slow unit spike trains (Fig 9B). Slow whisker-evoked spiking occurs in some L2/3 units in mice [20] but was not evident in our dataset in rats. Instead, firing of Slow Positive units in L5 was strongly related to drink port choice. Choice-related spiking [45] occurs in many cortical areas, including primary visual cortex [46], S1 of primates and rodents [11,47–49], and even subcortically [49,50]. In rodent S1, many L2/3 neurons exhibit choice-related spiking in near-threshold detection tasks [11,49]. Choice-related firing emerged significantly after the second impulse of the sequence and was consistent during the third impulse, 65 ms before the average nose poke withdrawal (Fig 8E). A neural decoder built from Slow unit spike trains predicted behavioral choice from mean firing rate in the stimulus period and in the 100 ms prior to nose poke withdrawal (Fig 9). Choice-related firing was absent in L4 Fast units, suggesting it did not represent reafference from fast whisker sensory signals (Fig 8E). Choice-related spiking could reflect reafference from slow head movements prior to nose poke withdrawal, potentially mediated by POm afferents to L5 [51] or an internal decision or motor preparatory signal. Its onset after the second impulse could reflect an early behavioral decision based on first and second impulse stimulus information or an early stimulus-independent “"guess” that biased subsequent stimulus-dependent drink port choice. Thus, Slow Positive units do not appear to integrate across whisker impulses but combine weak impulse responses with a distinct, slow signal related to behavioral choice. We found that during vibrotactile discrimination, most S1 neurons represent the velocity and timing of individual whisker impulses at rapid, 5–20 ms time scales. While there was some history dependence of whisker responses due to modest adaptation, we did not observe evidence of positive integration across whisker impulses in S1 firing rates. Thus, temporal integration for discrimination is likely to occur downstream of S1, in higher sensory or premotor regions. These may include S2, prefrontal cortex, and premotor cortex, as in primate vibrotactile discrimination [1]. We cannot rule out that S1 could learn to temporally integrate under conditions in which rats were more reliant on slow cues for behavioral discrimination. For whisker texture perception, our finding of short time scale coding in S1 suggests that S1 primarily encodes low-level kinematics of individual stick/slips and bends [6,21], which are integrated downstream to represent texture or other surface features. Female Long-Evans rats were >3 mo of age. All procedures were approved by the UC Berkeley Animal Care and Use Committee (protocol R309-0516BC) and comply with NIH guidelines. The computer-automated chamber contained a nose poke, flanked by a wall-mounted whisker stimulus panel (2 x 2 cm) that was carried on a hidden piezoelectric actuator (Piezo Systems PSI-5H4E). Whiskers were trimmed to 15 mm in length. The right-side C, D, and E row whisker tips rested against the panel while the rat was in the nose poke (Fig 1A). Nearby right and left drink ports contained infrared-LED beam sensors to detect nose entry and delivered calibrated water rewards. Trials were monitored by infrared video. Each trial was self-initiated by entry into the nose poke. After a variable delay (75–100 ms), a sequence of three rapid whisker deflections was delivered via the panel. The rat was required to remain in the nose poke for 250 ms to ensure full sequence delivery. The rat then withdrew from the nose poke and was rewarded (0.05–0.1 mL water) for choosing the drink port that was associated with the presented stimulus. Incorrect drink port choice or premature nose poke withdrawal triggered a time-out tone (4–6 s) and no reward. In a subset of sessions, high-speed video (119 Hz) was recorded. In this task, each whisker sequence consisted of four pulses. Two pulses were low-amplitude, slow pulses (S) that were 0.7 mm amplitude, 120 mm/sec peak velocity, 12.5 ms rise and fall time, and 25 ms total duration. Two were higher-amplitude, fast pulses (F) that were 1.2 mm amplitude, 216 mm/sec peak velocity, 9 ms rise and fall time, and 18 ms total duration. Trains of F-S-F-S or F-S-S-F pulses were presented (34 ms inter-pulse interval, total train duration 188 ms). In the “same-intensity” stimulus set, both FSFS and SFFS trains had identical pulse amplitude and, therefore, mean speed (mean speed 25.7 mm/sec for FSFS, and 26.4 mm/sec for SFFS). In the “different intensity” stimulus set, FSFS stimulus amplitude (and velocity) was increased to achieve a mean speed of 27.8 mm/sec, and SFFS stimulus amplitude (and velocity) was decreased to achieve a mean speed of 8.7 mm/sec. Training was performed in identical steps as above, using the “different-intensity” stimuli at the second training stage. No recordings were performed. Recordings were made with an array of four tetrodes carried in a custom 3D-printed chronic microdrive. Tetrodes (12.5 μm nichrome wire, gold plated to 0.2–0.3 MΩ impedance) were spaced 0.35 mm apart in a square configuration and moved together as a single bundle along a radial penetration. The tetrode drive was mounted in a surgical procedure under initial ketamine-xylazine anesthesia (90 mg/kg and 10 mg/kg), maintained by transition to 0.5%–3% isoflurane. A 4-mm craniotomy was opened over S1 (5.5 mm lateral, 2.5 mm caudal to bregma), the dura was removed, and the microdrive was positioned over the durotomy. The tetrodes were lowered into L2 of S1 and the microdrive was mounted with dental cement, sealing the craniotomy. Reference and ground electrodes were mounted in the skull. Postoperative analgesia was provided with Buprenorphine (0.05 mg/kg every 8 h) for 1–2 d post-surgery. Animals recovered 5–10 d prior to behavioral and recording sessions. Recordings were made during one to two behavioral sessions per day for each rat. Tetrode signals were amplified and filtered (Plexon, 100x gain, 0.3–8 kHz bandpass filter) and digitized at 32 kHz, using methods as in [21]. Neural data was acquired continuously. Tetrodes were advanced a half-turn (140 μm) every one to two recording sessions, at least 30 min before recording started. A new set of units was sampled in every session. If new units appeared spontaneously overnight, the tetrode was not advanced. Recording ended when the tetrode entered the white matter, as judged by absence of spiking activity when advancing the drive. Twelve to 22 d of recording were performed per animal. Recordings were made in C1–4, D2–4, and E3 whisker columns, as determined by hand mapping under isoflurane anesthesia prior to the recording sessions. An electrolytic lesion was made at the final recording location to determine recording depth. Lesions were recovered in cytochrome oxidase-stained histological sections (100 μm thick) cut in the “across-row” plane, 45° coronal to the midsagittal plane [52,53]. This allowed the whisker row identity (A–E) of the recorded column and laminar identity of recording sites to be confirmed. Laminar boundaries were determined by aligning lesions with layer-specific CO staining boundaries, and were as follows: L2/3: 200–650 μm, L4: 650–975 μm; L5A: 975–1285 μm; L5B: 1285–1575 μm; L6: 1575–2200 μm. Single units were isolated offline using Wave_clus in Matlab [54]. After an initial automated clustering step, manual evaluation of all clusters was performed and manual changes to the clustering were carried out as needed. Single units were required to meet an interspike interval criterion (<0.5% of intervals less than 1.5 ms) and a signal-to-noise (STN) criterion for spike height (STN>2, with STN defined as the difference from trough to peak in the mean waveform divided by the average standard deviation across all samples in the waveform). Fast-spiking and regular-spiking units were classified by spike width, which was bimodally distributed. Fast spiking units had width <0.375 ms trough-peak delay. Neural data were analyzed for five rats, including one rat for whom the fixed-panel control task showed substantial task performance in the absence of panel movement (filled circles in Fig 2D). This rat’s data were included because panel-evoked responses, stimulus decoding, and choice decoding did not differ from other rats (not shown). A neural decoder was constructed to predict stimulus identity (FFF, FMS, SMF, SSS) from single-trial spike trains of the recorded units. Each unit was represented by a one-versus-all (OVA) classifier that was trained by logistic regression to report the probability of each stimulus given a single-trial spike train (0–150 ms after stimulus onset, binned using either 10 ms bins or a single fixed time bin), selected randomly from recorded spike trains for that unit. Each classifier comprised four logistic functions, one for each stimulus. Logistic functions were fit using logistic regression and k-fold cross-validation and were specified by coefficients (one for each time bin, plus a bias term) that relate spike rate in each time bin to the probability of stimulus s being delivered. Model fitting was performed using a randomly chosen subset of the recorded trials (70%), and decoder performance was assessed on the remaining trials. The output of each unit classifier was normalized so that each unit had the same weight in population decoding. The population stimulus prediction sp was calculated by summing the probabilities of each stimulus over all units and selecting the stimulus with the maximal summed probability. Model fitting and population decoding were repeated 300 times, and average performance is reported. This framework is equivalent to determining sp as the stimulus that maximizes the conditional probability of the four stimuli given the neural population response, assuming that all single units are independent and the prior distribution of s is uniform. Rate-normalized and time-scrambled spike trains were generated by dividing each spike train by its -Euclidean norm and shuffling spike times within trials, respectively. A separate behavioral choice decoder was constructed similarly and was used for predicting right or left drink port choice on a given trial. Since this is a binary decision, a single logistic function was fit for each unit. The model was fit using spike train and behavioral choice data from all four stimuli. Decoder performance was assessed separately for FFF, FMS, SMF, or SSS stimulus trials in order to dissociate stimulus identity from the rat’s behavioral choice. The population choice prediction cp was selected as the choice with maximal summed probability across all units, given single-trial spike trains from trials with the chosen stimulus type. Model fitting and decoding procedures were the same as above. All decoding analysis was performed using Python and the scikit-learn machine learning toolbox [56].

10.1371/journal.pgen.1006152

A High Temperature-Dependent Mitochondrial Lipase EXTRA GLUME1 Promotes Floral Phenotypic Robustness against Temperature Fluctuation in Rice (Oryza sativa L.)

The sessile plants have evolved diverse intrinsic mechanisms to control their proper development under variable environments. In contrast to plastic vegetative development, reproductive traits like floral identity often show phenotypic robustness against environmental variations. However, it remains obscure about the molecular basis of this phenotypic robustness. In this study, we found that eg1 (extra glume1) mutants of rice (Oryza savita L.) showed floral phenotypic variations in different growth locations resulting in a breakdown of floral identity robustness. Physiological and biochemical analyses showed that EG1 encodes a predominantly mitochondria-localized functional lipase and functions in a high temperature-dependent manner. Furthermore, we found that numerous environmentally responsive genes including many floral identity genes are transcriptionally repressed in eg1 mutants and OsMADS1, OsMADS6 and OsG1 genetically act downstream of EG1 to maintain floral robustness. Collectively, our results demonstrate that EG1 promotes floral robustness against temperature fluctuation by safeguarding the expression of floral identify genes through a high temperature-dependent mitochondrial lipid pathway and uncovers a novel mechanistic insight into floral developmental control.

Various mechanisms have evolved to ensure proper organ formation under variable environments in order to complete one organism’s life cycle. In angiosperms, vegetative and reproductive organs show a differential plastic development between varied environments, with a low plasticity or high robustness for flower formation, but little is known about its intrinsic mechanism. Here we report that gene EG1 (EXTRA GLUME1) can enhance the floral robustness against temperature fluctuation in rice. EG1 encodes a predominantly mitochondria-localized functional lipase and its loss of function disrupts floral development in a high temperature-dependent manner. In consistent, both EG1 and its lipase activity are positively induced by high temperature. Transcriptomic and genetic analyses revealed that EG1 functions upstream of several floral identity genes, eg, OsMADS1, OsMADS6 and OsG1. Taken together, our results uncover a novel mitochondria-mediated lipid metabolic pathway to promote floral developmental robustness. Our findings may help to genetically improve floral traits of rice to maintain a stable yield when planted in different locations and/or under heat stress conditions.

The sessile plants have evolved various exquisite adaptive strategies to cope with environmental changes [1,2]. Among them, phenotypic plasticity is the ability of a single genotype capable of producing different phenotypes in response to varying environments [3–6]. For an integral high fitness, morphologies of vegetative organs of a single plant, such as roots, leaves and stems, require a high phenotypic plasticity [7–10], whereas that of reproductive organs, such as flowers, fruits and seeds, are always associated with low plasticity also known as phenotypic robustness/stability [11–15]. Thus, plants must coordinate the developments of these organs. Compared with progresses in understanding the molecular mechanisms of high phenotypic plasticity [10,16–19], very little is known about the molecular basis of phenotypic robustness [20,21]. Recent studies have shown that there are a group of specific genes regulating the degree of phenotypic plasticity and determining the reaction norm of a trait among various environments, which are termed plasticity genes [22–24]. However, most of the identified plasticity genes are high plasticity-associated [16,25–27], only few promote phenotypic robustness [28–30]. Members of HSP90 (HEAT SHOCK PROTEIN 90) family, as central hubs of numerous biological pathways, are required for maintenance of phenotypic robustness in both animals and plants [28,31–34]. MSH1 (MutS HOMOLOG1), a homolog of bacterial mismatch repair protein MutS, has been reported to repress the developmental plasticity of plant architecture, leaf morphology and flowering time in several dicot and monocot plants [29,35]. A nuclear protein RPL1 (RICE PLASTICITY1) in rice also appears to promote the relatively stable plant architecture and panicle structure between different environments [30]. Despite these discoveries, we still know very little about the molecular mechanisms of phenotypic robustness, especially that of plant reproductive traits. In addition to the known epigenetics-dependent transcriptional regulation and hormone signaling [20,29–31,35], lipid homeostasis is also known to influence phenotypic robustness [36,37]. Coordinated regulations of cellular lipid homeostasis are crucial to organisms’ adaptive robustness under severe temperatures [37–40]. Furthermore, lipid-related synthetases and lipases can also be regulated at transcriptional and posttranslational levels to influence lipid homeostasis [41,42]. Among them, mitochondria-associated lipid metabolism is key to the lipid homeostasis [43]. For instance, Arabidopsis seedlings with decreased cardiolipin in mitochondrial membrane are easier to turn yellow and necrotic under extended darkness or heat due to a failure of mitochondrial morphogenesis, showing a lowered stability [41,44,45]. However, it remains unclear whether mitochondria also mediate the phenotypic robustness in plant reproductive organs. Flower morphology, as a gold standard in plant taxonomy, has the most remarkable robustness within and between individuals of the same population [11,46], making it an ideal trait for studies on the molecular basis of phenotypic robustness against environmental fluctuation. Nevertheless, so far no gene has been identified to regulate the phenotypic robustness of floral identity, although several environment-dependent floral mutants have been reported [47–52]. We previously found that a rice floral mutant eg1 (extra glume1) exhibited a floral variation in different growth conditions [53], implying that EG1 is likely involved in floral robustness. Recent studies have shown that EG1 encodes a putative lipase regulating rice floral identity and meristem determinacy [53]. It also functions in JA (jasmonic acid) biosynthesis to promote the expression of floral identity gene OsMADS1 through an EG2/OsJAZ1, OsCOIb and OsMYC2- mediated JA signaling pathway [54], similar to its homologous genes AtDAD1 (DEFECTIVE IN ANTHER DEHISCENCE1) and AtDGL (DONGLE) in Arabidopsis [55,56]. In this study, we find that EG1 is a predominantly mitochondria-localized functional lipase and promotes floral robustness against temperature fluctuation in a high temperature-dependent manner. Collectively, our results reveal a novel molecular mechanism underlying floral phenotypic robustness. Previously, we found that eg1 displayed a floral identity variation possibly influenced by growth conditions [53]. To examine if this variability was mainly due to the environmental alterations, we analyzed the spikelet phenotypes of eg1-1 (in indica ZF802 background) and eg1-2 (in japonica ZH11 background) in two groups of separate environments (Fig 1) and found that the floral phenotypic variability of eg1 is likely caused by both genotype and environment. To define the phenotypic variability, we divided the spikelet phenotypes of eg1 into six groups, which were called variable phenotypes, including Wl (WT-like), eg (extra glume), pl (palea to lemma), sp (smaller pa), le (long empty glumes) and rs (reiterated spikelets) (Fig 1A, S1 Fig and S1 Table). The results showed that sp and rs of eg1-1 as well as Wl, le, eg and sp of eg1-2 exhibited significant plasticity between two environments, especially le of eg1-2, which displayed the highest plasticity (Fig 1B and 1C), suggesting that environment also contributes to the variations of floral phenotypes of eg1. To further examine the relationships of genotype (G), environment (E), genotype-environment interaction (GxE) and the variable phenotypes, we calculated their effects on phenotypes by two-factor ANOVA and found that Wl and pl were affected mainly by G, eg by both E and G but rarely by GxE, sp, le and rs by all three factors, and among them, le could serve as a marker for the phenotypic plasticity of eg1-2 due to its large proportion in a panicle and opposite phenotypes between two environments (Fig 1A). Taken together, these results showed that eg1 shows higher floral plasticity, suggesting that EG1 promotes the floral robustness in rice. To further examine the influence of genotype on the floral plasticity of eg1, we swapped the genetic backgrounds of two eg1 alleles. eg1-1 in a largely japonica background showed high phenotypic plasticity especially for le and rs phenotypes, similar to eg1-2 (ZH11), whereas eg1-2 in an indica-dominant background showed low floral plasticity similar to eg1-1 (ZF802) (Fig 2A), indicating that genetic backgrounds also influence the phenotypic plasticity of eg1 spikelets. To verify this finding, we further used CRISPR/Cas9 technology to construct eg1 alleles in Nipponbare (japonica) and Dular (indica) backgrounds. Two types of spikelet phenotypes were found in eg1-4 allele with Dular background and both showed low plasticity (Fig 2B), while eg1-5 and -6 alleles in Nipponbare background showed relatively higher plasticity than alleles in two indica backgrounds ZF802 and Dular, similar to that in japonica ZH11 (Fig 2C and S2 Fig), suggesting that the floral plasticity of eg1 alleles in japonica backgrounds tend to be higher than that in indica backgrounds. In another aspect, eg1 alleles in indica backgrounds had severer floral disturbance than that in japonica concerning Wl and rs phenotypes (Fig 2 and S1 Fig), suggesting that EG1 has functions in both floral robustness and identity, which are differentiated in two subspecies. To explore the possible causes of these differentiation, we compared the cis-elements and expressional level of EG1 in several japonica and indica varieties and discovered the correlative differences in both cis-elements and expressional levels between japonica and indica (two types) varieties (S3 Fig and S2 Table), which implied that transcriptional differences may be a crucial cause of functional differentiation of EG1 in subspecies. All these results indicated that both eg1 allelic variations and their genetic backgrounds regulate the floral plasticity of eg1. In order to find out the environmental factors mediating the plastic development of eg1 spikelets, we first compared the growing conditions for phenotypic analysis and found a marked difference in daily high temperatures of two environments (S4 Fig), suggesting that the temperature variation between two environments could be a major environmental factor influencing the eg1 plasticity. To verify this prediction, floral plasticity of wild-types and eg1 alleles were examined in two artificial growth chambers with 35°C light 12 hr / 20°C dark 12 hr and 25°C light 12 hr / 20°C dark 12 hr respectively, while other growth conditions were kept identical. The low plasticity of eg1-1 and nearly 70% le phenotypes of eg1-2 showed that the floral plasticity of eg1 in the chambers was similar to and even higher than that under natural growth conditions (Fig 3). These results showed that temperature is a major environmental factor mediating the floral plasticity of eg1. Previously, EG1 was shown to be localized in chloroplasts in transient expression assays [54]. However, EG1-like lipases appear to have variable subcellular locations [56–59]. To examine the subcellular localization of EG1 in vivo, two different EG1 and GFP fusion proteins driven by 35S promoter were first expressed in rice protoplasts and were found to be co-localized with both mitochondrial specific dye Mito Tracker Red and mitochondrial maker protein MTS-mOrange [60] but hardly with chloroplast auto-fluorescence, and an EG1-GFP fusion protein driven by native promoter was also detected in mitochondria (Fig 4A and 4B), suggesting that EG1 protein is mainly, if not all, localized in mitochondria. To compare this finding with the previous one, the reported vector pCAMBIA1301-Pro35S:EG1-GFP [54] was also examined in our transient system and a similar localization was detected (S5 Fig). To further verify the EG1 localization, subcellular fractionations of one-week seedlings of eg1-2 complementation lines with FLAG-EG1 (S6 Fig) were successively carried out and the EG1 fusion protein was predominately co-fractionated with mitochondria (Fig 4D), confirming the mitochondrial localization of EG1 in vivo. Taken together, these results showed that EG1 encodes a protein predominately localized in mitochondria. To explore the biochemical function of EG1, we tested its lipase activity [53] in vitro and found that both the full-length EG1 and truncated EG1 without predicted targeting peptides showed significant lipase activity (Fig 4E), indicating that EG1 encodes a functional lipase. Taken together, our results showed that EG1 functions as a predominately mitochondria-localized lipase. The dependence of the eg1 floral plasticity on environmental temperature raised a possibility that either EG1 or its product or both are likewise regulated by temperature. To examine these possibilities, some heat/cold responsive cis-elements were discovered in the 2 kb genomic sequence upstream of the start codon of EG1 (S3 Table), implying that its expression could be induced by extreme temperatures. To examine this possibility, one-week wild-type seedlings were treated under different temperatures and the EG1 transcript was found to accumulate gradually, to an extremely high extent under heat shock (42°C) as well as usual high temperature 35°C for rice (Fig 5A), but to some extent suppressed under cold stress (4°C) (S7A Fig), indicating the high temperature-induced expression of EG1. A similar result was obtained by using young inflorescences in which EG1 has a high expression (S7B Fig). To examine whether EG1 protein was also influenced by high temperatures, accumulation of FLAG-EG1 fusion protein in eg1-2 complementation lines, with a temperature-insensitive promoter (S7C and S7E Fig), was detected under different temperatures and found it was significantly induced at extreme high temperature 42°C than 25°C and 35°C (Fig 5B and S7D Fig), indicating a stabilization of EG1 protein under heat stress. Furthermore, we detected that lipase activity of EG1 fusion proteins increase as temperature rising (Fig 5C), consistent with the assumption of EG1’s function required under high temperatures. Additionally, we also examined the effect of high temperatures on EG1 subcellular localization, and found no obvious translocation in protoplast system (S8A Fig), while failed to detect EG1 protein in the subcellular fractions of complementation lines except under heat stress for its minute amount (S8B Fig), suggesting that temperature does not significantly influence the subcellular localization of EG1. The increased the transcriptional level, protein stability and lipase activity of EG1 under high temperatures, implying its more significant role under high temperatures. To verify this hypothesis, we observed the floral phenotypes of eg1 mutants under extremely high temperatures and found much severer spikelets in eg1 mutants especially in eg1-2, with multilayer lemma-like organs and undetermined inflorescence meristem, which have never been found in other temperature conditions (Fig 5D), showing a more significant function of EG1 at higher temperatures in floral robustness. eg1 was also found to grow faster than wild-type during primary growing days [54], and we detected this phenotype was much severer under extremely high temperature than others compared with wild-type, which was consistent with the floral phenotype (S9 Fig). Therefore, we concluded that EG1 functions in a high temperature-dependent manner to regulate the floral robustness. Since the floral plasticity of eg1 was influenced by both genotype and environment, to examine the molecular mechanism of genotype-environment interaction in floral plasticity, transcriptomes of inflorescences of two eg1 alleles (eg1-1 and eg1-2) and their wild-types (ZF802 and ZH11) in Beijing and Lingshui were analyzed. First, to evaluate the reliability of the transcriptomic data, we divided all transcripts into 33 modules by co-expression network analysis and analyzed their correlations with six variable phenotypes, and it turned out that the relationships among the variable phenotypes derived from these correlations were quite similar to their morphological correlations (S10 Fig), showing a good reliability of the transcriptomic data. Second, through overall comparisons of all transcriptomes, we discovered that the expression patterns of floral transcriptomes of eg1 alleles between two environments were significantly different from their wild-types (S11A Fig), indicating a role of EG1 in regulating expressions of environmentally responsive genes. The numbers of environmentally responsive genes in eg1 mutants were much larger than that of wild-types, in contrast to the similar numbers between two wild-types or two eg1 alleles (Fig 6A and S11B Fig), implying that EG1 negatively regulates the responses of its downstream genes to environment. To verify this finding, we analyzed the effects of G, E, and GxE on transcriptomes of eg1 and wild-type by two-way ANOVA and found that the number of genes significantly affected by E and GxE in eg1 were significantly larger than that in wild-types (Fig 6B, S11C Fig and S4 Table), displaying a switch of many genes from G-affected to E/GxE-affected ones (Fig 6C), indicating that EG1 represses its downstream genes not only to respond to, but also to interact with environment. Furthermore, the effects of the three factors on several important pathways varied significantly between eg1 and wild-type, including the pathways related to temperature response, lipid metabolism and floral development (Fig 6D and S5 Table), indicating that EG1 mediates a crosstalk of these pathways with environment. Since EG1 was reported to regulate JA biosynthesis [54], in order to analyze its effect on the floral robustness control, we examined the expressional patterns of JA biosynthesis and signaling associated genes in our transcriptome data, and found that the transcriptional responsive patterns to environment or transcriptional level of several JA signaling genes (JAZ7 and JAZ8) and JA biosynthesis genes (four methyltransferase genes) are varied in eg1 mutants (S12 Fig), implying a possible role of JA in the EG1-associated floral robustness regulation but different from previously reported [54]. Taken together, these results revealed that EG1 mediates the transcriptional responses of downstream genes and pathways to environmental fluctuation. The significant transcriptional effects on the floral development pathways based on the G, E and GxE analysis in eg1 mutant suggested that floral identity genes are likely involved in EG1-dependent floral robustness regulation. To examine this, expressional variation of thirteen floral identity genes between two environments were further analyzed, and seven (OsMADS1, OsMADS6, OsG1, OsMADS4, OsMADS7, OsMADS8, OsMADS58) of them showed both varied environment-dependent expression patterns and repressed expression levels in eg1 (S13 Fig), indicating their positive regulatory roles in the floral robustness. To further examine this possibility, genetic analysis between EG1 and three of them (OsMADS1, OsMADS6 and OsG1) were performed. OsMADS1 and OsMADS6 are two major genes regulating glume identity and floral determinacy in rice [61–69], and their expressions were significantly varied in eg1 (Fig 7E and S13 Fig), indicating that EG1 is required for their expressions. To examine their genetic relationships with EG1, a double mutant of OsMADS1 mutant allele nsr [61] and eg1-1 was generated and it exhibited longer and leafy lemmas/paleas similar to nsr, with all inner three whorls replaced by half-developed inflorescences or inflorescence primordia, which is severer than both single mutants (Fig 7B), indicating that OsMADS1 functions downstream of EG1 in lemma/palea identity and they may together regulate the determinacy of inner three whorls. In addition, the double mutant of eg1-1 and OsMADS6 mutant allele osmads6-5 [66] showed abnormal paleas, with all transformed into one or two small lemma-like glumes and mostly with inflorescence primordia inside the spikelets (Fig 7C), indicating that OsMADS6 functions downstream of EG1 in specifying palea but may also regulate floral determinacy together with EG1. osmads6-5 exhibited weaker floral disturbance in the F2 population when crossed with eg1-1 (ZF802), with lemma-palea mosaic paleas and usually normal inner whorls (Fig 7C), showing its floral phenotype is also greatly influenced by genetic backgrounds. To further examine the relationships among EG1, OsMADS1 and OsMADS6 especially in floral determinacy, nsr osmads6-5 and eg1-1 nsr osmads6-5 were successively generated. The floral meristems of these two mutants similarly generated continuous glume primordia or became inactive before inner three whorls developed (Fig 7D), which were more dedifferentiated than the inflorescence primordia of eg1-1 nsr and eg1-1 osmads6-5, supporting the findings that both OsMADS1 and OsMADS6 act downstream of EG1 and they together control the floral differentiation of inner three whorls. Compared with eg1-1, the rs plasticity of eg1-1 nsr osmads6-5 totally disappeared when these two MADS genes were both deficient (Fig 7B and 7D), supporting the important roles of OsMADS1 and OsMADS6 in the rs plasticity regulation. Additionally, the spikelet phenotype of eg1-1 nsr [54] and nsr osmads6-5 [68,69] were consistent to the previously reported, and eg1 was linked with the lemma-like (lel) structure, and not affected by the deficiency of OsMADS1 and OsMADS6 (Fig 7b–7d), implying that it is probably a special organ different from lemma and palea. Taken all these results together, we concluded that MADS1 and MADS6 together function downstream of EG1 to control the determinacy of inner three whorls of rice floret as well as to mediate the rs plasticity of eg1-1. Furthermore, the le phenotype (long empty glume) of eg1-2 has the highest plasticity among all variable phenotypes (see Figs 1–3), and OsG1 (Long Sterile Lemma) and OsMADS19/34 are two crucial genes suppressing the elongation of empty glumes in rice [70–72]. The expression levels and patterns of OsG1 but not OsMADS19/34 appeared to be aberrant in eg1-2 (Fig 7E and S13 Fig), implying that OsG1 may contribute to the le phenotype of eg1-2. To confirm this, g1-ele allele of OsG1 [72] was used to obtain a double mutant by crossing with eg1-2, and it turned out that almost all empty glumes of eg1-2 g1-ele were elongated to lemma-like organs similar to g1-ele, exhibiting much lower plasticity compared with eg1-2 (Fig 7F and 7G), indicating that OsG1 functions downstream of EG1 in regulating empty glume development and contributes to the plastic le phenotype of eg1-2. Taken together, these results show that OsMADS1, OsMADS6 and OsG1 all act downstream of EG1 to mediate the floral robustness regulation. To our knowledge, no plasticity genes have been confirmed to function in floral robustness in flowering plants [47–52]. Given the sessile nature of plant species, uncovering this class of genes and dissecting their molecular mechanisms are crucial for understanding the biology of flower development and evolution. In this study, we have shown that EG1 encodes a mitochondria-localized lipase functioning as a plasticity gene to regulate the rice floral robustness by a coordinated transcriptional regulation of temperature, lipid metabolism and flower development pathways. First, eg1 alleles showed floral variations under both natural and artificial conditions and five eg1 alleles produced increased floral plasticity. Second, RNA expression, protein stability and lipase activity of EG1 can respond to environment enhancing its function significantly in severe conditions such as heat stress. Third, EG1 appears to possess a “buffering” function of repressing environmental stimuli to interfere the target genes, and when environmental pressure becomes severer such as heat stress, the strengthened EG1 function induced by heat is enough to buffer the stronger and more deleterious effect of heat on the responsive transcriptional pathways. Last, EG1 influences the expression of numerous floral identity genes, which are the direct contributors of plastic development of eg1 spikelets. Thus, EG1 is the first plasticity gene regulating plant floral robustness against environmental fluctuation. Our finding indicates that flowers retain a system containing EG1 to sense and respond to environmental stimuli to maintain its stable development, and suggests a mechanism of transition from high plasticity to robustness in flower by recruiting plasticity-repressing genes, which ensures the coordinated development of organs with different plasticity in one organism. Lipid metabolisms are known to be involved in adaptive plasticity of organisms [36,37,41,42], and mitochondria, as the energy factory of eukaryotes and one of subcellular compartmentations of lipid metabolism, have been reported to be crucial to the adaptive stability of plant vegetative traits [41,43,44]. As EG1 is a predominately mitochondria-localized functional lipase, our results suggest that the mitochondria-mediated lipid metabolism plays an important role in the regulation of floral robustness against temperature fluctuation. However, it remains unclear how this could be carried out. It is likely that mitochondria-related lipid homeostasis could serve as a “buffer” to relieve the effect of environmental stimuli and mediate the temperature-dependent floral plasticity regulation (Fig 8). Recently, EG1 has been shown to influence JA synthesis, and JA signaling pathway regulates the transcription of floral identity gene OsMADS1 [54], showing that JA potentially mediates the crosstalk of EG1 and flower development related transcriptional factors. In fact, the most homologous gene of EG1 in Arabidopsis DGL has been verified to function in JA biosynthesis [55,56,73], though its chloroplast location was questioned [55], suggesting that a non-chloroplast localized lipase is likely to function in JA production (Fig 8). On the other side, we may fail to detect EG1 protein due to its minute amounts in chloroplasts. Furthermore, according to the much severer phenotype of eg1 and eg2-1D (a mutant allele of EG2/OsJAZ1) double mutant compared with two single [54], we noticed that JA signaling may be not the only pathway activated by EG1 to mediate the signal transduction from outside to inside of nucleus in the EG1-associated floral regulation, other lipid-related pathways and regulatory mechanisms are also possible (Fig 8). In our study, eg1 alleles showed a higher floral plasticity in japonica than that in indica varieties, while a severer floral disturbance in indica, revealing a functional divergence of EG1 in rice subspecies. No differences in coding sequences but cis-elements of EG1 in two subspecies were detected (S14 Fig), implying that this subspecific variation might be caused by an unknown differential responsive capability of promoters to environmental or endogenous stimuli. Besides, since genetic backgrounds are known to influence developmental outcomes via phenotypic modifiers [74,75], there also may be some indica-/japonica-specific modifiers to modulate EG1 function due to their different genetic backgrounds, which can be either epistatic genes or specific lipid substrates of EG1 required for floral developmental robustness. Genetic dissection of these subspecific modifiers of plasticity will provide further insights into the molecular mechanism of floral development. So far, all reported EG1 homologs in dicots have no apparent plasticity function in flowers [56–59], and EG1 homologs in monocots can be mainly divided into two clades, one similar to that in dicots and the other unique to monocot species based on the phylogenetic tree and predicted subcellular localizations (S13 Fig), we thus speculate that EG1 may have acquired a monocot-specific neofunctionalization in promoting floral robustness. Detailed genetic and biochemical analyses of these genes would provide additional clues to when and how the plasticity function and the subspecific divergence of EG1 arose in monocot species. In conclusion, our results revealed a novel function of EG1 in floral developmental robustness against environmental fluctuation by mediating the mitochondrial lipid metabolism. Our finding also provide a genetic means to maintain the stable flower development under environmental stress ensuring grain yield stability in rice and potentially in other monocot species. Further studies should unlock the molecular crosstalk between mitochondria and nucleus in regulating floral developmental robustness. Five rice eg1 recessive alleles were used in this study. eg1-1 and eg1-2 were from our previous work [53], in backgrounds of O. sativa L. spp. indica Zhefu 802 (ZF802) and O. sativa L. spp. japonica Zhonghua 11 (ZH11) respectively. eg1-4 was generated from indica Dular, and eg1-5 and -6 from japonica Nipponbare by CRISPR/Cas9 technology [76,77]. Besides, ZF802>ZH11 and eg1-1 (ZF802>ZH11) were isolated from an F2 population of ZF802 wild-type or eg1-1 backcrossed with ZH11 three times, and ZH11>ZF802 and eg1-2 (ZH11>ZF802) were isolated from an F2 population of ZH11 wild-type or eg1-2 backcrossed with ZF802 once. Other rice mutants, nsr [61] and g1-ele [72] were kindly provided by Dr. Zhukuan Cheng, and osmads6-5 also from our previous work [66]. Plants were grown in the natural conditions of Beijing (China) from Mar. to Oct. and Lingshui (Hainan province, China) from Dec. to Apr. (year 2014~2015). Weather data of 2014.7.28~2014.8.17 of Beijing and 2014.3.15~2014.4.4 of Lingshui, and 2014.2.1~2014.2.20 and 2014.4.1~2014.4.20 of Lingshui were shown respectively, which were ranged from ten days before spikelet meristem formation (~2 mm) to ten days after that. Artificial conditions in chambers were 35°C light 12 hr / 20°C dark 12 hr, 25°C light 12 hr / 20°C dark 12 hr and 40°C light 12 hr / 30°C dark 12 hr, respectively, with identical light intensity 50 μmol m-2 s-1 and relative air humidity 60%. Spikelets of eg1 were divided according to phenotypes of outer glumes (lemma, palea, empty glumes and extra glumes) following two rules: (1) the most significant mutant phenotype of eg1 is in the outer glumes; (2) outer glumes have linkages with inner organs: normal palea always linked with normal inner organs, lemma-like palea with increased stamens and pistils, smaller palea with decreased stamens, multilayer glumes linked with rs. Main panicles/inflorescences of plants at booting stage were used for phenotypic analysis. Percentages of all variable phenotypes in one panicle were calculated for comparisons, and in some cases more than one phenotype appeared in a single spikelet. Data were statistically analyzed by using one/two-way ANOVA (Excel). Effects of genotype, environment and genotype-environment interaction on all variable phenotypes were calculated using phenotypic statistic data of eg1 and wild-type grown in Beijing summer and Lingshui winter by two-way ANOVA. In CRISPR experiment, more than six independently homozygous lines were generated in each background, and their phenotypes were quite similar, thus only one/two lines were statistically analyzed and shown in our results. Full-length EG1 CDS was inserted into N- or C-terminal of GFP sequence of vector pBI221-35SPro:GFP, and full-length EG1 CDS with 1.5 kb native promoter into pCAMBIA1301-GFP. MTS-mOrange and Pro35S:COX11-GFP plasmids were kindly provided by Dr. Yaoguang Liu [60], and the reported vector pCAMBIA1301-35SPro:EG1-GFP was provided by Dr. Dabing Zhang [54]. All plasmids of high quality were prepared for protoplast transfection. Rice protoplast preparation from 2-week-seedlings grown in light and polyethylene glycol (PEG)-mediated transfections were performed as described by Bart et al. [78]. Images were captured by a confocal microscope (FluoView 1000, Olympus). Floral disorder of eg1-2 was complemented by genetic transformation using vector pTCK303-ProUBIQUTIN:FLAG-EG1. One-week seedlings of EG1 complementation lines were used for subcellular fractionation. Fractionations of mitochondria and chloroplasts were performed as described by Rodiger et al. [79]. After precipitating the organelle fractions, western blots were performed with α-FLAG (Sigma), mitochondria specific antibodies α-AOX1/2 and α-COXII (Agrisera), and chloroplasts specific antibodies α-RbcL and α-PsbA (Agrisera). Fractionation assays were performed with two independent complementation lines. Full-length EG1 CDS and a truncated EG1 CDS without predicted targeting sequence (135 bp) with engineered N-terminal SUMO tag were separately cloned into pET-30a (Novagen) to generate His6-SUMO-tagged fusion proteins. Mitochondrial targeting sequence was predicted with MitoProt II online [80]. DGL CDS without targeting sequence [81] was introduced into vector PMAL-C2X (NEB). All the fusion proteins were expressed in E. coli BL21 (DE3). The His6-SUMO-tagged fusion proteins were purified using Ni-NTA (Novagen) and eluted with buffer containing 25 mM Tris-HCl (pH 7.4), 150 mM NaCl and 250 mM imidazole. The MBP-tagged fusion proteins were purified using amylose resin (NEB) and eluted with buffer containing 50 mM Tris-HCl (pH 7.4) and 10 mM amylose. Imidazole in protein solutions was removed with desalting columns (Thermo Scientific) before lipase activity analysis. p-nPB (p-nitrophenyl butyrate, Sigma) was used as the substrate of lipase analysis in vitro. Colorimetric assays for lipase activity of fusion proteins were performed as described by Seo et al. [82] with some modifications. A solution containing 1.11 mg/mL p-nPB (dissolved with isopropanol) and B solution containing 50 mM Tris-HCl (pH 7.4) and 0.1% Arabic gum were first prepared. Reactive solution was composed of 1 volume A solution and 9 volumes B solution with 2% Triton X-100. About 20 μl purified proteins (~5 μg) and 180 μl reactive solution were used for each reaction. After incubated under different temperatures for 30 min, p-nitrophenol formation from p-nitrophenyl butyrate was determined spectrophotometrically at 405 nm by an ELISA microplate reader. DGL and porcine pancreatic lipase (Sigma) were used for positive controls, and p-nitrophenol (Sigma) for the standard curve. Cis-elements in the genomic sequence upstream of the start codon of EG1 was analyzed by PLACE online [83]. Wild-type ZF802, ZH11 and EG1 complementation lines were grown at 25°C for one week after germination, and wild-type ZH11 seedlings were grown in a consistent condition till 2 mm inflorescence meristem developed before being treated under different temperatures for different time. The transcripts of EG1 and FLAG-EG1 were analyzed by RT-QPCR or RT-PCR. The root and shoot phenotypes of seedlings were statistically analyzed after growing under consistent temperatures for six days after germination. Other condition parameters of physiological experiments were daylight 12 hr, light intensity 50 μmol m-2 s-1 and relative air humidity 60%. We use 42°C as an extremely high temperature for short treat but 40°C for long treat by considering the tolerance of plants. FLAG-EG1 protein in EG1 complementation lines was enriched by immunoprecipitation for its small amount and detected by western blot with α-FLAG (Sigma). Two mm inflorescence meristems of eg1-1, eg1-2 and their wild-types ZF802 and ZH11 planted in Beijing summer and Lingshui winter were used for RNA sequencing, and two biological replicates were performed. Total RNAs were isolated with TRIzol kit (Invitrogen). Illumina sequencing libraries were prepared according to the manufacturer’s instructions (Illumina Part # 15026495Rev. D), and sequenced with Illumina system Hiseq2500. Analysis of RNA-seq data was conducted following the standard protocol as described by Trapnell et al. [84]. The raw reads of RNA-seq were mapped to MSU_IRGP_V7 (japonica) and Oryza_indicaASM465 v1.23 (indica) by Tophat [85]. Cuffdiff [85,86] was used to identify the differentially expressed genes between different genotypes or different environments. Co-expression network analysis was performed using R packages WGCNA. Enrichment pathways of genes significantly and specifically affected by G, E or GxE was analyzed and the -log10 (P-values) was tested by Fisher's exact test with Bonferroni correction as described by Lu et al. [87]. Amino acids sequences of EG1 homologs in different plants were aligned with CLUSTAL W and maximum likelihood tree was constructed with MEGA6.0. Subcellular localizations of proteins were predicted with five sorts of software online. Among them, TargetP [88] (http://www.cbs.dtu.dk/services/TargetP/) has the best consistency compared with MitoProt II-v1.101 [80] (https://ihg.gsf.de/ihg/mitoprot.html), iPSORT [89] (http://ipsort.hgc.jp/), ProtComp 9.0 (http://linux1.softberry.com/berry.phtmtopic=protcompan&group=help&subgroup=proloc) and WoLF PSORT (http://www.genscript.com/wolf-psort.html) online. Predicted localizations with TargetP were shown in our results, a/b in which means the protein is more likely in “a” location than in “b” although both are possible.

10.1371/journal.pntd.0002315

Mycobacterium ulcerans Disease: Experience with Primary Oral Medical Therapy in an Australian Cohort

Mycobacterium ulcerans (MU) is responsible for disfiguring skin lesions and is endemic on the Bellarine peninsula of southeastern Australia. Antibiotics have been shown to be highly effective in sterilizing lesions and preventing disease recurrences when used alone or in combination with surgery. Our practice has evolved to using primarily oral medical therapy. From a prospective cohort of MU patients managed at Barwon Health, we describe those treated with primary medical therapy defined as treatment of a M. ulcerans lesion with antimicrobials either alone or in conjunction with limited surgical debridement. From 1/10/2010 through 31/12/11, 43 patients were treated with exclusive medical therapy, of which 5 (12%) also underwent limited surgical debridement. The median patient age was 50.2 years, and 86% had WHO category 1 and 91% ulcerative lesions. Rifampicin was combined with ciprofloxacin in 30 (70%) and clarithromycin in 12 (28%) patients. The median duration of antibiotic therapy was 56 days, with 7 (16%) receiving less than 56 days. Medication side effects requiring cessation of one or more antibiotics occurred in 7 (16%) patients. Forty-two (98%) patients healed without recurrence within 12 months, and 1 patient (2%) experienced a relapse 4 months after completion of 8 weeks of antimicrobial therapy. Our experience demonstrates the efficacy and safety of primary oral medical management of MU infection with oral rifampicin-based regimens. Further research is required to determine the optimal and minimum durations of antibiotic therapy, and the most effective antibiotic dosages and formulations for young children.

Mycobacterium ulcerans (MU) is responsible for disfiguring skin infections which are challenging to treat. The recommended treatment for MU has continued to evolve from surgery to remove all involved tissue, to the use of effective combination oral antibiotics with surgery as required. Our study describes the oral medical treatment utilised for consecutive cases of MU infection over a 15 month period at our institution, in Victoria, Australia. Managing patients primarily with oral antibiotics results in high cure rates and excellent cosmetic outcomes. The success with medical treatment reported in this study will aid those treating cases of MU infection, and will add to the growing body of knowledge about the relative roles of antibiotics and surgery for treating this infection.

Buruli ulcer, also known as Bairnsdale ulcer, Daintree ulcer, or Mossman ulcer is a necrotising infection of skin and subcutaneous tissue caused by Mycobacterium ulcerans (MU) [1]. The major burden of disease is found in tropical climates, but cases have been reported from 33 countries worldwide [2]. M.ulcerans infection has become endemic on Victoria's Bellarine Peninsula in South-eastern Australia [1] Surgical management was traditionally the standard treatment for M. ulcerans disease [1], [3], however recurrence of infection ensued in 17–32% of patients [4], [5]. Evidence of the effectiveness of antimicrobials used alone or combined with surgery [4]–[7], has led to an evolution of our standard treatment practice over the last 15 years to now comprise combination antimicrobial therapy with limited surgical debridement when required [5], [8]. Improved awareness in our region about MU lesions in the community has resulted in earlier referral of patients with smaller lesions to our department at Barwon Health [9]. We previously advocated for further studies using oral antibiotic regimens [7]. Here we describe results from an observational cohort of patients from South-eastern Australia managed with primary oral medical treatment for Mycobacterium ulcerans infection. This is an observational cohort study, approved by Barwon Health's Human Research and Ethics Committee. All previously gathered human medical data were analysed anonymously. Data on all confirmed M. ulcerans cases managed at Barwon Health has been collected prospectively since January 1998. From October 2010, our standard treatment practice for initial M. ulcerans lesions has comprised combination antimicrobial therapy, with limited surgical debridement performed to aid wound healing. The data extracted from medical records included; patient demographics and co-morbid conditions, details of the MU lesion, regimen and duration of antimicrobial therapy, and details of surgical procedures. Cases treated between October 2010 and December 2011 with 12 months follow-up were included in this cohort. A M. ulcerans case was defined as the presence of a lesion clinically suggestive of M. ulcerans plus any of (1) a culture of M. ulcerans from the lesion, (2) a positive Polymerase Chain Reaction (PCR) from a swab or biopsy of the lesion, or (3) histopathology of an excised lesion showing a necrotic granulomatous ulcer with the presence of acid-fast bacilli (AFB) consistent with acute M. ulcerans infection. The anatomical location of a M. ulcerans lesion was described as distal if it was on the elbow or below, or on the knee or below [7], [10]. Primary medical treatment was defined as treatment of a M. ulcerans lesion with either antimicrobials alone or antimicrobials in conjunction with limited surgical debridement. Drug dosages for adults included ciprofloxacin 500 mg twice daily, moxifloxacin 400 mg daily, rifampicin 10 mg/kg/day (up to a maximum of 600 mg daily), and clarithromycin 500 mg twice daily. Paradoxical reactions were defined by the presence of one or both of the following features: a) clinical: an initial improvement on antibiotic treatment in the clinical appearance of a M. ulcerans lesion followed by deterioration of the lesion or its surrounding tissues, or the appearance of a new lesion(s), and b) histopathology: examination of excised tissue from the clinical lesion showing evidence of an intense inflammatory reaction consistent with a paradoxical reaction [11]. Limited surgical debridement was defined as curettage of the lesion or a minor excision to remove excess granulation tissue and to debride ulcer margins, with or without the use of a split skin graft (SSG). Limited surgical debridement was undertaken primarily to remove necrotic tissue from the MU lesion in order to promote healing by secondary intention. Patients who underwent extensive surgery (defined as complete excision of the entire lesion including margins of non-necrotic tissue, with either direct closure or the use of a SSG or a vascularised skin and tissue flap for reconstruction or to cover the defect) were excluded from the formal analysis. Criteria for primary medical therapy in our practice includes; patient willingness to take antimicrobials, and no contraindications to antimicrobial therapy (for example; drug interactions, or severe liver disease). Criteria for complete surgical excision include factors such as; a lesion suitable for removal with direct wound closure, need for reconstruction to close a skin defect via flap or SSG, patient unable or unwilling to take antimicrobials, and patient or surgeon preference. Definitions of treatment success, treatment failure, disease recurrence, and immune suppression were as published previously [2], [5]. A complication of medical therapy was defined as an adverse event attributed to an antibiotic that required cessation of that medication. Data was collected using Epi-Info 6 (CDC, Atlanta) and analysed using STATA 12 (StataCorp, Texas, USA). Categorical values were compared using the Mantel-Haenszel test and median values were compared using the Mann-Whitney test. From 1/10/2010 through 31/12/11, there were 54 patients with MU infection managed at Barwon Health. From this cohort 11 patients (20%) were excluded from further analysis; 3 patients underwent primary complete surgical excision alone, and 8 patients were prescribed antimicrobials but also underwent complete surgical excision. There were no significant differences in baseline characteristics of those excluded from those included in the analysis (table 1). Forty-three (80%) patients were therefore included in this analysis. Baseline characteristics can be seen in table 1. All patients were primarily managed as outpatients. The majority of patients were male (65%), and the median age was 50.2 years (range 1.5–87.9 years). Lesions were ulcerative in 91% and WHO stage 1 in 86% of patients. All patients resided in areas of the Bellarine peninsula where MU is endemic; the majority residing in Point Lonsdale (36%) and Barwon Heads (29%). Four patients (9%) had known co-morbidities including diabetes (2), immune suppression (1), and malignancy (1). No patients were known to be HIV-infected though active screening was not performed. Antimicrobial regimens were all rifampicin-based. Rifampicin was combined with; ciprofloxacin in 30 (70%) patients, clarithromycin in 12 (28%) patients, and moxifloxacin in 1 (2%) patient. The median duration of therapy was 56 days (range 28 to 91 days). Seven of 43 patients (16%) received less than 56 days of therapy. Antibiotic-associated complications requiring cessation of one or more antibiotics occurred in 7 of 43 patients (16%). Two patients developed complications attributed to the combination of rifampicin and ciprofloxacin, 2 patients developed complications attributed to ciprofloxacin, 1 patient developed complications due to moxifloxacin, and 2 patients developed complications due to rifampicin. The most common complications were gastrointestinal upset in 4 patients, joint aches in 2 patients, hepatitis in 2 patients, tendonitis in 1 patient and thrombocytopenia in 1 patient. Of the 11 patients who underwent complete excision and were excluded from the study cohort, 2 of those 8 who took antibiotics developed complications. Nine of 43 patients (21%) developed paradoxical reactions a median of 34 days after antibiotic therapy initiation (IQR 20–92 days). Five of 43 (12%) medically managed patients also underwent limited surgical debridement, and 3 of these procedures involved a SSG for coverage of the defect. Overall, 42 of 43 patients (98%) were cured with primary medical therapy. Cosmetic outcomes were excellent in these medically managed MU cases (Figures 1 & 2). One patient failed primary medical therapy. This patient was a 17 month-old boy who presented with a nodular MU lesion on his arm of 1 cm diameter. He completed 56 days of rifampicin (10 mg/kg/day) with clarithromycin (13 mg/kg/day in twice daily dosing) both in liquid formulations with initial reduction in the size of the lesion. Four months after antimicrobials were completed the lesion enlarged and was debrided. Tissue from the debridement was culture positive for M. ulcerans. The patient was re-treated with rifampicin (10 mg/kg/day) and clarithromycin (13 mg/kg/day in twice daily dosing) both in liquid formulations, with subsequent ulceration of the nodule. A paradoxical reaction was diagnosed 4 weeks after re-treatment commenced based on the clinical deterioration of the lesion and was treated with prednisolone at a dose of 1 mg/kg for 4 weeks. The ulceration progressed and ultimately extended over an area of 12×5 cm and healed fully by secondary intention over a period of 9 months (16 months after initial treatment). Combination antimicrobial therapy is now routinely recommended to treat Mycobacterium ulcerans infection with or without the addition of surgical intervention [4]–[6], [12]. Here we describe our experience with primary medical management of MU infection in South-eastern Australia using oral rifampicin-based regimens and we demonstrate the efficacy and safety of this approach with healing of lesions within 12 months of therapy initiation in 98% of cases, an acceptable toxicity profile and good cosmetic results. In our study, the median duration of therapy was 8 weeks, although over 15% of patients completed less than 56 days of therapy. Previous data from Etuaful and others demonstrated viable MU organisms in tissue after 2 weeks of therapy, but not after 4, 8 or 12 weeks of therapy [13]. We reported that mycobacterial cultures were positive in excised specimens from the majority of patients (55%) who received less than 14 days of antibiotic therapy, but in only 1 of 8 patients (12.5%) who received more than 14 days of treatment [5]. Although WHO recommends combined antibiotic treatment for 8 weeks as first-line therapy for all M. ulcerans lesions [12], shorter durations of antimicrobials may be adequate in selected patients and further research should be performed to explore this possibility [7]. In the recent MU antimicrobial therapy literature, some differences exist in the inclusion or exclusion of patients who underwent surgery. In addition, there are likely geographic differences in surgical practice. In Chauty's medically treated cohort, 37% of patients underwent limited surgery, (and 13% underwent extensive surgery), while in Sarfo's study, 5% of patients underwent SSG while the remaining 95% healed without surgery [4], [7]. In the present study, 12% of medically managed patients underwent limited surgery (including minimal debridement and SSG) to promote wound healing and closure, and we excluded patients who underwent extensive surgery. The selection of patients for medical therapy has not been clearly defined, and size of the MU lesion may or may not be a useful criterion. While smaller (WHO category 1) lesions are likely to achieve cure with medical therapy, they are also likely to be suitable for surgical excision with direct closure. WHO category 1 lesions have been shown to heal with medical therapy after a median of 12–18 weeks [4], [6]. In contrast, large lesions (WHO categories 2 and 3) may require more disfiguring surgery to achieve healing, while medical therapy for these lesions may result in slow but gradual shrinkage and secondary healing. The healing of larger lesions (WHO categories 2 and 3) is variable, and occurs anywhere from 11–15.5 weeks [4], to 30 weeks after initiation of antimicrobials [6]. In our cohort, the majority of patients (81%) who underwent surgical excision had WHO category 1 lesions. In some other studies of medical therapy for MU, researchers have only included lesions less than 10 cm in cross-sectional diameter [6], [7], while we and others have included lesions of all sizes [4]. Antimicrobial therapy for MU appears to be safe and reasonably well tolerated. In a study of rifampicin and streptomycin therapy, treatment was well tolerated with only 3 of 160 patients (1.8%) developing side effects [4]. In a pilot study of oral chemotherapy for MU infection, the authors described rifampicin and clarithromycin in combination as well-tolerated with no adverse effects [7]. It is worthy to note that in their cohort of 30 patients, all were hospitalized during therapy, which would be expected to both boost adherence and manage minor side effects more easily [7]. In our cohort, 7 of 43 patients (16%) ceased one or more antibiotics due to side effects. This rate of side effects is in keeping with rates our group previously described for rifampicin, clarithromycin and ciprofloxacin, which were higher than rates reported in younger populations in Africa [4]. We believe the older median age of our cohort may explain reduced drug tolerability, but that overall the side effect profile is acceptable [5]. Antimicrobial therapy has been demonstrated to improve MU lesion healing and prevent relapse. In non-controlled trials, rifampicin-based regimens in conjunction with clarithromycin [7], or a fluoroquinolone [5], [11], have demonstrated success in managing MU infection. Our antimicrobial treatment success rate of 98% is in keeping with other studies of medical management of MU [4], [6], [7]. In Sarfo's study of 160 patients, and Chauty's cohort of 30 medically managed patients there were no recurrences or relapses during 1 year of follow-up after treatment initiation [6], [7]. In Nienhuis's study, there was a 1.4% recurrence rate [4], and in our cohort of 43 patients we describe one relapse in a patient with a nodular form of MU infection. Although the cure rates described in our Australian cohort are comparable to studies conducted in Africa [4], [6], [7], it is possible these outcomes could be influenced by identified small differences in M. ulcerans genomic sequences. M. ulcerans strains worldwide produce a very restricted repertoire of mycolactones, although Australian strains characteristically produce predominantly mycolactone C compared with predominantly mycolactone AB produced in Africa [14]. In addition, M. ulcerans has evolved by acquiring a plasmid (pMUM) and other independent genomic changes within strains from different areas to produce region- specific phenotypes and genotypes [15], [16]. However, recent analysis has revealed that Buruli ulcer in Africa and Australia is caused by one distinct lineage of mycolactone-producing Mycobacteria comprising a highly clonal group [17]. This close genetic relationship suggests that strain differences are minimal and there is currently no evidence to suggest that these strain differences significantly influence responses to therapy or cure rates. There are however some differences between patient populations in our study and those conducted in Africa. For example, the median age in our cohort was 50 years compared with 12 years in studies conducted in Ghana [4], [6]. In our study over 90% of lesions were ulcerated, compared with 36–70% in African cohorts [4], [6], [7], and none of our patients were HIV infected compared with 2% in Ghana [6]. Finally, there are possible unmeasured differences in nutrition and innate response to infection that may affect cure rates in different geographic regions. There are a number of potential reasons for the relapsed case seen in our study. Firstly, it may relate to the nodular form of disease as medical management is thought to be sufficient for ulcerative lesions but some believe that most non-ulcerative lesions require additional surgery [7]. It may be more difficult to sterilize lesions that don't ulcerate and discharge the underlying necrotic material containing large numbers of mycobacteria. Secondly, it is possible that the ulceration of this nodule was part of the natural history of nodular MU disease, as has been described by Nienhuis and others [18], whereby most nodules ulcerate during the healing process. However the fact that the area of ulceration was significantly larger than the initial nodule and the lesion remained culture positive 6 months after antibiotics commenced argue against this. Finally, it is possible that the liquid preparations of antimicrobials used in this case contributed to treatment failure. It has been shown that the bioavailability of rifampicin varies depending on the drug preparation [19]. Furthermore, experience suggests that the use of syrups in young children can adversely affect adherence potentially leading to under dosing. Thirdly, it may be that recommended doses of antimicrobials in young children are inadequate, especially as no specific research has been performed on drug levels in this age-group for MU treatment. Research in children hospitalized for treatment of tuberculosis has found that their serum rifampicin concentrations were considerably less than the suggested lower limit for 2-hour rifampicin concentrations in adults after receiving standard rifampicin dosages similar to those used in adults [20]. This study however utilized fixed-dose combination tablets formulated for paediatric use, which is not comparable to the preparation used in our patient. There are limitations to our study; firstly it's observational design and the exclusion of 20% of patients during the study period as they underwent excisional surgery. Although there were no significant differences in baseline characteristics between this group and those included, it is possible that there were other unknown confounders that may have introduced a selection bias and affected the validity of our findings [4], [6], [7]. Secondly, we are unable to detail the exact timing of healing of MU lesions, which would be valuable information. Finally, most of the lesions treated in our study (86%) were WHO Category 1 lesions and therefore the strength of our evidence for treatment success for lesions of larger size is weaker, although there have been descriptions in the literature of good success rates with medical therapy for lesions larger than 5 cm [6], [7]. Our experience demonstrates the efficacy and safety of primary oral medical management of M. ulcerans in an Australian cohort. Further research is required to determine the optimal and minimum durations of antimicrobial therapy and the most effective dosages and formulations of antimicrobials for young children.

10.1371/journal.pntd.0005716

A leprosy clinical severity scale for erythema nodosum leprosum: An international, multicentre validation study of the ENLIST ENL Severity Scale

We wished to validate our recently devised 16-item ENLIST ENL Severity Scale, a clinical tool for measuring the severity of the serious leprosy associated complication of erythema nodosum leprosum (ENL). We also wished to assess the responsiveness of the ENLIST ENL Severity Scale in detecting clinical change in patients with ENL. Participants, recruited from seven centres in six leprosy endemic countries, were assessed using the ENLIST ENL Severity Scale by two researchers, one of whom categorised the severity of ENL. At a subsequent visit a further assessment using the scale was made and both participant and physician rated the change in ENL using the subjective categories of “Much better”, “somewhat better”, “somewhat worse” and “much worse” compared with “No change” or “about the same”. 447 participants were assessed with the ENLIST ENL Severity Scale. The Cronbach alpha of the scale and each item was calculated to determine the internal consistency of the scale. The ENLIST ENL Severity Scale had good internal consistency and this improved following removal of six items to give a Cronbach’s alpha of 0.77. The cut off between mild ENL and more severe disease was 9 determined using ROC curves. The minimal important difference of the scale was determined to be 5 using both participant and physician ratings of change. The 10-item ENLIST ENL Severity Scale is the first valid, reliable and responsive measure of ENL severity and improves our ability to assess and compare patients and their treatments in this severe and difficult to manage complication of leprosy. The ENLIST ENL Severity Scale will assist physicians in the monitoring and treatment of patients with ENL. The ENLIST ENL Severity Scale is easy to apply and will be useful as an outcome measure in treatment studies and enable the standardisation of other clinical and laboratory ENL research.

Erythema nodosum leprosum (ENL) is a severe, painful complication of leprosy, which can occur before, during or after successful treatment of the infection. ENL is characterised by severe pain and the development of new painful skin lesions. Other organ systems are often affected. ENL may continue to affect people for many years leading to disability, significant loss of income and sometimes death. ENL often requires prolonged treatment with corticosteroids, thalidomide or other drugs that modulate the immune system. Thalidomide is the most effective treatment but is not widely available or affordable in many leprosy endemic settings. This results in many people taking high doses of corticosteroids for prolonged periods and being at risk of severe adverse effects. In order to assess the efficacy of different treatments for ENL it is important to be able to compare individuals before and after treatment. This is difficult to do in a complex, multi-system disorder such as ENL. To overcome this problem we have developed a scale, which is easy to use, to measure the severity of ENL. In this article, we describe the validation of this scale, which ensures that it is a useful measure of ENL severity.

Erythema nodosum leprosum (ENL) is a severe inflammatory complication of borderline lepromatous (BL) leprosy and lepromatous leprosy (LL). ENL affects up to 50% of individuals with LL and 5–10% of BL leprosy patients [1, 2]. A bacterial index of four or more is also a risk factor for developing ENL. ENL may occur before, during or after successful completion of anti-mycobacterial multi-drug therapy (MDT)[2]. ENL causes inflammation in many systems and is characterised by severe pain, tender cutaneous skin lesions, fever, joint and bone pain, iritis, orchitis, lymphadenopathy and neuritis [3]. Most patients have multiple episodes of painful inflammation extending over several years [2, 3]. ENL is associated with a deleterious impact on health related quality of life (HRQoL)[4], increased mortality[5] and severe economic hardship for affected individuals and their families[6]. The Erythema Nodosum Leprosum International STudy (ENLIST) Group[7] aims to improve the understanding of the mechanisms which cause ENL, improve the evidence to guide treatment decisions of individuals with ENL and improve access to effective treatments. The ENLIST Group includes clinicians and laboratory scientists with extensive experience in the treatment and investigation of the causes of ENL based at institutions in eight countries. The cause of ENL is unclear. It is associated with a complex array of immune activation and consequent inflammation, which requires immunosuppression. ENL skin lesions may show features of vasculitis and there is evidence of neutrophil and lymphocyte activation. The role of immune complexes in ENL remains unproven. Patients are treated with corticosteroids, clofazimine and thalidomide either alone or in combination, and less commonly other immunomodulatory agents, which are used for prolonged periods of many months or years [3]. Many patients require high doses of corticosteroids to control their disease and this leads to complications and deaths associated with long-term use of these drugs[8]. Thalidomide is usually effective but is not available in many countries or it is severely restricted because of the risk of teratogenicity. Other adverse effects occur with thalidomide and these have been reported to occur in 25–68% of individuals [9–12]. The neuropathy caused by thalidomide during its use to treat other conditions is approximately 20% but there are no good data for the frequency of thalidomide-induced neuropathy in patients with ENL[13]. The identification of other agents for controlling ENL has been identified as a priority for patients in countries where thalidomide is prohibited or highly restricted, unaffordable, ineffective or poorly tolerated [14]. The evidence for choosing the appropriate treatment for ENL is limited. There have been eight small, randomised treatment studies of ENL since the introduction of MDT [9–12, 15–17]. Only 269 patients were enrolled into these studies and just three studies with a total of 53 participants reported allocation concealment and blinding [15, 16]. Determining outcome measures for clinical studies in complex, multisystem disorders such as ENL is difficult. Quantitative severity scoring systems provide one possible outcome measure.There have been several scoring systems employed in studies of ENL however none have been validated [9, 18–21]. Unpublished (non-validated) scales have been shown to be a source of bias in randomised controlled trials [22, 23]. A Cochrane review highlighted the difficulty in comparing treatment studies in ENL and recommended the development of validated severity scales[14]. We developed a 16 item scale, the ENLIST ENL Severity Scale (EESS), for measuring the severity of ENL[24]. We applied and critically appraised three previously published scales for ENL that had not been validated and used regression analysis of data from our cross-sectional study of the clinical features of ENL[3] to enable us to develop the EESS. The scale incorporates assessments of pain and wellbeing using visual analogue scales (VAS), fever, skin signs, oedema, orchitis, ocular inflammation, joint and bone involvement, nerve assessments and urinalysis. We wished to validate the EESS and determine the minimal important difference (MID). MID is a concept used to determine whether an outcome is clinically relevant and relates to the smallest difference in score which is perceived as beneficial[25]. Participants who gave written, informed consent were recruited from seven centres: DBLM Hospital, Nilphamari, The Leprosy Mission Bangladesh; Oswaldo Cruz Institute, Rio de Janeiro, Brazil; the ALERT Center, Addis Ababa, Ethiopia; Bombay Leprosy Project, Mumbai, India; The Leprosy Mission Hospital, Purulia, India; Anandaban Hospital, The Leprosy Mission Nepal and the Cebu Skin Clinic, the Leonard Wood Memorial Center, Cebu, Philippines. Ethical approval was obtained from the Ethics Committee of the London School of Hygiene and Tropical Medicine (Reference 10370), The Leprosy Mission International Bangladesh Institutional Research Board, Brazilian National Ethical Review Board (CAAE: 12834613.4.0000.5248), AHRI-ALERT Ethical Review Committee (PO12/16), Ethics Committee of the Managing Committee of the Bombay Leprosy Project (BLP/OO/TRC/76A/2015), The Leprosy Mission Trust India Ethics Committee, the Nepal Health and Research Council (106/2011), Institutional Ethical Committee/Institutional Ethical Regulatory Board of Cebu Leprosy and Tuberculosis Research Foundation, Inc. (CLTRFI/LWM-IEC 2015–005). Individuals were eligible to participate in the study if they met any of the following criteria: For the purposes of the study ENL was defined as a patient with leprosy who had crops of tender cutaneous or subcutaneous lesions or was on treatment for ENL (whether it was active or not). The type of ENL was categorised as acute, recurrent or chronic which were defined as follows: Individuals who did not wish to give consent or were diagnosed with leprosy Type 1 reactions were excluded. Each participant was examined independently by a health worker (usually a doctor but sometimes a physiotherapist and in one centre an experienced leprosy research scientist) who had been trained to use the EESS and by an experienced leprologist who also applied the scale and categorised the ENL as “inactive” or “mild” or “moderate” or “severe”. We did not attempt to standardise the categorisation of ENL by the experienced leprologists. Neither assessor (nor the participant) was aware of the result of the other assessor’s examination. The time interval between the two assessments was kept as short as practicable. At a subsequent visit, the MID of the EESS was determined by applying the scale to individuals after treatment and asking both the participant and the examining leprologist to independently categorise the change as: “much better”, “somewhat better”, “no change” (or “about the same” for physicians), “somewhat worse” or “much worse”. The leprologist had performed one of the original assessments at the first visit but did not apply the EESS on the second occasion and was blinded to the result (as was the participant). The EESS, on this occasion, was applied by the same health worker as at the initial visit whenever possible. The MID methodology was only used for participants who had been categorised as having mild or moderate or severe ENL at the first set of assessments. All data including demographic, clinical and EESS were collected on data collection forms specifically designed for the study. The anonymised data were entered into a password protected Access database at each centre and subsequently merged. The data were analysed using Stata 14 (StataCorp. 2015 Stata Statistical Software: Release 14. College Station, TX: StataCorp LP). We wished to recruit 300 participants as this would provide more than 10 study subjects per scale item [26]. The internal consistency or reliability was assessed using Cronbach’s alpha. An alpha between 0.7 and 0.9 is considered acceptable[27]. The contribution of each item in the scale was assessed by calculating Cronbach’s alpha for the scale if that item were removed. The ability of the scale to discriminate between patients with different clinical severity categories was determined using analysis of variance. The threshold for accepting statistical significance was p<0.05. Inter-observer reliability was evaluated using Intra-Class Correlation of the total score of each examiner using a two-way analysis of variation (5% level of significance) and the strength of agreement criteria of Landis and Koch[28]. A Bland Altman plot of the difference between pairs of observations and the mean of those pairs was used to highlight any potential systematic differences between assessors Receiver operator characteristic curves were used to determine cut off points for mild, moderate and severe reactions. The ability of the scale to reflect the change in ENL was calculated as the mean of the change in severity associated with each of the reported outcomes “Much better”, “somewhat better”, “somewhat worse” and “much worse” compared with “No change” or “about the same” (for physician rated change). 447 individuals were recruited between 13th May 2015 and 16th July 2016. 336 (75.3%) were male and 110 (24.7%) female. 19 physicians classified the severity of the 210 participants with active ENL. The demographic and clinical features are summarised in Table 1. 54.3% of the 210 individuals with ENL were receiving treatment for their ENL at the time of enrolment. Of the nine drug regimes used, prednisolone (57.2%), prednisolone and clofazimine (24.6%), thalidomide (6.5%), thalidomide and prednisolone (4.3%) and clofazimine alone (2.9%) were the most common. Initially only 14 of the 16 items were considered for inclusion in the severity scale. The VAS Wellbeing item was excluded because we wished to maintain a strictly clinical focus. Orchitis was also omitted as we wished to try and produce a gender-neutral scale. The items showing the lowest levels of correlation were inflammation of the eyes due to ENL, urinalysis and the items related to sensory and motor nerve function. The internal consistency of the 14 potential items for inclusion in the scale was assessed using Cronbach's alpha producing an initial value of 0.7413. The series of analyses reported were based on the data from those individuals who had been classified as having mild or moderate or severe ENL (n = 210). Removing eye inflammation due to ENL and urinalysis increased the value of alpha to 0.7633. Removing the count of nerves with sensory NFI due to ENL and the count of nerves with motor NFI due to ENL further increased alpha to 0.7672. Removal of any further items brought a reduction in alpha, confirming the inclusion of the remaining 10 items in the scale. The derived 10 item scale was analysed separately for men to see if the addition of orchitis significantly altered the internal consistency. Using the data for men alone the 10 item scale has a Cronbach alpha of 0.7633 which increases to 0.7645 with the addition of orchitis. The increase in alpha did not result in greater internal consistency for men compared to men and women combined. Principal component analysis showed a general factor to which all 10 items contributed accounting for 33.4% of the total variance. A second “pain” factor contributed 16.0% of the total variance. It contrasted VAS pain, bone pain, inflammation of joints and nerve tenderness with items describing the number, extent and inflammation of skin lesions and lymphadenopathy. The 10 item scale discriminated well between patients with active ENL and those without. Fig 1 shows the distribution The difference in scores between the non-ENL group and those categorised as having mild ENL were significant (p<0.001, parametric and non-parametric test). A threshold value differentiating between those who were classified as having “moderate” ENL and those with “severe” ENL was not identified by either score or ROC curve. However, the difference in mean severity scores was statistically significant (p<0.001, parametric and non-parametric test). The EESS scores of participants with acute or recurrent or chronic ENL were not significantly different. The intra-class correlation coefficient assuming random effect for both patients and assessors and individual assessors is 0.797 (95% CI 0.742, 0.843). The strength of agreement is good[28]. A Bland-Altman plot (Fig 2) showed good agreement between the two assessors of each patient with no evidence of a systematic difference in terms of larger differences for higher severity scores. 15 (7.4%) of the 204 paired assessments fell outside the confidence limits but these were evenly distributed between positive and negative differences. To determine the cut off score between “mild” and more severe categories of ENL the ROC curve was plotted for patients classified as having “mild” ENL and those with either moderate or severe ENL (Fig 3). Mild ENL was determined to be an EESS score of 8 or less and more severe forms scoring 9 or more (Fig 4). The area under the curve for mild and combined moderate and severe ENL is 0.8372. This value indicates that the final scale is a good discriminator between the mild and more severe categories of ENL traditionally used by clinicians. 152 participants with ENL completed the two sets of assessments. The median interval between these two sets of assessments was 28 days, range 0 to 185.The changes in EESS scores and the participant rated and physician rated improvement are shown in Figs 5 and 6 respectively. Table 2 shows the mean change in scores for each participant-rated or physician-rated change and the difference from “no change” or “about the same” respectively. The MID was 4.9 for both participants and physicians which equates to a change in EESS score of 5. The change in mean EESS score from baseline to “much better” which resulted in an “important difference” was 8.4 for participants and 7.0 for physicians. This equates to a change in EESS score of 9. Leprosy reactions present a major challenge to the successful management of the disease. The adapted version of the EESS is a valid and reliable measure of the severity of ENL. We have been able to show that it discriminates between patients with mild ENL and more severe disease. The scale and accompanying guide (S1 Appendix) are easy for clinicians to use. A clinical tool to measure the severity of leprosy Type 1 reactions was designed and validated by some members of the ENLIST Group [29]. Type 1 reactions are a major cause of nerve damage in leprosy. The Type 1 reaction severity scale was first validated in Brazil and Bangladesh and in a subsequent study from Ethiopia[30]. It has been shown to reflect change in severity following treatment and has been used in clinical trials of corticosteroids, azathioprine and ciclosporin [31–34]. We believe that the EESS has the potential to be equally important in ENL. The EESS is gender-neutral following removal of the orchitis item which did not materially affect the internal consistency of the scale. The two items relating to NFI were removed from the final version of the EESS but clinicians should remain vigilant to new NFI occurring in the context of ENL, which we have previously shown to have a prevalence of 22.9% in individuals with ENL[3]. Nerve tenderness is a component of the scale but nerve tenderness does not always accompany NFI. Eye involvement due to ENL is not included in the scale and clinicians will need to be cognizant of this uncommon feature when using the EESS. We were able to demonstrate that the EESS discriminates between moderate and severe ENL but could not determine a cut off score for the two categories. We did not attempt to standardise physician assessment of the severity of ENL which is a limitation of the study. The overlap of scores in the group categorised as having “moderate” ENL and those with “severe” ENL is likely to have occurred due to variation in physician perception of the construct of ENL severity. The multisystem nature of ENL may mean that different physicians attach different weight to different symptoms or signs when categorising ENL severity. Variation between physician global assessments has also been reported to occur during comparison of different methods for the assessment of flares in systemic lupus erythematosus [35]. The MID of the EESS was determined using both participant and physician reported change. There are no agreed criteria for determining which group should be used for determining MID. It has been argued that physicians are the best judges of change in measures of disease activity or damage, whereas in functional or HRQoL measures it is the affected individual [36]. We felt it was important to assess the responsiveness of the EESS using both groups. The results were concordant with an MID of 5 for both groups. The greatest discordance, of two scale units, between the ratings of participants and physicians occurred for the change of “much better”. The EESS is the first validated, published scale of ENL severity and is responsive to change in ENL. We plan to use the EESS in future double-blind randomised controlled treatment studies of ENL and believe it will be an important tool for other clinical researchers. The scale will also be useful in providing a standardised way of describing the severity of ENL in patients who participate in immunological and genetic studies. It is equally important that the EESS be incorporated into routine clinical practice where we believe it will help physicians to assess, monitor and treat patients.

10.1371/journal.pgen.1007998

Evolution of an insect immune barrier through horizontal gene transfer mediated by a parasitic wasp

Genome sequencing data have recently demonstrated that eukaryote evolution has been remarkably influenced by the acquisition of a large number of genes by horizontal gene transfer (HGT) across different kingdoms. However, in depth-studies on the physiological traits conferred by these accidental DNA acquisitions are largely lacking. Here we elucidate the functional role of Sl gasmin, a gene of a symbiotic virus of a parasitic wasp that has been transferred to an ancestor of the moth species Spodoptera littoralis and domesticated. This gene is highly expressed in circulating immune cells (haemocytes) of larval stages, where its transcription is rapidly boosted by injection of microorganisms into the body cavity. RNAi silencing of Sl gasmin generates a phenotype characterized by a precocious suppression of phagocytic activity by haemocytes, which is rescued when these immune cells are incubated in plasma samples of control larvae, containing high levels of the encoded protein. Proteomic analysis demonstrates that the protein Sl gasmin is released by haemocytes into the haemolymph, where it opsonizes the invading bacteria to promote their phagocytosis, both in vitro and in vivo. Our results show that important physiological traits do not necessarily originate from evolution of pre-existing genes, but can be acquired by HGT events, through unique pathways of symbiotic evolution. These findings indicate that insects can paradoxically acquire selective advantages with the help of their natural enemies.

Parasitic wasps are important insect biocontrol agents. These insects are beneficial for the ecological service they provide, which largely contributes to the control of natural populations of their hosts. Paradoxically, they can be beneficial also for non-host individuals attacked by mistake, if these survive after parasitization. In nature, this can happen to lepidopteran larvae when attacked by a wasp harbouring a symbiotic virus that can mediate horizontal gene transfer. Indeed, this virus, injected along with the egg in the body of the host, can get integrated in the genome of the parasitized larva, carrying along exogenous genes. Because the non-host regularly suppresses the parasitoid egg and/or juveniles, any surviving individual with a stable insertion of new genes in the germ line will represent an evolutionary novelty, with expanded functional capacities, if the resulting gene domestication event confers new physiological traits. The immune function here discovered demonstrates that symbiotic associations can drive unique evolutionary pathways, maximizing the fitness of interacting organisms, which evolve as a complex unit with a shared gene pool.

Horizontal gene transfer (HGT) is a mechanism of accidental acquisition of genetic material by means other than reproduction, which in some evolutionary lineages, such as prokaryotes, is considered the major driving force in genome evolution [1]. In theory, all genes may undergo HGT, however current evidence on prokaryotes indicates that housekeeping genes, modulating cellular functions, are significantly more itinerant than regulatory genes [2, 3]. In eukaryotes, HGT occurrence in unicellular organisms has been frequently reported, while, in contrast, it has been considered rare in multicellular organisms, until the advent of high-throughput sequencing technologies, which have allowed the discovery of a considerable number of HGT events in these organisms. Indeed, although less common than in prokaryotes, HGT is far from being of marginal importance in genome evolution of multicellular eukaryotes [4–7)]. The HGT occurrence is frequent both in vertebrates and invertebrates, with bacteria and protists being the major gene donors, as they have established an extensive variety of symbiotic associations with higher organisms, which favours intimate contact and exchange of genetic material [6, 8]. This complements the major role played by transposable elements in shaping genome evolution [5–7, 9]. The study of HGT in eukaryotes has increasingly shed light on its frequency and on the functional categories of the genes involved [3, 5, 7, 9, 10]. However, only a few reports have addressed whether the transferred genes are neutrally included in the genome or functionally integrated into the biological pathways of the recipient organism (i.e., domestication). Cases of HGT from microorganisms and plants, potentially conferring a functional benefit, have been described in rotifers, nematodes and arthropods [11]. Genome and transcriptome analyses of plant-parasitic nematodes indicate that success in host colonization and exploitation could have been favoured by genes acquired from bacteria through HGT [12–16]. These genes encode plant cell wall-degrading or -modifying enzymes [11–17], as well as enzymes involved in vitamin synthesis [16], or invertases expressed in the nematode digestive system, and likely involved in the metabolism of plant sucrose [13]. Similarly, there is increasing evidence to suggest that HGT plays an active role in modeling insect genomes, providing novel enzymes for digestion and metabolism. Indeed, in coffee berry borer (Hypothenemus hampei) [18] and mustard leaf (Phaedon cochleariae) [19] beetles, genes acquired from gut bacteria may confer the capacity to hydrolyze host plant polysaccharides; similarly, genes of bacterial and fungal origin may mediate carotenoid biosynthesis in the pea aphid [20] and metabolism of sugar and amino acids in some phytophagous species of Lepidoptera [10]. Moreover, genes conferring the putative capacity to detoxify plant defense chemicals have been acquired by HGT, both in insects and mites [21]. Genes possibly implicated in controlling arthropod immunity or their interactions with other organisms also undergo HGT, as reported for Drosophila [5]. Several parasitic chalcidoid wasp species have acquired chitinase genes from fungi, which are expressed in the venom gland to produce virulence factors injected into the host at the oviposition [22]. The recurrent and independent transfer of bacterial genes encoding antibacterial toxins to distinct eukaryotic lineages suggests that these genes can enhance the fitness by adding new immune defense barriers [23]. Only recently it has been demonstrated that these domesticated sequences express active antibacterial effectors in the recipient organisms [24]. Indeed, in the tick Ixodes scapularis, this allows the control of the proliferation of the Lyme disease agent, Borrelia burgdorferi, in order to keep its abundance at optimal levels for an effective vectoring activity. However, an active role of these genes in the modulation of the immune response against infection is still elusive. Polydnaviruses (PDV) have recently been identified as a new source of sequences contributing to insect genome evolution by HGT [6, 25]. PDV are viral symbionts of braconid (Bracovirus-BV) and ichneumonid (Ichnovirus-IV) parasitic wasps attacking lepidopteran larvae, which are injected during the oviposition [26, 27]. They express virulence factors in parasitized hosts triggering immunosuppression and a number of physiological alterations to allow progeny survival and development [26–29]. The PDV are integrated into the wasp’s genome as provirus, while free virions are produced only in the calyx region of the wasp’s ovary. These contain multiple DNA circles, some of which, upon infection of host tissues, become integrated into its genome, starting the expression of virulence factors without undergoing replication [27, 28]. This latter integration property has been proposed as the main route of entrance of genetic material from the BV associated with the parasitic wasp Cotesia congregata (CcBV) into the genome of non-permissive hosts, which survive and can transmit CcBV sequences integrated in the germ line [25]. The viral origin of one of these insertions in the moth species Spodoptera exigua, named gasmin, is unequivocally corroborated by the presence of a bracoviral regulatory sequence [25]. A recent work has tried to analyse the microevolutionary forces driving the domestication of gasmin [30]. This was done by baculovirus-mediated expression of gasmin in the larvae of a population lacking a functional gasmin gene (i.e., the European population of S. exigua that bears a truncated non-functional gasmin gene). The reduced mortality of larvae infected by a baculovirus expressing gasmin was attributed to a putative protecting role exerted by this protein against viral infection and/or replication [30]. However, this presumed antiviral defense barrier was surprisingly associated with an enhanced susceptibility of S. exigua larvae to bacterial infection [30]. Collectively, the functional evidence provided was largely indirect and strongly influenced by baculovirus infection, which can have effects difficult to tease apart from those induced by gasmin. Then, the microevolutionary scenario cannot be unequivocally interpreted, since a clear conclusion on the effective physiological role of this viral protein and on the adaptive advantage conferred by the domestication of its coding gene is still lacking. In order to fill this gap, here we report a detailed molecular and functional characterization of a gasmin homologue, identified in S. littoralis, using a gene silencing approach instead of a baculovirus mediated expression in a gasmin-free environment, in order to get rid of all potential experimental artifacts that have apparently influenced previous studies. Our work demonstrates the important role of gasmin in the modulation of the cellular immune response. This sheds new light on the importance that HGT can have in the evolution of metazoan genomes, and on how innate defense barriers in insects can be paradoxically shaped by parasitic wasps and their associated viral symbionts, which are potent pathogens of the wasp’s host. The genomic sequence of Sl gasmin, including the intron (1126 bp), shows very high identity with S. exigua gasmin (KP406767), S. litura gasmin (obtained through data mining; MTZO01009970.1) and with viral BV2-5 (AJ632326) (98, 72 and 86% respectively) (Fig 1A). The coding DNA sequence (CDS) encodes a predicted protein of 346 amino acids (aa) (with a putative signal peptide of 21 aa) (Fig 1B and 1C), that shows 95% and 76% sequence identity with S. exigua gasmin and with the homologue viral protein encoded by CcBV (CcBV 25.3; CAG17487), respectively (Fig 1B and 1C). To learn more about the origin of the Sl gasmin, we retrieved all sequences showing high similarity, available in NCBI, and constructed a similarity tree using the predicted amino acid sequences (Fig 2). Phylogenetic analyses strongly support the monophyly of the S. exigua and S. littoralis protein sequences (BS = 100), and indicate that S. litura represents a sister to these sequences. Other relationships, including the placement of a homologue identified in the winter moth (Operophtera brumata), are poorly supported and available sampling does not permit the confident assignment of likely donors for any of the lepidopteran genes. While all of the lepidopteran proteins (with the exception of the S. exigua and S. littoralis sequences) are highly divergent from one-another, examination of the aligned protein sequences reveals that most indels are concentrated in regions that also show indels and high divergence among viral sequences, possibly consistent either with rapid evolution under altered selective constraints in various lepidopteran groups, or with multiple acquisitions and independent losses in insects. Further sequence sampling would be required to conclusively resolve this issue. To investigate the existence of selective pressure on S. littoralis and S. exigua sequences we used FEL (Fixed Effects Likelihood), a program which uses a maximum-likelihood approach to infer nonsynoymous (dN) and synonymous (dS) substitution rates on a per-site basis for a given coding alignment and corresponding phylogeny. As expected, negative purifying selection has been observed at 26 sites with P values≤0.05 (S1 File). The expression profile of Sl gasmin in different tissues of S. littoralis larvae was analyzed by qRT-PCR. The haemocytes (i.e., the circulating immune cells) were by far the most active site of transcription (One-Way ANOVA: F(3,24) = 49.07, n = 7, P<0.0001, df = 27) (Fig 3), suggesting a key-role of this gene in the immune response. Because this gene is of viral origin, and its viral homologue is expressed in parasitized hosts, we formulated the hypothesis that it confers a selective advantage both to the wasp bearing the donor viral symbiont and to the recipient moth species. Indeed, because the protection of wasp’s juveniles against host immune responses is mediated by broad immunosuppressive strategies, it is reasonable to assume that a reinforcement of the antimicrobial barriers, by preventing secondary infections of the host, is also beneficial for the wasp progeny. To corroborate this hypothesis, Sl gasmin transcription was assessed in response to microbial challenge. The injection of both Gram-positive (Staphylococcus aureus) and Gram-negative (Escherichia coli) bacteria in S. littoralis larvae, as well as of the yeast Saccharomyces cerevisiae, significantly enhanced the transcription of Sl gasmin in the haemocytes (One-Way ANOVA S. aureus: F(7,88) = 150.68, P<0.0001, n = 12, df = 95; E. coli: F(7,88) = 169.61, P<0.0001, n = 12, df = 95; S. cervisiae: F(7,88) = 68.072, P<0.0001, n = 12, df = 95), with slightly different temporal profiles (Fig 4). The expression level of the target gene increased very rapidly following bacteria injection, while S. cerevisiae challenge triggered a comparatively slower response soon after injection, followed by a gradual decrease to the basal level 12 h post-injection (Fig 4). The very high transcription level in the haemocytes and its enhancement following microbial challenge strongly suggested the important role for Sl gasmin in the immune response mounted by S. littoralis larvae. To characterize Sl gasmin at the functional level, we pursued a loss-of-function strategy, through RNAi-mediated silencing, which proved to be very efficient in S. littoralis [31, 32], like in many other lepidopteran species [33–39], in spite of several cases of refractiveness to RNAi reported for species/strains belonging to this order [39]. Oral treatment of experimental larvae with Sl gasmin dsRNA was effective in silencing the target gene from the first day of the 5th instar until the prepupal stage (Student’s t test. Day 1: t = 26.838, df = 14, P<0.0001; day 2: t = 11.362, df = 14, P<0.0001; day 3: t = 15.408, df = 14, P<0.0001; day 4: t = 18.456, df = 14, P<0.0001; day 6: t = 22.903, df = 14, P<0.0001; day 7: t = 15.258, df = 14, P<0.0001) (Fig 5). The relevance of the immune role played by Sl gasmin in vivo was assessed by scoring the impact of the silencing of its coding gene on the host septicaemia induced by Bacillus thuringiensis toxin Cry1Ca, using the experimental approach previously described to study the Bt killing mechanism [32]. Cry toxins produced by this entomopathogenic bacterium are active upon ingestion. When in the midgut, they interact with brush border membranes of epithelial cells and cause the osmotic lysis of these latter. The resulting tissue lesions allow the entrance of gut microflora into the insect haemocoel, causing septicaemia and insect death [32]. Then, we hypothesized that the reduced immunocompetence associated with Sl gasmin silencing should result in the enhancement of Cry1Ca killing activity, as a consequence of uncontrolled bacterial proliferation. Indeed, this was the case. Sl gasmin dsRNA treatment significantly enhanced the mortality of S. littoralis larvae treated with Cry1Ca toxin (Table 1), and this enhancement was associated with a significant increase of bacterial load in the haemolymph (Fig 6, see S1 Table for statistics). These data clearly support the key-role played by this gene in the modulation of the antimicrobial immune response in vivo, and nicely corroborate the proposed key-role of septicaemia in the Bt killing mechanism [32]. Thereafter, we performed specific experiments to understand which component of the immune response was controlled by Sl gasmin. We focused on the three main cellular responses against foreign invaders that are mediated by haemocytes: nodulation, encapsulation and phagocytosis [40–41]. On the occasion of each experiment performed to assess the impact of Sl gasmin silencing on immune response, we always checked the level of gene silencing on day 1 of 5th instar larvae (t = 37.873, P<0.0001, df = 46, n = 24) (Fig 7A). Gene silencing did not interfere with the nodulation response (i.e., multicellular aggregation of haemocytes that entrap a large number of microorganisms) induced against either bacteria (Fig 7B) or yeast (Fig 7C) (E. coli: t = 0.501, P = 0.6208, df = 26, n = 14; S. cervisiae: t = 0.386, P = 0.7028, df = 26, n = 14). Similarly, encapsulation (i.e., formation of a multilayered capsule of haemocytes around the non-self object) and melanization of chromatographic beads injected into experimental larvae were not affected by gene silencing (t = 0.674, P = 0.5056, df = 22 n = 14) (Fig 7D). In contrast, phagocytosis of bacteria was strongly inhibited in experimental larvae treated with Sl gasmin dsRNA, as their haemocytes were almost completely unable to internalize either Gram-negative (E. coli) (t = 13.610, P<0.0001, df = 19, n = 11) (Fig 7E) or Gram-positive (S. aureus) bacteria (t = 10.725, P<0.0001, df = 20, n = 11), within 10 min or 30 min of incubation, respectively (Fig 7F). We focused on such a short time in order to detect any precocious enhancement of one of the immune barriers most rapidly activated by foreign invaders [41–43]. To check if the disruption of phagocytosis was due to a negative effect of gene silencing on cytoskeleton architecture and dynamics, the haemocyte spreading/adhesion and the distribution/polymerization of actin were investigated, as readouts of their immunocompetence. Haemocytes extracted from both control and silenced larvae showed identical levels of aggregation and adhesion on glass surfaces (Fig 8A and 8D), large nuclei (Fig 8B and 8E), and a similar network of polymerized actin, underlying plasma membrane protrusions, such as lamellipodia (sheet-like protruding two-dimensional lobes of the plasma membrane at the leading edge of the cell) and filopodia (protruding one-dimensional microspikes of the cell membrane) extensions (Fig 8C and 8F). These findings demonstrated that the adhesion properties of haemocytes, which strongly depend upon proper actin cytoskeletal dynamics, were completely unaffected by gene silencing. To assess whether the Sl gasmin silencing had any impact on S. littoralis humoral immune responses, we analyzed the transcription profiles of genes encoding humoral effectors secreted by haemocytes in the haemolymph, including antimicrobial peptides (AMP) and lysozyme [44, 45], following immune challenge with different microorganisms. No negative effects of Sl gasmin silencing on the ability of haemocytes to mount a humoral response were detected. Indeed an increase of transcript level of humoral effectors upon immune challenge was observed both in silenced and control larvae (Fig 9, see S1 Table for statistics). Collectively, these results indicate that Sl gasmin exerts a key-role in the modulation of phagocytosis by larval haemocytes. The presence of a signal peptide in Sl gasmin suggested that it is likely secreted by haemocytes and exerts its role in the haemolymph plasma. To investigate this hypothesis, the putative presence of Sl gasmin in the haemolymph plasma and its changes in response to gene silencing were assessed by liquid chromatography-tandem mass spectrometry in the multiple reaction mode (LC-MRM/MS). Specific peptides belonging to Sl gasmin sequence were selected (S2 Table) and monitored by MRM analysis of the tryptic digest of haemolymph samples, leading to the unambiguous identification of the protein. The observed changes of Sl gasmin transcript level induced by gene silencing (One-Way ANOVA: F(4,40) = 401.6, P<0.0001, n = 9, df = 44) (Fig 10A) were perfectly mirrored by changes in the abundance of the protein in the haemolymph plasma (Kruskal-Wallis test: KW = 36.28, P<0.0001, n = 8) (Fig 10B). Moreover, the injection of E. coli into control larvae induced a steep increase of Sl gasmin transcription (Fig 10A) and of the encoded protein titer in the haemolymph plasma (Fig 10B). The presence of Sl gasmin protein in the haemolymph plasma, associated with (1) the increase of its titer triggered by immune challenge, and (2) the phagocytosis failure in its absence strongly suggested a functional role of Sl gasmin in mediating pathogen recognition and subsequent phagocytosis by haemocytes. Indeed, Sl gasmin could act as an opsonization factor, i.e., a molecule that coats pathogens and mediates their recognition and suppression by immune cells. To test this hypothesis, a rescue experiment of phagocytosis was carried out by incubating the haemocytes obtained from silenced larvae (unable to perform phagocytosis) in the plasma from control larvae: this restored phagocytosis of bacteria in vitro (t = 11.67, P<0.0001, n = 14, df = 26) (Fig 10C). In contrast, haemocytes obtained from control larvae did not show phagocytic activity in vitro when resuspended in the plasma obtained from silenced larvae, which contained much lower levels of Sl gasmin. This experiment allows to conclude that (1) Sl gasmin is an essential element required for pathogen recognition and subsequent phagocytosis by haemocytes, and (2) the haemocytes of both control and silenced larvae are fully functional. To further clarify the role of Sl gasmin in the haemolymph, the hypothesis of a direct interaction between the protein and the pathogen was investigated by a proteolytic shaving approach coupled with LC-MRM/MS, which has previously been successfully used to identify host proteins interacting with invading bacteria [46]. For this purpose, intact E. coli cells were incubated with the plasma of immune primed S. littoralis larvae, which contain high levels of Sl gasmin. After incubation, the bacterial surface was proteolytically shaved to release putative Sl gasmin peptides, demonstrating the occurrence of a direct interaction of the protein with bacterial cells. The released peptide mixtures were then analysed by LC-MRM/MS, monitoring specific peptides selected from Sl gasmin sequence (S2 Table). The level of Sl gasmin adhering to E. coli cells was substantially and significantly higher (t = 7.969, P<0.0001, n = 5, df = 7) than that recorded in control plasma used for incubating the bacteria (Fig 10D). Hence, Sl gasmin present in the haemolymph plasma sticks on the surface of invading bacteria and is essential to mediate bacteria recognition and subsequent phagocytosis reaction by haemocytes. Bacteria incubated with buffer allowed the exclusion of false positive results. A homologue of the bracoviral gene BV2-5 was recently found in the genome of S. exigua (named gasmin) [25]. Here we have found an additional homologue of BV2-5 in the genome of S. littoralis (Sl gasmin), and observed that is highly expressed in larval stages exposed to an immune challenge. The results of the present study, based on complementary functional and molecular evidence, unequivocally demonstrate that this gene encodes an opsonizing factor triggering a fast phagocytic response. As already suggested [25], gasmin has been likely acquired by a basal ancestor of Spodoptera genus and maintained in distant species which separated about 24 MYA (S. littoralis and S. exigua) [47]. This suggests that there is a strong positive selection favoring the fixation of this gene, which, however, cannot be found in some related species (S. frugiperda), where the gene has been lost during evolution [25]. The important immune function of this gene, acquired through a mechanism of HGT mediated by a bracovirus, which acts as a unique gene vector [6, 25], reinforces the defense toolkit of S. littoralis. The integrative properties of BV have likely allowed the horizontal transfer of this viral gene. Indeed, parasitism on non-host larvae escaping death and accidental integration of BV sequences into the germ line allowed the stable integration into the moth genome and its transfer over generations [25]. This is the likely scenario in which several bracovirus genes have been acquired by the monarch butterfly (Danaus plexippus), the silkworm (Bombyx mori), the beet armyworm (S. exigua) and the fall armyworm (S. frugiperda), even though an adaptive advantage for these recipient insects has not been unequivocally demonstrated [25, 30]. To broaden the survey of gasmin-like genes present in insects and bracoviruses, we retrieved all candidate sequences available in the public databases, which showed high similarity to Sl gasmin (in total 17). These genes show great variability both in bracoviruses and insects, and the phylogenetic analyses fail to conclusively resolve any relationships (particularly for the S. litura homologue). The topology of the recovered phylogenetic tree might be the result of HGT and gene loss events, or other large modifications which led to a rapid evolution of the gene under positive selection, as expected for transferred genes response to new genetic and environmental contexts [10]. The gasmin genes are no exceptions, as most of the retrieved sequences have largely diverged. However, the striking similarity and conservation of the genomic and protein sequences present in S. littoralis and S. exigua strongly suggest a common origin. Further sampling of virus and insect gasmin homologues would be required to conclusively resolve this issue. Collectively, the experimental evidence we present on Sl gasmin indicates that its transcription in the haemocytes is rapidly boosted by immune challenge, and the encoded protein is an opsonizing factor. Once in the haemolymph, Sl gasmin provides a better protection by mediating a fast phagocytic reaction by the haemocytes, which is abolished by RNAi-mediated gene silencing. However, this evidence raises the question about the role of the viral homologue, which is nearly identical and likely functionally equivalent. We may reasonably speculate that in the case of parasitized host larvae, where the BV genes are actively expressed, the viral homologue of Sl gasmin may limit the risk of accidental bacterial infection, which would be detrimental both for the host and the developing parasitoid progeny. Indeed, 24 h after oviposition by C. congregata, the viral gene BV2-5 is highly expressed in the fat body and in the haemocytes of parasitized Manduca sexta larvae, where, concurrently, antimicrobial peptides are up-regulated, while genes involved in the phenol oxidase cascade and cellular immune responses are strongly down-regulated [48, 49]. Then, our results and the evidence available in the literature both point out that the immunosuppression induced by the wasp is selective, and provides protection to its juveniles by disabling only the responses effective against large intruders, leaving unaltered or even potentiating the antimicrobial barriers, in order to prevent dangerous secondary infections by pathogens. This could have been one of the functional constraints driving the evolutionary pathway of integration of this gene in the BV genome. Indeed, reinforcing the clearing capacity of the host haemolymph by enhancing the phagocytic activity appears to be an effective strategy when, concurrently, other barriers are very rapidly disrupted by maternal secretions injected at the oviposition by the parasitic wasp. This is an interesting hypothesis worth of further research efforts. The effective importance in vivo of this mechanism of haemolymph clearance is supported by the reduced capacity to withstand Bt-induced septicaemia we observed in S. littoralis larvae exposed to RNAi-mediated silencing of Sl gasmin. This implies that the acquired opsonizing factor is important in the modulation of phagocytosis efficiency in vivo. Phagocytosis is one of the first, most rapid and effective barriers of protection against microorganisms. Indeed, haemocytes immediately start their phagocytic activity in response to intrusion of pathogens into the haemolymph, and each cell is able to engulf hundreds of bacteria [41]. Phagocytosis initiation can be either direct, via specific haemocyte-surface receptors, or indirect, via opsonins that label pathogens, and thus make them recognizable by haemocyte-surface phagocytic receptors [40, 41, 50]. Opsonization factors in insects are still poorly studied [51–53]. In Diptera, opsonization-dependent phagocytosis is mainly mediated by thioester-containing proteins (TEP), a group of proteins that includes the α2-macroglobulins and complement factors that mediate phagocytosis in vertebrates, which have been reported both for Anopheles and Drosophila species [50–52, 54, 55]. TEP are proteins that specifically promote the phagocytosis of different sets of pathogens [52]. In Lepidoptera, several opsonins secreted by different tissues have been described [53, 56–58] and, in all cases, opsonization factors showed the capacity to bind a broad range of microbes, such as both Gram-positive and Gram-negative bacteria, and yeasts. Here we present evidence that Sl gasmin is an additional opsonization factor present in Lepidoptera, which has been transferred through a BV associated with a parasitic wasp attacking a non-permissive ancestor of Spodoptera species. We speculate that this passage of Sl gasmin confers effective and fast protection to the wasp eggs and juveniles soon after parasitization. This is relevant from a functional point of view, as the insertion of the ovipositor into the host body can be a route of infection, besides any other risk of accidental infection. A future comparative analysis of opsonizing factors active in the immune response by Spodoptera species lacking gasmin-like factors will likely shed light on the adaptive mechanisms that have favoured the fixation of this gene during evolution. We can reasonably speculate that the addition of gasmin to the array of pre-existing opsonizing factors results in a faster reaction and a broader capacity to engulf different types of microorganisms. These are likely possibilities worth of further research efforts. The contrasting functional evidence provided for S. exigua gasmin, generated, as already said above, by a biased experimental approach [25, 30], is problematic to interpret both from a mechanistic and evolutionary point of view. It is not easy to imagine that a protein triggering cytoskeleton disruption associated with reduced viral infection does not affect a number of other functions, with possible trades-off counteracting the fixation of its coding gene. This seems to be the case, as the proposed putative benefit of protection against viruses brings in reduced protection against other pathogens; indeed, haemocytes from an European strain of S. exigua, which do not produce functional gasmin, as it bears a truncated gasmin gene, fails to engulf bacteria by phagocytosis when infected with the recombinant baculovirus expressing gasmin [30]. This does not fit with what we have observed in the present study and it is difficult to imagine that we have so different functions for proteins sharing a very high level of sequence identity. While we performed a functional analysis by RNAi mediated gene silencing in vivo, for S. exigua gasmin a recombinant baculovirus was used for in vitro functional studies in an insect population that does not have genes encoding gasmin. This could have generated uncontrolled phenotypic responses, partly due to the concurrent effect of viral infection and the overexpression of a heterologous protein, that we do not even know if it is secreted outside of the infected cells [25, 30]. Moreover, data derived from these in vitro studies are somewhat difficult to interpret from an in vivo functional perspective. In particular, it is not easy to justify why gasmin, considered as an immune disrupter, is up-regulated in response to an immune challenge [30]; conversely, the results we present here are highly consistent with the proper immune role of Sl gasmin. In conclusion, our results demonstrate that in insects an immune function has been reinforced by HGT of a viral gene. The route of acquisition of the immune gene Sl gasmin, which is mediated by a viral symbiont of a parasitoid wasp attacking moth larvae, sheds light on a completely novel evolutionary pathway [6, 25]. The integrative properties of BV, which are powerful natural genetic engineers, pave the way for the transfer of sequences among completely different evolutionary lineages. In the case of immune genes, these may undergo intense selection in the context of the complex immune interplay among parasitic wasps harbouring the BV, their hosts and associated microbiota, eventually favouring the most efficient traits conferring protection to the system. The “jump” of this genetic material in moth species generates the addition of a very effective function, paradoxically obtained with the help of a parasitoid. Then our study demonstrates that, in multicellular organisms, essential physiological functions can be acquired and/or shaped by HGT and do not always derive from evolution of pre-existing genes. S. littoralis larvae were reared on artificial diet (47.3 g/l wheat germ, 67.3 g/l brewer’s yeast, 189 g/l corn meal, 6.8 g/l ascorbic acid, 0.75 g/l cholesterol, 0.5 g/l propyl 4-hydroxybenzoate, 3 g/l methyl 4-hydroxybenzoate, 1.3 g/l wheat germ oil, 33.8 g/l agar and 3 g/l vitamin mix (1.2 g/Kg vitamin B1, 2.6 g/Kg vitamin B2, 2.5 g/Kg vitamin B6, 40 g/Kg choline, 10 g/Kg pantothenic acid, 32 g/Kg inositol, 0.25 g/Kg biotin, 2.5 g/Kg folic acid, 5 g/Kg 4-aminobenzoic acid, 0.5 mg/Kg vitamin B12, 10 g/Kg glutathione, 2.1 g/Kg vitamin A, 0.25 g/Kg vitamin D3, 24 g/Kg vitamin E, 0.25 g/Kg vitamin K, 25 g/Kg vitamin C in dextrose)), at 25 ± 1°C, 70 ± 5% R.H., and under a 16:8 h light/dark period. S. littoralis larvae were anaesthetized on ice and surface-sterilized with 70% ethanol prior to dissection. Larval haemolymph was collected from a cut of the leg and haemocytes were separated from plasma by centrifugation for 5 min, at 500 × g, at 4°C. Midgut and fat body were isolated after cutting the larval body lengthwise, and the remaining body carcass separately collected. These samples (i.e., haemocytes, midgut, fat body and carcass) were immediately transferred into TRIzol reagent (Life Technologies, Carlsbad, CA, USA) and kept at -80°C until total RNA extraction, that was performed according to manufacturer’s instructions. DNA was extracted from haemocytes using the protocol described elsewhere [59], with minor modifications. The concentration of extracted RNA or DNA was assessed by measuring the absorbance at 260 nm, with a Varioskan Flash Multimode Reader (Thermo Scientific, Waltham, MA, USA), and sample purity was evaluated by assessing 260/280 nm absorbance ratio. RNA quality was checked by electrophoresis on 1% agarose gel. A partial Sl gasmin cDNA (FQ973054.1) was identified by BLAST analysis [60, 61] in a public database of expressed sequence tags (EST) from S. littoralis female antenna, using as query the full length cDNA sequence of S. exigua gasmin (KP406767.1). To isolate Sl gasmin ORF, total RNA extracted from haemocytes of S. littoralis 6th instar larvae was subjected to retro-transcription (Ambion RETROscript® kit—Life Technologies). Given the very high level of sequence identity to gasmin (Fig 1), a cDNA was obtained by PCR, using Phusion High-Fidelity DNA Polymerase (Fisher Scientific, Pittsburgh, PA, USA) and primers designed to amplify the whole gasmin ORF (gasmin ORF forward primer ATGTTGCCTATTACCATACTAACG, gasmin ORF reverse primer ATACTGGAATTGGACATATTTGAGC). PCR conditions were programmed for 30 s at 98°C; 40 cycles of [10 s 98°C, 30 s 60°C, 1 min 72°C] and 15 min at 72°C. After amplification, the obtained PCR product was separated by gel electrophoresis and the visible band of the expected size was purified with a Quick gel extraction & PCR purification COMBO Kit (Invitrogen, Carlsbad, CA, USA). The PCR product was cloned into Zero-Blunt TOPO vector (Zero-Blunt TOPO PCR Cloning Kit, Invitrogen), according to the manufacturer’s instructions. After transformation of One Shot TOP10 chemically competent E. coli cells (Invitrogen), the transformants were incubated overnight at 37°C on LB plates containing 50 μg/ml kanamycin. Bacterial colonies containing the fragment of the appropriate size were selected by colony PCR, using Phusion High-Fidelity DNA Polymerase (Fisher Scientific) and M13 Forward (-20)/M13 reverse primers (Invitrogen), and grown overnight in LB medium containing 50 μg/ml kanamycin. The plasmid DNA was extracted from 4 ml of bacterial culture using a Charge Switch-Pro plasmid miniprep Kit (Invitrogen), as instructed by the manufacturer and sequenced. The presence of an intron into Sl gasmin sequence was determined by PCR amplification of the total DNA extracted from S. littoralis haemocytes using Phusion High-Fidelity DNA Polymerase (Thermo Fischer Scientific) with specific primers (gasmin ORF forward primer ATGTTGCCTATTACCATACTAACG; gasmin intron reverse primer CAGGTGTCCGCATTCCACTGA). The length of PCR products was checked by electrophoresis on 1% agarose gels, before sequencing. Extensive similarity searches of complete and high-throughput genome sequence databases hosted by NCBI using TBLASTN, as well as of non-redundant protein databases (BLASTP), using the inferred S. exigua and S. littoralis protein sequences, allowed the identification of numerous potential homologues. These were manually annotated to identify probable coding and intronic regions. Preliminary alignments of inferred amino acid sequences (351 aa) were performed using Muscle [62] and manually filtered to identify homologues, which could be aligned essentially contiguously with the Spodoptera protein sequences. Alignments were manually refined, and ambiguously aligned regions were excluded using the program GBlocks [63], leaving XX amino acid positions for phylogenetic reconstruction using PhyML [64] as implemented in the program SEAVIEW [65], under the JTT amino acid substitution matrix, with 4 gamma distributed substitution rate categories (see S2 File). Bootstrap proportions were estimated using parameters optimized on the original ML tree (rate distribution parameter (alpha) = 1.45). To identify signatures of selection the FEL (Fixed Effects Likelihood) method was used [66], implemented at the DataMonkey website (http://datamonkey.org/) [67]. Nucleotide sequences of S. exigua, S. littoralis and C. congregata virus circle 25, including the intron, were aligned and the alignment was manually curated to maintain in-frame codon alignment. Selection was tested in the S. exigua—S. littoralis branches. Total RNA used for transcriptional analysis was isolated as described above. Relative expression of studied genes was measured by one-step qRT-PCR, using the SYBR Green PCR Kit (Applied Biosystems, Carlsbad, CA, USA), according to the manufacturer’s instructions. S. littoralis β-actin gene (Z46873) was used as endogenous control for RNA loading. Primer Express 1.0 software (Applied Biosystems) was used to design the primers used (S3 Table). Relative gene expression data were analyzed using the ΔΔCt method [68–70]. qRT-PCR for measurement of Sl gasmin expression was carried out using specific primers (S3 Table), designed to detect a region of Sl gasmin mRNA not included in the sequence targeted by the dsRNA (see “dsRNA synthesis” paragraph). For validation of the ΔΔCt method, the difference between the Ct value of Sl gasmin and the Ct value of β-actin transcripts [ΔCt = Ct (Sl gasmin)-Ct (β-actin)] was plotted versus the log of ten-fold serial dilutions (2000, 200, 20, 2 and 0.2 ng) of the purified RNA samples. The plot of log total RNA input versus ΔCt displayed a slope less than 0.1 (Y = 1.149+0.0133X, R2 = 0.0493), indicating that the efficiencies of the two amplicons were approximately equal. To analyze Sl gasmin expression in response to microbial challenge, S. littoralis 5th instar larvae, surface-sterilized with 70% ethanol and chilled on ice, received an intra-haemocoelic injection of 2 × 107 E. coli, 3 × 108 S. aureus or 2 × 107 S. cerevisiae cells, suspended in 5 μl of PBS (Phosphate Buffered Saline: 137 mM NaCl, 2.7 mM KCl, 10 mM phosphate buffer, pH 7.4). Injections were performed through the neck membrane, using a Hamilton Microliter 1701RN syringe (10 μl, gauge 26s, length 51 mm, needle 3). At the injection and at different time points after injection, experimental larvae (n = 12 for each experimental treatment) were dissected and haemocytes were collected and processed for total RNA extraction as described above; the relative expression of Sl gasmin was assessed by qRT-PCR. Total RNA extracted from haemocytes of S. littoralis 6th instar larvae was retro-transcribed (Ambion RETROscript kit, Life Technologies) and a 789 bp long Sl gasmin cDNA fragment was obtained by PCR, using the Sl gasmin dsRNA forward primer (GCCGGCATGTTGTCTATTACC) in combination with the Sl gasmin dsRNA reverse primer (TCCTTCCAGCTTCTGAGTCA). This cDNA fragment was used as template for a nested-PCR reaction, performed with primers containing at their 5’ ends the T7 polymerase promoter sequence (T7-Sl gasmin forward TAATACGACTCACTATAGGGAG-TTCGAGGATACAAGCAGAG; T7-Sl gasmin reverse TAATACGACTCACTATAGGGAG-GGATGCTCAGGATATCTGTTAC). The resulting PCR product was used as template to synthesize Sl gasmin dsRNA (522 bp long), using the Ambion MEGAscript RNAi Kit (Life Technologies), according to the manufacturer’s instructions. Control dsRNA, 500 bp long, was obtained from a control template supplied by the kit used. dsRNA preparations were quantified by measuring their absorbance at 260 nm with a Varioskan Flash Multimode Reader, and purity was evaluated by assessing 260/280 nm absorbance ratios. Products were run on 1% agarose gels to confirm their integrity. S. littoralis 4th instar larvae (1st day) were anaesthetized on ice and 1 μl of Sl gasmin dsRNA or control dsRNA (see above), dissolved in PBS, was poured into the lumen of the foregut by means of a Hamilton Microliter 1701RN syringe (10 μl, gauge 26s, length 51 mm, needle 2). dsRNA treatments consisted of one oral administration of 150 ng per day, for 3 days (from 4th to 5th instar). After the last dsRNA administration and prior to any experiment, haemocytes from 3–4 treated larvae were used for qRT-PCR analysis, to confirm the occurrence of gene silencing. Purified Cry1Ca protein was produced in a recombinant B. thuringiensis strain EG1081 (Ecogen Inc.). Prior to use, Cry1Ca was dialyzed overnight, at 4°C in 50 mM sodium carbonate buffer, pH 9.0. After dialysis, toxin concentration was determined by the Bradford assay [71], using bovine serum albumin as standard. Silenced and control larvae were singly isolated in multi-well plastic trays (Bio-Ba-32, Color-Dec, Italy), containing artificial diet, covered with perforated plastic lids (Bio-Cv-4, Color-Dec), and maintained under the rearing conditions reported above. For the first 3 days, the upper surface (1 cm2) of the artificial diet (0.3 cm3) was uniformly overlaid with 50 μl of purified Cry1Ca toxin, dissolved in 50 mM sodium carbonate buffer at pH 9.0. Control larvae were reared on artificial diet overlaid with 50 μl sodium carbonate buffer. Experimental larvae were maintained on artificial diet, replaced every 24 h, and daily inspected for survival, until pupation. To determine the 50% lethal concentration (LC50) of Cry1Ca toxin, the bioassay was carried out at 5 different concentrations of toxin and using 16 larvae for each experimental condition and control. Probit analysis [72], to determine LC50 values at day 10, 90% fiducial limits and toxicity increase ratio (TI) for each experimental treatment, was performed with the POLO-PC program (LeOra Software, Berkeley, CA). The assessment of the bacterial loads was performed as previously described [32]. Briefly, 6 h after the last Sl gasmin or GFP dsRNA administration, newly molted 5th instar larvae of S. littoralis, were exposed for 3 days to 2.7 μg/cm2 (corresponding to the LC50 of Cry1Ca in Sl gasmin silenced larvae). Both experimental groups included internal controls maintained on a toxin-free diet. On day 7, larvae were transferred to an untreated diet and, 24 h later, the midgut and haemolymph were separately collected under a horizontal laminar flow hood as described above. Experimental samples were obtained by pooling 10 larvae. The experiment was repeated 3 times. Changes over time in the relative bacterial load in the midgut (n = 7 for each sampling point) and haemolymph (n = 8 for each sampling point) samples were determined by qRT-PCR, measuring the transcript level of 16S rRNA (AJ567606.1) to assess the impact of Bt toxin and Sl gasmin silencing on bacterial proliferation. The qRT-PCR was performed as described above, using S. littoralis β-actin as reporter gene (primers used are reported in S3 Table). The impact of gene silencing on cellular immune responses was assessed by scoring its effect on encapsulation, nodulation and phagocytosis. Encapsulation and nodulation responses were assessed as previously described [31, 32]. CM Sepharose fast flow chromatography beads (Pharmacia), suspended in PBS, were injected into the haemocoel of S. littoralis larvae using a Hamilton Microliter 1702 RN syringe (25 μl, gauge 22s, length 55 mm, needle 3). After 24 h, beads were recovered upon larval dissection and scored to evaluate their encapsulation rate, which was expressed with the encapsulation index (E.I. = [Σ (encapsulation degree × total beads of this degree)/ total beads × 4] × 100), that takes into account both the encapsulation degree of each recovered bead (0—no cells adherent to the beads, 1—up to 10 adherent cells, 2—more than 10 adherent cells but no complete layer around the bead, 3—one or more complete layers without melanization, 4—one or more complete layers with melanization) and the relative abundance of beads with a given encapsulation degree [73]. For the nodulation assay, 12 h after the last dsRNA administration, S. littoralis larvae, surface-sterilized with 70% ethanol and chilled on ice, received an intra-haemocoelic injection of 5 μl of a PBS suspension of 2 × 106 E. coli cells, or 2 × 107 S. cerevisiae cells. Injections were performed through the neck membrane, using a Hamilton 1701 RN SYR (10 μl, 26s gauge, 55 mm long, point style 3). A thoracic leg was cut 18 h after injection, and the exuding haemolymph was collected and immediately diluted into an equal volume of ice-cold MEAD anticoagulant buffer (98 mM NaOH, 145 mM NaCl, 17 mM EDTA, and 41 mM citric acid, pH 4.5). The haemocyte nodules occurring in the haemolymph samples were counted under a light transmitted microscope at 400× magnification (Axioskop 20, Carl Zeiss Microscopy, Germany), using a Bürker chamber. When an intense immune response gave rise to large aggregates of nodules difficult to count separately, the number of distinct nodules observed was arbitrarily doubled, because the percentage of nonwhite pixels measured (ZEN software; Carl Zeiss Microscopy) on the large aggregates was on average twice that measured on a bright-microscopy field containing discrete nodules and free haemocytes. To measure phagocytosis competence of S. littoralis haemocytes, an in vitro assay was performed as described in [32] with minor modifications. Briefly, haemolymph samples were collected from a cut of the leg into ice-cold PBS (1:1 v/v) and added with an equal volume of a PBS suspension of 2 × 106 fluorescein conjugated E. coli cells (K-12 strain BioParticles, fluorescein conjugate, Invitrogen) or 2 × 107 S. aureus (Wood strain, BioParticles fluorescein conjugate, Invitrogen). After incubation with E. coli (10 min) or with S. aureus (30 min), samples were loaded into a Bürker chamber, where total and fluorescent haemocytes were counted under a fluorescence microscope (Axioskop 20). Prior starting incubation experiments, vital staining with trypan blue was used to routinely check the viability of collected haemocytes. A haemolymph aliquot was mixed with 0.4% (w/v) trypan blue (2:1 v/v), prior to count viable and dead cells under a light transmitted microscope (Axioskop 20), using a Bürker chamber. Haemocyte samples with a viability rate lower than 98% were discarded. For rescue experiments with haemocytes from silenced larvae, haemolymph samples were extracted from S. littoralis larvae chilled on ice, 24 h after the last dsRNA administration (Sl gasmin dsRNA or control dsRNA). Samples were centrifuged 5 min at 500 × g, at 4°C. The plasma was kept on ice, haemocytes were resuspended in PBS and centrifuged as previously described. PBS was then removed and haemocytes from larvae treated with Sl gasmin dsRNA were resuspended in the plasma isolated from larvae treated with control dsRNA, while haemocytes from larvae treated with control dsRNA were resuspended in the plasma isolated from larvae treated with Sl gasmin dsRNA. Then, the phagocytosis by haemocytes was evaluated as described above. The humoral immune response, as affected by gene silencing, was assessed by measuring the transcript level of genes encoding antimicrobial peptides and lysozyme, in response to injections of different microorganisms, as previously described [32]. Briefly, 6 h after the last dsRNA administration, S. littoralis larvae, surface-sterilized with 70% ethanol and chilled on ice, received an intra-haemocoelic injection of 2 × 107 E. coli or S. aureus cells, or 3 × 108 S. cerevisiae cells, suspended in 5 μl of PBS. Injections were performed through the neck membrane with a Hamilton 1701 RN SYR (10 μl, gauge 26s, length 55 mm, needle 3). At the time of injection and 18 h after injection, larvae (n = 8 for each experimental sample) were dissected and haemocytes, midgut, and fat body were collected and processed for total RNA extraction, as described above. The relative expression of attacin 1 (FQ971100.1), gloverin (FQ965511.1), and lysozyme 1a (FQ961692.1) were thus assessed by q-RT-PCR as described above. Primers used are reported in S3 Table. Newly moulted 5th instar larvae of S. littoralis, treated with Sl gasmin dsRNA or control dsRNA, as described above, were surface-sterilized with 70% ethanol and chilled on ice. Larval haemolymph from individual larvae was collected from a cut of the leg and placed on glass slides for 10 min, to allow the haemocytes to settle and attach to the glass. Haemolymph was then carefully removed and haemocytes rinsed 3 times with PBS. Attached cells were fixed for 10 min in 4% paraformaldehyde in PBS, washed 3 times in PBS and permeabilized for 4 min with 0.1% Triton-X100 in PBS. Haemocytes were washed 3 times in PBS and then incubated for 20 min with 4 μg/ml TRITC-phalloidin (Tetramethylrhodamine B isothiocyanate-phalloidin). After 3 rinses in PBS, the samples were mounted in Vectashield Mounting Medium with DAPI (Vector Laboratories) and examined under a fluorescence microscope (ZEISS Axiophot 2 epifluorescence microscope). The observations have been performed in 3 different experiments and considering at least 10 randomly selected microscopic fields for each experimental condition. To detect the presence of Sl gasmin in the plasma, haemolymph was centrifuged as described above to remove the haemocytes, and the supernatant (plasma) was stored at -80°C. Samples were then dissolved in denaturant buffer (6 M urea, 10 mM EDTA, 300 mM Tris, pH 8.0) containing dithiothreitol (10-fold molar excess on the Cys residues) at 37 °C for 2 h, before the addition of iodoacetamide (IAM) to perform carboamidomethylation, using 5-fold molar excess of alkylating agent on thiol residues. The mixture was incubated in the dark at room temperature for 30 minutes and the product was purified by Chloroform/Methanol/H2O precipitation. Supernatants were removed and the pellets were dried. Digestion of the protein mixture was carried out in 10 mM ammonium bicarbonate (AMBIC), using trypsin at a 50:1 protein:enzyme mass ratio. The samples were incubated at 37°C for 16 h and dried after acidification (10% HCOOH in water). To eliminate any impurity, samples were suspended in 200 μl of 100 mM AMBIC, filtrated by centrifugal filter units (0.22 μm) and dried in a speed-vac concentrator. Samples were evaporated and suspended in 50 μl of 0.1% HCOOH in water. Peptide mixtures were analyzed by LC-MRM/MS analysis using a Xevo TQ-S (Waters, Milford, MA, USA) with an IonKey UPLC Microflow Source coupled to an UPLC Acquity System (Waters), using an IonKey device. For each run, 1 μl peptide mixture was separated on a TS3 1.0 mm × 150 mm analytical RP column (Waters) at 60°C, with a flow rate of 3 μl/min using 0.1% HCOOH in water (LC-MS grade) as eluent A, and 0.1% HCOOH in acetonitrile as eluent B. Peptides were eluted (starting 1 min after injection) with a linear gradient of eluent B in A, from 7% to 95% in 55 min. The column was re-equilibrated at initial conditions for 4 min. The MRM mass spectrometric analyses were performed in positive ion mode using a MRM detection window of 0.5–1.6 min per peptide; the duty cycle was set to automatic and dwell times were minimal 5 ms. Cone voltage was set to 35V. The selected transitions and the collision energy for each Sl gasmin peptide are reported in S2 Table. To determine whether Sl gasmin is able to bind to the surface of bacteria, plasma samples obtained from S. littoralis 5th instar larvae (n = 20) were added to an equal volume of MEAD and incubated with an equal volume of E. coli suspension in PBS (4 × 106 cells for each μl of haemolymph) for 1 h. The suspension was then centrifuged for 10 min, at 12,000 × g, at 4°C and the pellet resuspended in 2 ml of 10 mM phosphate buffer, 45 mM NaCl, pH 7.4. Centrifugation and resuspension were repeated and the bacterial pellet as well as supernatants were frozen in liquid nitrogen and stored at -80°C until use. In control experiments bacteria were incubated with PBS and MEAD (1:1:1 v/v/v). Samples were submitted to reduction, alkylation and tryptic digestion as described above. After the preparation step they were processed and analyzed by LC-MRM/MS, as previously described. LC-MRM/MS analyses were performed on 3 technical replicates for each biological replicate and the average of these multiple measurements was used for data analysis. The data obtained represent the average value of total ion current associated to each transitions for the selected peptides. Unless differently indicated, all reagents were provided by Sigma-Aldrich, Italy. Data were analyzed using Prism (GraphPad Software Inc. version 6.0b, San Diego, CA, USA) and SPSS (IBM SPSS Statistics, Version 21, Armonk, NY) software. The comparison between 2 experimental groups was done using the unpaired Student’s t test, while in the case of more than 2 experimental groups, One-Way ANOVA. Two-Way ANOVA was carried out on AMP and lysozyme 1A immune induction experiments, with RNAi treatment and bacterial injection as factors, while a Three-Way ANOVA was carried out for bacteria relative quantification, with dsRNA treatment, time and Cry1Ca toxin exposure as factors. When necessary transformation of data was carried out, to meet the assumption of normality. Levene’s test was carried out to test the homogeneity of variance. When significant effects were observed (P value<0.05), Bonferroni’s post-hoc test was used. When one of the assumptions was not met, even after the transformation of the data, Kruskal-Wallis one-way ANOVA (non-parametric ANOVA) test was employed.

10.1371/journal.pgen.1004888

Notch Signaling Mediates the Age-Associated Decrease in Adhesion of Germline Stem Cells to the Niche

Stem cells have an innate ability to occupy their stem cell niche, which in turn, is optimized to house stem cells. Organ aging is associated with reduced stem cell occupancy in the niche, but the mechanisms involved are poorly understood. Here, we report that Notch signaling is increased with age in Drosophila female germline stem cells (GSCs), and this results in their removal from the niche. Clonal analysis revealed that GSCs with low levels of Notch signaling exhibit increased adhesiveness to the niche, thereby out-competing their neighbors with higher levels of Notch; adhesiveness is altered through regulation of E-cadherin expression. Experimental enhancement of Notch signaling in GSCs hastens their age-dependent loss from the niche, and such loss is at least partially mediated by Sex lethal. However, disruption of Notch signaling in GSCs does not delay GSC loss during aging, and nor does it affect BMP signaling, which promotes self-renewal of GSCs. Finally, we show that in contrast to GSCs, Notch activation in the niche (which maintains niche integrity, and thus mediates GSC retention) is reduced with age, indicating that Notch signaling regulates GSC niche occupancy both intrinsically and extrinsically. Our findings expose a novel role of Notch signaling in controlling GSC-niche adhesion in response to aging, and are also of relevance to metastatic cancer cells, in which Notch signaling suppresses cell adhesion.

Aging is frequently associated with a decline in the size of stem cell pools, but little is known regarding the molecular mechanisms underlying this process. Here, we report that Notch signaling is increased in GSCs as they age, and this promotes their removal from the niche in an E-cadherin dependent manner. In contrast to GSCs, niche cells exhibit decreased Notch signaling with age; Notch signaling in these cells controls niche integrity, and consequently GSC retention. While Notch signaling in the niche is regulated by insulin signaling, Notch signaling in GSCs is controlled by Sex lethal, an RNA-binding protein. These results imply that Notch signaling is regulated in a cell-type-dependent manner, and coordination between GSCs and their niche facilitates the removal of cells from the niche during the aging process.

Age-associated depletion of stem cell pools has been reported for mammalian satellite stem cells, Drosophila male and female GSCs, and C. elegans GSCs [1]–[4]; however, the mechanisms underlying such depletion remain unknown. The stem cell niche houses stem cells and maintains their cell identity, by providing physical contact and stemness factors, respectively [5]. In addition to the niche, stem cell-intrinsic factors also regulate stem cell function [6], [7]. These signals are tightly coupled, and regulate stem cells to fit the current needs of the organism. During aging, diminished niche function leads to stem cell loss [1]; on the other hand, it is unknown whether stem cells influence their own attachment to the niche as they age. Moreover, it is also unclear how niche cells coordinate with stem cells in response to aging. Drosophila is a small organism with a short life span; such properties, combined with the availability of powerful genetic approaches, making this organism eminently suitable for investigations into cellular and organismic responses during aging. In addition, the Drosophila ovary houses well-characterized GSCs and their niche (Fig. 1A) [8]. These advantages make the Drosophila ovary an excellent model in which to study the communication of stem cells with themselves and the surrounding environment. One Drosophila ovary is composed of 16 to 20 ovarioles, which are the basic functional unit of egg production [9]. The anterior-most structure of the ovariole is called the germarium; the tip of the germarium contains the GSC niche, which is composed of terminal filament, cap cells, and anterior escort cells [10], [11]. GSCs make direct contact with cap cells, a major niche component, through E-cadherin-mediated cell-cell adhesion [12]; the GSC fusome, an organelle with a membranous-like structure, is juxtaposed to the interface between cap cell and GSC [13]. GSC division gives rise to a cystoblast, which subsequently undergoes four rounds of incomplete division to form a 16-cell cyst, in which the cells are interconnected with branched fusomes [9]. The 16-cell cyst is then surrounded by a layer of follicle cells, and eventually develops into a mature egg. The Notch signaling pathway is highly conserved, and plays critical roles in the regulation of stem cells in different systems [14], [15]. In Drosophila, it controls the maintenance of niche cap cells [16], [17]; however, whether it plays a role in GSCs is not clear. Drosophila has one Notch receptor (encoded by Notch, N), which is a single-pass transmembrane protein. Upon binding to one of its ligands, the cleaved intracellular domain of Notch translocates to the nucleus, where it regulates transcription of gene targets. The extracellular domain of Notch is composed of several epidermal growth factor (EGF)-like repeats, which must be highly glycosylated for proper function of the protein [18]. Fringe (encoded by fng) is a glycosyltransferase that adds N-acetylglucosamine onto this domain, thereby modulating the binding of Notch to its ligands. The insulin/insulin-like growth factor (IGF) signaling pathway is also evolutionarily conserved; this pathway controls various stem cell types in response to nutritional input and aging [8], [19]–[22]. We have previously shown that insulin signaling controls Fringe-mediated Notch activation in the GSC niche, and this is required to maintain niche integrity, thereby supporting the retention of GSCs in the niche [23], [24]. In addition, insulin signaling also directly controls GSC proliferation [25], [26], but it is not clear if Notch signaling is involved in this process. In this article, we demonstrate that insulin and Notch signaling have independent roles in regulating GSCs, in contrast to the earlier observation that insulin signaling regulates Notch activation in niche cells to support their maintenance [23], [24]. While insulin/IGF signaling is required for GSC proliferation and maintenance, Notch signaling underlies the diminished ability of GSCs to occupy the niche during aging. In young ovaries, Notch signaling in GSCs is low, allowing them to persist in the niche; as ovaries age, Notch signaling in GSCs is elevated, resulting in the loss of GSCs from the niche. In contrast to GSCs, Notch signaling in niche cells (which is required for niche integrity) is decreased with age, thus contributing to age-induced GSC loss. We therefore demonstrate that both intrinsic and extrinsic Notch signaling control the niche residency of GSCs during aging. Notch signaling has been previously reported to be required for the integrity of the Drosophila female GSC niche, and thus contributes to GSC maintenance [16], [17]. We observed that Notch signals are also present, albeit weakly, in GSCs (S1 and S2 Fig.)[24]; however, the function of Notch signaling in GSCs is unknown. To address this question, we used mitotic recombination to generate GSCs with mutations in N (indicated by the absence of GFP) (Fig. 1B and C). We first addressed whether Notch signaling is required for GSC division (Fig. 1D). We counted the number of wild-type versus mutant cystoblasts and cysts present in N mosaic germaria containing one wild-type and one mutant GSC. Crucially, the relative numbers of wild-type and mutant cystoblasts or cysts were unaffected by early germline death (S3 Fig.). Because each cystoblast or cyst is derived from one GSC division, the ratio of mutant to wild-type progeny is a measure of their relative division [25]. The number of progeny derived from control GSCs labeled with GFP was approximately equal to those without GFP in mock mosaic germaria, resulting in a relative division rate equal to approximately 1.0 at one week and two weeks after clone induction (ACI). Division was unaffected in GSCs homozygous for N55ell, a null allele [23] (Fig. 1D); these mutant GSCs exhibited decreased levels of Notch signaling as compared to their neighboring control GSCs (S4 Fig.). We next asked if Notch signaling controls GSC maintenance, by examining the number of germaria carrying N55ell mutant GSCs over time (S1 Table). At three weeks, 89% of FRT19A control germaria (n = 737) retained at least one wild-type control GSC generated from the first week, indicating that up to 11% of GSCs undergo turnover naturally, consistent with an earlier report [27]. Most mutant GSCs, however, were retained at 3 weeks ACI in N55e11 mutant mosaic germaria (115±7%, n = 392), indicating that they are resistant to the age-dependent attrition displayed by wild-type GSCs. These findings demonstrate that Notch signaling is not directly required for GSC division; instead, it may be important for long-term GSC maintenance. Notch activation mediated by Fringe (a glycosyltransferase modulating Notch for its binding to ligands) is required for the maintenance of Drosophila GSC niches [28], and is regulated by insulin signaling in response to diet [23], [24]. Insulin signaling is known to directly control GSC division [25]; indeed, insulin receptor (dinr339) mutant GSCs divided three times slower than wild-type GSCs (Fig. 1D). We also found that only 39.8±7% of dinr339 mutant mosaic germaria carried dinr339 mutant GSCs at three weeks ACI (S1 Table), implying a direct role of insulin signaling in GSC maintenance with age. However, fringe mutant GSCs exhibited decreased levels of Notch signaling and behaved similarly to N55e11 mutant GSCs (S4 Fig. and S1 Table), indicating that Fringe-mediated Notch signaling and insulin signaling play distinct roles in GSCs, and Notch signaling in GSCs and niche cells is regulated through different processes. Surprisingly, a 44% net increase in the proportion of N55e11 mutant GSCs was observed at three weeks ACI (Fig. 1E), while no obvious change was observed in the proportion of wild-type control GSCs in FRT19A control germaria (Fig. 1E and S1 Table). Because a niche usually houses two to three GSCs, this relative increase may reflect a loss of neighboring normal GSCs, rather than an increase in the number of N55e11 mutant GSCs. Indeed, the proportion of N55e11 mosaic mutant germaria carrying a mixture of GFP-negative and GFP-positive GSCs (i.e., partial GSC clones) decreased from 77% to 30%, while the proportion of germaria in which all GSCs were mutant (i.e., full GSC clones) increased from 23% to 70% by three weeks ACI (Fig. 1F). In FRT19A and FRT80B mock mosaic germaria, only 2% and 27% of the observed increase, respectively, was due to the natural loss of neighboring marker-positive GSCs. We confirmed these results using GSCs homozygous for N5419 (a genetic null allele [29]), and also observed similar phenomena in fng[13] and fng[L73] mutant GSCs (Fig. 1E and F, and S1 Table). These results indicate that GSCs with low Notch signaling are more likely to stay in the niche, while those with higher levels of Notch signaling are more likely to be excluded. To test the above hypothesis, we generated genetic mosaic females carrying GSCs mutant for NAXE2 (as identified by the absence of GFP), a hypermorphic allele [29]; these mutant GSCs exhibited increased levels of Notch signaling as compared to their neighboring control GSCs (S4 Fig.). As expected, NAXE2 mutant GSCs were lost faster than control GSCs (only 75% of NAXE2 mutant GSCs remained by three weeks ACI) (Fig. 1E and S1 Table). Similarly, constitutive activation of Notch signaling by overexpression of the Notch intracellular domain (NICD) in GSCs also accelerated their loss (discussed later). We consistently failed to detect apoptotic germ cells in germaria carrying N55e11 or fng[L73] mutant GSCs (see S3 Fig.), suggesting that the lost cells undergo differentiation. However, we cannot rule out the possibility that apoptotic cells were lost too rapidly from the germaria to be detected. We next investigated whether enhanced niche occupancy of N mutant GSCs is associated with increased Bone Morphogenetic Protein (BMP) signaling, a major pathway for GSC maintenance (Fig. 2). However, the levels of two well-established reporters of BMP signaling [30], phosphorylated (p) Mad (pMad) and Dad-lacZ, were unaffected in Notch signaling-defective GSCs, as compared to their neighboring control GSCs. Next, we hypothesized that increased retention of N mutant GSCs in the niche may arise from enhanced adhesive ability. E-cadherin is a cell-cell adhesion molecule required for GSC-niche adhesion [12]; we found that its expression was significantly increased in mutant GSC-niche junctions of N55e11 germaria as compared to neighboring normal GSC-niche junctions at one week (98±5 vs. 82±5 arbitrary units, P<0.05) and two weeks (91±7 vs. 71±6 arbitrary units, P<0.05) ACI (Fig. 3A–B). We also observed that the contact area between N55e11 mutant GSCs and the niche at one week ACI was greater than that between neighboring normal GSCs and the niche (17±1.0 vs. 13±0.4 µm2, P<0.01) (Fig. 3A′ and D). In contrast, E-cadherin expression and contact area were similar between GFP-positive and -negative GSC-niche junctions in mock mosaic germaria. Germaria containing fng[L73] or fng[13] mutant GSCs exhibited the same properties as those carrying N55e11 mutant GSCs, but not until two weeks ACI (S5 Fig.); this reflects the relatively moderate effects of these mutations on GSC competitiveness (S1 Table). Conversely, both E-cadherin expression and contact area were decreased at the boundary of N gain-of-function mutant (NAXE-2) GSCs and their niches at 3 weeks ACI (Fig. 3C and E); at this time point, NAXE-2 mutant GSCs were lost more rapidly from the niche than their neighboring control GSCs (see Fig. 1E). These results suggest that low levels of Notch signaling in GSCs increases GSC adhesion to the niche, thereby allowing the cells to out-compete their normal wild-type GSCs. To investigate the requirement of E-cadherin for Notch signaling-mediated niche occupancy of GSCs, we decreased E-cadherin expression in N55e11 mutant GSCs, and then determined the proportion of N55e11 mutant GSCs in mosaic germaria (Fig. 4). To generate N55e11 mutant GSCs expressing RNAi targeting E-cadherin (shgRNAi), we utilized a combination of FLP-mediated recombination and the binary UAS-GAL4 expression system (Fig. 4A). For these experiments, Drosophila females carried a tub-GAL80 transgene (encoding a suppressor of GAL4 binding to the UAS element) located distal to FRT19A on one X chromosome, and the N55e11 mutant allele distal to FRT19A on the other X chromosome. The product of the nos-GAL4 transgene was used to drive UAS-shgRNAi and UASp-mCD8-GFP specifically in germ cells [31]. The tubulin promoter enables ubiquitous expression of GAL80, and thus GFP and shgRNAi are suppressed in all cells of the germarium. After recombination, GSCs with two copies of the N55e11 mutant allele would lose tub-GAL80, allowing nos-GAL4 to drive expression of GFP and shgRNAi (Fig. 4B–D). Consistent with the above findings (see Fig. 1), the proportion of N55e11 mutant GSCs was increased as early as two weeks ACI, as compared with the control (Fig. 4E). Conversely, the proportion of N55e11 mutant GSCs expressing shgRNAi remained at a similar level to controls (Fig. 4E), indicating that Notch signaling controls GSC competitiveness in an E-cadherin-dependent manner. E-cadherin expression is reduced at the junction between GSCs and their niche as germaria age; overexpression of shg in GSCs delays age-dependent GSC loss [32], raising the possibility that Notch signaling may be increased in GSCs with age, and mediate suppression of E-cadherin expression. To test this hypothesis, we examined Notch signaling in the GSCs of young and aged germaria using E(spl)m7-lacZ (Fig. 5A–D). Expression of E(spl)m7-lacZ was approximately 3-fold greater in niche cap cells (n = 154) than in GSCs (n = 67) in one-week-old germaria; E(spl)m7-lacZ expression was 68% lower in aged niche cap cells (n = 140) than in young cap cells (n = 154, P<0.001), supporting the known role of Notch signaling in niche maintenance [23], which declines with age. However, we observed a 20% increase of E(spl)m7-lacZ expression in aged GSCs (n = 54, P<0.05) as compared to their younger counterparts. Similar observations were made using a different Notch signaling reporter, Su(H)Gbe-lacZ (S6 Fig.) [33]. We also performed microarray analysis of age-dependent changes in transcriptional profiles in GSCs. To this end, we used bam mutant females in which the number of GSCs in germaria is increased (Bam is required for GSC differentiation) [34], [35]. We isolated GSCs from 1-, 5-, and 8-week-old bam mutant germaria, and performed microarray analyses on three biological replicates (see ‘Materials and Methods' for details). We report that expression of several Enhancer of split (E(spl)) genes was increased with age in bam mutant GSCs, indicating an elevation of Notch signaling (Fig. 5E and E′). These results clearly demonstrate that Notch signaling is up-regulated in GSCs with age. Notch activation requires direct contact between the receptor and its ligands, Delta and Serrate, which are mainly produced in the GSC niche [24]. However, we observed that bam mutant GSCs which escaped from the niche still exhibited increased Notch signaling as they aged, suggesting that direct contact between GSCs and the niche is not required for such signaling. Expression levels of Delta (encoded by Dl) and Serrate (encoded by Ser) were previously reported to be too low to be detected in the germaria [24]; in this study, we did not observe increases in expression of Dl or Ser in GSCs or their niche (analyzed using two well-characterized reporter lines) (S7 Fig.)[24]. These results suggest that the increase in Notch signaling in GSCs with age is unlikely to be due to increased expression of Notch ligands. It is known that Notch signaling is negatively regulated in the developing wing and ovarian epithelial follicle cells by Sex lethal (Sxl; encoded by sxl), a female specific RNA-binding protein, via post-transcriptional control of N [36]. We therefore investigated whether Sxl also controls Notch signaling in GSCs (Fig. 6). We first generated mosaic germaria carrying GSCs with two copies of sxlf2 (identified by the absence of GFP), a hypomorphic allele [37]. These sxlf2 mutant GSCs exhibited increased Notch signaling (as determined by examining the expression of E(spl)m7-lacZ) as compared to their neighboring control GSCs (Fig. 6A–C). Consistent with our finding that elevating Notch signaling in GSCs induces their loss from the niche, sxlfs3 (anothour hypomorphic allele [38]) mutant GSCs were rapidly lost by two weeks ACI (S1 Table). In addition, the use of sxlRNAi under the control of a germ cell-specific driver (nos-GAL4) to knock down Sxl expression in GSCs also resulted in a 47% increase of Notch signaling in GSCs (n = 65) as compared to controls (n = 74), P<0.01), without affecting Notch signaling in the niche (Fig. 6D–F). These results indicate that Sxl suppresses Notch signaling in GSCs. Of note, we found that ∼20% of sxlRNAi -knock down and sxlf2 mosaic germaria, and ∼8% of sxlfs3 mutant germaria contained cystoblasts with defects in differentiation (S8 Fig.); this finding is consistent with the known role of Sxl in the control of GSC differentiation [39]. Further, we found that expression of Sxl in the anterior-most germ cells (including GSCs) of germaria decreased with age (1-week-old: 45±1 arbitrary units, n = 88 vs. 7-week-old; 34±1 arbitrary units, n = 103, P<0.01) (Fig. 6G–I). Western blot analysis was used to show that Sxl expression in old ovaries is reduced as compared to that in young ovaries (S9A Fig.). Similar observations were also made in isolated GSCs from young and aged bam mutant females (S9B Fig.). Our results suggest that the induction of Notch signaling with age in GSCs is at least partly mediated by Sxl. We further hypothesized that the elevation of Notch signaling in older GSCs may help to promote GSC loss from the niche. To test this hypothesis, we used an N RNAi line or a constitutively-active form of N (NNICD) driven by nos-GAL4 to alter Notch signaling in the GSCs in the adult, and then examined the number of GSCs at different ages (Fig. 7). Overexpression of NNICD in GSCs decreased E-cadherin expression at GSC-niche junctions and accelerated aging-dependent GSC loss (Fig. 7A, B, D, and E). We consistently observed that elevation of Notch signaling or knock down of E-cadherin in GSCs using sxlRNAi or shgRNAi, respectively, also promoted GSC loss. However, although inhibition of Notch expression increased E-cadherin expression at the GSC-niche junction (Fig. 7A, C, and D), this was not sufficient to suppress age-dependent GSC loss (Fig. 7E), indicating that multiple factors are required for this process. Although our clonal analysis revealed that Notch signaling is not required for the maintenance of GSCs in the niche (see Fig. 1), GSCs were rapidly lost (40% decrease) from one-week-old females subjected to NRNAi-knock down; however, by two- to three-weeks, maintenance in knock-down females was better than that of wild-type females, with levels eventually becoming comparable to that in controls. We suspect that varying expression levels of GAL4 among GSCs in the niche (25 germaria were analyzed) created a competitive environment (Fig. 7E), resulting in GSC loss. However, after the first week, Notch expression in NRNAi-knock-down GSCs in the niche had dropped to the lowest recorded level, resulting in the loss of competitive behavior, together with an increase in E-cadherin expression at GSC-niche junctions; consequently, these GSCs exhibited enhanced retention. Above three weeks in age, other factors (like a reduction in niche-derived stemness signals) may function in parallel with Notch signaling to control GSC aging. We also confirmed our results using a second germline driver, maternal-tubulin-GAL4 (mat -GAL4). Unlike nos-GAL4, mat-GAL4 is expressed evenly in GSCs within the germaria (25 germaria were examined) (S10A Fig.). Although flies carrying a single copy of mat-GAL4 did not survive longer than 4 weeks after eclosion at the non-permissive temperature (29°C), we were still able to observe that forced Notch signaling accelerated GSC loss with age, and disruption of N slowed GSC loss with age (S10B Fig.). Collectively, our findings reveal that Notch signaling promotes GSC aging, and hence such signaling needs to be maintained at a relatively low level in young GSCs for their anchorage to the niche. Here, we have described a novel role for Notch signaling in the regulation of stem cell-niche contact during aging, in parallel with factors that control stem cell self-renewal (Fig. 7G). In young germaria, Notch signaling is low in GSCs, but high in the niche, thereby ensuring that GSCs are maintained in the niche (E-cadherin levels are high). However, Notch signaling in older germaria is elevated in GSCs but decreased in the niche, thereby reducing GSC-niche adhesion (E-cadherin levels are low). Our results indicate that Notch signaling is differentially regulated in GSCs (via negative regulation by Sxl) and their niche (via insulin signaling, which is decreased with age [23]), and mediates coordination between stem cells and niche for stem cell maintenance in response to aging. These results have broad significance for stem cell therapy, particularly when considered with the well-known effects of aging on reproduction [40] and the high degree of conservation of Notch signaling function [14]. Notch signaling plays critical roles in stem cell maintenance, differentiation, and aging [15]; however, the mechanisms by which Notch signaling controls stem cell aging remain unclear. It has been shown that Notch signaling controls myogenesis and neurogenesis via the control of skeleton muscle stem cells and neural progenitors; however, a reduction in Notch signaling impairs these stem cell functions during aging [41], [42]. Unlike muscle stem cells and neural progenitors, we found that Notch signaling is up-regulated in GSCs with age, and this suppresses GSC adhesion to the niche. Increased Notch signaling activity is also observed in mesenchymal stem cells (MSCs) of patients with Hutchinson-Gilford Progeria Syndrome (HGPS), a premature-aging disease; this increase is caused by abnormal activation of Lamin A, which induces nuclear defects and DNA damage [43], [44]. Increased Notch activation in MSCs forces their differentiation toward osteoblasts, but prevents their differentiation into adipocytes. It is possible that GSCs and MSCs share similar gene expression profiles, which are required for activation of Notch signaling in response to aging; this is supported by the detection of germinal transcripts in MSCs [45]. The number of stem cells gradually decline during aging [32], suggesting that the remaining GSCs in the niche are more competitive than those that are lost. It has been proposed that under normal conditions, cycling tissue stem cells behave as an equipotent population, in which the balance between differentiation and proliferation is achieved by frequent and stochastic stem cell loss and replacement (a process termed neutral competition), thereby maintaining tissue homeostasis [46], [47]. However, this balance is disrupted during aging, and, assuming that stem cell loss rate is higher than the speed of replacement, this results in stem cell loss. Here, we argue that aging induces the differences among stem cells in the niche which affect niche residency. For example, aged GSCs in the niche accumulate different levels of DNA damage [48], which may result from the accumulation of excessive reactive oxygen species (ROS); this may ultimately lead to GSC loss [32]. In addition, we also observed that GSCs in the niche exhibited various levels of enhanced Notch signaling (S11 Fig.); this suggests that GSCs with higher levels of Notch signaling activity compared to their counterparts in the niche will be lost earlier, because of decreased GSC-niche attachment. Intriguingly, ROS have the ability to activate Notch signaling via the activation of metalloprotease ADAM17, which triggers the release of NICD [49]. Taken together, Notch signaling might serve as a quality control to removal “poor” GSCs from the niche. Notch signaling is used for cell-fate determination and appropriate stem cell function in several different contexts, while the manner in which it is regulated is different between the two sexes. This arises because Sxl, a female-specific RNA-binding protein, negatively regulates Notch signaling during larval and follicular epithelium development, by binding to N mRNA to control Notch translation [36]. In this study, we show that Sxl maintains low levels of Notch signaling in young female GSCs, and elimination of Sxl in GSCs elevates Notch signaling, triggering early GSC loss. In addition, the pattern of Sxl expression and Notch signaling activity in the germ line is reciprocal (see Fig. 6G and S1A Fig.); Sxl is strongly expressed in GSCs and cytoblasts, but weakly in differentiating germ cell cysts, while Notch signaling activity is low in GSCs and cystoblasts, but high in germ cell cysts. These results suggest that Sxl may regulate Notch signaling in female GSCs through a post-transcriptional process present in other cell types. Notably, Notch signaling activities were similar between male GSCs and their progeny (S12 Fig.); this raises the question, ‘how does Sxl modulation of Notch signaling shape the differences between male and female GSCs?’ In addition, it is not clear how Sxl expression is decreased in cells that require higher Notch signaling, such as aged GSCs. It has been shown that transcription of sxl is directly regulated by JAK/STAT signaling during embryogenesis [50]. However, unlike Sxl, JAK/STAT signaling is not required for female GSC maintenance [2], [3]. In a study by Vied et al. [4], Hh signaling was observed to suppress the function of Sxl in the female germline; however, Hh signaling activity is decreased during aging [51]. These results imply that Sxl expression in GSCs is controlled by neither JAK/STAT nor Hh signaling. Although Sxl is a female-specific protein, sxl transcripts with the third exon are present in males; these transcripts encode a non-functional Sxl protein with an early stop codon [52]. During development of females, sxl is maintained in the on-state via an auto-regulatory feedback loop [53]. In this loop, female Sxl proteins promote their own expression by directing the splicing machinery to skip the third exon of transcripts derived from the constitutive sxl late promoter, which is active in both males and females. Aging is frequently associated with down-regulation of pre-mRNA processing factors that are required for alternative splicing [54], raising the possibility that the decrease in Sxl expression in GSCs with age is due to reduced efficiency of splicing. Nevertheless, further investigation is required to pinpoint the mechanism behind this phenomenon. We observed that Notch signaling in the niche and GSCs was affected by age in different ways; during aging, Notch signaling was decreased in the niche, but increased in GSCs. A subset of niche cap cells make direct contact with GSCs [24], raising the possibility that Notch signaling in niche cells may affect Notch signaling in GSCs, and vice versa. However, disruption of Notch signaling in GSCs did not affect Notch signaling in niche cap cells (see S2 Fig.), and knock down of N or Dl in niche cap cells did not affect Notch signaling in GSCs (S13 Fig.). In addition, our mosaic clonal analysis revealed that within Nloss of function heterozygous mutant niches, Nloss of function homozygous mutant GSCs (lacking GFP) exhibit higher levels of Notch signaling as compared to their neighboring control GSCs (most of which will be heterozygous for Nloss of function alleles, as the recombination efficiency between two in trans FRT19 loci is less than 30%), indicating that the increase of Notch signaling in GSCs occurs in a cell-autonomous manner. Similar conclusions were reached through GSC clonal analysis using the Ngain of function allele. Further, induction of Notch signaling in GSCs enhances GSC loss, and based on our microarray analysis of bam mutant GSCs distal from the niche; we conclude that the regulation of E-cadherin by Notch signaling in GSCs is performed in a cell-autonomous manner. These results also suggest that the mechanisms that regulate Notch signaling in niche cap cells and GSCs are independent of one another. Indeed, while Notch signaling in GSCs is controlled by Sxl, knock down of sxl expression in niche cap cells did not affect Notch activation (S13 Fig.). Furthermore, we previously demonstrated that Fringe (which adds sugar to Notch) is required for Notch activation in niche cap cells [55]), while excessive fng (induced by Foxo under insulin insufficiency) suppresses Notch activation. Here, we observed that N and fng mutant GSCs exhibit similar behavior; however, Notch signaling is not decreased in insulin signaling-defective GSCs (S14 Fig.), suggesting that Notch signaling is differentially regulated between GSCs and their niche. Drosophila stocks were maintained at 22–25°C on standard medium, unless otherwise indicated. yw and w1118 strains were used as wild-type controls. The null inr339, bam1, bamΔ86, fng[13], fng[L73], N5419, N55e11, hypomorphic sxlf2 and sxlfs3, and hypermorphic NAXE2 alleles have been described previously [23], [29], [37], [38], [56]–[59]. UAS-RNAi lines against N (VDRC 27228), fng (VDRC 51799), shg (VDRC10392), and sxl (10853R-3) were obtained from the Vienna Drosophila RNAi center or Fly Stocks of National Institute of Genetics. The efficiencies of N RNAi, fng RNAi, and shgRNAi have been previously reported [24], [60], and the efficiencies of the shgRNAi and sxlRNAi lines were examined in the ovary (S15 Fig.). The UASt vector (which contains the hsp70 promoter) is suitable for expressing RNAi constructs in the female germline line. Although the SV40 sequence located at the 3′UTR region of the UASt vector does not stabilize transcripts for nuclear export and protein synthesis, it does not influence transcription and RNA targeting of RNAi in the nuclei of germ cells. Dad-lacZ was used to monitor BMP signaling [30], and E(spl)m7-lacZ, E(spl)mβ-CD2, Notch responsive element (NRE)-pGreenRabbit, NRE-pRedRabbit, NRE-pBlueRabbit, and NRE-pVenusRabbit reporters were used to monitor Notch signaling [24], [61]. E(spl)mβ-CD2 and NRE reporters were not detectable in GSCs. The nanos (nos)-Gal4-VP16 and meta-GAL4 lines have been previously described [31], [62]. Flies expressing RNAi or other transgenes driven by nos-GAL4 also carried tub-GAL80ts to control GAL4 expression [33]; these flies were cultured at 18°C during development, and then switched to 29°C to allow GAL4 expression. Vasa-GFP was used to identify germ cells [63]. Other genetic elements are described in Flybase (http://flybase.bio.indiana.edu). A fragment of the Notch coding region lacking the extracellular domain was removed from the PIZ-NΔECN construct (kindly provided by B. DeDecker, University of Colorado Boulder) using the NotI and KpnI sites; the fragment was subcloned into the UASpI vector (modified from pUASp, T. Murphy, NCBI) to create pUASp-NΔECN (referred to as NNICD in this study). To construct UASp-mCD8-GFP, the mCD8-GFP fragment was amplified from the UASt-mCD8-GFP construct using a pair of primers carrying FseI and AscI sites (sequences are available upon request). The fragment was subcloned into the UASpI vector (modified by L. Lee, University of Vanderbilt). Transgenic lines were generated as described previously [64]. Genetic mosaics were generated by Flipase (FLP)/FLP recognition target (FRT)-mediated mitotic recombination [65]. For conventional mosaic analysis, females of genotype neoFRT19AFLP122/ubiGFPFRT19A, N*FRT19AFLP122/ubiGFPFRT19AFLP122, tubGAL80FRT19AFLP122/neoFRT19A; hsflp/+; neoFRT80B/arm-lacZFRT80B, hsflp/+; fng*FRT80B/arm-lacZFRT80B, hsflp/+; FRT82Bneo/FRT82Barm-lacZ, or hsflp/+; FRT82Bdinr339/FRT82Barm-lacZ were generated from standard crosses (N* represents N55e11, N5419, or NAXE2; fng* represents fng[13] or fng[L73]). For MARCM (Mosaic Analysis with a Repressible Cell Marker) [66], females of genotype tubp-GAL80FRT19AFLP122/N*FRT19A; nos-GAL4vp16 UASp-mCD8-GFP/+, tubp-GAL80FRT19AFLP122/N*FRT19A; nos-GAL4vp16>UASp-mCD8-GFP&UAS-shgRNAi were generated. To generate GSC clones, two-day-old females were subjected to heat shock for 1 hour at 37°C, twice a day for three days. After heat shock, females raised at 25°C were transferred to fresh food daily until dissection. N loss-of-function mosaic mutant females were cultured at 18°C to avoid lethality. Homozygous mutant cells were identified by the absence of ß-gal or GFP in conventional mosaic analyses, but recognized by the presence of GFP in MARCM. Ovaries or testes were dissected, fixed, and immunostained as described previously [26]. The following primary antibodies were used: mouse anti-Hts (1B1) (Developmental Studies Hybridoma Bank, DSHB, 1∶50), mouse anti-Lamin (Lam) C (LC28.26) (DSHB, 1∶50), Rat anti-E-cadherin (ECAD-2) (DHSB, 1∶3), mouse anti-Sxl (M180) (DSHB, 1∶350), rabbit anti-pMad (Smad3, #1880) (Epitomics, 1∶200), mouse anti-β-gal (Promega, 1∶500), and rabbit anti-GFP (Torrey Pines, 1∶1,000). Mouse anti-NICD and NECD (DHSB) failed to generate specific signals in GSCs. AlexaFluor 488-, 568- or 633-conjugated goat species-specific secondary antibodies (Molecular Probes, 1∶1000) were subsequently used. ApopTag Fluorescein Direct In Situ Apoptosis Detection Kit (Roche) was used as described [26]; the positive controls for this assay were two-day-old yw females starved on a diet of sugar and water for two days. Samples were stained with 0.5 µg/ml DAPI (Sigma), mounted in 80% glycerol containing 20.0 µg/mL N-propyl gallate (Sigma), and analyzed using a Zeiss LSM 700 confocal microscope. GSCs were identified by the anterior position of their fusome (labeled by 1B1 staining), which is juxtaposed to cap cells (cap cell nuclear envelopes were labeled by LamC staining) [24]. Germaria analyzed for (i) GSC division, (ii) expression of E(spl)m7-lacZ, Dad-lacZ, E-cadherin, and pMad, or (iii) GSC-niche contact area contained only one wild-type GSC and one marker-negative GSC. To measure GSC relative division rates, the number of GFP-positive progeny (cystoblasts and cysts) was divided by the number of GFP-negative progeny in a given germaria. Due to each cystoblast carries a fusome undergoes four more rounds of division to form two, four, 8, and 16-cell cysts; the cells in each cyst remain interconnected by a branched fusome. Therefore, the numbers of fusomes represent the numbers of GFP-negative progeny derived from the GFP-negative GSC, and likewise for the fusomes carried by GFP-positive progeny. For quantification of fluorescence signals, all clearly-stained germaria were subjected to analysis. For measuring E(spl)m7-lacZ, Dad-lacZ, pMad, Sxl and LamC expression, Image J was used to measure the average fluorescence intensity (arbitrary units) in confocal Z-sections at the largest GSC cytoplasmic or nuclear diameter. For quantification of E-cadherin and niche-GSC contact area, five to six optical sections (0.6 µm) were taken along the Z-axis of the E-cadherin-expressing interface between cap cell and GSC. The average intensity of E-cadherin signals at the region of contact between a GSC and cap cells was measured using Image J. For niche-GSC contact area, Avizo software (Visualization Science Group) was used to reconstruct and calculate the surface area volume along the Z-axis of the fusome. Statistical analysis was performed using Student's t-test. Twenty pairs of anterior transparent ovaries parts were dissected from one- and seven-week-old flies, or ∼2×105 GSCs were isolated from one-and five-week-old bam mutant females flies; cells/tissues were then lysed in RIPA buffer (20 mM Tris-HCl pH 7.5, 150 mM NaCl, 1 mM EGTA, 1% NP-40) supplemented with 2X Complete Proteinase Inhibitor Cocktail, EDTA-free (Roche) on ice for 1 hour. Lysates (60 µg aliquots) were boiled in sample buffer (50 mM Tris-Cl, pH 6.8, 5% ß-mercaptoethanol, 2% SDS, 0.1% bromophenol blue, and 10% glycerol) for 10 minutes, separated by 10% SDS-PAGE, blotted onto a PVDF membrane, and then blocked with 1X Tris-buffered saline containing 0.1% Triton X-100 (TBST, pH 7.5) and 0.5% bovine serum albumin (BSA) for 1 hour at room temperature. The blots were incubated with anti-Sxl (M180) (DSHB, 1∶350) and anti-α-tubulin (Sigma T9026, 1: 5000) antibodies at 4°C overnight with shaking. After three 10 minute washes with 1X PBST, the blots were incubated with anti-mouse IgG-HRP secondary antibody (Millipore; 1: 5000) for 1 hour at room temperature, and then washed three times with 1X TBST. Signals were detected and measured using the ECL system (Perkin Elmer), and compared to a molecular weight standard (Thermo). Several hundred bam1vasa-GFP/bamΔ86vasa-GFP ovaries were dissected in Grace's insect medium (GIBCO) with 10% FBS, and were subsequently incubated with 0.45% Trypsin (Invitrogen) and 2.5 mg/ml collagenase (Invitrogen) for 20 minutes at 25°C with vigorous shaking. Cell suspensions were filtered twice through a 40 µm nylon mesh. Cells were collected by centrifugation at 1000×g for 7 minutes, re-suspended in 1 ml of Grace's insect medium with 10% FBS and 1 mg/ml of propidium iodide, and then immediately sorted by fluorescence-activated cell sorting (FACS) with the Becton Deckinson FACSCalibur using CELLQUEST software. Living GSCs expressing Vasa-GFP were sorted by gating GFP-positive and red-negative cells with the exclusion model. Sorted cells were collected and kept in Trizol reagent (Invitrogen) at −80°C until RNA extraction. RNA was extracted from 7×105 isolated bam mutant GSCs, and 0.2 µg of total RNA was amplified using a low Input Quick-Amp Labeling kit (Agilent Technologies, USA) and labeled with Cy3 (CyDye, Agilent Technologies, USA) during in vitro transcription. Cy3-labeled cRNA (1.65 µg) was fragmented to an average size of about 50–100 nucleotides by incubation with fragmentation buffer at 60°C for 30 minutes. Fragmented labeled cRNA was then pooled and hybridized to Agilent Fly custom 4×44 K Microarrays (Agilent Technologies, USA) at 65°C for 17 h. After being washed and dried with a nitrogen gun, microarrays were scanned with an Agilent microarray scanner (Agilent Technologies, USA) at 535 nm for Cy3. Scanned images were analyzed using Feature extraction 10.5.1.1 software (Agilent Technologies, USA); image analysis and normalization software were used to quantify signals and background intensity for each feature. Selected candidates were validated by q-PCR. In brief, total RNA extracted from isolated GSCs was reverse transcribed using the Transcriptor First Strand cDNA Synthesis Kit (Roche). Steady-state mRNA levels were determined using LightCycler 480 Probes Master combined with a Universal Probe library (Roche); each gene was analyzed using the primer pairs and probes listed below: E(spl)m1: probe#87, 5′-CGAAAGGAATAGCGTGCAG-3′ and 5′-AACTTCTCGTGCAGATTCTCG-3′; E(spl)m4: probe#14, 5′-CTCTGGAGTCCTGCGAGAA-3′ and 5′-GCTTCGAAGTCGTAGTCCTCAA-3′; E(spl)m6: probe #55, 5′-TCCAACTAGTCCAAAGGATGC-3′ and 5′-AACCATCGAGGGTCTCCAA-3′; E(spl)m7: probe#66, 5′-AGCGACAACGAGTCTCTGCT-3′ and 5′-TTACCAGGGACGCCACAC-3′; E(spl)m8: probe#70, 5′-AGCAATTCCACGAAGCACA-3′ and 5′-GAGGAGCAGTCCATCGAGTT-3′; E(spl)mgamma: probe#153, 5′-TCGATGTGACCAAGATGGAG-3′ and 5′-TATCTACCAGGGACGCCAGA-3′; E(spl)mdelta: probe#60, 5′-CATTGTAATTTATTTCATCAACTTTGC-3′ and 5′-TTAATGAGGCTAAGTGGAAGCTC-3′; RpL19: probe #128, 5′-GAGCGTATTGCCACCAGGA-3′ and 5′- CGATCTCGTCCTCCTTAGCA-3′; RpL32: probe #117, 5′-CGGATCGATATGCTAAGCTGT-3′and 5′- CGACGCACTCTGTTGTCG-3′.

10.1371/journal.pntd.0005518

Rapid mapping of urinary schistosomiasis: An appraisal of the diagnostic efficacy of some questionnaire-based indices among high school students in Katsina State, northwestern Nigeria

In sub-Saharan Africa, over 200 million individuals are estimated to be infected with urinary and intestinal schistosomiasis. In a bid to lay a foundation for effective future control programme, this study was carried out with the aim of assessing the diagnostic efficacy of some questionnaire-based rapid assessment indices of urinary schistosomiasis. A total number of 1,363 subjects were enrolled for the study. Questionnaires were administered basically in English and Hausa languages by trained personnel. Following informed consent, terminal urine samples were collected between 09:40 AM and 2:00 PM using clean 20 ml capacity universal bottles. 10μl of each urine residue was examined for the eggs of S. haematobium using x10 objective nose of Motic Binocular Light Microscope (China). The average age ± Standard Deviation (SD) of school children examined was 15.30 ± 2.30 years and 40.87% were females. The overall prevalence and geometric mean intensity of S. haematobium infection were 26.41% (24.10─28.85) and 6.59 (5.59─7.75) eggs / 10 ml of urine respectively. Interestingly, a questionnaire equivalence of the prevalence obtained in this survey was 26.41% (24.10─28.85) for Rapid Assessment Procedure based on self-reported blood in urine. The results of correlation analyses demonstrated significant associations between the prevalence of S. haematobium infection and contact with potentially infested open water sources (r = 0.741; P = 0.006). By regression model, cases of respondents with self-reported blood in urine are expected to rise to 24.75% if prevalence of the infection shoots up to 26.5%. The best RAP performance was obtained with self-reported blood in urine. Based on the overall prevalence value, the study area was at a “moderate-risk” of endemicity for urinary schistosomiasis. Chemotherapeutic intervention with Praziquantel, the rationale behind rapid assessment procedure for schistosomiasis, has been recommended to be carried out once in every 2 years for such communities.

Schistosomiasis is a water-borne neglected infectious disease of poverty that has consistently plagued over 200 million helpless inhabitants of the tropics, particularly, sub-Sahara Africa. Under the auspices of different nomenclatures and affiliations, many control programmes based on Praziquantel have been inaugurated over the past decades. Bearing in mind that globally, schistosomiasis exist in focal pockets within peri-urban and rural settings, and the lowest cost of a generic 600-mg tablet is approximately US$ 0.08, it is imperative to focus control resources more on high risk settings. In order to identify such settings, rapid means of mapping schistosomiasis prevalence are carried out either with questionnaires or biomedical testing with reagent strips. Rapid assessment procedure for urinary schistosomiasis, the focus of this study, builds substantially on the perception of respondents about the disease through visible blood in their urine (where applicable). We conducted this present survey in 6 communities of Katsina State, northwestern Nigeria by interviewing and examining the urine of 1, 363 high schools students for the eggs of Schistosoma haematobium. A unique discovery in this survey was that contact with unwholesome water bodies, where properly defined, was significantly associated with urinary schistosomiasis, both as a single index and when combined with itching experience.

Schistosomiasis, a water-borne neglected tropical disease (NTD), has been reported as the second most prevalent parasitic disease after malaria [1]. The causative agent of human schistosomiasis is a digenetic trematode blood fluke of the genus Schistosoma with a complex, indirect life cycle involving different species of freshwater snails [2, 3]. These snails serve as intermediate hosts to S. haematobium, S. intercalatum, S. japonicum, S. mansoni and S. mekongi which parasitize humans [4,5]. Humans become infected when the infective larvae mechanically penetrate their skin after contact in fresh water bodies located in environment characterized by poor hygiene and sanitation [6]. The distribution of schistosomiasisis is more abundant in the African region with 42 countries endemic for the infection. In sub-Saharan Africa, over 200 million individuals are estimated to be infected with urinary and intestinal schistosomiasis [7], with approximately 393 million people at risk of infection from Schistosoma mansoni, of which 54 million are infected while 436 million people are at risk of S. haematobium infection and 112 million are infected [8]. The most widely used approach for the diagnosis in endemic settings is the detection of schistosome eggs in either stool or urine specimens by light microscopy. The first step in targeting health interventions is to map the disease geographically and rank it according to the risk of infection and morbidity [4]. The use of geographical information systems highlighted the scarcity of data in endemic region such as Africa, and emphasizes the need for a rapid, non-invasive and inexpensive epidemiological assessment tool that can be fully integrated within existing administrative systems [9]. Simple school questionnaires were developed for S. haematobium and has since been validated in many ecological, epidemiological, and sociocultural settings across sub-Saharan Africa. It is well accepted and operationally feasible. It is faster and less expensive than the standard parasitological diagnosis [10]. The basis of this method was that, being a chronic disease, the only clear symptom school-age children could observe and easily remember was the presence of blood in urine [11]. They build directly on a community’s perception of disease, involve the active participation of teachers and schoolchildren, and represent a first step towards involving the community in control activities. Macrohaematuria, microhaematuria, and proteinuria are assessed by reagent strips [12]. In a bid to lay a foundation for effective future control programme for urinary schistosomiasis in the study area, we embarked on this cross-sectional survey with the aim of assessing the diagnostic efficacy of some questionnaire-based rapid assessment indices of urinary schistosomiasis. Written ethical clearance to conduct the survey was issued by the Ethical Committee of the Katsina State Ministry of Education, Dutsin-Ma Zonal Office. School heads and students gave oral informed consent to participate after appropriate briefing on the background and objectives of the study. Oral assent, aided by an interpreter, was provided by students after appropriate briefing on the background and objectives of the study. They demonstrated this by willingly providing their names for a written documentation during the interview. Information obtained from the subjects was kept confidential. Noteworthy is the fact that formal consent could not be obtained from the parents and guardians of the subjects partly because the cultural and religious situation of the study area was volatile. To buttress this point, there is history of attempted physical attack on healthcare officials in the study area. The study was undertaken in twelve (12) high schools from six (6) communities of Dutsin-Ma and Safana (809 km2) Local Government Areas (LGAs) of Katsina State, Northwestern Nigeria (see Fig 1). As at 2006 National Census, both neighboring LGAs were inhabited by 353,450 people [13]. Noteworthy is the fact that the study area, characteristically sandy with a rocky terrain (typical of western upland plateau) is drained by different water bodies, the largest being Zobe Dam. The study covered a bio-geographical Sudan Savannah area characterized by low to moderate endemicity for urinary schistosomiasis. By Agro-ecological classification, it belongs to the Sudan savanna vegetation zone of Nigeria [14]. The main economic activity there is farming, with millet as the subsistence crop. The predominant ethnic groups, Hausa and Fulani, complement crop production with trading and nomadism. Both LGAs have a mean annual rainfall and temperature less than 800mm and 30°C respectively [15]. A cross-sectional study design was adopted in this present survey. By estimating the prevalence of Schistosoma haematobium at 30% with power and sampling error of 90% and 5% respectively, a sample size of 912 was obtained. This calculation was made based on the standard of World Health Organization for sample size estimation [16]. Simple random sampling technique was employed to select the total number of 1,363 secondary school students who participated in the study between May and August, 2015. This sample size accounted for effect size and any anticipated non-response. School based questionnaire with questions relating to the knowledge of urinary schistosomiasis, sources of water, and local name associated with the disease was used in the survey. For information on urinary schistosomiasis, question asked was: “Do you know any student in this school who reportedly pass blood-stained urine?” For a positive response, the next question was: “What is the local language for this condition?” The individual questionnaire was designed, among other things, to obtain responses from subjects on water contact activities (fetching, swimming and play in shallow water), experiences of itching, haematuria, and pain while urinating. To elicit responses for some of these experiences, interviewees were asked the following questions: “Have you ever experienced: (i) pains while urinating? (ii) blood in your urine?” Whenever a response was positive for the latter, each subject was further questioned: “How have you been treating it?” Because the study population was divided along the language lines of English and Hausa, questionnaires were administered accordingly by trained personnel which included the investigators and selected teachers from the participating schools. In the study area, S. mansoni is co-endemic with S. haematobium. Consequently, some interviewees suffered a mixed infection. Consequently, there was unusual discovery of the eggs of the former (S. mansoni) in a few urine samples that were as well positive for the eggs of the latter (S. haematobium). They were distinguished using their unique identification keys, that is, the possession of a lateral and terminal spine by the eggs of S. mansoni and S. haematobium respectively [18]. However, it has been reported that under special epidemiological settings with a very high prevalence of urinary schistosomiasis but a very low prevalence of intestinal schistosomiasis, eggs of S. mansoni do occur in urine [19]. 10μl of each urine residue was examined for the eggs of S. haematobium using x10 objective nose of Motic Binocular Light Microscope (China). Each average egg count was recorded as number of eggs per 10 ml of urine sample using a multiplier factor of two. While prevalence was grouped into low (˂ 10%), moderate (≥ 10%–49%) and high (≥ 50% or more) [20], intensity of infection was categorized into light (˂ 50 eggs / 10 ml of urine) and heavy (≥ 50 eggs / 10 ml of urine) infections according to standard method [4]. All data obtained from the survey were entered into Microsoft Excel 2010 (USA) and analysed using SPSS 15.0 (Chicago, USA). The relationships between Rapid Assessment and Parasitological Indices were assessed using Spearman’s rank correlation and linear regression analyses. The prevalence for S. haematobium infection, as well as for micro-haematuria (shown in Figs 2–7) was calculated on a school-level basis. These school-level estimates were correlated with the following rapid assessment indicators [i.e. water contact (in %), self-reported blood in urine (in %), itching, urethral pain, and combined RAPs based on water contact/ self-reported blood in urine/ itching/ urethral pain (in %), and water contact/ itching (in %)] using the Spearman's rank correlation coefficient test. Both school-level prevalences of S. haematobium infection and micro-haematuria were used in a linear regression model as continuous dependent outcomes (in %) and were controlled for, in a univariable manner, each rapid assessment indicator. Prior to reporting our findings, necessary diagnostic tests were performed. Normality was achieved and no key model assumptions were violated. Statistical significance was considered at 95% confidence level (CL) with a P value of 0.05. The diagnostic performances of indices for identifying “low risk”, “moderate risk” or “high risk” schools were assessed by calculating sensitivities, specificities, and positive and negative predictive values. In the six (6) communities surveyed, 1,363 students with ages ranging from 10–25 years were interviewed and examined. The average age ± Standard Deviation (SD) of school children examined was 15.30 ± 2.30 years and 40.87% were females. Of the total number interviewed, 360 respondents were infected. It is worthy of note that the prevalence and geometric mean of S. haematobium egg counts in the study communities ranged from 15.59% to 37.28% and 3.27 to 16.42 eggs / 10 ml of urine respectively (see Table 1 and Fig 2). Hence, the overall prevalence and geometric mean intensity of S. haematobium infection were 26.41% (24.10–28.85) and 6.59 (5.59–7.75) eggs / 10 ml of urine respectively. The arithmetic mean intensity of infection was 27.90 (19.55–36.25) eggs / 10 ml of urine. Males recorded a higher prevalence [40.07% (36.69–43.56)] and geometric mean intensity of S. haematobium infection [7.52 (6.33–8.94) eggs / 10 ml of urine]. Furthermore, males were 9 times [COR (95% CI): 9.39 (6.54–13.49)] more likely to be infected with the cercariae of S. haematobium (see Table 1). Spearman’s rank correlation (see Table 2 and Figs 3–7) demonstrated significant associations between: S. haematobium prevalence and contact with potentially infested open water sources (r = 0.741; P = 0.006); prevalence of micro-haematuria and contact with potentially infested open water sources (r = 0.643; P = 0.024); self-reported blood in urine and prevalence of S. haematobium infection (r = 0.629; P = 0.028); prevalence of micro-haematuria and the combined RAP of water contact, self-reported blood in urine and painful urination (r = 0.741; P = 0.006) and prevalence of micro-haematuria and the combined RAP of self-reported blood in urine and painful urination (r = 0.727; P = 0.007). Significant associations (see Table 2) were also obtained with: prevalence of S. haematobium infection and the combined RAP of water contact, and itching (r = 0.636; P = 0.026); and prevalence of S. haematobium infection and the combined RAP of self-reported blood in urine, and painful urination (r = 0.587; P = 0.045). The questionnaire equivalence of the prevalence revealed a similar prevalence of 26.41% (24.10–28.85) for RAP based on self-reported blood in urine (see Fig 5). However, no significant association was recorded between the prevalence of light infection intensity and water contact (r = 0.573; P = 0.051); school community egg load and water contact (r = 0.364; P = 0.245); prevalence of micro-haematuria and the combined RAP of water contact, and itching (r = 0.357; P = 0.255). Noteworthy was the fact that the single RAP index of pain while urinating recorded the poorest performance since it had no statistically significant association (P ˃ 0.05) with any Parasitological indices (see Table 2). We discovered that, for a 26.5% rise in contact with infested water sources, the: prevalence of S. haematobium infection will increase by 1% (using the equation: y = 0.6667 * x– 16.6667 shown in Fig 3); prevalence of microhaematuria will only increase by 3.6% (see Fig 4). However, cases of respondents with self-reported blood in urine are expected to rise to 24.75% if prevalence of the infection shoots up to 26.5% (see Fig 5). Meanwhile, using equation 0.6 * x + 9, the prevalence of microhaematuria is anticipated to increase by 24.9% with a similar 26.5% rise in the number of respondents with combined experiences of contact with infested water sources, self-reported blood in urine and painful urination (see Fig 6). Similarly, for 26.5% increase in the number of respondents with combined experiences of contact with infested water sources and self-reported blood in urine, the prevalence of microhaematuria is expected to remain 24.9% (see Fig 7). All RAPs showed sensitivities which ranged from 33.33–98.08%. However, RAPs based on water contact, pain while urinating and self-reported blood in urine respectively recorded high sensitivities in descending order of magnitude. Meanwhile, specificity ranged from 45.46–94.32%. The least value was recorded for RAP based on water contact while the highest was obtained with the combined RAP of water contact, self-reported blood in urine and pain while urinating. In summary, the best RAP performance was obtained with self-reported blood in urine which had a sensitivity of 60.28% and specificity of 91.43%. Co-incidentally, this RAP also recorded the best combined values for Positive Predictive Value (PPV) and Negative Predictive Value (NPV) (Table 3). Currently, the most widely used clinical approach to determining the prevalence and intensity of infection due to S. haematobium is manual egg count by means of urine microscopy. Our data showed that by this gold standard [17, 21], the overall prevalence and geometric mean intensity of urinary schistosomiasis were 26.41% (24.10–28.85) and 6.59 (5.59–7.75) eggs / 10 ml of urine respectively, with males being 9 times [COR (95% CI): 9.39 (6.54–13.49)] more likely to be infected compared to females. It is pertinent to state that this gold standard employed vis a vis the rapid assessment procedures in this cross-sectional survey was rather cumbersome and frustrating. However, it is of interest that a questionnaire equivalence of the prevalence obtained in this survey revealed a similar prevalence of 26.41% (24.10–28.85) for RAP based on self-reported blood in urine. Again, findings in this survey showed that the best RAP performance was obtained with self-reported blood in urine with a sensitivity and specificity of 60.28% and 91.43% respectively. To corroborate the reliability of this RAP index, a report from Yemen shows that 72.2% of respondents who suffered heavy intensity of infection with Schistosoma haematobium visibly experienced blood in their urine [17]. Furthermore, high sensitivity and specificity have previously been reported in other urinary schistosomiasis endemic settings. For example, in a similar survey conducted in southwestern Nigeria, a specificity of almost 100% was obtained [22]. In northern Ghana, self-reported haematuria showed a sensitivity of 53% and a specificity of 85% [23]. Co-incidentally, in this survey, self-reported blood in urine also recorded the best results for Positive Predictive Value (71.62%) and Negative Predictive Value (86.51%). The latter simply means that, of all the subjects who tested negative for urinary schistosomiasis by microscopic examination, 86.51% were actually negative while 13.49% were positive, going by questionnaire-based rapid means of assessment using self-reported blood in urine (macro-haematuria). In addition, when this RAP was combined with water contact and pain while urinating (dysuria), a higher Positive Predictive Value (75.64%) was obtained. That is, this combination unmasked 4.02% of more subjects that were infected compared to the single RAP. This is indeed a cost-effective means of improving on the quality of data obtained in urinary schistosomiasis research. Moreover, the result of correlation analysis demonstrated a statistically significant association between self-reported blood in urine and prevalence of S. haematobium infection (r = 0.629; P = 0.028). Better still, when self-reported blood in urine was combined with RAP based on painful urination, a stronger association (r = 0.727; P = 0.007) was obtained between them and micro-haematuria. Meanwhile self-reported blood in urine, micro-haematuria and painful urination (dysuria) have been previously identified as morbidity markers of urinary schistosomiasis [17, 24, 25]. The implication of these is that we can use a RAP based on self-reported blood in urine to predict the parasitological prevalence of urinary schistosomiasis in either moderate or high endemic settings. It could as well produce a reliable result in areas where biomedical reagent strips are not available [26]. Previous studies carried out in some African countries (Cameroon, Congo, Democratic Republic of the Congo, Ethiopia, Malawi, Zambia and Zimbabwe) also showed that macro-haematuria had a very good diagnostic ability to detect “high-risk” schools while ruling out “low-risk” ones [10]. In a survey conducted in the Tanga region of the United Republic of Tanzania, average of 75% school-age children were reportedly accurate in their self—diagnosis of urinary schistosomiasis using the presence of blood in urine (haematuria) as a rapid diagnostic procedure [27]. To the best of our knowledge, contact with potentially infested, open, and unwholesome water sources is not in use as a rapid assessment indicator for urinary schistosomiasis. However, in this survey, the result of correlation analysis demonstrated significant association between prevalence of S. haematobium infection and contact with potentially infested open water sources (r = 0.741; P = 0.006). This did not come as a surprise because urinary schistosomiasis has been constantly reported as a water-borne disease [10, 22, 28]. When employed as a single RAP index in this present survey, it recorded a very high sensitivity (98.06%) and Negative Predictive Value (98.49%) but low values for the duo of specificity (45.46%) and Positive Predictive Value (39.22%). Meanwhile, when combined with self-reported blood in urine and dysuria, its sensitivity markedly reduced to almost half (49.17%) while its Negative Predictive Value dropped to 83.79%. Its specificity (94.32%) and Positive Predictive Value (75.64%), however, approximately doubled. It also demonstrated a statistically significant association (r = 0.643; P = 0.024) with the prevalence of micro-haematuria. More interestingly, when combined with other RAPs based on self-reported blood in urine and painful urination, a stronger relationship was obtained with the prevalence of micro-haematuria (r = 0.741; P = 0.006). The implication of these findings is that when subjects are carefully interviewed as regards their water contact activities, to a large extent, a good rapid diagnostic result for urinary schistosomiasis could be obtained. This is a good news to all high risk endemic settings where diagnostic kits and microscopes are very short in supply. Bearing in mind that indiscriminate mass chemotherapeutic intervention with Praziquantel is indeed not harmful [27], on the basis of this finding, it could be achieved successfully without anticipating any severe adverse reactions. In the context of Schistosomiasis Elimination Strategy and Potential Role of a Vaccine in Achieving Global Health Goals co-sponsored by Bill and Melinda Gates Foundation and the National Institute of Allergy and Infectious Diseases [29], self-reported blood in urine, as a single or combined RAP index, could play a major diagnostic role in unraveling new endemic foci for mass drug administration. As it stands, self-reported blood in urine will continue to be a relevant rapid diagnostic RAP index until schistosomiasis is eradicated. This survey is, however, subject to some limitations. To start with, adults were not included. Therefore, the result reported here may not be applicable to the whole population of the study area because previous findings have shown that the diagnostic efficacy of haematuria as a RAP index is inversely proportional to the age of subjects but stable in teenage children [30]. An extrapolation is only applicable after a painstaking assessment of the school enrollment, and the overall socio-cultural and epidemiological condition of the study area [11]. Moreover, previous report has shown that the accuracy of macro-haematuria as a yardstick for rapid assessment of urinary schistosomiasis may be better when a day-to-day variation in eggs excretion is considered [31]. However, this survey did not capture a serial assessment of each subject for macro-haematuria. Since report has shown that the identification of schools and communities endemic for schistosomiasis is a key issue in any control programme [32], the high performing RAPs in this study could be employed to discover more endemic foci in the existence of ongoing regular Mass Drug Administration (MDA) in Nigeria. Based on this overall prevalence value of 26.41% (24.10–28.85) obtained in this survey, it is obvious that the study area was at a “moderate-risk” of endemicity for urinary schistosomiasis [4, 20]. Meanwhile, chemotherapeutic intervention with Praziquantel, the rationale behind rapid assessment procedure for schistosomiasis, has been recommended to be carried out once in every 2 years for such communities [20]. Although, water contact was found to have a good diagnostic efficacy, The best RAP in this survey was self-reported blood in urine. Both RAPs performed better when combined with other RAP indices.

10.1371/journal.pgen.1004723

ZTF-8 Interacts with the 9-1-1 Complex and Is Required for DNA Damage Response and Double-Strand Break Repair in the C. elegans Germline

Germline mutations in DNA repair genes are linked to tumor progression. Furthermore, failure in either activating a DNA damage checkpoint or repairing programmed meiotic double-strand breaks (DSBs) can impair chromosome segregation. Therefore, understanding the molecular basis for DNA damage response (DDR) and DSB repair (DSBR) within the germline is highly important. Here we define ZTF-8, a previously uncharacterized protein conserved from worms to humans, as a novel factor involved in the repair of both mitotic and meiotic DSBs as well as in meiotic DNA damage checkpoint activation in the C. elegans germline. ztf-8 mutants exhibit specific sensitivity to γ-irradiation and hydroxyurea, mitotic nuclear arrest at S-phase accompanied by activation of the ATL-1 and CHK-1 DNA damage checkpoint kinases, as well as accumulation of both mitotic and meiotic recombination intermediates, indicating that ZTF-8 functions in DSBR. However, impaired meiotic DSBR progression partially fails to trigger the CEP-1/p53-dependent DNA damage checkpoint in late pachytene, also supporting a role for ZTF-8 in meiotic DDR. ZTF-8 partially co-localizes with the 9-1-1 DDR complex and interacts with MRT-2/Rad1, a component of this complex. The human RHINO protein rescues the phenotypes observed in ztf-8 mutants, suggesting functional conservation across species. We propose that ZTF-8 is involved in promoting repair at stalled replication forks and meiotic DSBs by transducing DNA damage checkpoint signaling via the 9-1-1 pathway. Our findings define a conserved function for ZTF-8/RHINO in promoting genomic stability in the germline.

Proper response to DNA damage and repair of DNA double-strand breaks (DSBs) is important to maintain genomic integrity and promote both accurate chromosome segregation and tumor suppression. Here we define the roles of a previously uncharacterized and conserved protein, ZTF-8, which is required for proper DNA damage checkpoint activation as well as DSB repair. Specifically, we provide a direct demonstration that ZTF-8 participates in both mitotic and meiotic DSB repair and in the meiotic DNA damage checkpoint via interacting with the 9-1-1 complex in the C. elegans germline. We propose that ZTF-8 is involved in promoting repair at blocked replication fork sites and meiotic DSBs in part by transducing DNA damage checkpoint signaling via the 9-1-1 DNA damage response complex.

Genome instability is a hallmark of cancer cells and a critical feature that enables tumor progression. Instability allows cells to break and reform chromosomes, generate new oncogene fusions, inactivate tumor suppressor genes, amplify drug resistance genes, and therefore increase their malignancy. This whole progression often accompanies the disruption of DNA repair genes as the failure in DNA repair permits an increased rate of chromosome breakage and mutagenesis [1]. For example, many mutations involved in DNA repair genes have been linked to the progression of diverse cancers including breast, ovarian, and skin cancer, as well as leukemia and lymphomas. These include germline mutations in breast cancer susceptibility 1 (BRCA1), BRCA2, BRIP1, RAD50, the Nijmegen breakage syndrome NBS1 gene and the Fanconi anemia FA genes [2]. Germline defects in three known RecQ helicases cause defined genetic disorders associated with cancer predisposition and/or premature aging. These include Bloom's, Werner's and Rothmund–Thomson syndromes, which are caused by defects in the BLM, WRN and RECQ4 genes, respectively [3]–[5]. Considering that many germline mutations in DNA repair genes specifically involved in double-strand break repair (DSBR) are linked to tumor progression [2], and that failure to properly repair programmed meiotic DSBs can impair chromosome segregation, understanding DSBR at a molecular and cellular level, in a genetically tractable multicellular system, is of extreme importance. Studies in the yeasts S. cerevisiae and S. pombe revealed various gene functions required for the DNA damage checkpoint pathway [6]. Most DNA damage response (DDR) genes were identified through the genetic analysis of mutants defective in either the transcriptional or cell cycle arrest responses to DNA damage. The DNA damage checkpoint proteins in S. pombe include those encoded by several radiation-repair (rad) genes. A phosphatidylinositol kinase family in S. pombe is both structurally and functionally related to human ATM and ATR [7]. However, the lack of an apoptosis pathway in yeast and the high degree of conservation for known components of the DDR pathway between worms and humans have positioned the nematode C. elegans as an excellent genetic system to study DNA damage induced cell cycle arrest and apoptosis [8]–[11]. Here we have identified a role for ZTF-8, a protein conserved from worms through humans, in the repair of both mitotic and meiotic DSBs and in the activation of the pachytene DNA damage checkpoint in the C. elegans germline. We show that ZTF-8 localizes to both chromatin and the nucleolus. Changes in its subcellular localization in response to DNA damage, as well as its ATL-1- and ATM-1-dependent chromatin localization, support a role for ZTF-8 in DDR and DNA repair. Moreover, ztf-8 mutants exhibit specific DNA damage sensitivity to γ-irradiation (γ-IR) and hydroxyurea (HU), and not to UV, nitrogen mustard (HN2) or camptothecin (CPT) treatment, suggesting a role in DSBR. This is further supported by the activation of an S-phase checkpoint and the accumulation of recombination intermediates during both mitotic and meiotic progression in ztf-8 mutant germlines. However, while the S-phase checkpoint is intact, impaired meiotic DSBR progression partially fails to trigger the CEP-1/p53-dependent DNA damage checkpoint in late pachytene, also suggesting a role for ZTF-8 in DDR. This is further supported by the interaction of ZTF-8 with MRT-2, the C. elegans homolog of the Rad1 protein found in S. pombe, Drosophila, and mammals, and a member of the 9-1-1 DDR complex, and the impaired localization of HUS-1 onto chromatin in response to exogenous DSB formation in ztf-8 mutants. Loss of ZTF-8 function resulted in partially impaired activation of germ cell apoptosis, a reduced brood size and the accumulation of RAD-51 foci, all of which were rescued in transgenic worms expressing human RHINO, suggesting that its functions are conserved between species. Taken together, our analysis supports a model in which ZTF-8 plays a role in repair at stalled replication forks and meiotic DSBs as well as in meiotic DNA damage checkpoint response via the 9-1-1 pathway. ztf-8 (open reading frame ZC395.8) was identified in an RNAi-based screen performed as in [12] designed to find meiotic candidates among genes with germline-enriched expression in C. elegans. ztf-8 encodes for a 687 amino acid protein that contains C2H2-type zinc-finger binding domains, and is highly conserved from worms to humans (Figure 1A; Figure S1). The ZTF-8 protein also carries a predicted DNA binding site (APSES) in its N-terminal region. The APSES motif has been reported to be required for the interaction between its putative mammalian homolog RHINO, a protein implicated in interacting with the Rad9-Rad1-Hus1 complex (9-1-1), and the ATR activator, TopBP1 [13], [14]. The high degree of conservation of the APSES motif throughout species supports a concurrent conservation of the DNA damage checkpoint machinery in different species. The ztf-8 deletion mutant (tm2176), obtained from the Japanese National Bioresource Project, carries a 524 base pair out-of-frame deletion encompassing most of exon 6 along with exons 7 through 11 (Figure 1A). This deletion results in a premature stop codon and the loss of a predicted zinc-finger motif, a predicted phosphorylation site, and three putative sumoylation sites. Analysis of wild type and ztf-8 mutant lysates on Western blots, utilizing an affinity purified ZTF-8-specific N-terminal antibody, revealed that the protein migrates at a higher molecular weight than expected, and may therefore be undergoing some form of modification (77 kDa and 35 kDa were the expected protein sizes for wild type and the ztf-8 mutant, respectively, Figure 1B). ztf-8 mutants exhibit a 25% reduction in brood size compared to wild type, indicative of sterility (Figure 1C). Brood size was not significantly reduced in the heterozygotes, suggesting that tm2176 is a recessive allele. ztf-8(tm2176) homozygous mutants do not exhibit any larval lethality suggesting that ZTF-8 does not have a role during larval development. However, homozygous mutants do exhibit weak, but significant, embryonic lethality (0.91% compared to 0.01% in wild type, P = 0.0078 by the two-tailed Mann-Whitney test, 95% C.I.) and high incidence of males (Him) phenotypes (0.26% compared to 0.03% in wild type, P = 0.0313), which taken together suggest a role in promoting accurate meiotic chromosome segregation. Finally, a trans-heterozygote for a deficiency encompassing the ztf-8 locus (ztf-8(tm2176)/sDf121) did not exhibit a significant decrease in brood size (P = 0.6542) or an increase in either embryonic (P = 0.1823) or larval lethality (P = 0.0724) compared to ztf-8(tm2176) homozygous mutants, suggesting that ztf-8(tm2176) is likely a null (Figure 1C). To gain some insight into the function of ZTF-8 we examined its localization by immunostaining dissected wild type hermaphrodite gonads with an affinity purified ZTF-8 specific C-terminal antibody (Figure 2A). ZTF-8 signal is observed in mitotic nuclei at the distal tip (premeiotic tip). This signal is then reduced upon entrance into meiosis (leptotene/zygotene stages = transition zone) and remains weak through the mid-pachytene stage. However, the ZTF-8 signal increases once again in late pachytene nuclei and persists through late diakinesis oocytes. This dynamic pattern of expression suggests regulation of ZTF-8 during meiotic prophase. At a higher resolution, ZTF-8 signal is observed as foci both on chromosomes as well as in the nucleolus (Figure 2B). Specifically, 34% of ZTF-8 foci at the premeiotic tip, and 78% at late pachytene, localize to DAPI-stained chromosomes, with the remaining foci being localized to the nucleolus (Figure 2C, n = 102 nuclei at the premeiotic tip, and 53 nuclei at pachytene, from 7–10 gonads). ZTF-8 localization is also observed in gut and embryonic nuclei and this signal is specific since it is absent in ztf-8 homozygous mutants (Figure S2). Taken together, these localization studies suggest both mitotic and meiotic roles for ZTF-8. Either exposure to genotoxic agents or DNA replication stress can lead to checkpoint responses in the C. elegans germ line. Specifically, the replication-dependent S-phase checkpoint is activated in response to stress, such as that stemming from HU treatment, DNA damage and abnormal DNA structures [15], and results in transient S-phase arrest, which is characterized by a premeiotic tip exhibiting enlarged nuclear diameters in the C. elegans germline [16]. In ztf-8 mutants, enlarged mitotic nuclei were observed at the premeiotic tip compared to wild type (Figure 3A). Activation of the DNA damage checkpoint in ztf-8 mutants is further supported by the elevated levels of ATL-1 (ATR homolog) and phosphorylated CHK-1 (pCHK-1) observed in these nuclei even without γ-IR exposure (Figure 3B and 3C). Given that ATL-1 is recruited to stalled replication fork sites [16], ZTF-8 is likely required for repair at stalled replication forks. This is supported by the further increase in nuclear diameter observed among mitotic nuclei at the premeiotic tip in the mutants following treatment with HU (3 fold induction in ztf-8 mutants compared to 1.9 fold in wild type), a ribonucleotide reductase inhibitor which blocks DNA synthesis by preventing expansion of the dNTP pool and results in replication fork stalling (Figure 3A). Lower levels of PCN-1, the C. elegans ortholog of mammalian PCNA, in HU treated worms (Figure 3D and 3E) is further evidence of an S-phase arrest, consistent with studies in human cells where PCNA is absent from S-phase nuclei following HU treatment [17]. PCN-1 signal was observed only in 69% of mitotically dividing nuclei in ztf-8 mutants compared to 92% in wild type, suggesting that slowing down S-phase in response to nucleotide depletion prevents association of PCN-1 onto replication sites. No significant difference is observed between ztf-8 mutants and wild type with markers for G2/M and mitosis such as CDK-1 phospho-TYR15 and phospho-histone H3 (pSer10), respectively (Figure S3). Altogether, these observations indicate that there is activation of the S-phase checkpoint resulting in cell cycle arrest in the ztf-8 mutants and that ZTF-8 function may be required for repair at stalled replication forks. To further examine the role of ZTF-8 in DNA damage repair, adult hermaphrodites were exposed to different types of DNA damage and embryonic lethality was monitored as in [18], [19] (Figure 4A). Exposure to HU, which results in a checkpoint-dependent cell cycle arrest, led to significant changes in hatching in ztf-8 mutants compared to wild type (100% and 96%, respectively, at 15 mM). Also, ztf-8 mutants showed increased larval lethality following HU exposure, further suggesting that ztf-8 may be required for repair following collapse of stalled replication forks. ztf-8 mutants exhibited reduced hatching frequencies compared to wild type following the induction of DSBs by γ-IR exposure. Specifically, only 52% and 34% hatching was observed among the progeny of ztf-8 mutants exposed to either 30 or 100 Gy, respectively, compared to 64% and 58%, respectively, in wild type. Interestingly, in γ-IR exposed mutants we observed chromatin fragments with RAD-51 foci, which mark sites undergoing DSBR [20], present in nuclei from leptotene/zygotene to pachytene (Figure 4B). These observations strongly suggest that ZTF-8 is required for DSBR following γ-IR exposure. Exposure to HN2, which produces DNA interstrand crosslinks (ICLs), UV, which induces cyclobutane pyrimidine dimers, and CPT, which results in a single ended DNA double-strand break when collision of a replication fork occurs at the lesion, did not significantly reduce hatching levels in ztf-8 mutants compared to wild type (Figure 4A). Taken together, these results indicate that the function of ZTF-8 in DNA repair exhibits a high degree of DNA damage specificity, being required for recovery from replication fork collapse and DSBR. To determine whether ZTF-8 is required for DSBR in both mitotic and meiotic nuclei, levels of RAD-51 foci were quantitated and compared between wild type and ztf-8 germline nuclei (Figure 4C, 4D and Figure S4). Since nuclei are positioned in a temporal-spatial manner along the germline in C. elegans, proceeding in a distal to proximal orientation from mitosis into the various stages of meiotic prophase I, levels of RAD-51 foci were assessed both in mitotic (zones 1 and 2) and meiotic nuclei (zones 3–7). In wild type, a few mitotic RAD-51 foci were observed at zones 1 and 2, and they are mainly derived from single stranded DNA gaps formed at stalled replication forks or resected DSBs resulting from collapsed replication forks [21]. During meiotic prophase, SPO-11-dependent programmed meiotic DSBs are induced. Levels of RAD-51 foci start to rise at the transition zone (zone 3) and reach their highest levels at early to mid-pachytene (zones 4 and 5). As repair is completed, levels of RAD-51 foci are progressively reduced in late pachytene (zones 6 and 7). In ztf-8 mutants, levels of RAD-51 foci were higher than those observed in wild type mitotic (20.7% of nuclei contained 1–3 RAD-51 foci compared to 7.8% for wild type in zones 1 and 2 combined, P<0.0001 by the two-tailed Mann-Whitney test, 95% C.I.) and meiotic germline nuclei (an average of 3.4 RAD-51 foci/nucleus were observed in ztf-8 germlines at zone 5 compared to 3.0 for wild type; P = 0.0045). Higher levels of RAD-51 foci persisted through late pachytene in ztf-8 mutants compared to wild type (2.4 RAD-51 foci/nucleus compared to 1.4, P = 0.0025, and 1.5 foci/nucleus compared to 0.6, P = 0.0081, in zones 6 and 7, respectively) suggesting either a delay in meiotic DSBR or an increase in the levels of DSBs formed during meiosis. This defect in DSBR does not stem from either impaired axis morphogenesis or chromosome synapsis since immunolocalization of either SMC-3, required for sister chromatid cohesion, or SYP-1, a central region component of the synaptonemal complex, was indistinguishable from wild type (Figure 5). To better distinguish the mitotic from the meiotic effects seen in DSBR we quantified the levels of RAD-51 foci in the germlines of ztf-8;spo-11 double mutants, which lack the formation of meiotic programmed DSBs (Figure 4D). Elevated levels of RAD-51 foci were still present throughout the germline compared to spo-11 single mutants, suggesting that DSBs of mitotic origin persist into the meiotic region in ztf-8 mutants. To test if repair of programmed meiotic DSBs is also impaired in ztf-8 mutants, we subtracted the number of foci of mitotic origin found in ztf-8; spo-11 double mutants from the total number of RAD-51 foci observed in ztf-8 single mutants (Figure 4D). Elevated levels of RAD-51 foci were still observed in the meiotic zones of ztf-8 mutants compared to wild type (e.g. zones 6 and 7) indicating that meiotic DSBR is also impaired in ztf-8 mutants contributing to the elevated levels of recombination intermediates detected in the germline. Taken together, these data support a role for ZTF-8 in promoting the normal progression of DSB repair in both mitotic and meiotic germline nuclei. Persistence of unrepaired DSBs can activate a DNA damage checkpoint resulting in increased apoptosis during late pachytene in the C. elegans germline [22]. Interestingly, the elevated levels of RAD-51 foci observed in ztf-8 mutant germlines were not accompanied by increased levels of germ cell apoptosis in this mutant compared to wild type (P = 0.399, by the two-tailed Mann–Whitney test, 95% C.I., Figure 6A). Moreover, the levels of germ cell apoptosis were lower in ztf-8 mutants compared to wild type following induction of exogenous DSBs by exposure to γ-irradiation (P = 0.004). These data suggest that either the DSBs marked by RAD-51 foci are repaired before nuclei are directed into an apoptotic fate or the DNA damage checkpoint machinery is impaired in ztf-8 mutants. To examine this further, we used a HUS-1::GFP transgenic line and monitored the localization of this 9-1-1 DDR component in ztf-8 mutants [8]. The weak HUS-1::GFP signal detected in ztf-8 mutants compared to wild type, even after the induction of exogenous DSBs by γ-IR, suggests that the DNA damage checkpoint operating in late pachytene is impaired in ztf-8 mutants (Figure 6B). However, the observation of higher levels of apoptosis in IR compared to non-IR treated ztf-8 mutants suggests that activation of the late pachytene DNA damage checkpoint, while impaired, is still not fully abrogated in ztf-8 mutants (Figure 6A, P<0.0001). In fact, the level of apoptosis observed in ztf-8 mutants is higher than that in a hus-1(op241) mutant, which is required for CEP-1/p53-dependent DNA damage-induced apoptosis (Figure 6A, P<0.0001, [8]) suggesting only a partial reduction in the activation of germ cell apoptosis. Consistent with previous observations, the level of apoptosis observed in irradiated ztf-8 mutants is significantly reduced in cep-1;ztf-8 mutants (Figure 6A, P<0.0001) and restored to non-IR levels, indicating that ztf-8 mutants are experiencing DNA damage-induced apoptosis. Taken together, these studies indicate a function for ZTF-8 in the 9-1-1 mediated meiotic DNA damage checkpoint. The increased levels of RAD-51 foci observed in mid to late pachytene suggest a role for ZTF-8 in DSBR via homologous recombination during meiosis. To determine whether ZTF-8 plays a role in meiotic crossover formation we examined crossover frequency and distribution in both an autosome (V) and a sex chromosome (X) in ztf-8 mutants compared with wild type (Figure 7A). 46.6 cM and 47 cM intervals, corresponding to 81% and 76% of the whole length (interval A to E) of chromosomes V and X, were examined utilizing 5 single-nucleotide polymorphism (SNP) markers along each chromosome as in [18]. Crossover frequency in this interval was weakly, but not significantly, reduced by 5.5% on chromosome V and 4.1% on the X chromosome compared to wild type (P = 0.7657 and P = 0.8872, respectively, by the two-tailed Fisher's exact test, C.I. 95%). Furthermore, the crossover distribution patterns were not altered in either the autosome or the sex chromosome. Crossover distribution was still biased to the terminal thirds of autosomes and somewhat evenly distributed along the X chromosome as demonstrated in [23]. These results suggest that ZTF-8 is not required for the regulation of either crossover frequency or distribution in the autosomes and the sex chromosome. Further evidence indicates that ZTF-8 does not affect crossover formation. First, the levels of ZHP-3 and MSH-5 foci, which are proposed to mark crossover sites, were indistinguishable between wild type and ztf-8 mutants (Figure 7B; [24]–[26]. Second, mostly six pairs of attached homologous chromosomes, at levels similar to wild type, were detected in late diakinesis oocytes in ztf-8 mutants, suggesting that crossover formation resulted in the formation of functional chiasmata (Figure 7C). Therefore, these data indicate that while ZTF-8 is required for normal DSB repair progression it is not required for completion of interhomolog crossover formation. Both brc-1 and fcd-2 mutants exhibit accumulation of RAD-51 foci but normal levels of crossovers, and are required for meiotic DSB repair using sister chromatids when homologous chromosomes are not available [27], [28]. To test if ZTF-8 is required for intersister repair, we employed a syp-3(ok758) null mutant background in which meiotic DSB formation still takes place but chromosomes no longer synapse and therefore interhomolog recombination is abrogated due to the lack of a stably held homologous chromosome that can be utilized as a template for repair (Figure 7D; [29]). While we did not observe any evidence of chromosome fragmentation, we found that 4.4% of oocytes at diakinesis exhibited misshapen, unstructured chromatin in the double mutants but not in the syp-3 mutants (0/32 in syp-3 and 2/45 in syp-3;ztf-8). Similar unstructured chromatin was observed in brc-1;syp-2 mutants, also impaired for chromosome synapsis, albeit at an approximately 6-fold higher frequency [27], suggesting only a modest contribution by ZTF-8 to intersister repair when interhomolog repair is abrogated during meiosis. To determine whether ZTF-8 localization might be altered in response to either replication arrest or DSBs we exposed wild type worms to either HU or γ-IR, respectively, and monitored ZTF-8 localization in the germline. Unlike the unexposed germlines, in which ZTF-8 signal is present in nuclei at the premeiotic tip and in late pachytene and only very weak signal is observed at transition zone (Figure 2A), brighter ZTF-8 foci were observed from premeiotic tip to transition zone with HU treatment (Figure 6C). ZTF-8 also formed bright aggregates or foci following γ-IR treatment in nuclei from the premeiotic tip to the pachytene stage. These bright foci started to appear 15 minutes after irradiation, increased in intensity at 30 minutes and started to disappear 120 minutes after irradiation, suggesting a transient nature to this change in localization (Figure 6C and Figure S5). While most of the large foci apparent after either γ-IR or HU treatment were localized to the nucleolus, some were also present associated with chromatin. Specifically, 19% (n = 45 nuclei) of the large foci were associated with the DAPI signal in premeiotic tip nuclei following HU treatment, and 26% (n = 115) and 21% (n = 75) in premeiotic tip and pachytene nuclei, respectively, following γ-IR, while these large foci were rarely observed at either stage in untreated worms. Given the higher levels of smaller foci observed associated with chromatin in non-IR worms, these results suggest that ZTF-8 may be relocalizing after exogenous DNA damage, becoming enriched at or near sites of damage. Consistent with our assessment for specificity in DNA damage sensitivity (Figure 4A) we did not observe altered localization of ZTF-8 following exposure to either HN2, UV or CPT (Figure S6) suggesting a specific response primarily to replication arrest and DSB formation. Interestingly, both ATL-1 and ATM-1 (ATM homolog) are required for the proper localization of ZTF-8. Specifically, ZTF-8 is observed forming larger foci in premeiotic tip nuclei in both atl-1 and atm-1 mutants compared to wild type (Figure 6D). However, at the pachytene stage, where crossover recombination is completed, these larger ZTF-8 foci were only observed in atl-1 mutants, but not atm-1, suggesting that ZTF-8 localization is dependent on ATL-1 during both mitotic and meiotic progression. Consistent with this idea, ZTF-8 acquires the same enlarged focal pattern in atm-1;atl-1 double mutants as that observed throughout the germlines of atl-1 single mutants. These observations suggest that proper localization of ZTF-8 requires both ATL-1 and ATM-1 during mitosis and early meiosis, but mostly only ATL-1 during late meiotic prophase. However, given the pleiotropic nature of the atm-1 and atl-1 mutants, we cannot exclude the possibility that the localization of ZTF-8 is altered as a result of alterations to the pattern of damaged DNA. To further examine the nature of the large ZTF-8 foci observed in response to DNA damage, we assessed whether ZTF-8 co-localizes with any known DDR or DNA repair proteins at either 10 or 30 minutes post γ-IR treatment, compared to untreated gonads. We failed to detect co-localization between ZTF-8 and MRE-11 (involved in DSB resection; [30], [31], RPA-1 (single-stranded DNA binding protein; [32]), RAD-51 (strand invasion/exchange protein; [33]) and RAD-54 (required for removal of RAD-51; [34]). However, utilizing a HUS-1::GFP transgenic line we found partial co-localization between ZTF-8 and HUS-1 predominantly in the absence of γ-IR exposure (Figure 6E). Specifically, 82% of HUS-1::GFP foci co-localized with ZTF-8 foci at the premeiotic tip (n = 62 nuclei from 5 germlines). However this level decreased to 9% after γ-IR treatment (n = 46 nuclei from 4 germlines). We observed a similar trend in pachytene nuclei. However, the presence of various stretches and clusters of foci for HUS-1::GFP precluded us from quantifying the degree of co-localization at this stage. Interestingly, 67% of the foci observed co-localizing at the premeiotic tip did not localize to DAPI-stained chromosomes, suggesting that their co-localization may be taking place outside of repair foci when exogenous DSBs are absent. Of note, the HUS-1::GFP transgene has been reported to only partially rescue the apoptotic defect observed in hus-1(op241) mutants [8], and the HUS-1::GFP signal is weak, especially in the absence of exogenous DSBs, the level of co-localization observed between HUS-1::GFP and ZTF-8 might represent an underestimate. Finally, ZTF-8 signal was reduced in hus-1 mutants compared to wild type at both the premeiotic and pachytene stages (Figure 6F), complementing our observation of a decrease in HUS-1 signal in ztf-8 mutants (Figure 6B). Altogether, these data indicate that HUS-1 and ZTF-8 are partly interdependent for their localization and suggest a potential, either direct or indirect, interaction between these proteins. To examine whether ZTF-8 interacts with any of the members of the 9-1-1 complex we applied a yeast two-hybrid approach. We tested the full length and three specific regions of ZTF-8 (N1–330, M270–598 and C400–687) for interactions with potential candidates (Figure 8A). ZTF-8N includes a putative sumoylation site, a zinc-finger domain and the predicted APSES DNA binding motif. ZTF-8M contains a zinc-finger domain and a putative phosphorylation site. ZTF-8C contains three putative sumoylation sites. Interestingly, an interaction was observed between MRT-2/Rad1, and the full length ZTF-8 (Figure 8B). A lack of detectable interaction between MRT-2 and any of the ZTF-8 truncations suggests that the N, M, and C regions alone may not be sufficient to sustain an interaction with MRT-2. Similar to the human RHINO protein [13], a mutation in the conserved APSES DNA binding domain (SSLCPNA to AAAAAAA) abolished the binding affinity to MRT-2 suggesting that the APSES domain is required for the interaction between the member of the 9-1-1 complex and ZTF-8 (Figure 8B and Figure S1). The human RHINO protein was shown to co-immunoprecipitate with TopBP1 and Rad9 suggesting a link to the 9-1-1 complex [13], although a direct protein interaction with any of the 9-1-1 complex members was not demonstrated. Full-length ZTF-8 does not interact via a yeast two-hybrid method with HPR-9 (Rad9 homolog), MUS-101 (TopBP1 DNA topoisomerase 2 beta binding protein), and HUS-1, suggesting that the connection with the 9-1-1 complex may be via MRT-2. CLK-2 (S. cerevisiae Tel2p ortholog), which was previously reported to exhibit synthetic sterility with ZTF-8 [35], also did not interact with the full-length ZTF-8 by this assay. Importantly, similar results were obtained using different combinations of yeast strains and plasmids, further supporting these observed interactions. However, a mild interaction was observed between CLK-2 and the ZTF-8M truncation. Given that only this ZTF-8 truncation also exhibits mild interactions with HPR-9 and MUS-101, these may be false positive (non-specific) interactions resulting from either the misfolding of this truncated protein or it being “sticky”. To examine if ZTF-8 and RHINO indeed share functional conservation, transgenic lines expressing RHINO were tested for their ability to rescue the phenotypes observed in ztf-8 mutant animals. Human RHINO rescued the reduced brood size, elevated levels of RAD-51 foci and impaired germ cell apoptosis observed in ztf-8 mutants (Figure 8C). Altogether, these data support a role for ZTF-8, the functional RHINO homolog, in promoting the proper activation of the DNA damage checkpoint by interacting with MRT-2/Rad1 a component of the 9-1-1 complex. Our studies also suggest that RHINO may be directly connected to the 9-1-1 complex in a similar manner and that it may play a role in maintaining genomic integrity during meiosis in humans. Impaired ztf-8 function results in a reduced brood size, mild embryonic lethality and increased levels of X-chromosome non-disjunction. The hypersensitivity of ztf-8 mutants to exogenous DSBs and replication arrest, coupled with the increased levels of recombination intermediates detected in both mitotic and meiotic germline regions, and the presence of chromatin fragments marked by RAD-51 foci, all strongly support a role for ZTF-8 in homologous recombination repair in C. elegans. In addition to its role in DSBR, our studies suggest that ZTF-8 acts in the DNA damage checkpoint pathway, consistent with the role of RHINO in human cells. This idea is further supported by the fact that proper localization of ZTF-8 requires both ATL-1 and ATM-1, which are kinases central for DNA-damage response [36]. ZTF-8 might be a direct target for phosphorylation by ATM/ATR given that S/TQ sites, shown to undergo such phosphorylation following DNA damage, are present in ZTF-8 (Figure S1) [37]. Furthermore, both the observed synthetic lethality with clk-2 and decreased HUS-1::GFP signal in ztf-8 mutants strongly implicates ztf-8 in DNA damage response. We demonstrated that ZTF-8 partially co-localizes with the 9-1-1 complex and interacts with MRT-2 in a manner dependent on the presence of the APSES domain. We propose that ZTF-8 is involved in promoting repair at stalled replication forks and meiotic DSBs in part by transducing DNA damage checkpoint signaling via the 9-1-1 pathway (Figure 9). Increased levels of RAD-51 foci were observed in both mitotic and meiotic zones in the germlines of the ztf-8 mutants (Figure 4D). What causes accumulation of RAD-51 foci in ztf-8 mutants? The activation of the S-phase cell cycle arrest in ztf-8 mutants indicates that the mitotic increase in the levels of RAD-51 foci likely stems from a role for ZTF-8 in repair at stalled or collapsed replication forks. This idea is supported by the increased nuclear diameter and HU induced embryonic and larval lethality observed in the mutants (Figure 3A and 4A). On the other hand, the elevated levels of RAD-51 foci during meiosis can be explained by the progression of unrepaired breaks of mitotic origin to the meiotic stages as well as defective DSBR during meiosis per se, as evidenced by comparing the levels of mitotic to meiotic (SPO-11-dependent) DSBs (Figure 4D). How does ZTF-8 work in the repair at stalled or collapsed replication forks and SPO-11-induced DSBs? We considered the possibility that ZTF-8 functions as a part of the Shu complex, which has been reported to suppress the HU sensitivity observed in mutants of SGS1, which encodes the budding yeast homolog of the BLM helicase and, similar to ZTF-8, is primarily localized to the nucleolus. However, unlike ztf-8 mutants where unrepaired DSBs persist, the number of RAD51 foci in a shu1 deletion strain is decreased compared to wild type [38]. Moreover, no important amino acid conservation is found between ZTF-8 and the Shu components or their human homologs [38], [39]. Given that ZTF-8 is largely localized to the nucleolus in nuclei at the mitotic zone, we examined whether it might play a role in maintaining G/C tracts, which have the potential to adopt secondary structures such as the G-quadruplex and thus induce DNA replication arrest. However, we did not detect significant changes in the sizes of the GC tracts found in either ztf-8 single or dog-1;ztf-8 double mutants (n = 41 for each), where DOG-1 is the C. elegans homolog of the FANCJ helicase previously implicated in poly(G)/poly(C) (G/C) tract maintenance during DNA replication [40], [41]. Recent studies found that the RNA:DNA hybrid structures known as R-loops formed between nascent mRNA and template DNA during transcription can impair replication and cause checkpoint activation during meiosis, mimicking similar phenotypes found in ztf-8 mutants and suggesting a possible involvement of ZTF-8 in antagonizing R-loop formation [42], [43]. Alternatively, ZTF-8 might be required for the translesion synthesis (TLS) pathway given that it carries a potential ubiquitin-binding zinc-finger (UBZ) domain found in TLS polymerases and that defective TLS repair results in accumulation of RAD-51 foci [44]. Although ztf-8 mutants are not sensitive to DNA interstrand crosslinks (Figure 4A), a role in the TLS pathway is still plausible as not all UBZ-containing TLS components, such as POLK-1/POLκ, are sensitive to ICLs [45]. This idea is further supported by observations that the 9-1-1 complex is required for recruiting translesion polymerases to stalled replication forks in S. pombe [46], [47], and therefore the absence of ZTF-8 might lead to the accumulation of unrepaired breaks. In fact, the two observed functions for ZTF-8 in DNA repair and DNA damage checkpoint activation correspond to the two distinct roles ascribed to the 9-1-1 complex: 1) checkpoint signaling through ATM and ATR to stimulate the DNA repair pathway; and 2) as a recruitment platform for the TLS machinery at stalled replication forks [48], [49]. Our results suggest that ZTF-8 is required for both functions of 9-1-1 and we hypothesize that the former might be its prevalent mode of action during meiosis and the latter its primary mode of action during S-phase. Late pachytene nuclei carrying unrepaired DSBs, as visualized by RAD-51 immunostaining in ztf-8 mutants, can activate a DNA damage checkpoint at that stage and be converted into apoptotic nuclei [18], [22]. Accumulation of ATL-1 further supports the activation of the DNA damage checkpoint via the CEP-1/p53 pathway in the ztf-8 mutants [16]. However, levels of germ cell apoptosis following the induction of exogenous DSBs were not as elevated as in wild type. A simple explanation is that ZTF-8 is required for proper function of either the apoptotic machinery or the DNA damage mediated apoptosis pathway. However, normal levels of physiological germ cell apoptosis are still present in the ztf-8 mutants, suggesting that ZTF-8 is not required for the apoptotic machinery (Figure 6A). Moreover, in cep-1;ztf-8 double mutants the level of apoptosis was comparable to that of cep-1 mutants supporting an impaired DNA damage checkpoint in ztf-8 mutants. In the C. elegans germline HUS-1 and CEP-1/p53 act in the same pathway and HUS-1 is required for the CEP-1/p53-dependent DNA damage induced apoptosis [8]. Our observations of a weaker HUS-1::GFP signal in ztf-8 mutants either in the presence or in the absence of exogenous DSBs, the interaction between ZTF-8 and the 9-1-1 complex via MRT-2, and the weak levels of apoptosis despite the elevated levels of unrepaired recombination intermediates highlighted by RAD-51 foci present in late pachytene, suggest that ZTF-8 is required for the intact DNA damage response signaling pathway. The kinetics of HUS-1::GFP localization are different from that of ZTF-8. ZTF-8 partially co-localizes with HUS-1::GFP, a component of the 9-1-1 DNA damage checkpoint, both in the nucleolus and on chromatin at mitotic and meiotic stages when no exogenous DSBs are present. ZTF-8 starts to form bright foci as early as 15 min after γ-IR treatment, but the number of bright foci starts to decrease 2 hr after irradiation while HUS-1::GFP not yet forms distinct foci at chromatin. Importantly, ZTF-8 does not co-localize with the HUS-1::GFP bright and distinct foci that appear on chromatin as early as 3 hr after γ-IR [8]. These observations are consistent with ZTF-8's relocalization after DNA damage and suggest that ZTF-8 is required for proper 9-1-1-mediated signaling, co-localizing with the complex until DSBs occur, upon which the 9-1-1 DNA damage complex re-localizes to DSB sites. Although ZTF-8 is important for the CEP-1/p53-dependent activation of the meiotic DNA damage checkpoint it is not required for mitotic cell cycle arrest. This is distinct from MRT-2 and HUS-1, which have been previously shown to exhibit both impaired mitotic cell cycle arrest and meiotic DNA damage checkpoint activation [8], [22]. Importantly, mitotic germ cells in ztf-8 mutants were proficient for G2 arrest following exposure to γ-IR (40 Gy) as observed by a 1.5-fold increase in nuclear diameter that was similar to the 1.4-fold increase observed in IR-treated wild type nuclei compared to the non-IR nuclei (n = 50–138 nuclei each for wt, wt +IR, ztf-8 and ztf-8+IR; P<0.0001 by the two-tailed Mann-Whitney test, 95% C.I.). Therefore, the absence of a detectable mitotic cell cycle arrest defect in ztf-8 mutants is not simply due to differences in the type of damage induced, namely stalled replication forks in S-phase compared to DSBs in meiotic prophase I. Instead, this further suggests that ZTF-8 may be required for a separate function of the 9-1-1 complex during S-phase, such as possibly in the TLS pathway, and not in checkpoint signaling. In summary, our study discovered ZTF-8, a previously uncharacterized protein, and its functions in the germline. We have revealed that ZTF-8 plays both mitotic and meiotic roles via its requirements for DSB repair and DNA damage checkpoint activation through an interaction with the 9-1-1 complex. A previous study identified an overexpression of RHINO in breast cancer cells and that its depletion by small-hairpin RNAi suppressed their cell growth [50]. However, no previous studies have been reported on the meiotic roles of RHINO although it is expressed in both testis and ovary in normal human tissues [50]. Here we demonstrated that ZTF-8 and RHINO share functional conservation. Therefore, the insights we provide as to how ZTF-8 is tied into the 9-1-1 complex for repair at stalled replication forks and meiotic DSBs, as well as for the activation of the CEP-1/p53-dependent germ cell apoptosis pathway, shed new light on how RHINO may be operating via the 9-1-1 complex in these different contexts during mammalian mitosis and meiosis. C. elegans strains were cultured at 20°C under standard conditions as described in Brenner [51]. The N2 Bristol strain was used as the wild-type background. The following mutations and chromosome rearrangements were used in this study: LGI: atm-1(gk186), cep-1(lg12501), hus-1(op244), hT2[bli-4(e937) let-?(q782) qIs48](I; III); LGIII: ztf-8(tm2176), sDf121, qC1[dpy-19(e1259) glp-1(q339) qIs26] (III); LGIV: spo-11(ok79), nT1[unc-?(n754) let-? qIs50](IV; V), nT1[qIs51] (IV; V); LGV: atl-1(tm853), syp-1(me17) [[52] [53] [51] [54] [55]. The ztf-8(tm2176) mutant was generated by the Japanese National BioResource Project for C. elegans and carries a 524 base pair out-of-frame deletion that removes most of exon 6 along with exons 7 through 11 (Figure 1). This deletion results in a premature stop codon and loss of the zinc-finger motifs located in the middle of ZTF-8, the four ΨKXE consensus sumoylation sites, and a putative phosphorylation site (http://www.phosphopep.org). To test if human RHINO can rescue the ztf-8 mutant phenotypes, RHINO cDNA was cloned into the pID2.02 plasmid, which contains unc-119(wt), then injected into ztf-8(tm2176);unc-119(ed9) mutants and screened for wild type (non-Unc) moving worms [56], [57]. Rescued non-Unc worms were irradiated to obtain single/low-copy integration of transgenes as described in [58]. Presence of RHINO was confirmed by PCR. Empty pID2.02 was injected into ztf-8(tm2176);unc-119(ed9) as a control. ZTF-8 homology searches and alignments were performed using Uniprot (http://www.uniprot.org/). Pfam and Prosite (release 20.70) were used for zinc-finger motif predictions [59]. Quantitative analysis of RAD-51 foci was performed as in [20]. Between 5 to 8 germlines were scored for each genotype. The average number of nuclei scored per zone for a given genotype was as follows, ± standard deviation: zone 1, n = 157.4 ±38.3; zone 2, n = 151.8±39.5; zone 3, n = 136.6±27.1, zone 4 = 212.8±69.1, zone 5 = 130.0±49.3, zone 6 = 124.5±56.4, zone 7 = 108.2±46.8. Statistical comparisons between genotypes were performed using the two-tailed Mann-Whitney test, 95% confidence interval (C.I.). Young adult homozygous ztf-8 animals were picked from the progeny of ztf-8/qC1 parent animals. To assess for IR sensitivity, animals were treated with 0, 10, 30 or 100 Gy of γ-IR from a Cs137 source at a dose rate of 1.8 Gy/min. For HN2 sensitivity, animals were treated with 0, 50, 100 or 150 µM of HN2 (mechlorethamine hydrochloride; Sigma) in M9 buffer containing E. coli OP50 with slow shaking in the dark for 20 hours. CPT (Sigma) treatment was similar but with doses of 0, 100, 500, or 1000 nM. After treatment with either HN2 or CPT, animals were washed twice with M9 containing TritonX100 (100 ml/L) and plated to allow recovery for 3 hours [18]. UV irradiation treatment was performed utilizing the XL-100 Spectrolinker UVC. Worms were exposed to 0, 100 or 150 J/m2 of UVC and plated to allow recovery for 3 hours. HU sensitivity was assessed by placing animals on seeded NGM plates containing either 0, 10 or 15 mM HU for 20–24 hours. Hatching sensitivity was examined in>24 animals 4 hours after HU treatment. For all other damage sensitivity experiments,>24 animals were plated, 7 per plate, and hatching was assessed for the time period of 20–24 hours following treatment. For L1 genotoxic assays, L1(P0) worms were plated on NGM plates with either 0 or 25 mM HU and incubated for 16 hours. The number of live adult progeny (F1) were counted as described in [19]. Each damage condition was replicated at least twice in independent experiments. Feeding RNAi experiments were performed at either 20°C or 25°C as described in [60]. Either the entire coding sequence of ztf-8 (Geneservice) or cDNA corresponding to its C-terminal 501 bp cloned into the pL4440 feeding vector were used for RNAi experiments. HT115 bacteria carrying the empty pL4440 vector were used as control RNAi. cDNA was produced from single-worm RNA extracts using the One step RT-PCR system (USB). The effectiveness of RNAi was examined by assaying the expression of the transcript being depleted in four individual animals subjected to RNAi by feeding. Expression of the myo-3 (K12F2.1) transcript was used as a control. Rabbit polyclonal antibodies against N- and C-terminal peptides of C. elegans ZTF-8 (ETLKEEGAHFYKHFKYKRYC and CHHSRSSYRGNRDDRGSRW, respectively) were generated by Yenzym antibodies, LLC. Antisera were affinity-purified using SulfoLink (Pierce) following the manufacturer's instructions. Whole mount preparations of dissected gonads, fixation and immunostaining procedures were carried out as described in [20]. Primary antibodies were used at the following dilutions: rabbit α-ZTF-8 (1∶200), rabbit α-ATL-1 (1∶500; [16]), rabbit α-RAD-51 (1∶2000; SDIX), mouse α-NOP-1 (1∶100; EnCor Biotech), rabbit α-PCN-1 (1∶10000; [61]), rabbit α-phospho Ser10 Histone H3 (1∶200; Upstate Biotechnologies), guinea pig α-SYP-1 (1∶200; [62]), rabbit α-SMC-3(1∶100; Chemicon), rabbit α-HDA-1 (1∶200; Santa Cruz), rabbit α-Histone H3 (1∶200; Cell Signaling), rabbit α-CDK1 pTyr15 (1∶50; Calbiochem) and rabbit α-pCHK-1 (1∶50; Santa Cruz). Secondary antibodies used were: Cy3 anti-rabbit, FITC anti-rabbit, Cy3 anti-guinea pig and FITC anti-mouse (Jackson Immunochemicals), each at 1∶200. Immunofluorescence images were collected at 0.2 µm intervals with an IX-70 microscope (Olympus) and a CoolSNAP HQ CCD camera (Roper Scientific) controlled by the DeltaVision system (Applied Precision). Images were subjected to deconvolution by using the SoftWoRx 3.3.6 software (Applied Precision). Germlines of age-matched (20 hours post-L4) animals were analyzed by acridine orange staining, as described in [24], utilizing a Leica DM5000B fluorescence microscope. Between 22 and 95 gonads were scored for each genotype. Statistical comparisons between genotypes were performed using the two-tailed Mann-Whitney test, 95% C.I. The full-length of the ztf-8 open reading frame, as well as C- (400–687), middle (270–598) and N- (1–330) terminal truncations were amplified by PCR. A cDNA library generated from mixed-stage C. elegans was used for the amplification with primers that contain Gateway compatible sequences and a gene specific sequence as indicated in Table S1. Gateway cloning, cDNA and ORFeome library screening, and X-Gal, -URA and 3AT assays for examining yeast two hybrid interactions were performed as in [63].

10.1371/journal.ppat.1000534

The Two-Domain LysX Protein of Mycobacterium tuberculosis Is Required for Production of Lysinylated Phosphatidylglycerol and Resistance to Cationic Antimicrobial Peptides

The well-recognized phospholipids (PLs) of Mycobacterium tuberculosis (Mtb) include several acidic species such as phosphatidylglycerol (PG), cardiolipin, phosphatidylinositol and its mannoside derivatives, in addition to a single basic species, phosphatidylethanolamine. Here we demonstrate that an additional basic PL, lysinylated PG (L-PG), is a component of the PLs of Mtb H37Rv and that the lysX gene encoding the two-domain lysyl-transferase (mprF)-lysyl-tRNA synthetase (lysU) protein is responsible for L-PG production. The Mtb lysX mutant is sensitive to cationic antibiotics and peptides, shows increased association with lysosome-associated membrane protein–positive vesicles, and it exhibits altered membrane potential compared to wild type. A lysX complementing strain expressing the intact lysX gene, but not one expressing mprF alone, restored the production of L-PG and rescued the lysX mutant phenotypes, indicating that the expression of both proteins is required for LysX function. The lysX mutant also showed defective growth in mouse and guinea pig lungs and showed reduced pathology relative to wild type, indicating that LysX activity is required for full virulence. Together, our results suggest that LysX-mediated production of L-PG is necessary for the maintenance of optimal membrane integrity and for survival of the pathogen upon infection.

The human pathogen Mycobacterium tuberculosis (Mtb) survives in the hostile intracellular environment, in part, by withstanding the actions of host-induced cationic antimicrobial peptides (CAMPs). Membrane phospholipid composition and the resultant charge could play an important role in Mtb survival within the host. Acidic phospholipids such as cardiolipin, phosphatidylinositol and its mannoside derivatives, phosphatidylglycerol, and a single basic species, phosphatidylethanolamine, are constituents of the Mtb membrane bilayer. We demonstrate that lysinylated phosphatidylglycerol (L-PG) represents another basic phospholipid and that the lysX gene, which encodes a two-domain protein with lysyl transferase and lysyl-tRNA synthase activities, is necessary for L-PG production. We show that L-PG is required for maintenance of an optimal membrane potential and resistance towards CAMPs. Phagosomes containing the lysX mutant showed an increased association with lysosomes, and the lysX mutant showed growth defects in mouse and guinea pig lungs, indicating that LysX activity is required for full virulence. Collectively, our results suggest that LysX activity, which is responsible for the production of L-PG, is necessary for maintenance of an optimal membrane potential such that the pathogen can grow optimally upon infection, presumably by withstanding the actions of CAMPs.

Mycobacterium tuberculosis (Mtb), the causative agent of tuberculosis, is a successful human pathogen that has infected more than one-third of the world's population. The success of Mtb as an infectious agent relies, in part, on its ability to modulate the expression of bacterial factors in response to infection so that it can successfully multiply within the hostile host environment [1]. The characteristic lipid-rich cell envelope of Mtb is one of the factors believed to contribute to its survival in vivo [2],[3]. It is generally believed that Mtb polar lipids (PoLs) consisting of acidic phospholipids (PL) such as cardiolipin (CL), phosphatidylglycerol (PG), phosphatidylinositol and its mannoside derivatives, in addition to basic PL such as phosphatidylethanolamine, are important constituents of the Mtb membrane [2]. Mtb PLs are known to function as important immune modulators [4] and have been shown to be released within phagosomes and transferred into lysosomes [5],[6]. It is interesting to note that PG, which is an abundant PL in other bacteria, is only a minor species in Mycobacteria, whereas CL is a major species [2],[3] with a high turnover rate [7]. The relative ratio of acidic to basic PLs is one of the determinants of net membrane charge. In some Gram-positive pathogens such as Staphylococcus aureus and Listeria monocytogenes, a fraction of the PG or CL molecules, or both, are lysinylated by the esterification of a glycerol hydroxyl group to lysine. Lysinylation imparts a net positive charge to these acidic PLs. This could, in turn, influence the ratio of acidic to basic PLs, resulting in an altered membrane charge. This could explain the bacterial susceptibility to cationic antibiotics (CAMAs) and peptides (CAMPs) [8],[9]. Although Mtb PLs have been well characterized for more than four decades, it is unknown if lysinylated PLs are a subset of the Mtb PLs and, if so, what the consequences associated with the absence of these lysinylated PLs might be. The present study demonstrates that the Mtb lysX gene, encoding a two-domain protein, is required for the production of lysinylated PG (L-PG), and the absence of L-PG is associated with changes in membrane potential, increased sensitivity to CAMAs and CAMPs, and growth defects in vivo. In order to detect lysinylated PLs in Mtb, actively growing cultures were incubated with 14C-lysine for 3 days; total lipids were extracted, and PoLs were resolved by thin layer chromatography (TLC). A distinct radiolabeled lipid was evident, indicating that lysinylated PLs are members of the Mtb PoL pool (Fig. 1A, lane i). In S. aureus, the mprF gene is responsible for L-PG production [8]. Homology searches of the Mtb genome identified Rv1640c as lysX, which encodes an mprF-like gene as a fusion to a lysyl-tRNA synthetase (lysU). The latter gene is distinct from the essential housekeeping tRNA synthetase (Rv3598c). The mprF gene in S. aureus encodes a protein with potential lysyl transferase activity [10]. In order to evaluate the function of lysX, we created a lysX mutant strain, Rv-80lys, by replacing the majority of the coding region comprising the mprF and lysU domains with a gentamycin resistance cassette using homologous recombination (see Methods section). A complementing derivative of this strain, Rv-81ami, was created by integrating a plasmid expressing the intact lysX gene under the control of the amidase promoter [11]. The lysX mutant strain was found to be defective in the production of L-PoLs (Fig. 1A, lane iv compared with lane i). L-PoL production was restored, however, in the lysX complemented strain Rv-81ami (Fig. 1A, see lane vii), confirming that the lysX gene product is responsible for the production of L-PoLs. Staining TLC plates with iodine (lanes ii, v, viii and xi) or ninhydrin (lanes iii, vi, ix and xii), on the other hand, did not detect L-PoLs, indicating that they may not be an abundant lipid species. We cultured Mtb in the presence of 14C-acetic acid and extracted total lipids, followed by TLC separation and subsequent quantification of L-PoL relative to total input radioactivity, and found that L-PoL accounts for approximately 0.3% of the total lipids (data not shown). In order to determine the nature of the L-PoL, preparative 2D-TLC was carried out to collect L-PoLs. Structural analysis of the L-PoL was carried out using a combination of MALDI-MS, amino acid analysis and NMR (Fig. 2A–D). The MALDI-MS analysis in negative-ion mode revealed m/z 681 ([M-H]−) to be the molecular ion peak (Fig. 2A). The 1H-NMR results confirmed the presence of an acetyl group at δCH3 2.1 ppm and δCH2 from the primary amine in lysine at δ 2.4 ppm (Fig. 2B), whereas the 31P-NMR spectrum showed a shift in the phosphorus resonance spectrum at δ −14.96 ppm (Fig. 2B inset). Fatty acid analysis demonstrated that the molecule was C18 (data not shown), and amino acid analysis following acid hydrolysis confirmed the presence of lysine (Fig. 2C). Together, these data demonstrate that lysine is covalently linked to PG with the predicted structure shown in Fig. 2D. The L-PoL in the text is referred to hereafter as L-PG. Similar structural analyses of the corresponding unlabeled PoL of the slower migrating radioactive lipid species of lysX (Fig. 1A, lane iv) could not be done, given that it was present in negligible quantities (not shown). The thermal decomposition products of lysine are well characterized [12]. We speculate that the PoL accumulation in the lysX strain is a consequence of lysine degradation. Further studies are required to clarify the nature of the lipid species accumulating in the lysX mutant. As previously noted, the Mtb lysX is a fusion gene encoding both mprF and lysU activities, with mprF located at the 5′ end of the lysX gene (see Fig. 1B). The Gram-positive bacteria that have been shown to produce L-PG, however, contain only mprF. The Mtb MprF and S. aureus MprF share three domains of unknown function, DUF470, DUF471 and DUF472. In order to evaluate whether L-PG production in Mtb requires the activities of both the LysU and MprF domains, we generated Rv-82med, a lysX complemented derivative that produces only the MprF domain (see Methods) and evaluated its ability to produce L-PG following the incubation of actively growing cells with radiolabeled lysine. The Rv-82med strain, much like Rv-80lys, was defective in L-PG production (see Fig. 1A, lane x and compare with lane iv). Quantitative real-time PCR analysis using primers and TaqMan probes targeted to the mprF region of lysX (compared to the 16S rRNA housekeeping gene) revealed that the expression of mprF in Rv-82med was comparable to that in Rv-03 wild type and Rv-81ami (data not shown). Together, these results indicate that Rv-82med expresses mprF and that the MprF domain alone is not sufficient for the production of L-PG in Mtb. Gram-positive organisms such as S. aureus and B. subtilis are sensitive to cationic antimicrobial antibiotics (CAMAs) such as vancomycin (Van) and polymyxin-B (PMB) and to cationic antimicrobial peptides (CAMPs) such as human neutrophil peptide (HNP-1) and lysozyme. On the other hand, Mtb is generally tolerant to these compounds. HNP-1 and lysozyme are produced in neutrophils and macrophages, respectively. It is generally believed that CAMPs induce cell death by interfering with the integrity of the negatively charged membrane. Furthermore, the ability of intracellular pathogens to resist the action of CAMPs produced by the host is, in part, responsible for pathogen proliferation upon infection [13]. The presence of lysine groups on the acidic PG would impart a net positive charge and, therefore, could affect the net ratio of positively charged to negatively charged PL species. Thus, the absence of L-PG could make the Mtb membrane relatively acidic, thereby sensitizing the bacterium to the action of CAMAs. To test this possibility, actively growing Rv-03, Rv-80lys, Rv-80ami and Rv-82med were exposed to Van and PMB, and growth and viability were measured (Fig. 3A). Van and PMB interfered with the growth and viability of Rv-80lys and Rv-82med (see Fig. 3A-ii and 3A-iii), inset showing viability after 3 days of exposure; and Figure S1 showing viability after 6 days of exposure). Comparisons of growth, measured as the change in optical density (OD), and viability, measured as the change in CFU, revealed that while the lysX mutant was relatively more sensitive to Van and PMB than Rv-03, it was able to recover when grown in the absence of antibiotics, indicating that Van and PMB do not exert potent bactericidal activity. All of the strains grew well in the absence of antibiotics, although the lysX mutant showed a small reduction in growth rate in the absence of antibiotics (Fig. 3A-i), inset shows an approximately 0.3 log reduction in viability). Visualization of Rv-80lys cells following nucleoid staining and bright field or fluorescence microscopy did not reveal any significant differences in cell morphology or nucleoid organization (data not shown). The increased sensitivity of the lysX strain to antibiotics suggests an enhanced association between the two. To test this possibility, actively growing lysX and Rv-03 cells were stained with fluorescent-vancomycin (Fl-Van), and the staining patterns were visualized by fluorescence microscopy. Earlier studies revealed that in stained Mtb cells, Fl-Van associates with the nascent growth zones, primarily at the poles and mid-cell septa [14]. These studies also indicated that not all Mtb cells could be stained with Fl-Van [14]. We found that a higher percentage of lysX cells were stained with Fl-Van compared to Rv-03 (see Figure S2). Approximately 52% of lysX cells showed staining patterns not only at the mid-cell and polar septa but also over the entire length; meanwhile, only 32% of wild type cells showed such a staining pattern (see Figure S2 legend for details). These results are consistent with the idea that Van is able to gain better access to Rv-80lys cells compared to Rv-03 cells. Next, we examined whether the lysX mutant was also sensitive to lysozyme and HNP-1. Similar to the results seen with the CAMAs, HNP-1 and lysozyme significantly reduced the viability of Rv-80lys compared to Rv-03 and Rv-81ami (Fig. 3B-i and 3B-ii), see legends for P values). The phenotype of Rv-82med was found to be similar to that of Rv80lys (Fig. 3B). Together, these results indicate that the absence of L-PG production is associated with increased sensitivity of the bacterium to the actions of CAMPs and CAMAs. Importantly, these experiments also showed that complementation of the lysX mutant restored the wild type growth phenotype under these conditions (Fig. 3A and 3B). We wished to test whether the absence of L-PG production in Rv-80lys cells was associated with changes in the properties of the PL bilayer (e.g., membrane potential). The membrane potential of the Rv-80lys cells was determined using a slow-response membrane potential-sensitive dye, DiOC2(3), and comparing with Rv-03 cells. This cationic cyanine dye exhibits green fluorescence (Ex = 488 nm and Em = 520 nm) in the monomeric state and red fluorescence (Ex = 488 nm and Em = 620 nm) in the aggregated or oligomeric state. As a negative control, the membrane potential was measured following exposure of the cells to the proton ionophore m-chlorophenylhydrazone (CCCP), which is known to eliminate the proton gradient across the membrane. As seen in Figure 4, the membrane potentials (measured as the ratio of red to green fluorescence) of the lysX mutant Rv-80lys and Rv-82med were 21% and 17%, respectively, higher than that of the Rv-03 and complemented Rv-81ami (P<0.002). The increased ratio of red to green fluorescence observed in lysX mutants suggests accumulation of the positively charged lipophilic dye on the negatively charged membrane. The red to green fluorescence ratio in all strains was decreased to similar levels (∼41%) in the presence of CCCP (P = 0.001). Presumably, this reduction reflects the completely depolarized state of the membrane. Together, these results indicate that the membranes of Rv-80lys and Rv-82med are hyperpolarized relative to Rv-03 and Rv-81ami. We next examined whether the Mtb lysX mutant Rv-80lys showed proficient growth in macrophages upon infection of the THP-1 macrophage cell line. The Rv-80lys showed a modest growth defect in macrophages compared to Rv-03 and complemented Rv-81ami (see Fig. 5, P = 0.01 for day 3, P = 0.006 for day 6). Similar results were also noted for Rv-82med (Fig. 5, P<0.014 for day 3 and P<0.03 for day 6 compared to Rv-03). Intracellular replication of Mtb is, in part, due to its ability to resist the delivery of its phagosomes to lysosomes [15]. This process can be visualized by examining the co-localization of Mtb with the lysosome-associated-membrane protein (LAMP-1). In order to address whether the lysX mutant had a phagosome-lysosome fusion defect, we infected macrophages with Mtb strains expressing green-fluorescent protein and visualized co-localization with LAMP-1. Increased association of phagosomes containing Rv-80lys with lysosomes was evident compared to Rv-03 and complemented Rv-81ami (Fig. 6, p<0.001). Rv-82med behaved like Rv-80lys, indicating that the full-length lysX gene is required for functional activity. These results are consistent with the hypothesis that the lysX mutant is not as proficient as the Rv-03 strain in preventing fusion of phagosomes with lysosomes, which could contribute to defects in intramacrophage replication. The production of inflammatory cytokines tumor necrosis factor-alpha (TNF-α), IL-6 and IL-10 is necessary to mount a protective immune response against Mtb infection [16],[17],[18]. TNF-alpha restricts the growth of Mtb in alveolar macrophages [18], and the multiplication of virulent Mtb in monocyte-derived macrophages (MDMO) is associated with suppression of TNF-α production during the early periods after infection. To test selected pro-inflammatory cytokine responses of macrophages, MDMO were infected with Rv-80lys and Rv-03 strains, and the production of TNF-α and IL-6 was measured (Fig. 7). As can be seen, the secretion of TNF-α was elevated after infection with the lysX mutant compared to the wild type and complemented strains, similar to MDMO cells exposed to PMA (Fig. 7A). Similarly, the secretion of IL-6 was also increased following infection with the lysX mutant compared to the wild type and complemented strains (Fig. 7B). To evaluate the phenotype of lysX in vivo, C57BL/6 mice and Hartley strains of guinea pigs were aerosol infected with lysX, and the viability of the pathogen was measured (Fig. 8). The lysX mutant showed only a modest growth defect in mice (Fig. 8A) but was clearly attenuated in guinea pigs (Fig. 8B) and showed reduced dissemination to the spleen (Figure S4). Gross pathology and histopathology of the lungs of infected mice at 28 days (Fig. 8C and E) and guinea pigs at 42 days (Fig. 8D and F) showed distinct differences between the wild type and the lysX mutant. Hematoxylin-eosin staining confirmed that the lungs infected with Rv-03, but not those infected with Rv-80lys, had extensive inflammation in both species and showed caseating granulomas in guinea pigs. The vast differences in the growth kinetics and pathology between the wild type and lysX mutant pathogens indicate that LysX activity is required for full virulence. Interestingly, the lysX complemented strain Rv-81ami behaved like the lysX mutant with respect to in vivo growth and pathology (data not shown), indicating that the complemented strain is not able to restore lysX function in vivo. The ability of Mtb strains to produce complex cell wall-associated lipids called phthiocerol dimycocerosates (PDIM) and to bind and reduce neutral red dye is associated with virulence. Avirulent and attenuated strains are defective in these processes. Furthermore, virulent Mtb strains propagated in the laboratory often lose these properties [2],[19],[20],[21]. The neutral red reduction and PDIM profiles of the lysX mutant were comparable to those of wild type cells (Figure S5), indicating that the observed in vivo growth defects of the lysX mutant are not due to a loss of PDIM and defect in neutral red reduction. The primary conclusion of our data is that L-PG is one of the basic PLs in Mtb and that the lysX gene, encoding the two-domain LysX protein, is responsible for its production. Although L-PG is a minor PL species of Mtb, its absence has several consequences, one of which is an alteration of the membrane potential. This underscores the role of LysX activity in maintaining optimal membrane function. Presumably, the absence of L-PG in the lysX mutant shifts the ratio of acidic to basic PLs, thereby hyperpolarizing the membrane. A consequence of the absence of L-PG is the increased sensitivity of the pathogen to lipophilic antibiotics such as PMB and Van. It is likely that hyperpolarization of the membrane in the lysX mutants due to its net negative charge promotes interactions with cationic peptides and antibiotics produced by the host immune system, which in turn could lead to the killing of the invading pathogens [13],[22]. It is known that host-induced CAMPs are one of the frontline defenses against invading pathogens. Therefore, sensitization of lysX mutant Mtb cells to the action of CAMPs suggests that maintenance of the optimal membrane potential is necessary for Mtb growth in vivo. In partial support of this claim, we found that the lysX mutant showed defects in intracellular replication (Fig. 5) and that infection of macrophages with lysX led to increased production of pro-inflammatory cytokines (Fig. 7). We also found that the lysX mutant showed increased co-localization with LAMP-1 vesicles (Fig. 6). Finally, we showed that the lysX mutant was attenuated in guinea pig lungs and had a modest growth defect in mouse lungs (Fig. 9). Together, these results are consistent with the hypothesis that LysX activity is required to maintain an optimal membrane potential and possibly to promote pathogen survival upon infection. Notably, the gross pathological differences between the lysX mutant and wild type were striking compared to the modest differences in growth in vitro and ex vivo (see Figs. 3, 5 and 8). The reduced bacterial burden and the reduced pathology and size of granulomas in the lungs of guinea pigs clearly suggest that LysX activity is required for bacterial multiplication and virulence. Evaluation of the host-induced cytokine response following different stages of infection with wild type and lysX mutant pathogens could provide valuable insights into lysX function. Our studies also showed that the lysX mutant, like wild type, retained the ability to produce PDIMs and reduce neutral red (Fig. S5). It remains to be evaluated, however, if other membrane and cell wall-associated lipids are modulated in the lysX background. The production of L-PG is believed to involve two biochemical steps: the generation of lysyl-tRNA by the LysU protein and the transfer of a lysine group from the lysyl-tRNA to PG by MprF, a membrane-bound lysyl-transferase protein [23]. The Gram-positive bacteria shown to produce L-PG carry a single housekeeping lysU gene that encodes a cytosolic LysU protein [8],[9],[24]. E. coli does not contain L-PG, but ectopic expression of the S. aureus mprF gene allows E. coli to accumulate L-PG in their membranes, suggesting that cytosolic LysU and membrane-bound MprF cooperate to produce L-PG [10],[25]. Mtb contains two lysU genes, one encoded by Rv3598c, which is an essential gene, and the other encoded by the lysU domain of lysX [26]. Since expression of the mprF fragment of lysX does not lead to the production of L-PG (Fig. 2), it appears that in Mtb, unlike in other bacteria, the cytosolic LysU and the membrane-bound MprF do not cooperate to produce L-PG. This raises the question as to why a dedicated lysU gene product, distinct from the housekeeping gene, is required for L-PG production in Mtb. One possibility is that the lysinylation reaction occurs on the membrane, and the local presence of LysU and MprF activities are required to transfer lysine from the lysyl-tRNA to the membrane-bound PG. If the cytosolic lysyl-tRNA could not diffuse through the Mtb plasma membrane, a separate activity would be needed to replace it. Nonetheless, such dedicated activities imply that PG lysinylation in Mtb is a tightly regulated reaction. The temporal expression profile of Mtb genes upon infection in mice shows that lysX is upregulated during acute and chronic infection [27]. Presumably, increased expression levels of lysX would ensure that sufficient levels of L-PG were produced to maintain the optimal ratio of acidic to basic PLs. This would, in turn, ensure that the optimal membrane potential required for Mtb proliferation upon infection is maintained. Another possibility, although unlikely, is that the demand for lysyl-tRNA required for lysinylation and protein synthesis cannot be met by a single housekeeping enzyme. Clearly, however, further studies are required to address this issue. While this manuscript was in preparation, Vandal et al. reported the characterization of several transposon mutants of Mtb that were hypersensitive to acidic pH, one of which was lysX [28],[29]. Their transposon mutants were hypersensitive to antibiotics and other stressors such as heat, SDS and DETA-NO. Although the lysX mutant was moderately sensitive to DETA-NO, its growth was not attenuated in murine lungs. It is unknown whether L-PG is produced in the lysX transposon mutant and whether the lysX mutant shows any residual activity. As shown in Fig. 1A, our lysX mutant was generated by removing most of the coding sequence responsible for producing the mprF and lysU activities. We demonstrated that L-PG was not produced in the lysX mutant and that maintenance of the membrane potential and resistance to CAMPs were dependent on LysX activity. Importantly, we showed that LysX activity was required for full virulence in mice and guinea pigs. These results underscore the importance of lysX function in Mtb survival upon infection. One limitation of our results, however, is that the complemented Rv-81ami was not able to restore the lysX defect in vivo, although it did restore defects in other assays reported in this study. One possibility is that the expression of lysX in-trans at an attB locus was not sufficient to restore the LysX activity to optimal levels, and small changes in activity could have consequences for the complementation phenotype in vivo. Further studies are required to address this issue. L-PG appears to be a minor lipid species, yet the loss of L-PG production affected membrane potential and Mtb growth in vitro and in vivo. It is interesting to note that PG, the purported substrate of L-PG, is also a minor lipid species in Mtb and other mycobacterial species [2],[3],[30],[31]. This raises the question of how the lysinylation of a minor PL species contributes to the observed phenotype. It is known that PG is a biosynthetic intermediate of CL, one of the major PL species of mycobacteria. Indeed, the enzymatic activities responsible for CL production from PG pools have been detected in mycobacteria [32]. PG also accumulates as a result of CL catabolism and, if unregulated, could be further processed to produce a diacylglycerol intermediate via the action of phospholipases [33],[34]. Accordingly, we speculate that the lysinylation step helps to prevent PG degradation such that the optimal membrane potential required for Mtb survival upon infection is maintained. Our results also suggest that changes in membrane potential are a potential mechanism for regulating CAMP sensitivity in Mtb and possibly in the mprF mutants of other bacteria; therefore, this could be exploited to develop novel antimicrobial compounds. It is tempting to speculate that by manipulating LysX activity, we could promote the action of other conventional antibiotics against Mtb. Mouse and guinea pig infection protocols were approved by the Animal Care and Use Committee at Johns Hopkins School of Medicine, Baltimore, MD for mice and at Texas A & M University, College Station, TX for guinea pigs, under the NIH contract (Tuberculosis Animal Research and Gene Evaluation Task force). The Mtb strains were cultured in Middlebrook 7H9 broth supplemented with 10% OADC (oleic acid, albumin, dextrose, catalase) and 0.05% Tween-80. As needed, 50 µg/ml hygromycin (hyg), 10 µg/ml kanamycin (kan) or 50 µg/ml gentamycin (gm) was added. For the determination of viable colonies and the scoring of recombinants, bacterial cells were plated on Middlebrook 7H11 agar plates containing the appropriate antibiotics. In some experiments, cultures were grown in the presence of L-[U-14C]-lysine (300 mCi/mmol, Amersham Pharmacia Biotech) or [1,2-14C] acetic acid (46 mCi/mmol, PerkinElmer), and total lipids were extracted and resolved by TLC. The radioactivity present in the L-PG spot was determined and normalized relative to total in put radioactivity. The lysX coding region was cloned downstream of the amidase promoter in an integration-proficient, hygromycin-resistant plasmid and electrotransformed into Mtb in order to generate the lysX merodiploid strain. The chromosomal copy of the lysX gene was disrupted in the lysX merodiploid background by homologous recombination as described previously [35]. Using this approach, 90% of the lysX coding region was replaced with a 900-bp gentamycin resistance cassette. This strain, designated as Rv-81ami, was the lysX complemented strain. Next, the resident integrated plasmid encoding the functional lysX gene was replaced with an empty kanamycin-resistant plasmid to generate the lysX mutant strain, designated as Rv-80lys, as described [14],[35],[36],[37]. A cartoon depicting the lysX mutant and complemented strain construction is shown (Fig. 9). All strains were confirmed by PCR and Southern blot analysis. For the generation of the lysX complemented strain expressing the mprF domain, a 1,950 bp lysX gene encoding the mprF domain was amplified by PCR using the primers MVM530lysF (5′-GGCGAATTCCATATGGGACTCCACTTAACTG-3′) and lys650MM606R (5′-AGC AGCAAGCTTCTAGAATCACGCCAACCGCTCGGGACTGC-3′) and cloned into the pJFR19 vector under the control of the amidase promoter [38]. The integrity of cloned insert was verified by DNA sequencing. This recombinant plasmid was used to replace the resident empty plasmid in the Rv-80lys mutant to generate Rv82-med. This strain was confirmed by PCR and Southern blot analysis. The human monocyte cell line THP-1 (American Type Culture Collection, Rockville, Maryland) was used. Cells were grown in RPMI 1640 (Invitrogen, CA) supplemented with 2 mM L-glutamine, 1 mM sodium pyruvate, 10% fetal bovine serum (Invitrogen) and 100 U/ml penicillin G (Sigma, MO). The viability of the macrophages was determined using trypan blue staining. Monocytes were differentiated into macrophages by exposure to 50 nM PMA (phorbol 12-myristate 13-acetate; Sigma) and 7.5 ng/ml IFN-γ (human interferon-gamma, Peprotech) for 24 h, followed by a 24 h incubation with 50 nM PMA alone. The macrophages were washed three times with RPMI 1640 medium and incubated in medium that was not supplemented with PMA or IFN- γ for the next 24 h. Single cell suspensions of Mtb strains in RPMI 1640 media were used to infect 4.5×105 macrophages in triplicate in a 24-well plate at a multiplicity of infection of 1∶2–4. After 3 hours of infection, macrophages were lysed in 0.09% SDS, and viability was determined to get a t0 count. No statistical differences in viability among these strains were noted at t0. Subsequently, macrophages at 3 and 6 days post-infection were also processed in order to determine Mtb viability. The lysX strains were no more sensitive than the other strains in terms of the concentrations of SDS used to lyse the macrophages and process them for viability determination (see Figure S3). THP-1 macrophages (5×105) attached to glass coverslips were infected with GFP-producing Mtb (Rv-03, Mtb Rv-80lys, Rv-81ami and Rv-82med). The macrophages were fixed, blocked and incubated with H4A3 monoclonal antibodies to LAMP-1, followed by a rhodamine-conjugated goat anti-mouse IgG, as described [38]. Bacterial co-localization with LAMP-1-positive vesicles appeared as yellow spots. The experiments were done in duplicate, and representative images are shown. Peripheral blood mononuclear cells (PBMC) were isolated from healthy volunteers by differential gradient centrifugation on Ficoll-Paque Plus (Amersham Biosciences). Adherent monocytes were isolated by seeding 5×106 cells in 24-well plates in MDMO-media (RPMI 1640 supplemented with 10% heat-inactivated human serum) and incubating for 90 min at 37°C in 5% CO2. Following the removal of non-adherent cells, MDMO-media was added, and cells were incubated at 37°C for 4 days to mature into macrophages and then used for infection with Mtb strains. 5×105 macrophages were infected in triplicate in 24-well plates at a multiplicity of infection of 1∶5 as described previously [38]. At indicated periods of infection, supernatants were removed, and the TNF-α and IL-6 levels were measured using ELISA assays (eBioscience, Inc., CA) according to the manufacturer's instructions. In some experiments, MDMO were stimulated with 150 nM PMA, and the secretion of TNF-α was measured. The extraction of total lipids from whole cells using chloroform∶methanol (2∶1 v/v) and the separation of polar lipids in a solvent system containing chloroform∶methanol∶water (65∶25∶4) in the first dimension and chloroform∶methanol∶acetic acid∶water (80∶12∶16∶4) in the second dimension were performed as described previously [7],[32],[39]. Polar lipids were visualized by exposing the plates to iodine vapors or staining them with ninhydrin in order to detect amino acid-containing lipids. In some experiments, autoradiography was used to detect radiolabeled lipids. For PDIMs analysis hexane∶diethylether∶acetic acid (80∶20∶1, vol/vol/vol) solvent system was used. MALDI-MS was performed using an UltraFlex TOF/TOF (Bruker Daltonics, Billurica, CA) as described previously [39]. The L-PG sample in acetonitrile was mixed 1∶1 with 2,5-dihydroxylbenzoic acid matrix for spotting onto the target plate. The L-PG was hydrolyzed with 3 N HCl in methanol for 4 h at 80°C. The sample was dried and treated with silylation reagent (TRI-SIL, Pierce Biotechnology, Rockford, IL) for 30 min at room temperature. The trimethylsilylated derivatives were analyzed by GC/MS. Specifically, the sample was applied to a DB-5 column at an initial temperature of 60°C for 1 min, then increased to 130°C at a rate of 30°C/min, and finally increased to 280°C at a rate of 5°C/min. 1H and 31PNMR were performed at a concentration of 2 mg L-PG sample per 0.6 mL of CDCl3 on a Varian Inova 400 MHz instrument. The purified L-PG was incubated overnight at 100°C in 6 N HCl in a heat-block. Samples were cooled, evaporated to dryness, resuspended in water and subjected to amino acid analysis. Neutral red chemical staining of Mtb wild type and lysX mutants was carried out following the protocol described by Soto et al. [40]. In order to evaluate the growth inhibitory effects of cationic compounds, Van (1 ug/mL), PMB (100 units/uL), human neutrophil peptide-1 (25 ug/mL) or lysozyme (0.5 mg/mL) was added to the growth media. The human neutrophil peptide stock was dissolved in 0.1% tri-fluoro acetic acid (TFA), and the cultures contained 0.025% TFA. No growth inhibition was noted at this concentration of TFA. The cultures were initially diluted to an OD600 of 0.05, dispensed into a 96-well microplate (100 uL per well) and incubated at 37°C with rotation at 60 rpm. At the indicated time periods, the change in the optical density (A600) was measured, and viability was determined. Low dose aerosol infection experiments of mice (C57BL/6 female mice) and guinea pigs (Hartely strains) for evaluating the growth and viability of lysX strain were essentially as described previously [41]. Cytoplasmic membrane potential changes were determined using the slow response, membrane potential-sensitive cyanine dye DiOC2(3) (Sigma). Briefly, actively growing cultures of Mtb strains (OD600 = 0.8) were incubated with 3 µM DiOC2(3) for 5 h. Spectrofluorometry was used to detect the red fluorescence (488 nm/620 nm) associated with aggregates of DiOC2(3), which exhibits green fluorescence (488 nm/520 nm) in the monomeric state. The assay was performed using white 96-well microtiter plates (Perkin Elmer; Waltham, MA) and a Cary Eclipse spectrofluorometer (Varian; Palo Alto, CA). A negative (depolarized) control of 100 µM m-chlorophenylhydrazone CCCP (Sigma), a proton ionophore that destroys the proton gradient and eliminates the bacterial membrane potential, was included. The membrane potential was measured as the ratio of red fluorescence (associated with membrane potential changes) to green fluorescence (a cell size-dependent, membrane potential-independent signal). Preliminary optimization studies revealed that 5 h incubation was optimal for the measurements. Differences between groups were analyzed by multiple comparison procedures (Student-Newman-Keuls Method) with a simple one-way ANOVA or Mann Whitney Rank Sum Test, using SIGMASTAT (SPSS Science, Inc., Chicago, IL). A P value of less than 0.05 was considered significant.

10.1371/journal.pcbi.1005600

Rosetta:MSF: a modular framework for multi-state computational protein design

Computational protein design (CPD) is a powerful technique to engineer existing proteins or to design novel ones that display desired properties. Rosetta is a software suite including algorithms for computational modeling and analysis of protein structures and offers many elaborate protocols created to solve highly specific tasks of protein engineering. Most of Rosetta’s protocols optimize sequences based on a single conformation (i. e. design state). However, challenging CPD objectives like multi-specificity design or the concurrent consideration of positive and negative design goals demand the simultaneous assessment of multiple states. This is why we have developed the multi-state framework MSF that facilitates the implementation of Rosetta’s single-state protocols in a multi-state environment and made available two frequently used protocols. Utilizing MSF, we demonstrated for one of these protocols that multi-state design yields a 15% higher performance than single-state design on a ligand-binding benchmark consisting of structural conformations. With this protocol, we designed de novo nine retro-aldolases on a conformational ensemble deduced from a (βα)8-barrel protein. All variants displayed measurable catalytic activity, testifying to a high success rate for this concept of multi-state enzyme design.

Protein engineering, i. e. the targeted modification or design of proteins has tremendous potential for medical and industrial applications. One generally applicable strategy for protein engineering is rational protein design: based on detailed knowledge of structure and function, computer programs like Rosetta propose the sequence of a protein possessing the desired properties. So far, most computer protocols have used rigid structures for design, which is a simplification because a protein’s structure is more accurately specified by a conformational ensemble. We have now implemented a framework for computational protein design that allows certain design protocols of Rosetta to make use of multiple design states like structural ensembles. An in silico assessment simulating ligand-binding design showed that this new approach generates more reliably native-like sequences than a single-state approach. As a proof-of-concept, we introduced de novo retro-aldolase activity into a scaffold protein and characterized nine variants experimentally, all of which were catalytically active.

Since the 1990s, computational protein design (CPD) has been a powerful tool of protein engineering. For example, CPD has been successfully utilized to increase thermostability of proteins [1–3] and to design new or altered binding specificities for metals [4], DNA [5] or other ligands [6, 7]. Additionally, CPD was applied to even more challenging tasks like the design of novel protein-protein interfaces [8, 9], de novo enzymes [10] or artificial folds not found in nature [11, 12]. Classical CPD methods, referred to as single-state design (SSD), optimize the amino acid sequence for the residue positions of a single backbone by means of an objective function [13]. A substantial contribution to the enormous success reached by SSD is due to refinements of the corresponding knowledge-based or statistical energy terms and the incorporation of backbone flexibility [14]. However, SSD is always a simplification because proteins populate conformational ensembles [15]. Moreover, certain design objectives such as negative design [16–18], multi-specificity design [19], the design of specific protein interfaces [20, 21] or the mimicking of backbone flexibility [22] require the concurrent assessment of several conformational or chemical states. This is why multi-state design (MSD) methodology is an emerging field in CPD [23] that extends the application spectrum and promises high success rates. Even the design of stable proteins profits from using backbone ensembles [24]. Typically, the optimization strategy of MSD consists of an “outer routine” that suggests possible amino acids sequences and an “inner routine” that assesses the fitness of a given sequence by performing rotamer optimization on each of the considered states and combines the individual scores [25]. This combined score enables a sequence selection driven by the energetic contribution of multiple conformational and/or chemical states. For example, in order to increase specificity of protein-protein interactions, one can utilize negative design and penalize those sequences that favor undesired interactions [16]. One of the first applications of MSD was the design of topologically specific coiled-coil structures consisting of 11-fold amino acid repeats whose stability was assessed by using terms of a standard molecular-mechanics potential energy function [26]. Later on, the binding pocket of a ribose-binding protein was successfully redesigned by means of MSD based on a standard force-field [27]. Meanwhile, many of the common optimization algorithms used in SSD have been adapted for MSD, including Monte Carlo (MC) with simulated annealing [28], genetic algorithms [29], the FASTER approach [25], dead-end-elimination [30], and cluster expansion [31]. Rosetta [32] is currently the most flexible and most widely used CPD software suite and offers several multi-state applications; noteworthy are MPI_MSD [33] and RECON [34]. MPI_MSD provides a generic multi-state design implementation based on a genetic algorithm that optimizes a single sequence on multiple states given a fitness function. RECON starts by individually optimizing one sequence for each state; subsequently the computation of a consensus sequence is promoted by incrementally increasing convergence restraints. However, the current implementations of both methods are limited to certain design tasks and cannot make use of fine-tuned protocols like those required for enzyme design [35] or anchored design of protein-protein interfaces [36]. In order to overcome this limitation, we have developed MSF and our integration of this modular framework into Rosetta facilitates the transfer of already proven single-state protocols to an MSD environment. Here, by using MSF, we first corroborate the superiority of MSD for enzyme design based on two in silico benchmarks for ligand binding. Applying the same protocol, we then designed nine experimentally active retro-aldolases. MSF is a programming framework that allows the user to develop and execute Rosetta protocols in an MSD environment. The modular software architecture of MSF significantly reduces the development efforts involved; see Fig 1. MSF requires as input a set of states s1, …, sn, e. g. structural conformations, and a population of sequences seq1, …, seqm, which will be subsequently altered by the sequence optimizer. The evaluator determines n state-specific scores for each seqi according to the chosen Rosetta protocol. These n × m scores are the input of a user-defined fitness function, which combines the scores to determine the fitness of each sequence and communicates these values to the sequence optimizer. The task management is as follows for all protocols: one process controls the sequence optimizer and a user-defined number of evaluator-processes execute the protocol in parallel, which guarantees high scalability. Technical details and availability are described in S1 Text; MSF will be part of an upcoming weekly release of Rosetta. As has been shown, a genetic algorithm (GA) successfully samples sequence space in MSD calculations [16, 27, 33]. Therefore, we have implemented the sequence optimizer based on the well-proven GA of Rosetta. Briefly, a GA imitates the process of natural selection by maintaining a population of design sequences that are evolving for a number of generations, while the selection pressure of the fitness function eliminates less optimal solutions. The final output of MSF is a population of optimized sequences. By contrast, a standard SSD implementation that utilizes MC optimization generates one sequence. Both MSF and MPI_MSD rely on the Rosetta GA. However, MPI_MSD does not support the integration of existing SSD protocols such as enzyme design that requires the additional optimization of catalytic constraints. Thus, our aim was to offer a framework that minimizes the development effort of supplying SSD protocols with MSD capability. The architecture of MSF strictly separates the tasks of optimization and the application-specific assessment of states. The resulting modularity allows an informed Rosetta user to implement MSD for existing protocols in a straightforward manner. Most importantly, the functionality of the protocols is unchanged and all options remain available. In addition to protocol porting, the user has to set up an application-specific fitness function, which defines the design goal. If it is the goal to alter conformational, binding, or catalytic specificity, the fitness function often has to consider positive and negative design. For the assessment of one positive state s+ and one negative state s-, the following function has been proposed [25]: fitness+,−(seqi)=Δscore+(seqi)−wΔscore−(seqi) (1) Here, Δscorel (seqi) is the difference of scores calculated for seqi and seq0; seq0 is the optimal sequence determined in an SSD for the states sl ∈ {s+,s−} and w is a weighting factor. Similar approaches, which were based on the computed transfer free energy from the target state to the ensemble of competing states [16] or on differences of Rosetta energies [33] guided the MSD of protein interfaces. Equally to MPI_MSD, our framework MSF supports the specification of a broad range of fitness functions. For the initial implementation of MSF, we have integrated enzdes and AnchoredDesign, two widely used Rosetta protocols. enzdes provides ligand binding and enzyme design functionality by repacking and redesigning residues around the binding/active site and by optimizing catalytic contacts. AnchoredDesign creates a protein-interface by transferring a key interaction identified in a natural binding partner of the target protein to a surface loop of the scaffold protein. Afterwards, the surface of the scaffold is redesigned with backbone flexibility to generate a novel binding partner of the target [36]. To validate AnchoredDesign in the MSF context, we redesigned the interface of the factor B serine protease domain from Homo sapiens (PDB ID 1dle). For this single example, the MSD approach performed better that the corresponding SSD protocol; see S2 Text for details. In order to demonstrate the potential of MSF for a large number of cases, we focused on enzdes by performing in silico and in vitro experiments. For the in silico assessment, the fitness of the sequences was computed according to Eq 2 based on the Rosetta total score (ts) averaged over all states. In the following, we designate software protocols as program:protocol. For example, Rosetta:enzdes (or for the sake of brevity enzdes) and Rosetta:MSF:GA:enzdes (MSF:GA:enzdes) are the names of the SSD and MSD implementations of enzdes. The most obvious usage of MSD is its application to an ensemble representing the native conformations of a protein. In solution, a protein’s structure is dynamic and nuclear magnetic resonance (NMR) offers an experimentally determined estimation of protein dynamics. Interestingly, in previous analyses SSD protocols performed better on crystal structures than on NMR templates [22, 37]. We speculated that this performance loss can be compensated, if MSD is applied to a whole ensemble and we decided to assess a ligand-binding design. Thus, for a first performance comparison of the SSD algorithm enzdes, and the MSD algorithm MSF:GA:enzdes, we chose an NMR ensemble of the human intestinal fatty acid binding protein (hIFABP) with bound ketorolac (PDB ID 2mji). This ensemble consisting of ten conformations was prepared for ligand-binding design (see Materials and Methods) and the design shell contained 21 residue positions in the vicinity of the ligand. Our protocol allowed Rosetta to find a low energy sequence by arbitrarily choosing residues for these positions. For each of the individual conformations conf(l), 1000 randomly seeded runsl (i) of enzdes (SSD) were started. Design quality was monitored by computing for each number of runs i the score tsSSDhIFABP(i). This is the mean total score deduced from corresponding conformations (Eq 6) given in Rosetta Energy Units (REU). MSF:GA:enzdes (MSD) was applied to the full ensemble and the GA was started. Analogously to the SSD experiment, the mean total score tsMSDhIFABP(j) was computed for each generation j (Eq 7). As a second measure of design quality, we determined the native sequence similarity recovery (nssr). Commonly, the performance of design algorithms is assessed by means of the native sequence recovery (nsr) [38–40], which is the fraction of identical residues at corresponding positions of the native and the designed sequence. The concept of nsr is blind for a more specific comparison of residues beyond identity, which may impede a detailed assessment. In contrast, for the computation of nssr, all residue pairs reaching a BLOSUM62 score > 0 are considered similar and contribute to the nssr value (Eqs 4 and 5). The plots shown in Fig 2 indicate that the SSD and the MSD algorithm converged after 1000 runs or 800 generations, respectively, both with respect to sequence recovery and ts values of the chosen sequences. The mean nsr values of enzdes and of MSF:GA:enzdes were 20.00% and 27.14%, and the mean nssr values were 41.90% and 46.66%. Only two of the ten enzdes designs reached an nssr value (47.62% and 61.90%, respectively) that was higher than the mean nssr of MSF:GA:enzdes. In summary, MSF:GA:enzdes performed better than enzdes suggesting the usage of MSD if sequences have to be designed for an ensemble. Altogether, the energies of models generated in SSD were on average 7.11 REU lower than those in MSD. However, a comparison of ts scores is no ideal means to compare SSD and MSD performance. In MSD, a sequence is a compromise that has to satisfy the constraints associated with all conformations in an acceptable manner. In contrast, SSD customizes a low energy sequence for each conformation. Thus, it is no surprise that the mean ts values of SSD sequences are superior to those of the MSD results. On the other hand, due to these specific adaptations based on single, less-native conformations, the SSD sequences are receding from the native ones, which are considered as close to optimal [41]. This undesired effect is less pronounced for MSD sequences computed on the whole native ensemble. We conclude that nsr and nssr are more suitable than ts values for a comparative benchmarking of SSD and MSD approaches. A standard dataset for the assessment of ligand-binding and enzyme design is the enzdes scientific sequence recovery benchmark. It consists of 51 representative proteins in which the ligand is bound with an affinity of 10 μM or lower [42]. During benchmarking, a given CPD algorithm redesigns residues of the design shell enclosing each ligand and the algorithm’s ability to recapitulate the native sequence (nsr and nssr values) is measured. However, for an assessment of de novo design algorithms, this approach may be misleading, because the required remodeling of a chosen protein is more demanding than the recapitulation of its native binding pocket. We created a more realistic benchmark that is devoid of a perfect backbone/rotamer preorganization and is more suitable for the assessment of de novo design algorithms. For feasibility reasons, we randomly selected 16 proteins prot(k) of the above 51 benchmark proteins. The corresponding ligands were removed and for each of the 16 apoproteins, an ensemble of 20 conformations was created using the Backrub server [43], which generates near-native conformational ensembles [44, 45]. Next, by superposition of each conformation with the corresponding crystal structure, the ligands were transferred to the binding pockets. Thus, the resulting dataset BR_EnzBench featured for each of the 16 prot(k) 20 backbone conformations that are near to native but lack the implicit pre-organization induced by a bound ligand in a crystal structure. We used BR_EnzBench to compare the performance of SSD and MSD for de novo ligand-binding design. All design shell residues were initially mutated to alanine and the conformations were energy-minimized to further increase the difficulty for CPD algorithms to recover the native sequence. To prevent a hydrophobic collapse of the alanine-only design shells, minimization was performed with backbone constraints. Thus, the CPD problem to be solved within the scope of this benchmark was to design a binding pocket by sequence optimization of the all-alanine design shells. For SSD with enzdes, all conformations of each protein were considered independently and for each conformation, 1000 randomly seeded designs were performed. Design quality was assessed by means of the three parameters nsr, nssr, and ts. The respective values were averaged for each of the 16 prot(k) (Eqs 10 and 11) and are listed in Table 1. Additionally, the convergence of the design process was followed by monitoring the mean performance for each number i of design runs (Eqs 8 and 9); these values are plotted in Fig 3. To conduct multi-state design by means of MSF:GA:enzdes, for each prot(k), the 20 conformations were divided into four ensembles ensmk each containing five conformations. Note that the conformations that are combined in each of the ensembles ensmk are unrelated, due to the stochastic approach of the Backrub algorithm. The GA was started on a population consisting of 210 sequences and stopped after 600 generations, because convergence was reached. Analogously, nsr, nssr, and ts values (Eqs 14 and 15) were determined for each MSD run and averaged for each of the 16 proteins. These results were added to Table 1. As above, the convergence of the GA was followed be monitoring the mean performance for each generation j (Eqs 12 and 13); these values are also plotted in Fig 3. The protein-wise comparison (Table 1) indicates that in 10 out of the 16 cases, the nsr and in 13 out of all 16 cases, the nssr values of MSF:GA:enzdes designs are superior to the corresponding values of enzdes designs. MSF:GA:enzdes recovers on average a higher percentage of native residues (Δ nsr = 2.65%) and a higher percentage of similar residues (Δ nssr = 6.79%). Thus, with respect to the more adequate similarity measure nssr, MSD performs 15% better than SSD for this benchmark (p = 0.004, Wilcoxon signed rank test). In addition, multi-state designs have slightly better energies (Δ ts = 2.51 REU), which is in contrast to the hIFABP results and is most likely due to the smaller ensemble size. Fig 3 reflects the differences in convergence speed of both algorithms and indicates that the better performance has its price: the MC optimization utilized by enzdes leads to acceptable design solutions even after a low number of runs. In contrast, the GA of MSF:GA:enzdes is slower and more than hundred generations are required to surpass the performance of the SSD algorithm. For this set of parameters, MSF:GA:enzdes required approximately five times the number of core hours needed by enzdes; further details of computational costs are given in S2 Text. The sequence recovery reached for the hIFABP ensemble and for BR_EnzBench strongly suggests that MSF:GA:enzdes is superior to enzdes in more complex design applications. However, it was unclear to us, whether the different concepts (single-state versus multi-state) or the different optimizers (MC versus GA) contributed most to performance. Choosing an MSD approach increases computational cost, which has to be substantiated by making plausible that the choice of the optimizer is less important. The performance of MSF:GA:enzdes on BR_EnzBench was assessed ensemble-wise by determining the values nssrMSD(ensmk), which were averaged (Eq 12). As these ensembles contain not more than five unrelated conformations each, the nssrMSD(ensmk) values (Eq 16) vary due to the small sample size and one can sort for each prot(k) the four ensmk on their nssrMSD(ensmk) value. The result is a ranking ensrank=uk (1 ≤ u ≤ 4) of the four ensembles and we created the set ES1 that contained the 16 ensembles (one for each prot(k)) with the lowest nssrMSD(ensmk) value. Analogously, we compiled the sets ES2—ES4; consequently, ES4 consisted of those 16 ensembles that had the highest nssrMSD(ensmk) value; for details see Materials and Methods. For these four sets ESi, we determined boxplots of the corresponding nssrSSD and nssrMSD values; see Fig 4. The boxplots characterizing the SSD results are nearly identical; this finding indicates that the conformations allocated to the four sets ES1—ES4 give rise to a similar SSD performance. Moreover, the boxplots representing the nssrSSD (ES1) and nssrMSD (ES1) values are nearly identical (median values 47.60% and 47.76%), which indicates that the optimizer GA is not generally superior to MC. Additionally the continuous increase observed for the nssrMSD (ES1) - nssrMSD (ES4) - but not for the nssrSSD (ES1) – nssrSSD (ES4) values - supports the notion that it is the combination of conformations that strongly affects MSD performance. We thus conclude that the MSD approach - and not the optimizer - contributes most to the performance of MSF:GA:enzdes. Because Rosetta has a certain bias in recapitulating native residues [46], we assessed and compared the bias introduced by enzdes and MSF:GA:enzdes. For the assessment of the enzdes outcome, we selected the 13440 sequences representing the best designs on BR_EnzBench and determined nssrSSD (aaj) values. This distribution represents for all amino acids aaj the fraction of similar residues recovered at all design shell positions. Analogously, the distribution nssrMSD (aaj) was computed that indicates the fraction of similar residues recovered by MSF:GA:enzdes; for details see Materials and Methods. The two distributions, which are plotted in Fig 5, indicate similar recovery rates that are below the optimal value of 100% for all residues. Generally, sequence recovery for large polar or charged residues (D, E, H, K, N, R, S) is low, which contributes to Rosetta’s weakness in accurately designing hydrogen bonds and electrostatics [47]. Interestingly, enzdes is slightly better in recovering polar and charged residues, whereas MSF:GA:enzdes clearly recovers a higher fraction of hydrophobic residues (A, F, I, L, P, V, W, Y). This general trend is most evident in the two benchmark proteins with the most extreme differences in their individual nssrSSD and nssrMSD values: ARL3-GDP (PDB ID 1fzq) is a distinct GTP binding protein [48] from Mus musculus and both the ligand and the native binding pocket are considerably polar. Fig 6A shows that enzdes correctly recovers the residues interacting with the guanine group (colored in teal) of GDP, while MSF:GA:enzdes is less successful. On the other hand, in the glucose binding protein (PDB ID 2b3b) from Thermus thermophilus, four tryptophan residues provide tight binding to glucose by shape complementarity. Fig 6B shows that MSF:GA:enzdes recovers three critical tryptophan residues (colored in teal) in most designs, whereas enzdes prefers small polar residues that do not provide tight packing. We conclude that the representation of a protein by means of an ensemble improves hydrophobic packing but not the formation of polar interaction networks. Their design is considerably more difficult than hydrophobic packing due to the partially covalent nature of a hydrogen bond and the geometric requirements for orientations and distances [47, 49]. Molecular dynamics (MD) simulation is a well-established and reliable method for modeling conformational changes linked to the function of proteins [50]. Thus, MD provides an alternative to the Backrub approach for the generation of ensembles to be utilized in MSD. We were interested in assessing the designability of conformations resulting from unconstrained MD simulations of length 10 ns. In analogy to BR_EnzBench, we compiled the dataset MD_EnzBench consisting of 1000 conformations generated for each of the 16 benchmark apoproteins by means of YASARA [51]. Again, all design shell residues were replaced with alanine prior to design; see Materials and Methods. To assess the structural variability of MD_EnzBench conformations, Cα-RMSD values of design shell residues were determined in a protein-specific all-against-all comparison and then averaged. Analogously, the structural variability of BR_EnzBench conformations was determined. Interestingly, the variety of the binding pockets generated by the MD simulations is much larger than that generated by Backrub: the mean RMSD of MD_EnzBench is 0.62 Å and that of BR_EnzBench is 0.17 Å, which indicates that a 10 ns MD simulation generates an ensemble with higher structural diversity than the Backrub server. As a control of design performance, the 16 × 20 nssrSSDBR_EB(i=1) values of single enzdes designs generated for 20 protein-specific conformations from BR_EnzBench were summarized in a boxplot, which had a mean value of 43.88%. To assess the designability of the MD_EnzBench conformations, for each of the 1000 protein-specific conformations, one sequence was designed by means of enzdes and the resulting nssr values were averaged protein-wise. Fig 7 shows 100 boxplots each representing 16 × 10 nssr values resulting from ten conformations generated by the MD simulation in a 100 ps interval for each of the 16 prot(k). The mean of these nssr values is 42.53%, which testifies to a satisfying design performance, given that only one sequence was designed for each MD conformation. Moreover, the boxplots indicate that performance did not decrease for conformations generated at later phases of the MD simulation: the median nssr, and the first and third quartile of the most left and the most right boxplots are 42.10% [35.40%, 45.89%] and 42.24% [34.78%, 50.00%], respectively. In summary, these findings suggest that ensembles generated by MD feature higher conformational flexibility and appropriate de novo designability. The most convincing proof of concept for any CPD algorithm is the design of functionally active proteins. A non-natural reaction that is frequently chosen for enzyme design is the amine-catalyzed retro-aldole cleavage of 4-hydroxy-4-(6-methoxy-2-naphtyl)-2-butanone (methodol) into 6-methoxy-2-naphthaldehyde and acetone [52]. This multi-state reaction comprises the attack of an active site lysine side chain on the carbonyl group of the substrate to form a carbinolamine intermediate that is subsequently dehydrated to a protonated Schiff base. The latter is then converted to the reaction products by acid/base chemistry [53, 54]. The most active de novo retro-aldolase designs have been established on a jelly roll and several (βα)8-barrel proteins [55–57]. For comparison purposes, we selected the indole-3-glycerolphosphate synthase from Sulfolobus solfataricus (ssIGPS), a previously used thermostable (βα)8-barrel scaffold. The native ligand was removed and the apoprotein was subjected to conformational sampling. Using the protocol validated with MD_EnzBench, three individual MD simulations were performed for 10 ns. A clustering of MD snapshots based on RMSD values helps to choose near-native conformations [58]. Thus, we used Durandal [59] to cluster the snapshots (conformations) generated with each MD run and picked four conformations from the largest cluster. These 3 × 4 conformations and the crystal structure of the apoprotein constituted the structural ensemble for the subsequent enzyme design. Enzyme design generally starts with the assembly of a theozyme, which is a model for the proposed active site that is based upon the geometric constraints dictated by the expected transition state(s). To design retro-aldolase catalysis, we used a previously designed theozyme containing the carbinolamine reaction intermediate as transition state surrogate covalently bound to the catalytic lysine [56]. In addition, this theozyme contained an aspartate or a glutamate residue to function as general acid/base as well as a serine or a threonine residue to provide additional hydrogen-bonding interactions. Rosetta:match was applied to all conformations and created several thousand matched transition states (mTS) with catalytic triads Ki-[D,E]j-[S,T]k located at markedly different residue positions. A critical step of MSD is the compilation of the ensembles that are concurrently used as states. For enzyme design, ensembles ensmTS of mTS are needed and we compiled them the following way: first, mTS judged as binding the transition state only weakly were discarded. Second, mTS derived from different conformations were added to the same ensmTS, if identical catalytic triads were located at matching residue positions. Thus, each ensmTS contained a certain number of conformations accommodating the same catalytic triad. Third, the consistency of each ensmTS was assessed by superposing the transition states and by comparing the corresponding conformations. We chose 23 ensmTS consisting of 4 to 13 conformations (states) and their design and repack shells were defined by merging the output created by enzdes:autodetect for all conformations. MSF:GA:enzdes was executed with each ensemble until energetic convergence; see S3 Text for details of the protocol. In brief, to assess the designs we compared active-site geometry as well as total and interaction energies and the best 100 variants were subjected to MD simulations of 10 ns length. For each variant, we analyzed in detail catalytic site geometries of 100 snapshots (see Materials and Methods) and nine variants named RA_MSD1 to RA_MSD9 were chosen for biochemical characterization; see S3 Text. Because the catalytic efficiency and the conformational stability of initial designs are generally poor [60], further optimization is commonly performed by using either Foldit or other software tools to revert unnecessary mutations back to the native sequence of the scaffold [56], or by means of directed evolution [57]. However, we did not introduce subsequent stabilizing mutations into the sequences of RA_MSD1 to RA_MSD9 prior to a first experimental characterization. In doing so, we wanted to demonstrate the potential and also the limitations of multi-state designs. For a comparison of these novel designs with previous ones, we compiled a list of 42 retro-aldolases RA* from the literature (see S3 Text) that were also created in the ssIGPS scaffold by means of Rosetta. These RA* sequences differ on average at 15 positions from the native ssIGPS sequence; in contrast, our nine RA_MSD* sequences contain on average 21 amino acid substitutions. Moreover, RA* sequences deviate on average from RA_MSD* sequences at 24 positions, and 18 substitutions distinguish the most similar pairs of variants (RA41 versus RA_MSD9 and RA90 versus RA_MSD8). Even a previous (RA114) and a new design (RA_MSD1), which share the same catalytic residues K210 and S110, differ at 25 positions. Thus, although we utilized the same TS and the same scaffold that was used for the design of RA114—RA120 [56], our MSD approach has generated a set of entirely novel catalytic sites located in the same shell as used for previous designs; see Fig 8. The genes for RA_MSD1—RA_MSD9 were synthesized and expressed in Escherichia coli as fusion constructs with the gene for the maltose binding protein (MBP). The fusion proteins were purified with metal chelate affinity chromatography via their N-terminal hexa-histidine tags, resulting in high yields (50–150 mg protein/l expression culture). RA_MSD5 could be produced in soluble form also without MBP, whereas the other designs precipitated in the absence of the solubility enhancer. All designs showed modest catalytic activity with low substrate affinity, leading to conversion rates in the presence of 500 μM S-methodol ranging from 3 × 10−7 to 1.7 × 10−5 s-1 (Table 2). For the best designs, namely RA_MSD5 and RA_MSD7, the linear part of the substrate saturation curve was used to determine kcat/KM values of 3.47 × 10−2 and 1.41 × 10−2 M-1s-1 (S1 Fig; Table 2), which are similar to the values obtained for RA114 - RA120 [46]. Moreover, the RA_MSD5 designs with and without MBP displayed virtually the same kcat/KM values, excluding an influence of the solubility enhancer on activity. Due to the intentionally omitted step of secondary protein stabilization following the initial design process, eight of our nine designs were insoluble without MBP. We wanted to test whether protein stabilization would result in higher activity. Accordingly, we attempted to improve the stability of RA_MSD2, which has the lowest activity of all designs (Table 2), by using the fully automated in silico method offered by the PROSS webserver [61]. The six conformations of RA_MSD2 were individually submitted to PROSS and the corresponding output sets that contained 6 to 21 stabilizing mutations were merged to five consensus sequences; see S3 Text, Table B1. Variants RA_MSD2.4 and RA_MSD2.5 that contained the highest number of stabilizing mutations, could be produced in soluble form without MBP and were purified with high yield (about 25 mg protein/l expression culture). Activity measurements showed, however, that the additional stabilizing exchanges did not drastically improve the conversion rate of RA_MSD2; see Table 2. In summary, our results show that MSD (based on a structural ensemble) is comparably successful as SSD (based on a single structure) for establishing retro-aldolase activity on a thermostable (βα)8-barrel scaffold, indicating that this particular reaction requires only a limited degree of conformational flexibility. However, catalysis is often linked to conformational transitions which can only be captured by MSD approaches. Moreover, in contrast to SSD, MSD offers a broader functionality and is for example also suited for more challenging tasks like negative design. Two subsets of the scientific sequence recovery benchmark of Rosetta [42] were generated that contain 20 specifically prepared conformations of 16 proteins prot(k) with bound ligand. In order to exclude an erroneous conformational sampling, missing residues were reconstructed by using YASARA:loop_modeling [62] and the respective native sequences. Additionally, all ligands were removed prior to the conformational sampling of the resulting apoproteins. The dataset BR_EnzBench was created by using the BackrubEnsemble method of the Backrub server [43] to compute a conformational ensemble of 20 structures for each apoprotein. The second benchmark dataset MD_EnzBench was deduced from MD simulations of length 10 ns generated with YASARA (version 14.7.17) and the YAMBER3 force field, which has been parameterized to produce crystal structure-like protein coordinates [51]. For each of the 16 apoproteins, 1000 conformations were sampled at an interval of 10 ps. After sampling, the native ligands were re-introduced in all conformations of both subsets by means of PyMOL:superpose [63] and the respective apoproteins. For the corresponding holoproteins of BR_EnzBench and MD_EnzBench, the same design and repack shells were utilized. These were determined protein-wise for each of the BR_EnzBench conformations by means of Rosetta:enzdes:autodetect and merged. In all conformations, design shell residues were replaced with alanine and prior to design, all conformations were energy-minimized by means of Rosetta:fastrelax with backbone constraints. Parameters of MD simulations, Rosetta:fastrelax, and the composition of design and repack shells are listed in S2 Text. The first generation of the 210 sequences consisted of the given seed sequence and 209 mutants each with a randomly introduced single point mutation. During each generation cycle, half of the population was replaced with sequences seqi generated by means of single point mutations and recombination. The replaced sequences were those with worst fitness values fitness(seqi), which were computed for MSF:GA:enzdes according to: fitness(seqi)=1n∑l=1ntsl(seqi) (2) Here, n is the number of states (e. g. conformations s1, …, sn of a given prot(k)) and tsl is the Rosetta total score for a sequence given a state l. In all equations, ts values are given in REU. For a given pair of residues aa1, aa2 the nssr value was deduced from the scores of the BLOSUM62-matrix [64] as follows: nssr(aa1,aa2)={1ifBLOSUM62(aa1,aa2)>00else (3) For a given pair of sequences seq1, seq2 of length n, the nssr value was determined as the mean value deduced for residue pairs seq1[i], seq2[i]: nssr(seq1,seq2)=1n∑i=1nnssr(seq1[i],seq2[i]) (4) For a given set of design solutions ds = {seq1,…,seqm} and a native sequence seqnat, the value nssr(ds,seqnat) was computed according to: nssr(ds,seqnat)=1m∑i=1mnssr(seqi,seqnat) (5) The data set with PDB ID 2mji contains ten conformers of hIFABP and the bound ligand ketorolac; this ensemble has been deduced by means of solution NMR [65]. The set was downloaded from PDB and the ligand was parameterized using Rosetta:molfile-to-params [66]. Next, each of the ten conformations was energy-minimized via Rosetta:fastrelax with constraints. To obtain consistent design and repack shells, the shells determined by Rosetta:enzdes:autodetect for each conformation were merged. For SSD, enzdes was applied to each of the ten initial conformations conf(l) (1 ≤ l ≤ 10). Using the default MC optimization and the parameter set ps_enzdes, sequences seql (i) were generated by means of 1000 randomly seeded runsl (i) (1 ≤ i ≤ 1000). In order to control the convergence of the design process and for performance comparison, the seql*(i) with the best total score (ts) were chosen from seql (1,…,i) for each l and each i. Finally, the mean of the ten ts values was determined as a measure of design quality tsSSDhIFABP(i) reached in i SSD runs: tsSSDhIFABP(i)=110∑l=110ts(seql*(i)) (6) For MSD, all ten conformations conf(l) were considered as states and MSF:GA:enzdes was executed for 800 generations (i. e. design cycles) on a population consisting of 210 sequences with parameters ps_msf_enzdes. The initial population was seeded with the native sequence. The sequences representing a generation j were ranked with respect to ts values and the ten top scoring sequences seqlt(j) (1 ≤ t ≤ 10) were stored in order to allow for the subsequent performance comparison. Finally, the mean of the 10 × 10 ts values was determined as a measure of design quality tsMSDhIFABP(j) reached in j MSD generations: tsMSDhIFABP(j)=1100∑l=110∑t=110ts(seqlt(j)) (7) Further details of the analysis can be found in S2 Text; it lists parameters of Rosetta:fastrelax and the design protocol, and the composition of the design and repack shell. For SSD, enzdes was applied to each of the 20 initial conformations conf(l) (1 ≤ l ≤ 20) of each prot(k) (1 ≤ k ≤ 16) from BR_EnzBench. Using default MC optimization and the parameter set ps_enzdes (see S2 Text), sequences seqk,l (i) were generated by means of 1000 randomly seeded runsk,l (i) (1 ≤ i ≤ 1000). In order to control the convergence of the design process and for performance comparison, those seqk,l*(i) having the best ts value were chosen from seqk,l (1,…,i) for each k, l, and i. Finally, mean performance reached in i SSD runs was measured by means of the score ∈ {nsr,nssr}, where nsr is the native sequence recovery and nssr is the native sequence similarity recovery: scoreSSDBR_EB(i)=1320∑k=116∑l=120score(seqk,l*(i),seqnatk) (8) tsSSDBR_EB(i)=1320∑k=116∑l=120ts(seqk,l*(i)) (9) Here, seqnatk is the native sequence of prot(k), and ts is the total score. To score SSD performance reached for one prot(k), the final score values were averaged over all conformations: scoreSSDBR_EB(k)=120∑l=120score(seqk,l*(1000),seqnatk) (10) tsSSDBR_EB(k)=120∑l=120ts(seqk,l*(1000)) (11) To assess the performance of MSD, each of the 20 conformations of a prot(k) was assigned to an ensemble ensmk (1 ≤ m ≤ 4) consisting of five conformations each. These five conformations were considered as states and MSF:GA:enzdes was executed for 600 generations on a population consisting of 210 sequences with parameter set ps_msf_enzdes (see S2 Text). The initial population was seeded with an all-alanine sequence. The sequences representing a generation j were ranked with respect to ts values and the five top scoring sequences seqk,mt(j) [1 ≤ t ≤ 5] were stored in order to allow for the subsequent performance comparison. Finally, mean performance values reached in j MSD generations were determined according to: scoreMSDBR_EB(j)=1320∑k=116∑m=14∑t=15score(seqk,mt(j),seqnatk) (12) tsMSDBR_EB(j)=1320∑k=116∑m=14∑t=15fitness(seqk,mt(j)) (13) Here, seqnatk is the native sequence of prot(k), score ∈ {nsr,nssr} is a sequence recovery, and fitness(seqk,mt(j)) is the mean ts score (Eq 2, n = 5) of a given sequence over the five conformations belonging to ensemble ensmk. To score MSD performance reached for one prot(k) after 600 generations, the final score values were averaged over all ensembles: scoreMSDBR_EB(k)=120∑m=14∑t=15score(seqk,mt(600)) (14) tsMSDBR_EB(k)=120∑m=14∑t=15fitness(seqk,mt(600)) (15) The 20 conformations of a given protein prot(k) from BR_EnzBench belong to one of four ensembles ens1k - ens4k. The performance values nssrMSD(ensmk) were determined for each prot(k) and each ensmk according to: nssrMSD(ensmk)=15∑t=15nssr(seqk,mt(600),seqnatk) (16) Here, seqnatk is the native sequence of prot(k). The values nssrMSD(ensmk) were used for a ranking ensrank=uk (1 ≤ u ≤ 4) of the four ensembles such that ensrank=1k is the one with the lowest nssrMSD(ensmk) value and ensrank=4k that with the largest one. Having ranked the ensembles of all prot(k), sets of ensembles were created such that the set ES1=∪k=116ensrank=1k contained those ensembles that performed worst and ES4=∪k=116ensrank=4k those that performed best and the intermediates with rank = 2 and rank = 3 performed accordingly. For these four sets ESi, boxplots of the corresponding nssrSSD and nssrMSD values were determined. In order to assess the amino acid composition of the enzdes outcome, the 42 seqk,l (1,…,1000) with optimal ts values were identified for each of the 20 conformations l of all prot(k) ∈ BR_EnzBench. For these 16 × 840 sequences seqSSDk, the values nssr(seqSSDk[i],seqnatk[i]) were determined (Eq 3) by comparing design shell and native (nat) residues i. The distribution nssrSSD (aaj) represents for all amino acids aaj their recovered similarity at all design shell positions. To assess the amino acid composition for the MSF:GA:enzdes outcome, the 16 × 4 × 210 sequences seqMSDk of the final populations (i. e. all seqk,m (600)) generated for the four ensemble groups of each prot(k) ∈ BR_EnzBench were used to determine the values nssr(seqMSDk[i],seqnatk[i]). The distribution nssrMSD (aaj) represents for all amino acids aaj their recovered similarity at all design shell positions. The scaffold protein indole-3-glycerol phosphate synthase from S. solfataricus (ssIGPS, PDB ID 1a53), was downloaded from PDB and the ligand IGP was removed. To generate a structural ensemble, three MD simulations were performed with the apoprotein for 10 ns by means of YASARA and the YAMBER3 force field. Using Durandal:smart-mode:semi-auto[0.03. 0.20], the snapshots of each trajectory were clustered individually and four conformations were chosen from the largest cluster. These 12 conformations and the crystal structure of 1a53 were used for matching the transition state (TS) and grafting the theozyme of the retroaldol reaction [56] by means of Rosetta:match. Each of the resulting matched transition states (mTS) consisted of a catalytic triad Ki-[D,E]j-[S,T]k at three residue positions i, j, k that occured in one of the 13 conformations. Ensembles ensmTS of mTS used as input for MSF:GA:enzdes were generated as follows: first, mTS were discarded that were classified as weak TS binders or TS destabilizers. For example, matches with catalytic residues near the protein surface were eliminated. Second, mTS were grouped according to the composition and localization of the catalytic triad and those ensembles were selected that were compatible with most of the 13 conformations. Third, ensmTS were assessed with respect to the structural similarity of the superposed theozymes. In total, 23 ensembles ensmTS containing 4 up to 13 conformations were chosen. For each ensmTS, the design and repack shells were defined by merging the outcome of Rosetta:enzdes:autodetect for all corresponding conformations and MSF:GA:enzdes was executed on a population of 210 sequences that were seeded with the native sequence of ssIGPS. At convergence, the design process was stopped, which was the case after 97 to 710 generations. S3 Text lists more details of the design procedure like parameters of MD simulations and of Rosetta:match, and the specification of the TS. After MSD of retro-aldolases, the designs were filtered by ts values and active-site geometry. The best 100 designs were selected for 10 ns MD simulations in water and for one conformation of each design ensemble, 100 snapshots were generated. Two simulations were performed; the first one was based on the enzyme/TS complex. As a control, the second MD simulation was based on the enzyme/substrate complex and the substrate methodol was created by deleting the lysine-substrate bond of the TS. For each trajectory, catalytic distances, angles and torsion angles were plotted as boxplots and used to assess the designs; see S3 Text. Variant RA_MSD2 was chosen for solubilization experiments and all six conformations conf(l) of the corresponding ensemble were submitted to the PROSS server [61], which was used with default settings allowing for mutations at all positions. For each input conf(l), PROSS provided seven mutated sequences mut_seql(i) (1 ≤ i ≤ 7) containing an increasing number of putatively stabilizing mutations. For each i (degree of stabilization), an MSA that contained all sequences mut_seql(i) computed for all conf(l) was generated and weblogo [67] was used to determine a sequence logo. Finally, consensus residues deduced from the sequence logos were accepted as mutations at sites that did not interfere with the catalytic center. All sequence logos are shown in S3 Text. The genes encoding the designed retro-aldolases were optimized for E. coli codon usage and ordered as synthetic gene strings from Life Technologies. Cloning was performed via BsaI restriction sites into pET28a (Stratagene) and pMalC5T (New England Biolabs) plasmids specifically modified for this method of cloning. Both vectors fuse an N-terminal his6-tag to the target proteins, pMalC5T additionally adds MBP. The cloning method is derived from golden gate cloning [68]. Details of plasmid construction and cloning procedure will be published elsewhere. E. coli BL21 Gold cells were transformed with the resulting plasmids. The cells were grown in Luria broth with 50 μg/ml kanamycin or 150 μg/ml ampicillin for pET28 constructs and pMAL constructs, respectively. At a cell density of OD600 = 0.5 protein production was induced by addition of 0.5 mM isopropyl-β-thiogalactopyranoside. After growth over night at 20°C the cells were harvested by centrifugation (Avanti J-26 XP, JLA 8.1000, 15 min, 4,000 rpm, 4°C). Cell pellets were resuspended in 50 mM Tris/HCl buffer (pH 7.5) with 300 mM NaCl. Cells were lysed by sonication (Branson Sonifier W-250D, amplitude 65%, 3 min, 2 s pulse/2 s pause). Cell debris was removed by centrifugation (Avanti J-26 XP, JA 25.50, 30 min, 14,000 rpm, 4°C) and soluble proteins were purified by nickel chelate affinity chromatography (GE Healthcare, HisTrap FF crude). The proteins were eluted with 50 mM Tris/HCl (pH 7.5) containing 300 mM NaCl using a gradient of 10–500 mM imidazole. Fractions containing sufficiently pure protein were pooled and excess imidazole was removed by dialysis against 50 mM Tris/HCl (pH 7.5) buffer containing 100 mM NaCl. Protein concentrations were determined by absorbance spectroscopy (NanoDrop One, Thermo Fisher) using extinction coefficients determined by the Expasy:ProtParam webtool. Retro-aldolase activity of the designs (30–50 μM) was measured at 25°C in 50 mM Tris/HCl (pH 7.5), 100 mM NaCl and 5% (v/v) dimethyl sulfoxide (for substrate solubility) by following the formation of the fluorescent product 6-methoxy-2-naphthaldehyde from non-fluorescent S-methodol (70% ee). The substrate was synthesized as described in S3 Text. Fluorescence was measured in a Mithras LB 940 plate reader (λex = 355 nm, λem = 460 nm) using black 96 well micro plates. The concentrations of product were determined with the help of a calibration curve. For the determination of conversion rates, each measurement was repeated four times, for kcat/KM determinations all points were measured as triplicates. The wild-type scaffold protein ssIGPS and the solubility tag MBP served as negative controls and did not show any detectable activity. Further control measurements showed that conversion rates in the presence of 5% (v/v) dimethyl sulfoxide were identical to those in 3% acetonitrile, which has been used for the characterization of other retro-aldolase designs [46].

10.1371/journal.pcbi.1006548

Dilution and titration of cell-cycle regulators may control cell size in budding yeast

The size of a cell sets the scale for all biochemical processes within it, thereby affecting cellular fitness and survival. Hence, cell size needs to be kept within certain limits and relatively constant over multiple generations. However, how cells measure their size and use this information to regulate growth and division remains controversial. Here, we present two mechanistic mathematical models of the budding yeast (S. cerevisiae) cell cycle to investigate competing hypotheses on size control: inhibitor dilution and titration of nuclear sites. Our results suggest that an inhibitor-dilution mechanism, in which cell growth dilutes the transcriptional inhibitor Whi5 against the constant activator Cln3, can facilitate size homeostasis. This is achieved by utilising a positive feedback loop to establish a fixed size threshold for the Start transition, which efficiently couples cell growth to cell cycle progression. Yet, we show that inhibitor dilution cannot reproduce the size of mutants that alter the cell’s overall ploidy and WHI5 gene copy number. By contrast, size control through titration of Cln3 against a constant number of genomic binding sites for the transcription factor SBF recapitulates both size homeostasis and the size of these mutant strains. Moreover, this model produces an imperfect ‘sizer’ behaviour in G1 and a ‘timer’ in S/G2/M, which combine to yield an ‘adder’ over the whole cell cycle; an observation recently made in experiments. Hence, our model connects these phenomenological data with the molecular details of the cell cycle, providing a systems-level perspective of budding yeast size control.

Proliferating cells need to coordinate the initiation of genome replication and cell division with cell growth. In particular, the average time between two division events must precisely allow for a doubling in cell volume. Any systematic deviation from this balance would lead to progressive changes in cell size over consecutive generations and to a breakdown of biochemical processes. Here, we study two molecular mechanisms by which budding yeast cells might achieve this coordination. Through mathematical modelling, we show that the dilution of an inhibitor of cell cycle progression by cell growth can facilitate size homeostasis. But this mechanism fails to reproduce the size of mutant cells in which parts of the control machinery have been altered. By contrast, the titration of an activator against a constant number of genomic sites recapitulates these data and achieves size homeostasis. Since the control network of cell cycle progression in budding yeast is structurally similar to mammalian cells, our model could indicate a common mechanism for size control.

Balanced growth of proliferating cells requires some coordination between the increasing size of a growing cell and its probability of undergoing DNA synthesis and division. In particular, the average time between two successive cell divisions must allow for a doubling in cell mass (or volume, which we will use interchangeably in the following). Any systematic deviation from this balance would lead to progressive changes in size over consecutive generations, eventually leading to the breakdown of biochemical processes. However, despite mounting evidence for active size control in various cell types and across different organisms [1], if and how cells measure their size and relay this information to the cell cycle remains controversial [2]. An elegant way to coordinate cell division and growth is to restrict passage through a certain cell cycle stage to cells that are larger than a particular target size [1]. Such ‘size checkpoints’ have been proposed to underlie size control at the Start transition in budding yeast [3–5], and at the G2/M transition in fission yeast [6–8] and slime mould plasmodia [9–11]. The critical size required to pass these transitions depends, among other things, on the ploidy of the cell and its nutritional status [2]. To establish a size checkpoint, cells need to generate a size-dependent biochemical signal. Yet, most cellular macromolecules increase in abundance proportionally to cell volume, so that their concentration remains constant and the reactions they are involved in are independent of size [12]. Several proteins that defy this general rule have been indicated in size control. The mitotic activator Cdc25, for instance, increases in concentration with size in fission yeast [8], while Whi5, an inhibitor of Start in budding yeast, is diluted by cell growth [13]. This suggest a general mechanism, in which size control emerges from the interplay between size-dependent and size-independent cell cycle regulators. Here, we study this intriguing possibility, focusing on the budding yeast cell cycle. The budding yeast Saccharomyces cerevisiae divides asymmetrically, with size control mainly operating in the new-born daughter cell when it commits to enter the cell cycle anew at the Start transition [3–5]. Passage through Start is driven by activation of the transcription factor SBF [14]. In early G1-phase, before Start, SBF is kept inactive by its stoichiometric inhibitor Whi5 [15,16]. To enter the cell cycle, the cyclin-dependent kinase Cdk1 (encoded by the CDC28 gene) in conjunction with its regulatory binding partner Cln3 phosphorylates Whi5, which partially liberates SBF from inhibition and induces the synthesis of other G1 cyclins (Cln1 and Cln2). Cln1/2:Cdk1 complexes then accelerate the phosphorylation of Whi5 and activation of SBF, thereby promoting the Start transition [15–17]. Recent experiments show that during G1 the concentration of Cln3, the activator of Start, is constant, while the concentration of Whi5 decreases, suggesting that an inhibitor-dilution mechanism facilitates size control [13]. However, previous theoretical considerations and experimental data suggested a different mechanism based on the titration of an activator that increases in molecule number during growth–as would be the case if its concentration is kept constant–against a fixed number of nuclear sites [18–20]. To test these hypotheses, we developed a mechanistic mathematical model of the budding yeast cell cycle. At its core, the model comprises a simple description of gene expression in which both size-dependent and size-independent synthesis of proteins emerge seamlessly from the assumption of differential affinity of genes for ‘transcription machinery’. This allows size-dependent proteins to maintain a fixed concentration during growth without the need for complex, gene-specific regulation and for size-independent proteins to maintain a fixed number of molecules per cell. Together, such size-dependent and -independent proteins can generate size-dependent biochemical signals for progression through the cell cycle. Using this model, we show that an inhibitor-dilution mechanism can facilitate size homeostasis and correctly account for changes in protein synthesis observed in experiments that perturb the number of gene copies of cell cycle regulators as well as the overall ploidy of the cell. However, the model fails to reproduce changes in cell size seen in some of these mutants. Intriguingly, a combination of inhibitor dilution and the titration of an activator against genomic sites correctly recapitulates these changes in cell size. Such a model also produces cell size patterns consistent with a ‘sizer’ mechanism in G1 and a ‘timer’ period comprising S, G2 and M-phase, which combine to yield an ‘adder-type’ behaviour over the entire cell cycle; an observation recently made experimentally [21]. Hence, our model unites various experimental findings that were previously thought incompatible. Experimental evidence suggests that size control emerges from the interplay of regulatory proteins whose synthesis rates depend on cell size and their size-independent counterparts [8,13]. To simulate the expression of such proteins we propose a simple mathematical model based on the differential binding of transcription machinery (TM) to genes (Fig 1A). We model cell growth by assuming that components of the TM are themselves synthesised from size-dependent genes, which makes the production of TM autocatalytic, and that products of size-dependent genes control the increase in cell volume. These simple assumptions result in an exponential rise in both the amount of TM and cell size over time (Fig 1B), as is characteristic for budding yeast both in single cells and at the population level [2,21]. We note that the accumulation of TM in our simulations is compatible with experimental data on RNA polymerase II, which has been implicated in global transcriptional control [22]. Moreover, cell growth in the model depends on proteins that are themselves made by TM, which naturally leads to a direct proportionality between cell volume and transcriptional capacity. More precisely, as cells produce more and more TM their volume growth rate increases by the same extent, such that the number of TM molecules per unit cell volume remains constant. The fact that larger cells contain more TM translates into an increased occupation of size-dependent genes by TM, while size-independent genes are already fully occupied in small cells due to their high affinity for TM (Fig 1C). Consequently, the transcriptional output from size-dependent genes increases with cell size, allowing their proteins to maintain a constant concentration during exponential cell growth (Fig 1D). By contrast, expression from size-independent genes remains almost constant, such that their proteins are diluted by cell growth. Note that protein transcription is a highly complex, non-equilibrium process involving the binding of transcription factors, chromatin remodelling and multiple layers of regulation [23,24], e.g. cell cycle and nutrient-dependent control. We propose that the basic size-related regulation shown here operates alongside these other mechanisms to compensate for changes in cell size. Furthermore, the two protein classes in Fig 1A represent extremes on either end of the binding-affinity spectrum. Intermediate expression patterns, including proteins that switch from being size-dependent to size-independent during cell growth, can arise in between these extremes (S1 Fig). Our gene expression model predicts that the two principal gene types react differently to a ploidy increase, i.e., a doubling of their copy number and of the rest of the genome. In particular, size-dependent genes compensate for ploidy by splitting TM between the two gene copies and the genome, whereas their size-independent counterparts compete efficiently for TM with other genes and increase in expression (Fig 1E). However, an additional gene copy in the absence of a ploidy increase leads to a higher expression of either gene type (Fig 1F). Hence, protein synthesis depends on the copy-number-to-ploidy ratio for size-dependent genes and strictly on the gene copy number for size-independent genes. In summary, our model uses a simple mechanism to explain why size-independent proteins are diluted by cell growth, whereas size-dependent proteins keep a constant concentration, without the need for complex, gene-specific regulation. Next, we asked whether the differential expression of cell cycle regulators according to the above model would allow budding yeast cells to control their size. In budding yeast, size control acts at Start [3–5], where cells commit to cell cycle entry. Hence, we developed a cell cycle model centred on this transition (Fig 2A). In this model, passage through Start is facilitated by the activation of SBF, which is opposed by the stoichiometric inhibitor Whi5. Through the phosphorylation of Whi5, Cln3 liberates SBF from inhibition, thus driving cell cycle entry (S2A Fig). Based on experimental observations [13], we assume that Whi5 is a size-independent gene, while all other proteins in our model are size-dependent. Consequently, cell growth in G1 dilutes the inhibitor of Start, Whi5, while the activator Cln3 is maintained at constant concentration (Fig 2B), as has been observed experimentally [13]. Our model shows that this inhibitor-dilution mechanism can establish a size threshold for Start, where SBF is relieved from Whi5 inhibition only after sufficient growth has occurred (Fig 2C). This transition is rapid and switch-like because of positive feedback via Cln1 and Cln2, which are expressed in response to SBF activation and further phosphorylate Whi5 [25,26]. The positive feedback loop creates a bistable switch, which implements the threshold response to graded changes in Whi5 concentration caused by cell volume growth, providing a sensitive size-sensing mechanism. After Start has been passed, growth is restricted to the bud [4], and it continues until the end of the cycle, when the degradation of Clb1 and Clb2 initiates the separation of mother and daughter cell (Fig 2D). Intriguingly, our model readily shows size homeostasis over multiple generations (Fig 2D, lower panel). In particular, daughter cells, which we follow in our simulations because they show strong size control, reach the same size as their mothers, suggesting that Whi5 dilution can indeed couple cell division to cell growth. In order to actively regulate cell size, i.e., to reduce size differences between cells, the inhibitor-dilution model requires that larger than average cells are born with lower than average Whi5 concentration so that they progress faster through G1, while smaller than average cells have higher Whi5 concentration, which gives them more time to grow. It has been proposed that this negative correlation between cell size at birth and Whi5 concentration results from the synthesis of a fixed amount of Whi5 during a period of fixed duration, which encompasses S-, G2- and M-phases in budding yeast [13]. By design our model accounts for this synthesis pattern, restricting Whi5 synthesis to the post-Start period (Fig 2E). We find that new-born cells do indeed exhibit a size-dependent Whi5 concentration (Fig 2F). This allows for the adjustment of G1 duration to a cell’s birth size (Fig 2G). In summary, our model demonstrates that size-independent synthesis of Whi5 and its dilution during G1 can allow cells to maintain their size over multiple generations by creating a cell-size threshold for Start. Furthermore, the synthesis of a fixed amount of Whi5 per cell cycle can adjust for size differences by tuning G1 duration. To further explore the model’s ability to reproduce size control, we compared it to experiments that vary the copy number of CLN3 and WHI5, as well as the cell’s overall ploidy [13]. These data were originally used to prove that Whi5’s synthesis rate is independent of cell size, while Cln3’s synthesis rate increases in larger cells ([13] and Fig 3A). These experiments also highlight that Whi5 synthesis is largely independent of ploidy, with only a slight decrease seen between haploid and diploid cells that harbour the same number of WHI5 copies (Fig 3A, left panel). Yet, when the copy number of its gene is doubled, the Whi5 synthesis rate changes proportionally. Cln3 expression, in contrast, does change with ploidy, i.e., the slope of the synthesis rate decreases in diploid cells with one copy of CLN3 compared to their haploid counterparts (Fig 3A, right panel). However, an increase in CLN3 copy number does not affect the Cln3 synthesis rate as long as the ratio between copy number and ploidy is kept constant. Crucially, diploid cells (with two copies each of WHI5 and CLN3) were shown to be roughly twice the size of haploid cells (with one copy each of WHI5 and CLN3). We simulated these copy-number mutants using the inhibitor-dilution model, which includes the features of gene expression shown in Fig 1. The resulting simulations correctly predict the changes in protein synthesis rates for both Whi5 and Cln3 (Fig 3B). In particular, they recapitulate the copy-number dependence of Whi5 synthesis rate and the copy-number-to-ploidy dependence of Cln3 synthesis rate. The model also correctly predicts the two-fold size increase between haploid and diploid cells. However, our simulations fail to reproduce the size increase observed between haploid and diploid cells with the same number of WHI5 copies (Fig 3C). More precisely, since protein synthesis rates for both Whi5 and Cln3 are similar in haploid and diploid cells with one WHI5 copy (S3A Fig), the model predicts a similar size threshold for Start (Fig 3D). In fact, considering the slight decrease in Whi5 synthesis rate observed in experiments [13], diploid cells with one WHI5 should show a slight decrease in size compared to haploid cells according to our model. Reference [13] attributes the observed increase in cell size between haploid and diploid cells (with one or two copies of WHI5) to a delay in S/G2/M progression for diploid cells. Testing this hypothesis, we find that it only partially accounts for the observed size changes (S3B Fig). In particular, diploid cells with one WHI5 are predicted to be smaller than haploid cells with two WHI5, suggesting a larger influence of Whi5 synthesis rate than ploidy (S3C Fig). However, in experiments the opposite is observed ([13] and Fig 3A). Moreover, a delay in S/G2/M progression together with the observed increase in Whi5 synthesis rate would lead to a more than two-fold difference between haploid and diploid cells (S3C Fig), in contradiction to experimental data [13]. Taken together, the inhibitor-dilution model thus correctly captures protein synthesis rates in copy-number and ploidy mutants but fails to reproduce the observed size increase for some diploid cells. Previous theoretical and experimental studies attributed the effects of ploidy on cell size to an alternative control mechanism relying on the titration of a protein with constant concentration against a fixed number of nuclear sites [9–11,18–20]. In particular, it has been suggested that Cln3 is titrated against SBF bindings sites on the genome [20]. Based on these suggestions, we augmented the inhibitor-dilution model with a titration mechanism to test whether these two concepts can be brought into unison (Fig 4A). In the pure inhibitor-dilution model, SBF, Whi5 and Cln3 interact in a strictly concentration-based manner (S2A Fig). By contrast, the titration model assumes that SBF occupies a fixed number of sites on the genome. In early G1 (i.e., in small daughter cells), these sites are filled with Whi5-inhibited SBF complexes to which Cln3 can bind tightly in a stoichiometric fashion. Once bound, Cln3 slowly hypo-phosphorylates Whi5 and dissociates in the process. However, it can rapidly rebind to unphosphorylated SBF:Whi5 (S2B Fig). As the cell grows larger, the number of Cln3 molecules per cell increases (Cln3 is a size-dependent protein, whose concentration is maintained constant in G1) (Fig 4B). This leads to a gradual accumulation of Cln3:Cdk1 heterodimers on Promoter:SBF:Whi5 complexes until all sites are filled, at which point free Cln3:Cdk1 kinase complexes emerge in the nucleus. Free Cln3:Cdk1 then promotes rapid hyper-phosphorylation of SBF-bound and free Whi5, facilitating the Start transition (Fig 4B). Similar to the inhibitor-dilution model, the titration model readily yields size homeostasis in consecutive generations (Fig 4C) by coupling the passage through Start to cell size (Fig 4D). When simulating changes in gene copy number, we observe that, similar to inhibitor dilution, the titration model correctly predicts protein synthesis rates (Fig 4E). However, the titration mechanism also captures the increase in size between haploid and diploid cells with the same number of WHI5 copies (Fig 4E). In particular, diploid cells harbour twice the number of SBF binding sites, which require a higher amount of Cln3, and therefore a larger cell size, to be filled (S4A Fig). Note that our model overestimates the size of diploid cells with one copy of CLN3 (Fig 4F). The cause for this discrepancy is that the absence of a second CLN3 copy in diploid cells only reduces Cln3 synthesis rate by ~15% (compare diploid cells with 1xCLN3 and 2xCLN3 in Fig 3A, right panel), whereas the model predicts a reduction by ~50% (Fig 4E, right panel). After accounting for this, cell size predictions are much more accurate (S4B and S4C Fig). It is not yet clear why a single CLN3 can partially compensate for the second copy’s expression rate in diploid cells. Further experimental evidence for a titration mechanism comes from an observed increase in cell size upon transformation of otherwise wild-type cells with a high copy number plasmid containing perfect SBF binding sites [20]. These decoy sites were proposed to change the size threshold for Start by binding Cln3 such that an increased number of Cln3 molecules, and hence a larger cell size, is required to initiate the transition. Simulating this setup, our model does indeed show such an increase in size (S4D Fig), providing further support for the existence a titration mechanism. In summary, a combination of Whi5 dilution and Cln3 titration against SBF binding sites is not only able to capture protein synthesis rates but also the size of WHI5- and CLN3-mutant haploid and diploid cells and of cells harbouring an increased number of SBF binding sites. Historically, three different strategies have been proposed to maintain cell size homeostasis: the sizer, where a cell cycle transition is triggered once the cells reaches a critical target size; the timer, whereby the cell cycle takes a constant amount of time; and the adder, postulating that cells add a constant volume each generation [27,28]. Each of these concepts may apply to the complete cell cycle or only to a certain cell cycle phase, and all of them generate characteristic size patterns that can be probed experimentally (Fig 5A). An ideal sizer mechanism suggests that the final volume at the end of the sizer period is independent of the initial volume, such that the added volume shows a linear slope of minus one, i.e., small cells need to grow more to reach the critical size. By contrast, exponentially growing cells that employ a perfect timer show a slope of plus one in the added volume as small cells grow less during the same time increment. Note that a slope of exactly one is only observed if cells double their mass within the phase that uses a timer, e.g. if the timer is employed over the whole cell cycle of a symmetrically dividing cell. Finally, an adder leads to a slope of zero since the added volume is assumed to be constant. We wanted to understand how these concepts connect to the mechanistic model of cell cycle control presented above. Simulations of our titration model reveal that G1-phase behaves like an imperfect sizer with smaller cells adding more volume during G1 (slope of -0.64; Fig 5B, right panel) and cell size at S-phase entry showing a slight positive correlation with birth size (Fig 5B, left panel). S/G2/M-phases, by contrast, exhibit a timer (see also S5A Fig). The combination of a mechanistic sizer and a mechanistic timer yields a phenomenological adder with the added volume being virtually independent of cell size at birth (R = -0.02; Fig 5B, right panel). However, the added volume is not directly sensed by the system in any way. Instead the negative slope of the sizer compensates for the positive slope of the timer. The results above raise the question as to why cells employ two seemingly different strategies in G1 and S/G2/M-phases, a sizer and a timer, respectively. Presumably, S, G2 and M-phase are completed fast, with a size-independent timing, to allow the mother cell to start the next budding event, while size control is relegated to the daughter cell’s G1 phase. In addition, a timer period of constant length in combination with a size-independent Whi5 synthesis allows the cell to produce a constant, size-independent amount of Whi5 per cell cycle (Figs 5C and 2E). This constant Whi5 amount is part of the mechanism that tunes G1 length with respect to birth size. Hence, the S/G2/M timer helps to set up the G1 sizer. We also note that our simulations show an imperfect sizer with a slight positive correlation between the cell volume at Start and the birth volume (Fig 5B, left panel), as has been found experimentally [21,27]. Whereas an ideal sizer requires the size threshold for Start to be independent of birth size, we find that cells which are larger at birth progress through Start at a slightly larger size (Fig 5B and 5D). According to our model, the main reason for this threshold change is the distribution of Whi5 molecules at cell division. In particular, larger cells are born with slightly higher amounts of Whi5 (Fig 5E), since some of the Whi5-containing complexes are distributed according to the volume ratio of mother and daughter cell (Fig 5F). It is primarily by this mechanism that birth size affects the size threshold for Start in our model, as shown in S5B Fig, where we manually set the Whi5 amount at birth to a constant value (birth size-independent) and find that the model behaves as an almost ideal G1 sizer. In summary, our model shows that size control in budding yeast uses an S/G2/M timer that helps to produce a constant amount of Whi5 per cell cycle and to facilitate a sizer in the G1 phase of daughter cells. Both mechanism combine to yield a phenomenological adder over the whole cell cycle. However, the size-dependent distribution of Whi5 at cell division can cause an imperfect adjustment to size differences at birth. Balanced growth, achieved by coupling cell division to the increase in cell mass, is crucial to cell survival as progressive changes in size over generations would eventually lead to a breakdown of biochemical processes. In this study, we developed a mechanistic mathematical model for size control in budding yeast based on the differential expression of cell cycle regulators in growing cells. We show that the interplay of size-dependent and size-independent synthesis of these regulators can establish a size threshold at Start and facilitate size homeostasis. It has long been recognised that the amounts of most proteins in a cell increase with cell size [29,30], such that protein concentrations remain constant and reaction rates are unaffected by growth [12]. This has also been observed for the majority of cellular mRNAs, suggesting that adaptation to volume growth occurs at the transcriptional level [22,29,31–33]. Based on these observations, we propose a general mathematical model for gene expression in growing cells which assumes that a limiting component of the transcription machinery, which we named TM and that may correspond to an RNA polymerase or factors influencing chromatin accessibility [18], is produced in an autocatalytic manner by transcribing its own mRNA. Under conditions where nutrients and precursors are not limiting, this leads to an exponential increase in TM. If we assume that the increase in cell volume depends on proteins that are themselves transcribed by TM, the exponential rise in TM directly translates into an exponential increase in cell volume and it naturally leads to a direct proportionality between both, i.e., protein synthesis rates per unit volume remains constant. This scaling is an emergent property of the system and does not require complex regulation or a dedicated mechanism that measures size and tunes transcriptional capacity accordingly. In very large cells, genes become saturated, at which point transcription rates remain constant and cell growth transitions into a phase of linear increase. These features of the model are consistent with a large body of experimental literature showing exponential growth of cell volume and transcription for small cells which plateaus when cells exceed a certain size [12,21,22,27,34]. Given this model of gene expression, two different types of genes emerge in our simulations based on their affinity to TM. Genes that bind TM with high affinity are saturated early, in small cells, and thus show size-independent protein synthesis. Consequently, they give rise to size-independent proteins, whose amount is constant, leading to a decreasing concentration in growing cells. Whi5 is an example of such a protein [13]. Due to their high affinity, size-independent genes compete efficiently for TM and an increase in their copy number, due to gene or genome duplication, directly translates into an increased synthesis and concentration. We hence propose that, in the context of size control, size-independent genes can act as gene-copy-number sensors. Beyond size regulation, the genes might encode proteins that need to be present in a fixed proportion to the genome content, e.g., transcription factors or histones. By contrast, size-dependent genes bind TM with lower affinity, such that their occupation by TM increases proportional to cell volume. Through this mechanism, their proteins can maintain a constant concentration until the gene is saturated. We propose that the majority of proteins uses this type of control, Cln3 being a concrete example [13]. Due to their characteristics, size-dependent genes share TM among themselves, such that an overall ploidy increase does not result in an increase in protein concentration. Size-dependent genes can hence act as sensors for the copy-number-to-ploidy ratio, the gene dosage. Variations of the affinity constants between the two extremes may lead to intermediate expression patterns, including genes that can switch from size-dependent to size-independent expression within a given range of cell sizes. We propose that this simple mechanism of gene expression operates alongside other forms of transcription control, which involves non-equilibrium processes and stochastic phenomena [23,24], to compensate for cell growth. By incorporating the gene expression model into a model of the yeast cell cycle, we show that size-independent synthesis of the inhibitor Whi5 and size-dependent synthesis of the activator Cln3, a mechanism termed inhibitor dilution [13], can indeed establish size control at Start. It is important to note that, because Whi5 is a stoichiometric inhibitor of SBF without catalytic activity [15,16], we have assumed in our inhibitor-dilution mechanism that Whi5 and SBF form a tightly bound complex. In addition, we assumed that phosphorylation of Whi5 by Cln3 breaks up the complex and liberates SBF. Considering that SBF maintains a constant concentration, as has been shown experimentally for one of its subunits [13], Whi5 is therefore in fact countered by two size-dependent activators, Cln3 and SBF. Given these molecular interactions, our results suggest that, in the inhibitor-dilution paradigm, the rising number of SBF molecules in a growing cell eventually overcomes inhibition by exceeding the constant number of Whi5 molecules (see S1 Text for details). Cln3 merely sets the threshold at which SBF activation occurs by keeping a fraction of Whi5 molecules in a phosphorylated (inactive) state. Because this fraction does not change appreciably with cell size, Cln3 is not directly involved in creating the size-dependent signal that facilitates Start in our version of the inhibitor-dilution model. Hence, between the inhibitor-dilution model and the titration-of-nuclear-sites mechanism there exists an intriguing symmetry, in which Whi5 and nuclear sites are very much alike. Both are constant in number and proportional to DNA content and both titrate away an activator. We also show that the gradual increase in SBF activity in response to cell volume growth that is caused by Whi5 dilution is converted into an all-or-nothing decision by a bistable switch located at Start. This switch is created by a positive feedback loop on SBF activity and it establishes a strict size threshold of Start. Hence, positive feedback and bistability are used to implement a size checkpoint in G1. While inhibitor dilution is able to maintain size homeostasis and reproduce the size increase seen in diploid cells, it fails to explain why an increase in ploidy at a constant number of WHI5 copies leads to larger cells. Such a change does not alter the expression of Whi5 and Cln3 and hence should not affect cell size at Start. Even the delay in S/G2/M progression observed experimentally [13] is unable to reproduce these size changes in our model, suggesting that ploidy influences cell size beyond an effect through Whi5 and Cln3 expression and S/G2/M duration. Such an effect could be mediated by an as-yet-unknown inhibitor of Start which is produced in a size-independent manner similar to Whi5. In this case, an increased expression of this inhibitor in diploid cells, due to a higher copy number of its gene, would cause the observed size increase. However, a much more appealing hypothesis is that the genome itself acts as an inhibitor of Start. In particular, the binding of SBF to a limited number of genomic sites, which was proposed based on experiments [20], essentially converts SBF into a variable that does not change in number with cell size, as only the SBF that is bound to the genome would affect the Start transition. Since the number of Whi5 molecules is constant as well, the amount of Whi5:SBF complexes, assuming tight binding between both, is not changing with cell size. However, the number of Cln3 molecules increases, such that Cln3 titrates against Whi5:SBF complexes on the genome. At a particular threshold size, Cln3 exceeds the number of Whi5:SBF complexes, leading to a sharp increase in free Cln3 that can trigger Start. Positive feedback is again used to convert this increase into an all-or-none decision. In this context, a diploid cell is larger because it contains twice the number of SBF binding sites, requiring more Cln3 molecules to trigger Start. Hence, the genome itself, through providing SBF binding sites that titrate Cln3, acts as a Start inhibitor, using a form of distributed control (binding sites distributed throughout the genome) instead of a single gene product such as Whi5. We show that this Cln3 titration model is consistent with WHI5 and CLN3 mutant phenotypes and with experiments in which additional SBF binding sites are expressed and consequently cell size at Start is increased [20]. Also note that size-independent synthesis of Whi5 in the titration model is beneficial because increasing Whi5 production in large cells would impair their progress through Start, thereby compromising size control. Moreover, the proportional increase in Whi5 synthesis with gene copy number allows for a constant ratio between Whi5 and SBF molecules on binding sites in cells with increasing ploidy, providing an intriguing hypothesis for why Whi5 is synthesised in a size-independent manner. A recent study of cell cycle commitment in buddy yeast called into question the dilution of Whi5, arguing instead that a size-dependent increase in the concentrations of G1/S transcription factors helps to set the size threshold for Start [35]. In S6 Fig, we show that such a model is indeed able to achieve size homeostasis but is incompatible with data on Whi5 synthesis rates and the size of some mutant strains (see S2 Text for details). Hence, our model in combination with careful measurements of not only protein concentrations but also protein synthesis rates and cell sizes in mutants could help to resolve such discrepancies. In recent years, studies of bacterial size control have argued for an ‘adder-type’ mechanism, whereby cells add a constant increment of cell mass per cycle [36,37]. A similar type of behaviour was found between two budding events in S. cerevisiae [34]. Yet, it remained unclear whether cells actively sense the added mass and use this information to regulate cell cycle events, a scenario later referred to as a mechanistic adder [21]. From our simulations, we indeed observe the presence of an adder over the whole cell cycle, with no correlation between the added cell mass and the volume at birth. However, this behaviour does not result from a direct mechanism, but rather from a combination of a mechanistic sizer in G1 and a mechanistic timer in S/G2/M, which is in excellent agreement with a recent study arguing that the adder phenomenon emerges from independent pre- and post-Start controls [21]. Similar to these and other experiments [5,21], our titration model shows an inverse proportionality between G1 length and birth size, and an imperfect sizer mechanism. We propose that adaptation is imperfect because of a volume-dependent distribution of Whi5. An ideal sizer, where Start size is independent of birth size, requires that each daughter cell receives a constant amount of Whi5. However, Whi5 complexes that diffuse freely in the nucleus or cytoplasm would be distributed based on the size of the daughter cell, with large cells receiving a larger increment of Whi5 that keeps them in G1 for longer. In our model, this results in a weak birth-size dependence of the Start threshold and imperfect size control. This might be one reason why cells do not rely on a pure inhibitor-dilution mechanism, which would exacerbate the influence of Whi5 distribution, but instead use a combination of Whi5 dilution and Cln3 titration. In addition, Cln3 is a highly unstable protein [38], and thus provides a snapshot of the current transcriptional capacity and volume of a cell, while Whi5 was produced in the previous cycle, inevitably introducing some form of memory of past growth conditions. In summary, our study provides a mechanistic model of gene expression and cell cycle regulation in budding yeast that readily shows size homeostasis. Since the control network of Start in budding yeast is structurally similar to restriction point control in mammalian cells, similar mechanisms could be at work during mammalian size control. Our models for budding yeast size control comprise sets of ordinary differential equations (ODEs). These ODEs describe the dynamics of genes and proteins in terms of their molecule number rather than concentration, which is used by most biochemical models that do not account for cell volume growth. In the following, we explain each of the two models (inhibitor dilution and titration of nuclear sites) in detail, starting with a generic description of gene expression that underlies both models. Both size-control models were prepared in the Systems Biology Toolbox 2 [44] for MatLab (version 9.1.0 R2016b) and simulated with the CVODE routine [45]. Bifurcation diagrams were calculated using the freely available software XPP-Aut [46]. Models are provided as S1–S3 Files in the Supplement and different versions are available at www.cellcycle.org.uk/publication. Model files were also deposited in BioModels [47] and assigned the identifiers MODEL1803220001 and MODEL1803220002. Parameter values and initial conditions are listed in S1–S4 Tables and S5 Table shows the changes required to simulate ploidy mutants in Figs 3, 4, S3 and S4. In order to simulate cells of different sizes (e.g. in Fig 2F and 2G), we varied the specific growth rate, with higher growth rates producing larger cells. In particular, the specific growth rate (μ) in our model follows from Eqs 9 and 10 as μ=1Vtd(Vt)dt=kVoSy∙GDTMVt∙GCNGDt. Since GDt ≫ GIt and almost all of the TM is bound to genes for the cell sizes we study here, the amount of transcriptionally active size-dependent genes can be approximated by the total number of TM (GDTM ≈ TMt). Moreover, we can calculate the transcriptional capacity per unit cell volume as TMtVt=kTmSykVoSy. Taken together this gives μ≈kVoSy∙kTmSykVoSy∙GCNGDt, demonstrating that by changing both kVoSy and kTmSy by the same factor, we can change the specific growth rate, while still maintaining the same transcriptional capacity per unit cell volume and thus similar protein expression. Accordingly, for simulations in Fig 2F and 2G, kVoSy and kTmSy were multiplied by a factor f ∈ [0.75, 1.25]. For simulations in Figs 5 and S5, we followed a single cell lineage over a large number of divisions to correlate cell sizes at different cell cycle stages. To obtain different cell sizes, we again varied the growth rate as described above, assuming that it changes at cell division. In particular, we assumed that the specific growth rate in the next cycle (μn+1) is partly inherited from the mother cell’s growth rate (μn) and partly influenced by stochasticity, e.g., by the random distribution of molecules at cell division, using μn+1=0.5∙μn+0.5∙μ¯∙(1+N(0,0.04)), where μ¯ is the average growth rate and N(0,σ) a normally distributed random variable with mean 0 and variance σ.

10.1371/journal.pgen.1005634

EGFR/Ras Signaling Controls Drosophila Intestinal Stem Cell Proliferation via Capicua-Regulated Genes

Epithelial renewal in the Drosophila intestine is orchestrated by Intestinal Stem Cells (ISCs). Following damage or stress the intestinal epithelium produces ligands that activate the epidermal growth factor receptor (EGFR) in ISCs. This promotes their growth and division and, thereby, epithelial regeneration. Here we demonstrate that the HMG-box transcriptional repressor, Capicua (Cic), mediates these functions of EGFR signaling. Depleting Cic in ISCs activated them for division, whereas overexpressed Cic inhibited ISC proliferation and midgut regeneration. Epistasis tests showed that Cic acted as an essential downstream effector of EGFR/Ras signaling, and immunofluorescence showed that Cic’s nuclear localization was regulated by EGFR signaling. ISC-specific mRNA expression profiling and DNA binding mapping using DamID indicated that Cic represses cell proliferation via direct targets including string (Cdc25), Cyclin E, and the ETS domain transcription factors Ets21C and Pointed (pnt). pnt was required for ISC over-proliferation following Cic depletion, and ectopic pnt restored ISC proliferation even in the presence of overexpressed dominant-active Cic. These studies identify Cic, Pnt, and Ets21C as critical downstream effectors of EGFR signaling in Drosophila ISCs.

Studies suggest that epidermal growth factor receptor (EGFR) signaling activation is a causal driver of many stem cell-derived epithelial cancers, including colorectal cancer. As in the human intestine, epithelial renewal in Drosophila intestine is orchestrated by intestinal stem cells (ISCs). EGFR signaling also plays an important role in regulating ISC proliferation in flies. However, the mechanism by which EGFR/Ras/MAPK signaling promotes ISC proliferation is poorly understood. Here we demonstrate that the transcriptional repressor, Capicua (Cic), mediates these functions of EGFR signaling. We found that the critical role of Cic as a negative regulator of cell proliferation in the fly midgut is consistent with its tumor suppressor function in mammalian cancer development. The direct target genes of Cic were identified by ISC-specific mRNA expression profiling and DNA binding mapping (DamID) method. Cic represses cell proliferation via regulating string (stg), Cyclin E (CycE), and the ETS domain transcription factors Ets21C and pointed (pnt). Using genetic tests we show that these interactions are meaningful for regulating stem cell proliferation. Combining our knowledge of Cic with what was previously known about CIC in tumor development, we propose that human CIC may regulate Ets transcription factors and cell cycle genes in Ras/MAKP-activated tumors.

EGFR/Ras/MAPK signaling has diverse functions in regulating cell proliferation, growth, differentiation and survival in most animal cells [1]. Abundant studies also indicate that epidermal growth factor receptor (EGFR) activation is a causal driver of many cancers, including breast, lung, brain, and colorectal cancer [2]. Similarly, activating mutations in KRAS and BRAF, which are essential downstream effectors of the EGFR, are among the most common mutations found in a very wide range of human cancers [3,4]. However, despite much study, many questions remain to be answered to fully understand the impact of EGFR and its downstream effectors during normal cell function and in carcinogenesis. As many epithelial cancers arise through dysregulation of the stem cell self-renewal and homeostatic maintenance of the epithelium [5], understanding the precise functions of EGFR signaling in epithelial homeostasis is very important. The Drosophila midgut is an outstanding model system to study the basis of epithelial homeostasis due to its simple structure, similarity to the mammalian intestine, and powerful genetics. As in the mammalian intestine, epithelial turnover in the fly midgut is carried out through a dynamic process mediated by intestinal stem cells (ISCs). ISCs undergo cell division to renew themselves and give rise to transient cells called enteroblasts (EBs), which can further differentiate into either absorptive enterocytes (ECs) or secretory enteroendocrine (EE) cells. When damaged or aged cells are lost from the fly’s gut epithelium, ISCs respond by dividing to replenish the epithelium [6,7,8]. During this response multiple Drosophila EGFR ligands, namely spitz (spi), vein (vn), and keren (krn) are induced in progenitor cells (EBs and ISCs), visceral muscle (VM) and ECs respectively. Thereby, the EGFR signaling pathway is activated in ISCs. This promotes ISC growth, division and midgut epithelial regeneration [9,10,11]. ISCs defective in EGFR signaling cannot grow or divide, are poorly maintained, and are unable to support midgut epithelial replenishment after enteric infection by the bacteria Pseudomonas entomophila (P.e.) [11] or Erwinia carotovora carotovora 15 (ECC15) [12]. Interestingly, the critical role of EGFR signaling in the Drosophila intestine is consistent with its role during mammalian gut homeostasis and colorectal cancer development [10,11,12,13]. EGFR signaling is required for murine ISC growth [14,15], and the deletion of Lrig1, a negative feedback regulator of EGFR signaling, causes excessive ISC proliferation [16]. Furthermore, adenoma formation in Apcmin/+ mice was severely impaired in a genetic background with partial loss of function of EGFR (Egfrwa2) [17]. Despites its importance, the mechanism by which EGFR/Ras/MAPK signaling promotes ISC proliferation is poorly understood in this cell type. Indeed, despite decades of intensive study, the precise linkage between EGFR/Ras/MAPK signaling and cell growth and division is surprisingly obscure for animal cells in general [3]. Textbook models highlight a prevailing model in which EGFR/Ras signaling controls cell proliferation via a Ras-Myc-CyclinD-Rb pathway [18,19]. While this may have relevance in some human cancers it is clearly not the case in normal Drosophila cells, and so other mechanisms should be sought and characterized. One potentially important downstream effector of EGFR signaling is the HMG-box transcriptional repressor Capicua (Cic). This highly conserved DNA binding factor has been shown to act downstream of receptor tyrosine kinase (RTK)/Ras/MAPK signaling in Drosophila eye and wing imaginal discs, embryos, and ovaries [20,21,22,23] where it regulates diverse RTK-dependent processes including cell proliferation, specification, and pattern formation. Cic orthologs from invertebrate and vertebrate species share two well-conserved regions: the HMG-box, presumed to mediate DNA binding at target promoters [21] and a C-terminal domain [24]. The C-terminal region of Drosophila Cic contains a “C1” motif important for repressor activity, and a “C2” motif that functions as a MAPK docking site responsible for downregulation of Cic following the activation of RTK signaling [25]. It has been proposed that MAPK phosphorylates Cic in its C2 motif, and that phosphorylated Cic is either degraded or re-localized to the cytoplasm [25]. Cic downregulation controlled by Torso and EGFR signaling varies in different Drosophila tissues [24]. For example, Torso RTK signaling, which also works via the Ras/Raf/MAPK pathway, apparently increases the rate of Capicua degradation by promoting its accumulation in the cytoplasm [26]. EGFR signaling has been reported to regulate Cic protein in distinct ways in different tissues. Wing and eye discs cell clones mutant for Egfr or Ras showed elevated levels of Cic protein [20,27]. In the ovary, in contrast, Cic protein localized to the cytoplasm in cells in which EGFR signaling was active, but in nuclei in cells in which EGFR signaling was inactive [25]. A recent study suggested that Cic actually undergoes a two-step process in releasing its target gene repression: slower changes in nuclear localization occur after a faster reduction of Cic repressor activity [28]. In cultured human cells, EGF stimulated dissociation of human CIC from importin-α4 (also known as KPNA3), an adaptor required for the nuclear import of many proteins. But full length GFP-CIC was nuclear even after EGF stimulation, and the N-terminal half of the CIC protein was found to be nuclear, even though it does not bind to importin-α4. Hence the biological significance of the CIC:importin association remains unclear [29]. CIC, the human homolog of Drosophila Cic, has been implicated in several human diseases including spinocerebellar ataxia type 1 (SCA1) neuropathology, oligodendroglioma (OD) [30] and Ewing-like sarcoma [31]. Human CIC is frequently mutated in samples from cancer genome studies such as The Cancer Genomic Atlas (TCGA) (S1 Fig) [32,33]. For instance CIC mutation was reported in 6 out of 7 brain tumors [30], 3 out of 11 breast cancers [34] and 6 out of 72 colorectal cancers [35]. The Drosophila work suggests that in these cases CIC loss might have the same downstream consequences (e.g. cell transformation) as oncogenic activation of the EGFR, RAS or BRAF, but this has not been rigorously evaluated. During RNAi screening we discovered that depletion of Cic in Drosophila’s intestinal stem cells (ISCs) activates these cells for rampant proliferation [11]. Based on previous studies in other fly organs we hypothesized that Cic might act as an obligate repressor downstream of EGFR signaling, itself a central driver of normal ISC proliferation in both flies and mice, as well as in many human colorectal cancers, which are frequently mutant for RAS, BRAF, or CIC. However, until now this hypothesis had not been tested and the underlying mechanisms via which Cic might control ISC proliferation remained undefined. In this report we demonstrate that Cic acts as a critical negative downstream regulator of EGFR signaling to control ISC proliferation. We show that EGFR/Ras activity controls Cic nuclear localization, and we present RNA-Seq and DamID-Seq datasets that together constitute a genome-wide survey of potential Cic target genes in Drosophila ISCs. Our analysis indicated that Cic not only directly regulates cell cycle regulators such as string (cdc25) and Cyclin E, but also the ETS transcription factors pnt and Ets21C, all of which must be de-repressed to activate ISCs for growth and division. To investigate a potential role for Cic in regulating ISC proliferation, we used the esg-Gal4-UAS-2XEYFP; Su(H)GBE-Gal80, tub-Gal80ts system (henceforth referred as esgts; Su(H)-Gal80) to express UAS-cic-RNAi specifically in ISCs. After 4 days of cic-RNAi induction, a dramatic increase in the number of YFP positive cells (Fig 1A and 1B) and a large increase in ISC mitoses were observed (Fig 1C). Most of the PH3+ cells were YFP+ [YFP+, PH3+ cells = 99.37% (nmidguts = 10 midguts, ncells = 994), YFP-, PH3+ cells = 0.63% (n = 10, ncells = 7)], indicating that Cic regulates ISC proliferation cell autonomously. When we used another ISC-specific driver Dlts (tub-Gal80ts UAS-GFP; Dl-Gal4) to knock down cic in ISCs specifically, we not only detected the same overporoliferation phenotype (S3A, S3B and S3E Fig) but also found that most of mitotic cells were GFP+ (S3F Fig). Increased GFP+ cells and mitoses were also noticed when the esgGal4 UAS-GFP tub-Gal80ts system (henceforth referred as esgts) was used to express UAS-cic-RNAi in ISCs and their undifferentiated daughters, the EBs (S2A–S2B and S3 Figs). To further validate the specificity of this RNAi experiment, GFP-marked ISC clones homozygous for the loss-of-function allele cicfetU6 [22] were generated using the MARCM system [36] (S2C–S2H Fig). The size of marked ISC clones was quantified at intervals after clone induction by measuring GFP-labeled clone areas. cic mutant clones were larger than control clones at all time points assayed (Fig 1D). In addition, the numbers of cells per clone were increased in the cic mutant clones (Fig 1E). To further confirm Cic’s function in the midgut, we generated viable transheterozygotes using three different loss-of-function alleles of cic. cicfetE11 is a P-element insertion mutant, while both cicfetT6 and cicfetU6 are homozygous lethal EMS alleles [22]. In addition to the EGFR-related extra wing vein phenotype reported previously [27], these transheterozygote mutants showed increased mitoses in their midguts (Fig 1I). As the ISCs are the predominant dividing cell type in Drosophila midguts, these data further indicate a role for Cic as an obligate repressor of ISC proliferation. To investigate the respective requirements of Cic in the ISC and EB cell types, the EB-specific driver Su(H)ts [Su(H)-Gal4,UAS-CD8-GFP; tub-Gal80ts] was used to knock down cic in EBs. Increased mitoses were observed after depleting cic in EBs (S3C–S3E Fig). However, in this case only a few GFP+ EBs were observed in mitosis, while most of the dividing cells marked by PH3 were GFP-negative (S3F Fig). These GFP-negative mitotic cells are likely ISCs. These data indicated that Cic has both cell autonomous and non-cell autonomous functions in regulating ISC proliferation. In this study we followed up on Cic’s cell autonomous effects on ISC proliferation, and the non-cell autonomous effect was not investigated further. To determine whether increased Cic function yields a phenotype similar to that of EGFR loss-of-function, we generated transgenic flies harboring UAS-cicΔC2-HA or UAS-cic-HA. CicΔC2 is a Cic derivative carrying a deletion of the MAPK docking site-C2 motif, and has been shown to be a dominant repressor that escapes inactivation by MAPK [25]. Either cic or cicΔC2 were over-expressed in progenitor cells using esgts, and then the flies were fed Pseudomonas entomophila (P.e.) for 12 hours to generate an enteric infection. ISCs from control midguts, which expressed GFP only, showed regeneration-associated proliferation [8]. In contrast both cic and cicΔC2 overexpressing midguts displayed an inhibition of regeneration after 12 hours P.e. infection (Fig 1J). To test if cic or cicΔC2 overexpression could influence turnover of the midgut epithelium we used the esgts F/O system (esg-Gal4; tubGal80ts Act>Cd2>Gal4 UAS-flp UAS-GFP) [11] to mark all the ISC progeny produced during 12 days of cic overexpression. Normally, the posterior midgut epithelium renews it self within about 12 days [8]. Therefore, control midgut epithelia were almost completely replaced by large GFP+ clones that formed during 12 days. However, in the gain-of-function Cic conditions, growth of GFP-marked clones was significantly decreased, indicating that gut epithelial renewal was greatly suppressed (Fig 1H–1J). EGFR activates ISCs for growth and division via Ras/Raf/MAPK signaling. When an activated form of the EGFR (λTOP) [37] or activated Ras (RasV12G) [38] is ectopically expressed in progenitor cells, ISC division is dramatically induced. Conversely, EGFR suppression by inducing Egfr-RNAi, Ras-RNAi, or MEK-RNAi in progenitor cells almost completely inhibits ISC division and growth [11,12]. Furthermore, inhibition of EGFR signaling suppresses the activation of ISC divisions after P.e. infection [10,11]. As demonstrated above, Cic knockdown and overexpression phenocopy these EGFR overexpression or knockdown phenotypes, respectively, suggesting that Cic may act as a downstream effector in the EGFR signaling in ISCs. To test the function of Cic in EGFR signaling we performed epistasis tests. After 2 days of clone induction with the esgts F/O system, control midguts generated only 2-cell clones, whereas clones overexpressing an activated variant of the EGFR, (λtop), grew very large and showed increased ISC division. However, when cic or cicΔC2 was co-overexpressed along with λtop, clone sizes and ISC mitoses were significantly reduced (Fig 2A–2C and 2I). Overexpression of Cic or CicΔC2 could also partially inhibit the ISC growth effects of RasV12S35, an activated allele that can activate RAF/MAPK signaling but not PI3K signaling [38] (Fig 2D, 2H, and 2J). Furthermore, we used esgts to induce Egfr-RNAi, or Ras-RNAi in combination with cic-RNAi. The cic, Egfr or cic, Ras double RNAi animals exhibited increased ISC mitosis relative to controls expressing Ras-RNAi or Egfr-RNAi only (Fig 2E–2G and 2K), indicating that reduced ISC proliferation caused by the inactivation of EGFR signaling can be restored by Cic knock-down. These epistasis data further support the hypothesis that Cic acts as a negative downstream effector of EGFR to regulate ISC proliferation. To understand how EGFR signaling controls Cic in ISCs, we expressed HA-tagged Cic or CicΔC2 protein in midgut progenitor cells (ISCs and EBs). As expected, HA-tagged Cic or CicΔC2 proteins were only detected in nuclei under normal conditions (Fig 3A–3A’ and 3B–3B’). However, HA-tagged Cic protein accumulated nearly exclusively in the cytoplasm when RasV12S35 was co-expressed with it (Fig 3E–3E’). In contrast, CicΔC2 remained in the nucleus even following ectopic RasV12S35 expression (Fig 3F–3F’). A similar but milder re-localization of Cic protein from the nucleus to the cytoplasm was observed following P.e. infection (Fig 3C and 3C’), a treatment known to increase MAPK signaling in the gut [11]. It is interesting to note that CicΔC2 did not completely suppress RasV12S35 induced ISC proliferation, even though it remained localized to nuclei in RasV12S35 expressing cells (Fig 2H and 2J). However, nuclear CicΔC2 lost its characteristic punctate localization in the presence of RasV12S35 expression, and became more diffusely localized in the nucleoplasm (Fig 3F–3F’). These results suggest that, although EGFR signaling controls Cic nucleo-cytoplamic localization via the C2 motif, there may be a second MAPK-dependent mechanism to regulate Cic repressor activity, involving dissociation from chromatin, that is C2-independent. Cic has been studied in several cell types from both Drosophila and humans. In human melanoma cells, CIC represses mRNA expression of the PEA3 subfamily of ETS transcription factors, namely ETV1, ETV4 and ETV5 [29]. In early Drosophila development post-transcriptional down-regulation of Cic by the Torso and EGFR pathways regulates terminal and dorsal-ventral patterning, respectively, by allowing expression of Cic target genes such as huckebein (hkb), intermediate neuroblasts defective (ind), and argos (aos) [39]. However, a genome-wide mapping of Cic target genes has not yet been reported. To identify Cic target genes involved in ISC growth and proliferation we profiled Cic binding throughout the genome using the “TaDa” (Targeted DamID)” technique. The TaDa method involves low-level expression of a GAL4-inducible Dam methylase-fusion protein in a specific cell type, enabling cell-specific profiling without cell isolation [40,41]. Here, we induced a low level of Dam-only or Dam-Cic fusion protein in progenitor cells (ISC & EB) using the esgts system and a 24-hour induction. Genomic DNA was extracted from isolated midguts, digested with Dpn I, which cuts only methylated GATCs, and amplified. The amplified gDNA fragments were subjected to high-throughput sequencing, rather than tiling microarrays as previously reported [40]. We identified 2279 binding sites that were highly enriched (log2 fold change > 3, false discovery rate<0.01%) when comparing Dam-Cic to Dam alone samples (S1 Table). These sites were non-randomly distributed in the genome, and were significantly over-represented ~500 bp 5’ to Transcription Start Sites (TSS; Fig 4A). Cic DamID was also performed on progenitor cells from P.e. infected midguts. After a 24 hours induction of Dam or Dam-Cic transgenes via the esgts system, flies were fed P.e. bacteria for 16 hours. The number of highly enriched (log2 fold change > 3, FDR < 0.1%) peaks was reduced to 1903. In addition, the fold change of peaks (Dam-Cic vs Dam-alone) after P.e. infection was significantly decreased (Figs 4B–4C and S4A). The frequency of peaks adjacent to TSS was also significantly reduced in the P.e.-infected midgut sample (Fig 4A). We believe that this decrease was due to the change of Cic localization from the nucleus to cytoplasm, which was caused by the activation of EGFR/Ras/MAPK signaling after infection. To further understand how Cic regulates ISC proliferation we performed gene expression profiling using amplified mRNA from FACS-sorted esg+ progenitor cells that expressed cic-RNAi, and controls. As a way to identify potentially direct target genes of Cic, the RNA-Seq and DamID-Seq data sets were cross-compared. Amongst 439 transcriptionally up-regulated genes (>1.5 fold change, 90% CI) (S2 Table), a large fraction [134 genes, (S3 Table)] had Cic binding sites as defined by DamID (Fig 4E). We next examined the enrichment of the DamID peaks in the transcriptionally induced genes, ranked by absolute expression change in cic knockdown progenitor cells (see Materials and Methods). Cic binding peaks that were significantly reduced upon P.e. infection (< 2 fold change) were enriched in up-regulated genes from the RNA-Seq dataset (Fig 4F). Hence, the set of genes present in the overlapping set are likely to be direct target genes of Cic. Many cell cycle regulators and genes involved in DNA replication were upregulated in Cic-depleted progenitor cells (Fig 4D). In addition, a large portion of cell cycle control genes that were upregulated upon cic-RNAi, including string (stg, Cdc25) and Cyclin E (CycE), had Cic binding sites (Fig 4D, 4G, and 4H). To further assess the reliability of this approach we examined the occupancy of Cic on its previously characterized direct target gene-aos [39]. Our DamID-Seq data showed that aos contained two Cic binding sites within its enhancer, and that their occupancy was significantly reduced after P.e. infection (S4B Fig). The significant induction of aos transcription was verified both by RNA-Seq and qRT-PCR data from FACS-sorted progenitor cells expressing cic-RNAi (S4C Fig). Having confirmed the reliability of our approach for identifying genes that are repressed by Cic in ISCs, we focused on genes likely to contribute to ISC proliferation. We were interested in stg and CycE because they are transcriptionally induced in proliferating ISCs [42], required for ISC divisions, and also sufficient to induce sustained ISC division when co-overexpresssed [42]. To further test whether Cic regulates the transcription of stg and CycE we measured their normalized expression ratios in gain- or loss-of-function Cic midguts via RT-qPCR (Fig 4I–4K). The stg and CycE mRNAs were significantly increased in Cic-depleted midguts, and decreased in midguts expressing the dominant active CicΔC2. Strong inductions of stg and CycE were also observed in Cic-depleted progenitor cells or ISCs purified using FACS (Fig 4J). Moreover, both the stg and CycE loci had multiple strong Cic-Dam-ID binding peaks containing TGAATG(G/A)A motifs, and binding these peaks were reduced by P.e. infection (Fig 4G and 4H). Consistently, the induction of stg and CycE transcription upon P.e. infection was significantly repressed by CicΔC2 overexpression (Fig 4K). These data support the notion that Cic controls ISC cell cycle progression by directly repressing transcription of stg and CycE via binding sites in their regulatory regions. It has been suggested that Cic might regulate the transcription of certain members in a subfamily of ETS transcription factors [29,31]. Consistent with this, we identified the Drosophila ETS transcription factors pnt and Ets21C as potential Cic direct target genes by both RNA-Seq and DamID-Seq (Figs 5 and S5). These genes contain Cic binding sites, were highly expressed in midgut progenitor cells, and were significantly induced upon infection or cic depletion or mutation. Notably, induction of pnt and Ets21C was detected in FACS-sorted ISCs depleted of Cic (Fig 5B). Moreover, the induction of pnt and Ets21C expression by P.e. infection was suppressed when the dominant active form, CicΔC2 was overexpressed (Fig 5D). Similar effects were observed when Cic was either depleted or overexpressed in whole midgut samples (Figs 5C and S5C). These data suggest that Cic also regulates pnt and Ets21C transcription in Drosophila midgut ISCs, by directly binding to these loci. As in the case of stg and CycE, this regulation appeared to be modulated by P.e. infection, most likely in a MAPK-dependent manner. The HMG box of Human Cic binds to TGAATG(G/A)A octamers in vitro [31]. This motif was also verified as a Cic binding sequence in several Cic target genes in Drosophila embryos and wing discs [39]. Notably, the TGAATG(G/A)A motif was observed in 692/2279 Cic binding sites in our DamID-Seq dataset (p-value = 3.045475× 10−11). Each of the four Cic target genes discussed above contained more than one TGAATG(G/A)A motifs in its Cic binding sites. Moreover, TGAATGAA motifs found in the pnt locus also mapped to Cic binding sites that we determined from Drosophila embryo ChIP-Seq (Fig 5E). This suggests that Cic may bind to the pnt locus via TGAATGAA octamers, and that the occupancy of Cic at the pnt locus may also be conserved in different Drosophila cell types. To further evaluate this hypothesis we performed electrophoretic mobility shift assay (EMSA). Cic showed specific binding to two DNA fragments from the pnt locus that were identified as prominent in vivo Cic binding peaks by DamID-Seq and ChiP-Seq (Fig 5F and 5G). Importantly, the EMSA interaction was lost when the HMG-box in Cic was mutated, or when the TGAATGAA motifs were mutated. These data strongly support the idea that Cic directly regulates pnt transcription by directly binding to TGAATGAA motif in pnt locus. Pnt is believed to be a downstream effector of EGFR signaling in developing Drosophila eyes [43,44,45]. The pnt locus produces two alternative transcripts that encode two different protein isoforms: PNTP1 and PNTP2 [44]. PNTP1 was proposed to be a constitutive activator of transcription, whereas PNTP2 has a PNT (pointed) domain that was reported to be phosphorylated by MAP kinase in vitro [45]. The mutant protein, PNTP2T151A, which cannot be phosphorylated in vitro, was unable to rescue pnt phenotype in eyes but instead enhanced the mutant phenotype, suggesting that the PNT domain is an auto-inhibitory domain that can be inactivated by MAPK-dependent phosphorylation [45]. Furthermore PNTP2 is thought to induce transcription of PNTP1, which might thereby encode the final nuclear effector of the EGFR pathway in eye discs [43]. In the midgut, we found an interesting interaction between Pnt and Cic: pntP1 and pntP2 were both induced when Cic was depleted, and both decreased when Cic was overexpressed (Figs 5B, 5C, and S5A). The expression of transcripts encoding both isoforms was also increased in P.e. infected guts (Figs 5D and S5B). This raises the possibility that pnt might be an important downstream effector of Cic in controlling ISC proliferation. To test this we over-expressed either pntP1 or pntP2 in progenitor cells using the esgts or Dlts driver systems. After 4 days of transgene induction a dramatic increase in ISC division was evident in response to either pntP1 or pntP2 (Figs 6A–6B, 6I, and S6A–S6B). Conversely, mutant clones that were generated using a pnt null allele (pnt Δ88) [46] did not grow past the 2-cell stage (S6D Fig). Moreover, when we depleted pnt in progenitor cells by expressing a pnt-RNAi that recognizes both isoforms, or generated homozygous pnt null mutant ISCs via MARCM, ISC proliferation after P.e. infection was suppressed (Figs 6C–6D, 6J, and S6E). Next, we investigated the functional significance of the inhibition of pnt expression by Cic. Whereas loss of Cic function induced massive ISC proliferation, inhibiting both isoforms of pnt in this context suppressed this over-proliferation (Figs 6G–6H, 6K and S6F–S6G). Conversely, when we over-expressed either pntP1 or pntP2 in ISCs that also overexpressed CicΔC2, the inhibitory effect of CicΔC2 on proliferation was bypassed and the cells divided (Figs 6F, 6L and S6C). Hence, a significant fraction of the ISC over proliferation caused by Cic knockdown can be attributed to Cic’s effects on pntP1 and pntP2 Interestingly, mutant clones generated using a pntP1 specific mutant allele, pnt Δ33 [45,47], or a pntP2 specific mutant allele, pnt Δ78 [45,47], grew normally. However ISCs mutant for the pnt null allele pnt Δ88 did not expand (S6D Fig). In addition, pnt Δ33 and pnt Δ78 homozygous clones in which cic was depleted by RNAi had similar numbers of cells to cic-depleted control clones (i.e. they overgrew), whereas pnt Δ88 null mutant clones contained significantly fewer cells (S6G Fig). These data not only support our conclusion that pnt is required for ISC proliferation as a target of Cic, but show that PNTP1 and PNTP2 have redundant function in regulating ISC proliferation. Furthermore, pntP2 homozygous mutant ISCs did not appear to have any defect in proliferation upon P.e. infection (S6E Fig). Overall these results indicate that pntP2, the isoform proposed to be activated directly by MAKP phosphorylation [45], is not specifically required in ISC proliferation. Pnt is the Drosophila ortholog of the human ETS2 transcription factor and has a conserved ETS-type DNA binding domain, while Ets21C is the Drosophila ortholog of the human proto-oncogene ERG. In addition to having Cic binding sites, RT-PCR and RNA-Seq data showed that Ets21C was highly induced upon P.e. infection (Figs 5A and S5C). Moreover RNAi mediated depletion experiments indicated that Ets21C was also required for ISC proliferation in response to P.e. infection (Fig 6J). Over-expression of Ets21C caused a strong increase of ISC division (Fig 6E and 6I) suggesting that transcriptional induction of Ets21C could promote ISC proliferation. Furthermore, ectopic expression of Ets21C in progenitor cells could bypass the strong growth-suppressive effect of depleting MEK (Fig 6M). These data indicated that Cic controls ISC proliferation in part by regulating Ets21C transcription. Finally, we tested whether Yan, an inhibitory ETS type transcription factor, reported to be MAPK responsive and to compete with Pointed for binding to common sites on the DNA [45,48,49], had an opposite function in ISCs. Although yan mRNA is expressed in the midgut (Fig 5A), yan depletion from ISCs did not produce a detectable effect (S6F Fig). Two independent yan-RNAi lines were used, both of which were proven to be effective by qRT-PCR (S6H Fig). In summary these observations suggest that EGFR signaling controls ISC growth and division by regulating the activity of Cic, Pnt and Ets21C but not Yan, and that Cic directly represses pntP1, pntP2 and Ets21C in this context. It is well established that EGFR signaling is essential to drive ISC growth and division in the fly midgut [10,11,12]. However, the precise mechanism via which this signal transduction pathway activates ISCs has remained a matter of inference from experiments with other cell types. Moreover, despite a vast literature on the pathway and ubiquitous coverage in molecular biology textbooks, the mechanisms of action of the pathway downstream of the MAPK are not well understood for any cell type. From this study, we propose a model summarized in Fig 7. Multiple EGFR ligands and Rhomboid proteases are induced in the midgut upon epithelial damage, which results in the activation of the EGFR, RAS, RAF, MEK, and MAPK in ISCs. MAPK phosphorylates Cic in the nucleus, which causes it to dissociate from regulatory sites on its target genes and also translocate to the cytoplasm. This allows the de-repression of target genes, which may then be activated for transcription by factors already present in the ISCs. The critical Cic target genes we identified include the cell cycle regulators stg (Cdc25) and Cyclin E, which in combination are sufficient to drive dormant ISCs through S and M phases, and pnt and Ets21C, ETS-type transcriptional activators that are required and sufficient for ISC activation. Although we found more than 2000 Cic binding sites in the ISC genome, not all of the genes associated with these sites were significantly upregulated, as assayed by RNA-Seq, upon Cic depletion. One possible explanation for this is that Cic binding sites from DamID-Seq were also associated with other types of transcription units (miRNAs, snRNAs, tRNAs, rRNAs, lncRNAs) that were not scored for activation by our RNA-Seq analysis. Indeed a survey of the Cic binding site distributions suggests this (S5 Table). This might classify some binding sites as non-mRNA-associated. However, it is also possible that many Cic target genes may require activating transcription factors that are not expressed in ISCs. Such genes might not be strongly de-repressed in the gut upon Cic depletion. In other Drosophila cells MAPK phosphorylation is thought to directly inactivate the ETS domain repressor Yan, and to directly activate the ETS domain transcriptional activator Pointed P2 (PNTP2) [45,50]. In fact Pnt and Yan have been shown to compete for common DNA binding sites on their target genes [45,48,49]. Thus, previous studies proposed a model of transcriptional control by MAPK based solely on post-translational control of the activity of these ETS factors. However, we found that Pnt and Ets21C were transcriptionally induced by MAPK signaling, and could activate ISCs if overexpressed, and that depleting yan or pntP2 had no detectable proliferation phenotype. In addition, overexpression of PNTP2 was sufficient to trigger ISC proliferation, suggesting either that basal MAPK activity is sufficient for its post-translational activation, or that PNTP2 phosphorylation is not obligatory for activity. Moreover, pntP2 loss of function mutant ISC clones had no deficiency in growth (S6D Fig) even after inducing proliferation by P.e. infection, which increases MAPK signaling (S6E Fig). These observations indicate that the direct MAPK→PNTP2 phospho-activation pathway is not uniquely or specifically required for ISC proliferation. Or results suggest instead that transcriptional activation of pnt and Ets21c via MAPK-dependent loss of Cic-mediated repression is the predominant mode of downstream regulation by MAPK in midgut ISCs. In addition to activating ISCs for division, EGFR signaling activates them for growth. Previous studies showed loss of EGFR signaling prevented ISC growth and division, and that ectopic RasV12 expression could accelerate the growth not only of ISCs but also post-mitotic enteroblasts [11]. Similarly, our study shows that loss of cic caused ISC clones to grow faster than controls, by increasing cell number as well as cell size (Figs 1H and S2C–S2H). For instance, increased size of GFP+ ISCs and EBs was observed when cic-RNAi was induced by the esgts or esgtsF/O systems (Figs 1B, 6G and S2B). Therefore, in our search for Cic target genes we specifically checked probable growth regulatory genes such as Myc, Cyclin D, the Insulin/TOR components InR, PI3K, S6K and Rheb, Hpo pathway components, and the loci encoding rRNA, tRNAs and snRNAs. We found that Cic bound to the InR, Akt1, Rheb, Src42A and Yki loci. However, of these only InR mRNA was significantly upregulated in Cic-depleted progenitor cells (S4 Table). In surveying the non-protein coding genome, we found that Cic had binding sites in many loci encoding tRNA, snRNA, snoRNA and other non-coding RNAs (S5 Table), though not in the 28S rRNA or 5S rRNA genes (S4 Table). Due to the method we used for RNA-Seq library production, our RNA expression profiling experiments could not detect expression of these loci, and so it remains to be tested whether Cic may regulate some of those non-coding RNA’s transcription to control cell growth. It is also possible that Cic controls cell growth regulatory target genes indirectly, for instance via its targets Ets21C and Pnt, which are also strong growth promoters in the midgut (Figs 6A–6B, 6E and S6A–S6B). But given that no conclusive model can be drawn from our data regarding how Cic restrains growth, one must consider the possibility that ERK signaling stimulates cell growth via non-transcriptional mechanisms, as proposed by several studies [51,52,53,54]. The critical role of Cic as a negative regulator of cell proliferation in the fly midgut is consistent with its tumor suppressor function in mammalian cancer development (S1 Fig). Also consistent with our findings are the observations that the ETS transcription factors ETV1 and ETV5 are upregulated in sarcomas that express CIC-DUX, an oncogenic fusion protein that functions as a transcriptional activator [31], and that knockdown of CIC induces the transcription of ETV1, ETV4 and ETV5 in melanoma cells [29]. Moreover the transcriptional regulation by ETS transcription factors is important in human cancer development (S7 Fig). Their expression is induced in many tumors and cancer cell lines. For example, ERG, ETV1, and ETV4 can be upregulated in prostrate cancers [55], and ETV1 is upregulated in post gastrointestinal stromal tumors [56] and in more than 40% of melanomas [57]. In addition, the mRNA expression of these ETS genes was correlated with ERK activity in melanoma and colon cancer cell lines with activating mutations in BRAF (V600E), such that their expression decreased upon MEK inhibitor treatment [58]. Furthermore, overexpression of the oncogenic ETS proteins ERG or ETV1 in normal prostate cells can activate a Ras/MAPK-dependent gene expression program in the absence of ERK activation [59]. These cancer studies imply that there is an unknown factor that links Ras/Mapk activity to the expression of ETS factors, and that some of the human ETS factors might act without MAPK phosphorylation, as does Drosophila PntP1. Combining our knowledge of Cic with what was previously known about CIC in tumor development, we propose that CIC is the unknown factor that regulates ETS transcription factors in Ras/MAKP-activated human tumors. In summary, our study has elucidated a mechanism wherein Cic controls the expression of the cell cycle regulators stg (Cdc25) and Cyclin E, along with the Ets transcription factor Pnt, and perhaps also Ets21C, by directly binding to regulatory sites in their promoters and introns. Using genetic tests we show that these interactions are meaningful for regulating stem cell proliferation. Therefore, we suggest that human CIC may also lead to the transcriptional induction of cell cycle genes and ETS transcription factors in RAS/MAPK activated- or loss-of-function-CIC tumors such as brain or colorectal cancers. esgts: esg-Gal4/Cyo; tubGal80ts UAS-GFP/TM6B [60] esgts F/O: esg-Gal4 tubGal80ts UAS-GFP/Cyo;UASflp>CD2>Gal4/TM6B [8] Tubts: tub-Gal4; tubGal80ts/TM3, ser [61](provided from Valeria Cavaliere lab) esgts; Su(H)-Gal80: esg-Gal4-UAS-2XEYFP; Su(H)GBE- Gal80, tub-Gal80ts (Gift from Steven Hou’s lab) UAS-λTOP/FM7 [37] UAS-RASv12s35 [38] UAS-Ras RNAi [11] UAS-Egfr RNAi [11] UAS-cic-RNAi/Cyo (VDRC KK103805) UAS-cic-RNAi/Cyo (VDRC KK103012) UAS-pnt.P1 (Bloomington Drosophila Stock Center 869) UAS-pnt.P2 (Bloomington Drosophila Stock Center 399) UAS-pnt-RNAi (Bloomington Drosophila Stock Center 31936) UAS-pnt-RNAi (Bloomington Drosophila Stock Center 35808) UAS-yan-RNAi (Bloomington Drosophila Stock Center 26759) UAS-yan-RNAi (Bloomington Drosophila Stock Center 34909) UAS-yan-RNAi (Bloomington Drosophila Stock Center 35404) UAS-Ets21C-RNAi (VDRC KK103211) FRT82B cicfetu6 / TM3, Sb, Se (gift from Jimenez lab, Barcelona) w; cicfetT6 /TM3, Ser (gift from Nilson lab, Canada) w; cicfetE11 / TM6b (gift from Nilson lab, Canada) w; +; UAS-cic-HA w; UAS-cic-HA; + w; +;UAS-cic ΔC2-HA w; UAS-cic ΔC2-HA; + FRT82B pnt Δ33 [45,47] (gift from Joseph Bateman lab, Wolfson Centre for Age-Related Diseases) FRT82B pnt Δ78 [45,47] (gift from Joseph Bateman lab, Wolfson Centre for Age-Related Diseases) FRT82B pnt Δ88[45,47] (gift from Joseph Bateman lab, Wolfson Centre for Age-Related Diseases) The cic ΔC2 was amplified from the pCasper4—cic ΔC2 plasmid. The cic or cic ΔC2 cDNAs were inserted into pUASg-attB-HA [62] vector and used to generate transgenic flies. To generate UAS-cicDam transgenic flies, Cic was amplified from a cDNA library prepared from midgut. This cic cDNA was inserted into the pUASTattB-LT3-NDam plasmid (from Andrea brand lab), and transgenics were produced. Ectopic expression of transgenes in the adult midgut was achieved using the temperature sensitive inducible UAS-Gal4 system [63], TARGET. Crosses were set up and maintained at 18°C, the permissive temperature. 3–7 day old flies were shifted to 29°C for different times as indicated. Gut infections were performed by feeding flies live P.e. in 5% sucrose on Whatman filter paper and yeast paste at 29°C. The MARCM system was used to generate ISC clones. In order to reduce heat shock dependent stress, the clones were induced by heat shocking 3–5 days old flies at 34°C for 20 minutes. The heat shocked flies were then kept at 25°C. Clone size was measured after 10, 20, 30 days of clone induction. The size of the clones was quantified by Fiji software measuring GFP+ area from z-projected confocal microscopy images. Female adult flies were dissected in 1×PBS. Midguts were fixed in 1×PBS with 4% paraformaldehyde for 30 minutes at room temperature. Samples were washed in PBS with 0.1% X-100 (PBST) for 3x10 minutes each. Then the tissues were blocked in PBS with 0.1% X-100, 2.5%BSA, 10% NGS for at least 30 min at room temperature. All samples were incubated with primary antibody overnight at the following dilutions: rat anti-HA (1:200; Roche), guinea pig anti-Cic (1:1000, generated by author), rabbit anti-PH3 (1:1000, Millipore). After washing 3 times 10 minutes each in PBST, samples were incubated with secondary antibodies for at least 2 hours at room temperature at a dilution of 1:1000. DNA was visualized with DAPI (0.1mg/ml, Sigma), diluted 1:200. Images of Figs 1A–1B and 2E–2H were acquired by Delta vision microscope and the rest of the fluorescence images were taken by Leica SP5 confocal microscope. Images were then processed using Fiiji and Adobe Photoshop software. RNA was extracted from 10–12 female midguts using the RNAeasy kit (QIAGEN). RNA isolation from sorted cells was performed as previously described [64] and 100ng RNA (non-amplifed) used for reverse transcription. cDNA was synthesized by QuantiTect reverse transcription kit (QIAGEN). RT-qPCR was performed on a Light Cycler 480 II using SYBR Green I (Roche). Experiments were performed in biological triplicate. Relative fold differences in expression level of target genes were calculated as ratios to the mean of the reference genes rp49 [65] and tubulin [23]. Primer sequences are given in Supplementary Material and Methods. RNA isolation and amplification from sorted cells was performed as previously described [64]. Four independent biological replicates were used for sequencing. Raw reads were checked for quality using Fastqc and subsequently aligned using Tophat2, version 2.0.9, against the Flybase genome version 6. Mapped reads were counted using HTSeq-count version 0.5.4p5 [66] with mode „union“. Genes showing a cpm value below 1 in four samples per treatment were considered as poorly expressed and filtered out before conducting differential expression analysis using edgeR, version 3.2.4 [67]. Since our replicates were generated independently, we used a paired design and corrected the resulting p-values by the Benjamini-Hochberg method [68]. Subsequently, genes with a fold change of 1.5 and an adjusted p-value lower than 0.1 were considered as significantly deregulated. Rp49 –Forward: TCGATATGCTAAGCTGTC Rp49 –Reverse: GGCATCAGATACTGTCCCTTG β-tubulin-Forward: ACATCCCGCCCCGTGGTC β-tubulin-Reverse: AGAAAGCCTTGCGCCTGAACATAG pnt-Forward: ACGCCCTATGATGCTCAATC pnt-Reverse: TATCCAGACCCAAGGTGCTC pntP1-Forward: CGACTGCGAACAATCTGGT pntP1-Reverse: TTGCTGGTGTTGTAGCCTGT pntP2-Forward: TTAGCCAATTGAACGGCATC pntP2-Reverse: GCACAGATCCTTGCATCCAT Ets21C-Forward: CCGGGCACTCAGGTACTACT Ets21C-Reverse: CATACTGGAGGCCGGATCT aos-Forward: AGAACCCATGGCTTACATGC aos-Reverse: CGTCGCGGGTGTTAAGTTAC yan-Forward: CTGCTGGTCATCGTGCTTAG yan-Reverse: GACCTCAGTGTGAGCAGCAA stg-Forward: CAGCATGGATTGCAATATCAGTA stg-Reverse: CAACGTCGTCGTCGTAGAAC CycE-Forward: ACAAATTTGGCCTGGGACTA CycE-Reverse: GGCCATAAGCACTTCGTC

10.1371/journal.ppat.1006786

eCD4-Ig promotes ADCC activity of sera from HIV-1-infected patients

Antibody-dependent cell-mediated cytotoxity (ADCC) can eliminate HIV-1 infected cells, and may help reduce the reservoir of latent virus in infected patients. Sera of HIV-1 positive individuals include a number of antibodies that recognize epitopes usually occluded on HIV-1 envelope glycoprotein (Env) trimers. We have recently described eCD4-Ig, a potent and exceptionally broad inhibitor of HIV-1 entry that can be used to protect rhesus macaques from multiple high-dose challenges with simian-human immunodeficiency virus AD8 (SHIV-AD8). Here we show that eCD4-Ig bearing an IgG1 Fc domain (eCD4-IgG1) can mediate efficient ADCC activity against HIV-1 isolates with differing tropisms, and that it does so at least 10-fold more efficiently than CD4-Ig, even when more CD4-Ig molecules bound cell surface-expressed Env. An ADCC-inactive IgG2 form of eCD4-Ig (eCD4-IgG2) exposes V3-loop and CD4-induced epitopes on cell-expressed trimers, and renders HIV-1-infected cells susceptible to ADCC mediated by antibodies of these classes. Moreover, eCD4-IgG2, but not IgG2 forms of the broadly neutralizing antibodies VRC01 and 10–1074, enhances the ADCC activities of serum antibodies from patients by 100-fold, and significantly enhanced killing of two latently infected T-cell lines reactivated by vorinostat or TNFα. Thus eCD4-Ig is qualitatively different from CD4-Ig or neutralizing antibodies in its ability to mediate ADCC, and it may be uniquely useful in treating HIV-1 infection or reducing the reservoir of latently infected cells.

Antibodies can bind HIV-1-infected cells by recognizing the viral envelope glycoprotein (Env) on the cell surface. Antibody-bound cells then recruit natural killer cells to eliminate these infected cells. Here we demonstrate the unique properties of eCD4-Ig, a potent and exceptionally broad antibody-like HIV-1 entry inhibitor. Like antibodies, eCD4-Ig can efficiently mediate killing of infected cells. However, unlike most antibodies, eCD4-Ig can promote Env conformational changes that then allow abundant but otherwise inert antibodies in patient sera to mediate killing of infected cells. This property may be especially useful in efforts to cure HIV-1 by reactivating virus in latently infected cells and then killing these virus-producing cells.

Natural killer (NK) cells and other Fc-gamma receptor (FcγR)-expressing cells can eliminate HIV-1 infected cells through antibody-dependent cell-mediated cytotoxicity (ADCC) [1–3]. These ADCC activities depend on the antibody isotype [4,5]. For example, IgG1 and IgG3 mediate efficient ADCC, whereas the IgG2 and IgG4 do so poorly. Growing evidence suggests that ADCC is an important component of protective immune responses against HIV-1 infection [1,2,6–8]. High levels of ADCC antibodies have been correlated with better disease prognosis and slower disease progression in HIV-infected individuals [1,3,9]. Further, analysis of the only human vaccine trial reporting some protection, RV144, identified ADCC-mediating antibodies as a correlate of protection [2,6]. Studies of passively administered antibodies in animal models have suggested that ADCC contributes to protection from simian-human immunodeficiency viruses (SHIV) [10]. The ability of some antibody-treated macaques to maintain low viral loads after cessation of treatment suggests that, in some contexts, antibodies can help eliminate infected cells and boost host immune responses so they can subsequently control viral replication [11]. Efforts are now underway to reduce the viral reservoir by combining latency-reversing agents, which stimulate virion production in latently infected cells, with antibodies that may accelerate elimination of these cells [12,13]. Engineered biologics with IgG1 Fc domains can also mediate ADCC [14]. We have developed one such inhibitor, eCD4-Ig, that is as potent as most HIV-1 broadly neutralizing antibodies (bNAbs), but which uniquely neutralizes 100% of tested HIV-1, HIV-2, and SIV isolates, all with 80% inhibitory concentrations (IC80s) less than 10 μg/ml [15]. eCD4-Ig is a fusion of CD4-Ig with a tyrosine-sulfated coreceptor-mimetic peptide appended to its carboxy-terminus [15,16]. The two sulfopeptides of the eCD4-Ig dimer and at least one of its CD4 domains engage Env to provide high avidity binding and prevent enhancement of infection observed with CD4-Ig. Its close emulation of the HIV-1 receptor CD4, and the HIV-1 coreceptors CCR5 and CXCR4, likely account for its exceptional breadth. The breadth and potency of eCD4-Ig has also been demonstrated in vivo: In macaque studies, adeno-associated virus (AAV) vector-delivered eCD4-Ig achieved complete and significant protection against multiple high-dose challenges with either SHIV-AD8 or SIVmac239 [15]. These in vivo studies utilized an eCD4-Ig variant with rhesus CD4 and IgG2 Fc domains, indicating that the ADCC activities of IgG1 may not be essential for prophylaxis. However, eCD4-Ig may also be useful in controlling an established infection or reducing the viral reservoir, where ADCC activities are likely more critical. We therefore investigated its ability to mediate ADCC alone and in combination with antibodies or patient sera. We observed that, despite lower occupancy of Env, eCD4-Ig mediated markedly more efficient ADCC than did CD4-Ig. Further, V3-loop and CD4-induced (CD4i) antibodies mediated more efficient ADCC in the presence of an IgG2 form of eCD4-Ig because eCD4-Ig promotes exposure of these epitopes. Most critically, eCD4-IgG2, but not bNAbs, markedly enhanced the ADCC activities of sera from six patients where weak ADCC activity was detected. Thus, unlike bNAbs, eCD4-Ig uniquely works with the host immune system to eliminate infected cells. If eCD4-Ig is to be used as part of a treatment or cure regimen, its ADCC activity is likely to be important. We therefore compared the ADCC activities of CD4-Ig and eCD4-Ig with IgG1 Fc domains (CD4-IgG1, eCD4-IgG1), and eCD4-Ig with an IgG2 Fc domain (eCD4-IgG2). As we have previously reported, the presence of the carboxy-terminal sulfopeptide did not interfere with the ADCC activity of eCD4-IgG1, rather it enhanced it [15]. Specifically, target cells infected with a CCR5-using (R5) isolate (YU2), a CXCR4-using (X4) isolate (NL4-3), or a dual-tropic (R5X4) isolate (89.6) were more efficiently lysed by CD16a+ NK cells when incubated with eCD4-IgG1 than with CD4-IgG1 (Fig 1A). Expectedly, eCD4-IgG2 did not efficiently mediate ADCC. The greater ADCC activity of eCD4-IgG1 relative to CD4-IgG1 is somewhat surprising, because cells expressing YU2 or BG505 Env bound fewer or similar numbers of eCD4-Ig relative to CD4-Ig (Fig 1B and 1C). These observations suggest that eCD4-IgG1 presents its Fc domains in an orientation more favorable to ADCC than does CD4-IgG1. eCD4-Ig binds both the receptor- and coreceptor-binding sites on HIV Env [15], and induces conformational changes that may affect access to antibody epitopes. We first investigated the ability of eCD4-Ig bearing a murine Fc domain (eCD4-mIg) to alter binding of several classes of HIV-1 neutralizing antibodies. The murine Fc domain allowed us to distinguish between binding of eCD4-Ig and HIV-1 antibodies with human Fc domains. HEK293T cells expressing BG505 Env lacking the cytoplasmic tail (BG505ΔCt) were pre-incubated with eCD4-mIg and then the resulting complexes were analyzed by flow cytometry for their ability to bind human antibodies. We observed that eCD4-mIg markedly increased binding of the CD4i antibodies 17b, E51, and A32 despite the presence of a potentially competing sulfopeptide (Fig 2A). However, as expected, CD4-Ig did so more efficiently (S1A Fig). Similarly, the epitope of the V3-loop antibody 447-52D was exposed by both CD4-mIg and eCD4-mIg, whereas the nearby epitope of the antibody F425-B4e8 was not further exposed (Fig 2B and S1B Fig). eCD4-mIg prevented binding of CD4-binding site antibodies (VRC01, 3BN117), V2 glycan apex antibodies (PGDM1400, PGT145), and antibodies recognizing the interface of Env subunits gp120 and gp41 (PGT151, 35O22) (Fig 2C–2E). In contrast, it did not affect binding of the V3-glycan antibodies (10–1074, PGT128) or an MPER class antibody (10E8) (Fig 2F and 2G) [17]. Our data suggest that eCD4-Ig might work synergistically with CD4i and V3-loop antibodies, but may antagonize CD4-binding site, apex, and interface antibodies. The ability of eCD4-Ig to expose the epitopes of V3-loop and CD4i antibodies is of interest because these antibodies are abundant but generally non-neutralizing in infected individuals [18,19]. We therefore investigated whether eCD4-IgG2, which does not mediate ADCC, could promote the ADCC activities of antibodies from these classes. We used ADCC-inactive eCD4-IgG2 so that we could monitor the effects of eCD4-Ig on ADCC mediated solely by the HIV-1 antibodies. eCD4-IgG2 markedly enhanced ADCC mediated by the V3-loop antibodies 447-52D (Fig 3A) and F425-B4e8 (Fig 3B) in cells infected with the HIV-1 isolates 89.6 and YU2, and to a lesser extent with NL4-3. It also promoted ADCC activity of the CD4i antibodies 17b (Fig 3C) and A32 (Fig 3D), whereas it attenuated ADCC mediated by the CD4-binding site antibody VRC01 (Fig 3E). eCD4-IgG2 did not alter ADCC mediated by 2G4, a control antibody recognizing Ebola virus GP1,2 (Fig 3F). We also analyzed the ability of these antibodies to neutralize the same viruses (S2 Fig). Notably, eCD4-IgG2 could enhance the ADCC activity of antibodies that were non-neutralizing. For example, neither antibody recognizing the V3-loop of Env (447-52D and F425-B4e8) neutralized the YU2 isolate, but in the presence of eCD4-IgG2, both could mediate ADCC. Thus eCD4-IgG2, which does not itself mediate ADCC, nonetheless can enhance the ADCC activity of non-neutralizing antibodies. The ability of eCD4-IgG2 to enhance the ADCC activity of V3-loop and CD4i antibodies raises the possibility that it may similarly promote ADCC in serum from HIV-1-infected patients, where these antibodies are abundant. To test this possibility, we initially screened de-identified sera from 15 patients for their ability to mediate ADCC against the YU2 isolate, and to neutralize the same isolate. Six of these sera mediated ADCC against YU2 but only one efficiently neutralized the same virus, with 50% neutralization at a 1:824 dilution (S3 Fig). We further evaluated ADCC activity of the six ADCC-active sera, as well as one ADCC-inactive serum (9121652) and serum from an HIV-1-negative individual. ADCC activity was measured in the presence or absence of eCD4-IgG2, or IgG2 forms of the bNAbs VRC01 and 10–1074 (Fig 4A). In all six cases, eCD4-IgG2 markedly promoted the ADCC activity of these patient sera, typically by 100-fold. eCD4-IgG2 did not alter the ADCC activity of HIV-1-negative serum or of HIV-1-positive but ADCC-inactive serum. In contrast, IgG2 forms of VRC01 and 10–1074 which, like eCD4-IgG2, do not mediate ADCC (Fig 4B), had no effect on the ADCC activity of patient sera. ADCC activity in the presence of eCD4-IgG2 was significantly greater than with sera alone (p = 0.0001), or in the presence of VRC01 or 10–1074 (both p = 0.0001; Fig 4C). Thus eCD4-IgG2, but not IgG2 forms of broadly neutralizing antibodies, can dramatically enhance the ADCC activities of sera from infected individuals. We additionally tested the effect of serum in combination with eCD4-IgG1 or CD4-IgG2 (Fig 4D). As expected, CD4-IgG2 synergized with serum, but not as efficiently as eCD4-IgG2. Also expectedly, eCD4-IgG1 combined with serum to afford highly potent ADCC activity, likely a combination of potent ADCC-activity of 1 μg/ml of eCD4-IgG1 itself (see Fig 4B) and its ability to synergize with patient serum. We conclude that patient serum is likely to improve the already potent ADCC activity of eCD4-IgG1. Collectively the data of Figs 2–4 show that, among potent HIV-1 entry inhibitors, eCD4-Ig can uniquely unmask epitopes of antibodies present in the sera of HIV-1-infected patients, likely because it alters the conformation of Env on the cell surface. One potential application of eCD4-Ig as a therapeutic would be as part of a “shock and kill” approach to a sterilizing cure. This approach seeks to eliminate the latent viral reservoir by reactivating latent provirus using a latency reversing agent (LRA) in combination with an agent that accelerates killing of reactivated cells. eCD4-Ig might be uniquely useful in this regard as both the IgG1 and IgG2 forms can neutralize newly produced virions as well as target reactivated cells for killing by ADCC, either directly by eCD4-IgG1 or by sensitizing cell-surface expressed Env to other circulating antibodies. To assess the potential of eCD4-Ig in this context, we investigated its ability to direct killing of reactivated latently infected cell lines (Fig 5). OM-10.1 (Fig 5A and 5C) or ACH-2 (Fig 5B and 5D) cells were reactivated using either the cytokine TNFα or the HDAC inhibitor, vorinostat (suberoylanilide hydroxamic acid, SAHA) and killing was measured using a flow cytometry-based ADCC assay. In all cases, eCD4-IgG1 showed potent ADCC activity. Additionally, while HIV+ serum 9121658 had some ADCC activity alone, ADCC activity was increased by the addition of 1 μg/mL eCD4-IgG2. Increases in these activities were significant when both cell types were reactivated with TNFα, and when ACH-2 cells were reactivated with vorinostat. A similar trend was observed with OM-10.1 cells reactivated with vorinostat, but these changes were not significant. As we previously observed in the luciferase-based assay (Fig 4D), eCD4-IgG2 was more potent at promoting serum ADCC activity than was CD4-IgG2 and this difference was significant in both cell lines reactivated with TNFα. These results suggest that eCD4-Ig may make an especially potent ‘kill’ for a shock and kill approach to eliminating the reservoir of latently infected cells. eCD4-Ig is a potent and exceptionally broad inhibitor of HIV-1 entry [15]. Here we examined its ability to mediate ADCC in several contexts. We first compared it with CD4-Ig, using an R5, an X4, and an R5X4 isolate. In each case, eCD4-IgG1 more efficiently mediated cell killing relative to CD4-IgG1 (Fig 1A). This enhanced ADCC activity is consistent with its higher avidity binding to Env, but still somewhat surprising because fewer eCD4-Ig than CD4- molecules bound the Env trimer (Fig 1B). These observations suggest that eCD4-Ig binds Env in a different and perhaps more stable orientation than does CD4-Ig, likely because sulfopeptide-binding fixes the orientation of the Fc domain. Alternatively, eCD4-Ig may bridge two or more Env trimers, perhaps more efficiently crosslinking FcγRIIIa on NK cells. In either case, the manner in which eCD4-Ig mediates ADCC appears to be qualitatively different from CD4-Ig. We also identified an important difference between eCD4-Ig and broadly neutralizing antibodies. Specifically, eCD4-Ig, like soluble CD4 and less-potent CD4-mimetic compounds [20–24], induces conformational changes in Env (Fig 2) that promote ADCC mediated by V3-loop and CD4i antibodies (Fig 3). These antibodies are frequently found in HIV-1-positive individuals [18,19], although by themselves they do little to control an infection. Most importantly, we show that, in contrast to antibodies, eCD4-IgG2, an ADCC-inactive form of eCD4-Ig, increased the ADCC activity of serum from infected patients much more effectively than any previously described agent (Figs 4D and 5), presumably by enlisting the help of antibodies that would not otherwise bind Env. Of note, eCD4-IgG2 did so more effectively than CD4-IgG2 (Figs 4D and 5), perhaps because the coreceptor-mimetic sulfopeptide contributes to the stability of Env in its CD4-bound conformation. The ability of eCD4-Ig to synergize with serum antibodies is potentially important if eCD4-Ig is used to control an established HIV-1 infection. Our data further suggest that, despite its lack of direct effector functions, eCD4-IgG2 can mediate these activities should the safety of eCD4-IgG1 be a concern. However, as we show in Fig 4D, ADCC-active eCD4-IgG1 can similarly combine with these antibodies to mediate even more robust killing of infected cells. Indeed, given its exceptional breadth, its potent intrinsic ADCC activity, and its ability to synergize with otherwise weakly active endogenous antibodies to further mediate ADCC, eCD4-IgG1 is likely to be more effective in vivo than any bNAb at eliminating infected cells. It is also possible that therapeutic vaccines that actively raise V3-loop and CD4i antibodies, perhaps by locking soluble Env trimers in the CD4-bound state [25], may further improve eCD4-IgG1-mediated cell killing. One context in which the unique ADCC properties of eCD4-Ig might be especially useful is in efforts to reduce or eliminate the reservoir of latently infected cells in a “shock and kill” approach to cure HIV-1 [12,13]. This approach relies on an agent such as a histone deacetylase inhibitor, for example vorinostat, or a TLR7 agonist to reactivate the latent HIV-1 provirus in infected cells, which can then be eliminated through ADCC or other mechanisms. The properties of eCD4-Ig shown here suggest that it might be more effective for this purpose than other CD4-mimetic compounds or bNAbs. Specifically, we show that eCD4-IgG2 can synergize with patient serum to kill cells reactivated with either vorinostat or TNFα, as can eCD4-IgG1 alone (Fig 5). Unlike small CD4-mimetic peptides and small molecules, eCD4-Ig can also potently neutralize virus, its IgG1 form can mediate ADCC directly, and both IgG1 and IgG2 forms can boost the ADCC activities of patient sera more effectively than any other agent. eCD4-Ig is likely more efficient and consistent at inducing the CD4-bound conformation than these small compounds, likely accounting for the inability of at least one such compound to promote serum-mediated killing of reactivated ACH-2 cells [24]. eCD4-Ig has key advantages over bNAbs as part of a shock and kill strategy. In general bNAbs do not promote the ADCC activities of patient sera, and their more limited breadth would preclude killing of every reactivated cell. In summary, we have shown that eCD4-Ig is more effective than CD4-Ig at mediating ADCC, and that, unlike broadly neutralizing antibodies, it can dramatically enhance weak ADCC activity of sera from infected patients. These properties may be useful in efforts to eliminate the viral reservoir. HEK293T (ATCC, Manassas, VA) and TZM-bl cell lines were grown in DMEM supplemented with 10% fetal bovine serum. TZM-bl cells were obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH, contributed by Dr. John C. Kappes, Dr. Xiayun Wu and Tranzyme Inc [26–30]. Expi293F cells were grown in Expi293 Expression media (Life Technologies, Carlsbad, CA). ADCC target and effector cells have been described previously and were a generous gift from Drs. Michael Alpert and David Evans. Briefly, CEM.NKR-CCR5-LTR-Luc ADCC target cells, harboring an HIV-1 Tat-inducible luciferase, were derived from CEM.NKR-CCR5 CD4+ T Cells obtained from the NIH AIDS Research and Reference Reagent Program (ARRRP), Division of AIDS, NIAID, NIH, contributed by Dr. Alexandra Trkola and have been previously described [31–35]. Targets cells were grown in R10 media, specifically RPMI supplemented with 10% FBS, 25 mM HEPES, 2 mM L-glutamine, and 0.1 mg/mL Primocin (InvivoGen, San Diego, CA). KHYG-1 derived NK cell line expressing human CD16a (V158 variant) have been previously described [31,32]. NK cells were grown in R10 media additionally supplemented with 1 μg/mL cyclosporine, and interleukin-2 (IL-2). OM-10.1 and ACH-2 cells were obtained from the NIH AIDS reagent Program, Division of AIDS, NIAID, NIH: OM-10.1 Cells from Dr. Salvatore Butera [36–40], ACH-2 from Dr. Thomas Folks [41,42]. Both OM-10.1 and ACH-2 cells were grown in RPMI supplemented with 10% FBS, 25 mM HEPES, 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin. The variable heavy and light chains of 447-52D, F425-B4e8, 17b, E51, A32, PGDM1400, PGT145, PGT151, and 2G4 were cloned into human IgG1 expression vectors as previously described [43]. Vectors expressing VRC01 heavy and light chains were obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH, from Dr. John Mascola [44,45]. 3BNC117 and 10–1074 IgG1 expression plasmids were provided by Dr. Michel Nussensweig. 10E8 expression vectors were obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH, from Dr. Mark Connors [46]. PGT121, PGT128 were provided by Dr. Dennis Burton. 35O22 expression vectors were obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH: Cat# 12584 mAb 35O22 heavy chain expression vector (CMVR) and Cat# 12585 mAb 35O22 light chain, from Drs. Jinghe Huang and Mark Connors [47]. Heavy and light chains for 447-52D and F425-B4e8 were synthesized by Integrated DNA Technologies (IDT, Newark, NJ) and cloned into IgG1 expression vectors. Plasmids encoding IgG2 forms of VRC01 and 10–1074 were generated by replacing genes encoding human IgG1 constant regions with those of human IgG2. Expression vectors for the BG505 gp160-Δcytoplasmic tail were provided by Drs. John Moore and PJ Klasse. eCD4-IgG1 has been previously described [15,16]. Briefly, eCD4-Ig is an Fc-fusion protein of CD4 domains 1 and 2 with the addition of a sulfated CCR5-mimetic peptide at the C-terminus. A plasmid encoding eCD4-IgG2 was generated by replacing sequence encoding the human IgG1 Fc domain in the eCD4-IgG1 expression plasmid with that of IgG2. Antibodies and eCD4-Ig were produced in Expi293 cells (Life Technologies, Carlsbad, CA). Cells were grown to a density of 3x106 cells/mL prior to transfection with Expifectamine according to manufacturer’s instructions (Life Technologies, Carlsbad, CA). 140 μg total DNA was transfected in 250 mL Expi293 cells. eCD4-IgG1 and eCD4-IgG2 plasmids were cotransfected at an 4:1 ratio with plasmid encoding human tyrosine protein sulfotransferase 2 (TPST2). Antibodies were produced by transfection of two plasmids encoding heavy and light chain, respectively, at a 1:1 ratio. 20 hours post-transfection, Expifectamine enhancers were added according to manufacturer’s instructions. 5 days post-transfection, media was collected for protein purification. Debris was cleared by centrifugation for 10 min at 4000g and filtered using 0.45-μm filter flasks (Thermo Scientific, Waltham, MA). Antibodies and Fc-fusion proteins were purified from supernatants using HiTrap MabSelect SuRe columns (GE Healthcare Life Sciences, Pittsburgh, PA). After protein binding, columns were washed extensively with PBS before elution with IgG Elution Buffer (Thermo Scientific Waltham, MA). Eluate pH was immediately adjusted with Tris-HCl 1M pH 9.0 Neutralization Buffer (G-Biosciences, Saint Louis, MO). Buffer was exchanged with PBS and protein was concentrated to 1 mg/mL by Ultrafiltration (Amicon Ultra, Millipore Sigma, Billerica, MA) at 4000 g. De-identified HIV-1 positive serum was obtained from Boston Biomedical Inc. (BBI, Boston, MA), and has been previously described [48]. De-identified uninfected serum was purchased from Sigma (Saint Louis, MO). All materials were handled in accordance with the regulations set forth by the Scripps Office for the Protection of Research Subjects. Viruses were produced in HEK293T cells following transfection of p89.6, pNL4-3, and pYU-2 HIV-1 molecular clones using calcium phosphate. Supernatants were harvested 48 hours post-transfection, filtered through 0.45-μm filters, aliquoted and frozen at -80°C. p89.6 molecular clone was obtained from the ARRRP, Division of AIDS, NIAID, NIH, deposited by Dr. Ronald G. Collman, MD [49–51]. pNL4-3 molecular clone was obtained from the ARRRP, Division of AIDS, NIAID, NIH, from material deposited by Drs. Suzanne Gartner, Mikulas Popovic, Robert C. Gallo, and Malcom Martin [52]. The pYU-2 HIV-1 molecular clone was obtained through the NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH, from Dr. Beatrice Hahn and Dr. George Shaw [53,54]. Viruses titers were quantified by p24 ELISA (Advanced Bioscience Laboratories, Rockville, MD). ADCC assay was performed as previously described [31,32]. Briefly, CEM.NKR-CCR5-LTR-Luc target cells were infected by spinoculation with 89.6 (400ng p24) and NL4-3 (200ng p24) 4 days and YU-2 (500ng p24) 3 days prior to assay. Infection amounts for each assay were determined by virus titration on the target cells. On the day of the assay, infected target cells were mixed with NK effector cells at a 10:1 ratio in the presence of antibodies or eCD4-Ig. Cell and inhibitor mixes were incubated for 8 hours at 37°C. ADCC activity was measured by luciferase using BriteLite Plus (Perkin Elmer, Waltham, MA) and measured using a Victor X3 plate reader (Perkin Elmer, Waltham, MA). ADCC against latently infected OM-10.1 and ACH-2 cells was measured using a flow cytometry-based assay. Specifically, cells were reactivated with 10 ng/mL TNFα (Life Technologies, Carlsbad, CA) or 1 μM SAHA (Sigma-Aldrich, Saint Louis, MO) for 24 hours prior to assay. Reactivated cells were incubated with NK-V158 cells at a 10:1 ratio in the presence of eCD4-Ig, CD4-Ig, or antibodies at a concentration of 1 μg/mL or HIV-1+ serum at a 1:1000 dilution. Target cells were fixed and permeabilized using the Fix and Perm kit (Life Technologies, Carlsbad, CA) and stained for intracellular p24 expression using FITC-conjugated clone KC57 (Beckman Coulter, Brea, CA). Data were collected using Accuri C6 Flow Cytometer and data analyzed with the C6 Software (BD Biosciences, San Jose, CA). Percent ADCC was normalized to infected cells in the presence of effectors but no inhibitor, and calculated with the following formula: [(%p24+ cells in targets + effectors − inhibitor) − (%p24+ cells in targets + effectors + inhibitor)]/(%p24+ cells in targets + effectors − inhibitor) × 100. TZM-bl neutralization assays were performed as previously described [15,16,43]. Briefly, eCD4-Ig or antibody titrations were incubated with infectious viruses for 1 hour at 37°C. TZM-bl cells were diluted in DMEM to 100,000 cells/mL and added to the virus/inhibitor mix. Cells were then incubated for 40 hours at 37°C. Viral entry was determined by luciferase readout with BriteLite Plus (Perkin Elmer, Waltham, MA) and read on a Victor X3 plate reader (Perkin Elmer, Waltham, MA). HEK293T cells were transfected with plasmids expressing HIV-1 envelope glycoprotein variants lacking cytoplasmic residues 732–876 (HXBc2 numbering). Cells were collected 48 hours post transfection with non-enzymatic dissociation buffer (Sigma-Aldrich, Saint Louis, MO). Cells were washed with flow cytometry buffer (PBS with 2% goat serum, 0.01% sodium azide) before incubation with eCD4-mouse-Ig for 1 hour on ice. Cells were subsequently incubated with human IgG1-containing antibodies. Antibody binding was determined with FITC-conjugated goat anti-mouse, and APC-conjugated goat anti-human secondary antibodies (Jackson ImmunoResearch, West Grove, PA). Between each antibody incubation, cells were washed twice with flow cytometry buffer. After incubation with secondary antibody, cells were washed once with flow cytometry buffer, once with PBS, and then resuspended in 1% paraformaldehyde in PBS. Binding was analyzed with an Accuri C6 Flow Cytometer and data analyzed with the C6 Software (BD Biosciences, San Jose, CA).

10.1371/journal.pcbi.1004569

An Ovol2-Zeb1 Mutual Inhibitory Circuit Governs Bidirectional and Multi-step Transition between Epithelial and Mesenchymal States

Reversible epithelial-to-mesenchymal transition (EMT) is central to tissue development, epithelial stemness, and cancer metastasis. While many regulatory elements have been identified to induce EMT, the complex process underlying such cellular plasticity remains poorly understood. Utilizing a systems biology approach integrating modeling and experiments, we found multiple intermediate states contributing to EMT and that the robustness of the transitions is modulated by transcriptional factor Ovol2. In particular, we obtained evidence for a mutual inhibition relationship between Ovol2 and EMT inducer Zeb1, and observed that adding this regulation generates a novel four-state system consisting of two distinct intermediate phenotypes that differ in differentiation propensities and are favored in different environmental conditions. We identified epithelial cells that naturally exist in an intermediate state with bidirectional differentiation potential, and found the balance between EMT-promoting and -inhibiting factors to be critical in achieving and selecting between intermediate states. Our analysis suggests a new design principle in controlling cellular plasticity through multiple intermediate cell fates and underscores the critical involvement of Ovol2 and its associated molecular regulations.

Cumulative evidence reveals remarkable lineage plasticity of somatic cells. Epithelial-to-mesenchymal transition (EMT) represents a prime example of such plasticity where an epithelial cell is converted into a mesenchymal cell. This process is used in normal development to generate crucial cell types, and is hijacked by cancer cells for invasion and metastasis. Recent studies also suggest the importance of EMT in generating stem cell properties. The reversibility of EMT and its sensitivity to varying environmental stimuli pose interesting challenges to understand the intricate regulatory networks that direct cellular state transitions and their dynamics. Here we use a systems biology approach to probe into the complexity of the EMT process. We report a new molecular regulation that expands the known regulatory network, and show that this new network is capable of generating multiple intermediate states, which we provide experimental evidence for. We present modeling and experimental results to highlight the significance of a delicate balance between EMT-promoting and -inhibiting factors for achieving and/or selecting an intermediate state, and to suggest the biological significance of the multiple intermediate states. This work further elucidates the complex strategies that control epithelial cell behavior and cancer/stem cell plasticity.

Epithelial-to-mesenchymal transition (EMT) represents an extreme form of cellular plasticity where an epithelial cell is converted into a mesenchymal cell. Complete EMT is essential during embryogenesis to generate crucial developmental cell types [1], whereas partial EMT occurs in committed epithelial tissues with yet unknown functional significance [2]. Recently, EMT has been shown to promote stem cell properties, as differentiated epithelial cells that have undergone a round of EMT gain multipotency and self-renewal capability [3–5]. Furthermore, reversible EMT plays important roles in pathological processes such as cancer metastasis and wound healing. EMT endows cancer cells with the ability to migrate and invade adjacent tissues through changes in adhesion and behavior. Upon arrival to the destination site, EMTed cancer cells can revert to the epithelial phenotype via mesenchymal-to-epithelial transition (MET) to settle and differentiate into secondary tumors [1]. Previous studies have identified key transcription factors (TFs) and microRNAs (miRNAs) that are involved in the regulation of EMT. In particular, mutual inhibition loops formed between Zeb1 and miR-200 [6], and between Snail and miR-34a [7] are critical components in the regulatory network [8]. Mathematical modeling suggested that these mutual inhibition loops govern a tri-stable system, in which cells can be stabilized at an epithelial (E) state, a mesenchymal (M) state, or an intermediate state exhibiting expression of signature genes of both E and M in a variable fashion [9,10]. The intermediate state identified by these models is proposed to associate with cancer cells that exhibit collective migration during tumorigenesis [9], implicating the clinical relevance of the ternary switch in cell plasticity. In recent experimental studies, we showed that transcription factor Ovol2 restricts EMT by directly inhibiting EMT-inducing factors including Zeb1, and that these regulations are critical for proper morphogenesis and for maintaining epithelial lineages in mammary gland and skin epidermis [11,12]. However, the precise role of Ovol2 in the context of the well-studied core molecular network that controls EMT dynamics remains to be elucidated. In addition, it is unclear how EMT-inhibiting transcriptional factors like Ovol2 and EMT-promoting transcription factors like Zeb1 interact integratively to regulate the intermediate state. In this work, we first provide new experimental evidence suggesting a direct regulation of Ovol2 by Zeb1, which together with previous reports of Ovol2 inhibition of Zeb1 [11–13] demonstrates the existence of an Ovol2-Zeb1 mutual inhibition circuit. We then present a mathematical model that includes this new regulation, revealing two, rather than one, intermediate states with distinct propensities to differentiate into E and M states. We show that the Ovol2-Zeb1 mutual inhibition circuit is essential for the existence and robustness of both intermediate states in this model, and experimentally validate a specific prediction of the model, namely that Ovol2 is able to reprogram any given states to an E state. Furthermore, we describe experimental results suggesting that mammary epithelial cell line MCF10A represents one of the intermediate states that exhibit a bidirectional potential to differentiate into both E and M states. Together, our findings uncover a new layer of complexity of the dynamic, multi-step transitions between E and M states and unravel key regulatory mechanisms that control such transitions. Mutual inhibition loops between EMT-inducing TFs and miRNAs (e.g. Zeb1-miR200 and Snail-miR34a) are critical for robust control of EMT/MET [8]. Our previous studies showed that Zeb1 is directly inhibited by Ovol2 in mammary and skin epithelial cells [11,12]. Zeb1 and Ovol2 are expressed in a mutually exclusive pattern in clinical and cell line samples [11,13], raising the possibility that Zeb1 may also inhibit Ovol2 expression. Indeed, sequence analysis revealed the presence of two conserved Zeb1 binding consensus sequences in the human and mouse OVOL2/Ovol2 promoters, one near the transcriptional start site (-335 bp and -111 bp for human and mouse genes, respectively) and the other further upstream (-1546 bp and -1167 bp for human and mouse, respectively) (Fig 1A). Using chromatin immunoprecipitation (ChIP) assay, we detected Zeb1 binding to the downstream but not upstream site (Fig 1A). Furthermore, forced expression of Zeb1 in MCF10A human mammary epithelial cells significantly decreased OVOL2 expression at a transcriptional level, whereas Ovol2 overexpression led to reduced level of ZEB1 transcript as expected (Fig 1B). These results are consistent with direct repression of OVOL2/Ovol2 expression by Zeb1, and together with previously published data suggest the existence of an Ovol2-Zeb1 mutual inhibition loop. To dissect the role of the Ovol2-Zeb1 loop in EMT dynamics, we incorporated this regulation, as well as the negative regulation of TGF-β signaling by Ovol2 [11] into a framework that has been successfully used to formulate a 3-state EMT system [14]. The new model thus contains three mutual inhibition loops: Zeb1-miR200, Snail-miR34a and Ovol2-Zeb1 (Fig 2A). To examine how the system might be stabilized at various stages of EMT, we performed bifurcation analysis with respect to external TGF-β as an EMT inducer. Interestingly, four distinct stable steady states corresponding to four cell phenotypes emerged with the addition of the Ovol2-Zeb1 loop (Fig 2B). In particular, two intermediate states appeared between a terminal E state and an M state (Fig 2B). We named the intermediate state closer to the E state I1, and the one closer to the M state I2. The dynamic feature of the four-state system is consistent with the recently proposed sequential cell-state transition in which more than one intermediate states may exist [15], and it is also compatible with existing EMT models [9,10,14] in terms of possible ternary switch in the system (I1-I2-M, E-I1-I2, E-I1-M or E-I2-M, depending on specific external stimulations). Importantly, our model predicted that elevated production of Ovol2 is able to reprogram all other states to the terminal E state (Fig 2C). We will return to this notion later. Previous work stipulates that unstimulated MCF10A cells are in an epithelial state and when stimulated by increasing concentrations of TGF-β they transition into first an intermediate (partial EMT) state and subsequently an M state [14]. However, by comparing the expression of epithelial (E-cadherin or Ecad) and mesenchymal (vimentin or Vim) markers between MCF10A and two breast cancer cell lines well-characterized for their cellular states (MCF7 = E state, MDA-MB231 = metastatic human breast cancer cells corresponding to an M state), we found MCF10A cells to be likely in a state that is intermediate between typical terminal E and M cells (Fig 3A, compare green population to others; S1 Fig). This is consistent with a recent study showing that MCF10A cells tend to collectively migrate [16], a feature that has been associated with the intermediate phenotype [9,17,18]. We surmise that the natural state of these cells is I1, because a majority of them show low to no Vim expression, suggesting more similarity to the terminal E than M cells. To experimentally test whether Ovol2 is able to reprogram I1-state cells into an E state (Fig 2C), we overexpressed Ovol2 in MCF10A cells using a lentiviral expression system in which the Ovol2-expressing cells can be distinguished from uninfected cells by bicistronic expression of GFP (S2 Fig). This led to significantly increased expression of Ecad, and decreased expression of Vim as assessed by quantitative population analysis using flow cytometry (Fig 3B, blue population). Comparison with Ecad/Vim profiles in Fig 3A reveals the similarity between Ovol2-reprogrammed cells and terminal E cells (MCF7). In contrast, overexpression of EMT inducers Snail or Zeb1 directed MCF10A cells to an M phenotype (Fig 3B, red and orange populations). To compare these observations with our mathematical model, we performed stochastic simulations with the basal model upon fluctuations in gene/protein expression (see details in Methods). We started simulation with the initial condition at I1 state (Fig 3C, green). Remarkably, at high basal production rates of Zeb1 and Ovol2 (Fig 3C, red and blue), the simulation produced similar Ecad/Vim expression patterns of these populations as those observed in our experiments. Consistent with previous reports that EMT promotes stemness [3–5], the expression of a well-known cancer stem cell marker CD44 [19] decreased upon Ovol2-induced transition to E and increased upon Zeb1/Snail-induced transition to M (Fig 3D). These observations demonstrate the bidirectional differentiation potential of MCF10A cells towards two opposite directions (i.e., I1 to E or I1 to M) and provide evidence for the opposing roles of Ovol2 and Snail/Zeb1 in the dynamic EMT system of MCF10A cells. As bifurcation analysis also predicted the ability of Ovol2 to reprogram M-state cells into an E state (Fig 2C), we tested the effect of Ovol2 overexpression on MDA-MB231 cells. Indeed, forced expression of Ovol2 was able to convert these cells to exhibiting a pattern of Ecad/Vim expression that is reminiscent of the terminal E state (Fig 3E). This finding is consistent with previous reports of Ovol2 overexpression inducing epithelial features in M-state cells [11,13]. Results of stochastic simulations for control M-state cells and Ovol2-overexpressed M-state cells using the corresponding conditions are in good agreement with this observation (Fig 3F). Interestingly, a time series experiment revealed that the downregulation of Vim precedes the induction of Ecad, suggesting that upon Ovol2 expression these cells first lose their memory of the M state and then acquire the E phenotype (Fig 4). Consistent with previous findings [14], a high dose of TGF-β resulted in a complete conversion of cells to what appears to be the M state, whereas low dose of TGF-β induced the appearance of two new populations in a heterogeneous culture that are likely I2 (previously termed P state for partial EMT in Zhang et al. [14]) and M states (Fig 5A and 5B and S3 Fig). Of note, in both mathematical modeling and experiments, the I2 state appears less stable than I1, as it 1) shows more vulnerability when facing fluctuations (S3 Fig); 2) entails a narrow range of Ovol2 concentration in the absence of strong TGF-β signaling (Fig 2C); and 3) is barely distinguishable from the M or I1 state experimentally and in simulations (Fig 5A and S3 Fig). The degree of separation of the different cell populations in our study is less remarkable than that reported [14], possibly due to the dynamic nature of I2 and the subtle differences in experimental conditions. Taken together, these experimental results support our computational discovery of a four-state dynamic system. Moreover, they highlight the ability of Ovol2 in reprogramming both I1- and M-state cells into a terminal E state, as predicted by our model. The EMT phenotypes and the cell state transitions that we have discovered through modeling and experiments are summarized in Fig 5C. First, we explored the roles of Ovol2 in regulatory control of the four states. Through bifurcation analysis with respect to external TGF-β and basal production rate of Ovol2 representing examples of EMT-inducing and -suppressing signals that are responsive to changes of the tissue microenvironment, we found these two signals to produce various combinations of cell phenotypes (Fig 6A(I)). Clearly, Ovol2 basal production rate exerted positive and negative effects on the robustness of E and M states, respectively (Fig 6A, blue and pink areas), and this is consistent with the demonstrated role of Ovol2 in preventing EMT and inducing MET ([11], [13] and this study). The effect of Ovol2 on the two intermediate states can be either positive or negative. Stability of I1 can be maintained when the strengths of Ovol2 and TGF-β signals are approximately proportional, with low levels of both giving rise to the most robust condition (Fig 6A, cyan area). In contrast, stability of I2 requires a minimum rate of Ovol2 basal production, but its robustness increases with higher levels of both Ovol2 and TGF-β (Fig 6A, orange area). In a specific case, when TGF-β signal was increased by 10-fold, higher Ovol2 basal production rate was required to retain the stability of both I1 and I2 states and to prevent the cells from differentiating into M state (Fig 6B, blue and orange triangles). Conversely, when Ovol2 basal production rate was increased, higher TGF-β signal strength was required to retain two stable intermediate states and to prevent the cells from differentiating into E state (Fig 6C, blue and orange diamonds). Overall, our analysis suggests that Ovol2 production tends to stabilize E state and destabilize M state, and that the two intermediate states are favored in two distinct conditions (high- versus low-signals) but both require the proper balance between EMT-inducing (e.g. TGF-β) and -suppressing signals (e.g. signals that induce Ovol2 expression). Next we reduced the strength of the Zeb1-Ovol2 mutual inhibition loop to determine its specific role in the four-state system. This led to no significant effect on the robustness of E state, moderately positive effect on that of M state (Fig 7, middle column, blue and pink areas), but significantly reduced robustness of the two intermediate states (Fig 7, middle column, cyan and orange areas). A complete blockage of the mutual inhibition loop resulted in a very small I1 region, and complete disappearance of the I2 region (Fig 7, right column). The role of the Ovol2-Zeb1 loop appeared distinct from that of the miR34a-Snail and miR200-Zeb1 mutual inhibition loops, as at least one of the miRNA-TF loops becomes dispensable for I2 (but not I1) when the other is intact (Fig 8, left and middle columns). This said, a complete blockage of both miRNA-TF loops abolished the intermediate states (Fig 8, right column). Partial blockage of each mutual inhibition loop gave rise to complex effects on the robustness of the two intermediate states (Table 1 and S5 Fig), but these effects are consistent with the finding of redundancy between the two miRNA-TF loops in terms of maintaining I2 state. Interestingly, partial blockage of miR200-Zeb1 resulted in a merge of the I1 and I2 regions, forming a large, continuous intermediate region (S5 Fig). Collectively, these results suggest that while all three mutual inhibition loops contribute to the existence and robustness of the two intermediate states, the strength of the Ovol2-Zeb1 loop is more critical. We next used stochastic modeling to examine how likely a population of cells at the two intermediate states differentiates into E or M state when gene/protein expression fluctuates. We chose a condition under which the four states are stable (Fig 6A, green star). Simulations were performed under this condition for two populations of cells originating from I1 and I2 respectively (Fig 9A and 9F). When fluctuations were small, cells stayed in the basins of attraction of their initial steady state by the end of the simulation (Fig 9A and 9D). Large fluctuations triggered both E(I)MT and M(I)ET of the cells originally in I1, resulting in a heterogeneous population containing E and M phenotypes, whereas the same level of fluctuations triggered E(I)MT of the cells originally in I2 (Fig 9B and 9E and S6 Fig). Thus, the I1 and I2 cells have distinct differentiation propensities, with I1 cells more likely differentiating into E state, whereas I2 cells more likely into M state. These simulations also allowed us to infer that E and M states are more stable than the intermediate states, and that I1 is more stable than I2 under the particular conditions tested (Fig 9C and 9F). We also asked whether changing Ovol2 production rate can affect the differentiation propensities from the I1 state. We reduced the basal production rate of Ovol2 by 20% and performed stochastic simulations starting from I1 as in Fig 9A and 9B. We found that this reduced Ovol2 production rate enabled more cells to settle at M state, and less cells to settle at I1 or E state (compare Fig 9B and 9H). We speculate that this is due to the reduced stability of E and I1 states, and/or the reduced energy barrier from I1 to I2 and M states (Fig 9I), providing a possible thermodynamic explanation for the role of Ovol2 in preventing EMT and inducing MET. Conversely, increased basal production of Ovol2 enabled some of the cells from I2 state to settle at E state instead of M state in the presence of large fluctuations (Fig 9J, 9K and 9L; compare panels E and K). These results suggest that the differentiation propensities of the two intermediate states can be regulated by tuning the level of Ovol2 expression. Our study provides both modeling and experimental evidence for a new intermediate state that lies between E and M states in addition to the recently observed intermediate state [9,10,14]. Previous studies based on epigenetic modifications predicted that multiple intermediate states may exist between terminal E and M states, and they may contribute to phenotypic plasticity in a continuous manner [15]. Additionally, Huang et al. classified 43 ovarian carcinoma cell lines into four subgroups, including an E-like intermediate and an M-like intermediate states in the EMT spectrum, based on expression patterns of signature EMT genes [20]. To our knowledge, our work is the first unequivocal demonstration of two intermediate states in EMT. Previous theoretical study revealed four types of stable states during T cell differentiation [21]; a common feature of that and our study is the inclusion of multiple (a minimum of three) positive feedback loops (including mutual inhibition). We anticipate that as the complexity of modeling increases by adding more regulatory elements, even more intermediate states may be observed, with the most extreme scenario being a spectrum of metastable or stable cell phenotypes lying between the terminal E and M states. A unique and interesting feature of our four-state model is that intermediate states are not necessarily metastable; instead they can be stable with no (I1) or high (I2) EMT-inducing/inhibitory forces. It is the balance between these two opposing forces that is critical for maintaining the intermediate states. Given the assumptions we make in our mathematical model, we have shown that the Ovol2-Zeb1 mutual inhibition loop is necessary for maintaining a four-state system. On one hand, this highlights the unique importance of the role of Ovol2 in EMT. On the other hand, the model leaves open the possibility that a four-state system could be governed by other unknown TFs that might be involved in a similar mutual inhibition loop. As discussed above, with additional positive feedback loops (>3), it is conceivable that additional intermediate states will emerge. As such, our model provides a framework for identifying and analyzing multiple intermediate phenotypes in EMT, and suggests a general and unique role of TF-based mutual inhibition loop in this system. With a proposed Ovol-Zeb1 mutual inhibition loop, a recent modeling study suggested important roles of Ovol2 in controlling the previously established three-state EMT system [22]. This is in agreement with our findings that Ovol2 is critical for both intermediate states. What is the advantage of having intermediate state(s)? Such state can be a “hybrid” state, where cells exhibit both, albeit partial, “E” and “M” phenotypes. Indeed, during mammary gland morphogenesis, some epithelial cells at the tips of growing ducts express mesenchymal markers while simultaneously retaining epithelial integrity, suggesting that they are in a naturally occurring “hybrid” state [23]. In this case, a “hybrid” phenotype would enable the cells to undergo collective migration, by which they invade the surrounding stroma as a coherent epithelial front to facilitate branching morphogenesis. The same may be true for metastatic cancer cells as they acquire a mesenchymal phenotype to invade the surrounding tissue and colonize distant sites as epithelial tumors. Alternatively, an intermediate state can be a “naïve” state, where cells are devoid of typical epithelial and mesenchymal features. Along this line, we note our experimental observation that MCF10A cells seem to first lose their initial phenotypes (E or M), and then gain their destination phenotypes (M or E) during the factors-directed state transitions (Fig 3). Traveling through a “naïve” state could be a useful mechanism to erase memories of old lineages, thus creating a window of opportunity for expanded differentiation plasticity as desired for multipotent stem cells. Why multiple intermediate states then? Chuong and Widelitz proposed the interesting idea that stem cell states can be regulated depending on the physiological needs of tissues to generate different numbers of intermediate stops on their journey to differentiation [24]. The same concept may be applicable to the EMT system, as having multiple intermediate states offers additional facets of regulation to more precisely control the temporal and spatial flux of epithelial cells through their differentiation/dedifferentiation pathway to adapt to various tissue environments or topology. Regardless of whether cells adopt an intermediate fate to gain "hybrid" behavior (e.g., collective migration) or to dedifferentiate to a naive state for lineage plasticity, the more intermediate states there are, the more thermodynamic traps that would be. Thus having more than one intermediate state would create a more controllable energy barrier so that cells do not easily fall into the mesenchymal state, which we know from previous studies is largely irreversible [14]. Non-genetic heterogeneity and spontaneous conversion among subpopulations have been documented for hematopoietic stem cells and breast cancer cells [4,25–27]. Theoretical analysis of these dynamic processes often involves the assumption that gene regulatory networks can generate multiple stem-like states that are adjacent in state space [25,26,28,29]. Our model presents a good example in which a network of three mutual inhibition loops is capable of giving rise to two adjacent states that may be associated with subpopulations of cells having distinct propensities for differentiation or tumorigenesis. The unstable nature of the I2 state under conditions examined and the phenotypic similarities between I1 and I2 states prevented us from further characterizing the molecular differences between the two intermediate cell populations and their corresponding cellular behaviors. As such, the functional significance of having two intermediate states has yet to be experimentally established. We have demonstrated bidirectional transitions of MCF10A cells upon Zeb1/Snail overexpression (I1-M transition), TGF-β treatment [I1-(I2)-M transition] and Ovol2 overexpression (I1-E transition) (Fig 5B). It is tempting to ask which extracellular signaling molecules can trigger Ovol2 upregulation and the subsequent transition to E state under physiological conditions. Among the possible candidates is a BMP signal as it is known to induce MET [11,13] and to positively regulate Ovol2 expression during embryonic stem cell differentiation [30]. Identification of Ovol2-inducing external signals that can induce MET of MCF10A cells will enable a finer analysis of the dynamic process of MET as well as further experimental validation of our mathematic model. In summary, our work identifies transcriptional factor Ovol2 and its mutual inhibition relationship with Zeb1 as critical additions to the known EMT regulatory network. Specifically, these new regulatory elements are important for attaining and maintaining the two intermediate states. Furthermore, their experimental perturbations allowed us to observe the bidirectionality of transitions from the intermediate states. Together, our study offers a framework for understanding the molecular strategies and design principles by which epithelial stem, progenitor, or cancer cells achieve multipotency or collective migration. MCF10A, MCF7, MDA-MB-231 were purchased from ATCC. MCF7 and MDA-MB231 cells were maintained in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum. MCF10A cells were grown in DMEM/F12 (1:1) medium with 5% horse serum, epidermal growth factor (10 ng/mL), cholera toxin (100 ng/mL) and insulin (0.023 IU/mL). For TGF-β treatment, cells were incubated with titrated concentrations of human TGF-β1 protein (R&D systems) in complete culture medium for 10 days. The culture medium was replaced daily and cells were passaged just before reaching full confluency. Recombinant lentiviruses expressing Ovol2 using the pHIV-ZsGreen lentiviral construct was described previously [11]. For Snail and Zeb1 expression, human SNAIL cDNA and mouse Zeb1 cDNA were cloned into the XhoI/NotI and EcoRI sites of pHIV-ZsGreen, respectively. Production and infection of lentiviruses were carried out as previously described [11]. Transduction unit of viral solution was estimated by measuring GFP-positive population using a flow cytometer. For Ecad/Vim profiling, cells were fixed with 4% paraformaldehyde, permeabilized with 0.2% Triton X-100, and stained with the following primary and secondary antibodies and reagents: anti-E-cadherin (Life Technologies, 1:500), anti-vimentin (Cell Signaling Technologies, 1:500), allophycocyanin (APC)-labeled anti-mouse IgG (Santa Cruz), Cy3-labeled anti-rabbit IgG (Jackson Immuno). For CD44 staining, live cells were stained with phycoerythrin (PE)-conjugated anti-CD44 antibody (Biolegend). Live-cell sorting for GFP+ cells was performed on a BD FACSAria equipped with FACS DiVa6.0 software operating at low pressure (20 psi) using a 100-μm nozzle. Cell clusters and doublets were electronically gated out. Cells were routinely double sorted and post-sort analysis typically indicated purities of >90% with minimal cell death (<10%). ChIP was performed with an anti-Zeb1 antibody (Santa Cruz) according to the previously described protocol [11]. The following primers were used to detect Ovol2 promoter regions: proximal site (forward, 5’-GTGATAGGGGTATGAAGCAGAGG-3’ reverse, 5’-CACCAGGAAACTTGGGAGTG-3’) and distal site (forward, 5’-AGCCCAGAAATCCGTTACCA-3’ reverse, 5’-CTCACTGCTGGAGGTTGTCT-3’). Total RNA was isolated using the TRIzol Reagent (Invitrogen) followed by cleaning up and RNase-free DNaseI treatment using the RNeasy mini kit (QIAGEN). cDNA was prepared using Retroscript Kit (Applied Biosystems) according to manufacturer’s instructions. Real-time PCR was performed using a CFX96 qPCR system and SsoAdvanced SYBR Green Supermix (Bio-rad). Comparative analysis was performed between the genes of interest normalized by the house keeping genes GAPDH and ACTB. The following primers were used: OVOL2 (forward, 5’-AGCTGTGACCTGTGTGGCAAG-3’ reverse, 5’-ACGAATGCCTGTGTGTGTGC-3’), ZEB1 (forward 5’-TTGCTCCCTGTGCAGTTACA-3’ reverse 5’-CGTTTCTTGCAGTTTGGGCA-3’), GAPDH (forward, 5’-GGACCTGACCTGCCGTCTAGAA-3’ reverse, 5’-GGTGTCGCTGTTGAAGTCAGAG-3’), and ACTB (forward, 5’-CTTCTACAATGAGCTGCGTG-3’ reverse, 5’-GGGTGTTGAAGGTCTCAAAC-3’). For semi-quantitative PCR, the following primers were used: CDH1 (forward, 5’-AAAGGCCCATTTCCTAAAAACCT-3’ reverse, 5’-TGCGTTCTCTATCCAGAGGCT-3’), SNAI1 (forward, 5’-CCTCCCTGTCAGATGAGGAC-3’ reverse, 5’-CCAGGCTGAGGTATTCCTTG-3’), VIM (forward, 5’-GACGCCATCAACACCGAGTT-3’ reverse, 5’-CTTTGTCGTTGGTTAGCTGGT-3’). We used ordinary differential equations (ODEs) to model the regulatory network shown in Fig 2A. The framework of the model stems from a recently published EMT model [14], and the modeling details are described therein. This framework employs mass-action dynamics to model microRNA-mRNA interactions with considerations of the microRNA binding sites on their targets. This modeling strategy was introduced by Lu et al. [9,31]. As other transcription factor regulations, interactions involving Ovol2 were modeled with Hill functions. Numerical bifurcation analysis was performed with PyDSTool [32]. To consider fluctuations in gene expression, we added multiplicative white noise to some of the ODEs. To determine which phenotype (basin of attraction) a cell adopts at the end of the simulations, we set the noise terms to zero and let the simulation continue until it reached steady state (S6 Fig). Lists of equations, parameters and assumptions can be found in supplementary materials. Stochastic simulations were performed with XPPAUT [33].

10.1371/journal.pcbi.1004337

Impact of School Cycles and Environmental Forcing on the Timing of Pandemic Influenza Activity in Mexican States, May-December 2009

While a relationship between environmental forcing and influenza transmission has been established in inter-pandemic seasons, the drivers of pandemic influenza remain debated. In particular, school effects may predominate in pandemic seasons marked by an atypical concentration of cases among children. For the 2009 A/H1N1 pandemic, Mexico is a particularly interesting case study due to its broad geographic extent encompassing temperate and tropical regions, well-documented regional variation in the occurrence of pandemic outbreaks, and coincidence of several school breaks during the pandemic period. Here we fit a series of transmission models to daily laboratory-confirmed influenza data in 32 Mexican states using MCMC approaches, considering a meta-population framework or the absence of spatial coupling between states. We use these models to explore the effect of environmental, school–related and travel factors on the generation of spatially-heterogeneous pandemic waves. We find that the spatial structure of the pandemic is best understood by the interplay between regional differences in specific humidity (explaining the occurrence of pandemic activity towards the end of the school term in late May-June 2009 in more humid southeastern states), school vacations (preventing influenza transmission during July-August in all states), and regional differences in residual susceptibility (resulting in large outbreaks in early fall 2009 in central and northern Mexico that had yet to experience fully-developed outbreaks). Our results are in line with the concept that very high levels of specific humidity, as present during summer in southeastern Mexico, favor influenza transmission, and that school cycles are a strong determinant of pandemic wave timing.

An influenza pandemic virus emerged in North America in 2009. Although the virus spread worldwide within months of its emergence, the timing of peak pandemic activity varied by nearly a year across global populations. In fact, the most intense period of pandemic activity occurred earlier in temperate countries of the southern hemisphere (e.g., South Africa, Argentina and Australia) than in countries that reported the first infections (Mexico and the US). This indicates that the timing of peak pandemic activity did not directly align with the arrival of the virus in a location, suggesting the influence of social and environmental factors on pandemic influenza transmission. In this study we examined two large pandemic outbreaks (“waves”) that occurred in summer and fall of 2009 in Mexico. The summer wave occurred in the tropical southeastern states of Mexico, whereas the larger fall wave occurred in the central and northern states. We find that the distinct pandemic waves were likely caused by complex interactions between regional differences in specific humidity, population level susceptibility, and school term cycles. Improving our understanding of the factors modulating regional transmission patterns will improve prediction of the spatial and temporal progression of future pandemics.

There is strong evidence that seasonal influenza activity around the world is modulated by environmental variability [1]. Temperate regions are characterized by annual winter epidemics [2,3] that may result from seasonal decreases in specific humidity and subsequent increases in virus survival and transmission [4–6]. Seasonal influenza activity in the tropics is not as clearly phased [7,8], but tends to peak in seasons with high levels of specific humidity and rainfall [1]. Yet, the extent to which these same factors affect the transmission of pandemic influenza remains largely unknown. The emergence of the novel A/H1N1pdm influenza virus in early 2009 and its subsequent global pandemic spread [9] provides a unique opportunity to examine these links. The influenza A/H1N1pdm virus spread globally within weeks of the first report of a laboratory-confirmed case on April 19, 2009 in southern California [10]. The global dissemination of the virus was facilitated by global transportation networks from its putative source in North America [11], and by June 11, 2009 laboratory-confirmed infections had been identified in over 70 countries [12]. Although the virus had spread worldwide within months of its emergence, the country-to-country timing of peak pandemic activity varied by >40 weeks [13]. In fact, the most intense period of pandemic activity occurred earlier in temperate countries of the southern hemisphere (South Africa, Argentina and Australia) than in countries that reported the first A/H1N1pdm infections (Mexico and the US) [13]. These observations indicate that the timing of peak pandemic activity did not directly align with the arrival of the virus in a location, suggesting the influence of social and environmental factors on pandemic influenza transmission. Several studies have examined the relative contribution of environmental drivers and school mixing on influenza transmission with conflicting results. School cycles have been shown to play a significant role in the transmission of seasonal influenza by modulating contact rates between children [14,15]. School closures have been linked to reductions in 2009 pandemic A/H1N1 transmission [16–18], while regional variability in school terms and weak child-mobility have been associated with the staggered occurrence of fall pandemic activity in US cities [17,19]. Specific humidity and prior immunity may have played a role in explaining a third wave of pandemic activity during the winter of 2009–2010 in the southeastern US [20]. Similarly, the spatiotemporal patterning of the 2009 pandemic in Canada was associated with school terms and specific humidity [21,22]. In Chile, latitudinal variation in the timing of peak pandemic activity was associated with specific humidity but not with winter vacations, as pandemic activity was already subsiding when schools closed [23]. Altogether, the transmission impact of environmental forcing, population-scale interventions and school cycles has been broached but yet to be fully understood for pandemic influenza. Here we fit Susceptible-Exposed-Infected-Recovered (SEIR) models to epidemiological data from Mexico using Markov Chain Monte Carlo (MCMC) approaches and investigate the importance of environmental forcing, school cycles, and interventions in generating the observed spatially-heterogeneous waves of 2009 pandemic influenza activity. Mexico is a compelling case study as it encompasses both temperate and tropical regions, interventions and school vacation periods were largely uniform across the country, and spatially-detailed epidemiological records are available [16]. We obtained daily laboratory-confirmed A/H1N1pdm influenza case data for April—December 2009 from a prospective epidemiological surveillance system that was established in response to the pandemic by the Mexican Institute for Social Security [16,24]. These data were previously used to document the regional patterns of pandemic activity in Mexico across 31 states and the Federal District (hereafter we refer to these as 32 “states”) [16]. Circulation of the pandemic virus was intense during 2009 in Mexico, subsided by January 2010, and was followed by a period of 2-years of sporadic transmission [25]. Hence we capture the full extent of the first year of pandemic activity in this analysis. A mandatory policy of school closure was strictly enforced for a 14-day period during April 27—May 11, 2009 across all states in Mexico [16]. In addition, our study period encompasses three school vacation periods, synchronous across Mexico, including a spring (April 5–18), summer (July 3—Aug 24) and winter break (beginning Dec 22; Fig 1). We retrieved and compiled daily specific humidity, precipitation and temperature data for the study period from the Global Land Data Assimilation System [26] for the most populous city in each state. There were three spatially-heterogeneous pandemic waves in Mexico in 2009, including a spring wave from April 1—May 20, a summer wave from May 21—August 1, and a fall wave from August 2—December 31, as previously defined [16]. For each state, we calculated the cumulative case proportion of each wave (the number of cases during the wave relative to the total number of cases during the study period) and their association with average temperature, average humidity and cumulative precipitation conditions during each wave period using Pearson correlation. We then classified each location as dominated by a spring, summer or fall wave based on the week with the maximum number of cases. To further assess the dynamical effects of environmental variability and school cycles on pandemic influenza transmission, we developed a deterministic SEIR compartmental model at the state level. As a first step, we fit separate models for each state, and as a second step, we explore a meta-population framework allowing for coupling between states. The simplest formulation of our model (independence between states) is expressed by the following equations: dSi/dt=−λiSidEi/dt=−λiSi−θ−1EidIi/dt=θ−1Ei−α−1IidRi/dt=α−1Ii where Si, Ei, Ii, Ri are, respectively, the number of susceptible, exposed, infected, and recovered individuals in Mexican state i, θ is the mean latency period, α is the mean infectious period, and λi is the force of infection [27]. In the main analysis presented here, the mean latency period (θ) was fixed at 1.4 days based on past estimates [28], while the mean recovery rate (α) was fixed at 1.6 days, consistent with a mean serial interval of 3 days [29]. Further sensitivity analysis was performed in which the mean latency period (θ) was allowed to vary from 1–1.8 days, while the mean recovery rate (α) was allowed to vary from 0.6–2.6 days, consistent with a mean serial interval of 2.4–3.6 days [29]. We allowed transmission to vary spatially and temporally as a function of specific humidity, school terms and interventions. Following [6] and [20], we let the basic reproduction number, R0,i(t), vary with time as a function of daily state-specific specific humidity qi(t): R0,i(t)=f(qi(t)) We chose this formulation for R0(t) because environmental forcing was the main factor under investigation in this study. The functional relationship between R0 and specific humidity was defined by fitting a Piecewise Interpolating Polynomial (PCHIP) curve through three critical points (p1, p2, p3). PCHIP was used because it provided optimal control over the relationship by forcing the curve to pass through the specificed points without overshooting the defined minima or maxima. The critical points were defined by 4 free parameters (w1, w2, w3, w4) (Table 1). We specified that R0 values at p2 and p3 to be greater than or equal to p1 (Fig 2). Altogether, this specification of the points allowed for U-, J-, and L-shaped cruves, in addition to flat lines (see S1 Text). Next, we let the effective reproduction number, Re(t), depend both on R0(t) the proportion of the susceptible population, Si(t), as well as school terms, v(t) and interventions, z(t). For school terms and interventions, we used step functions to represent changes in influenza transmissibility, as follows: v(t)={1,ifschool termγ1,ifschool vacation0.55≤γ1≤0.90z(t)={1,ifno interventionγ2,ifintervention0.55≤γ2≤0.90 where γ1, and γ2 are independent and bounded parameters (Table 2). Altogether, the effective reproduction number (Re) in state i, follows: Re,i(t)=v(t)z(t)R0,i(t)Si(t)N−1 We also developed several flavors of meta-population models to explore travel effects. Based on past work indicating the importance of local diffusion on influenza transmission [19] we allow for neighboring states to affect the local force of infection. In the absence of mobility data from Mexico to calibrate a more detailed population model, we also allow for the greater Mexico City area to affect the risk of transmission in other areas, as the Mexican capital is a hub for both air and bus travel, the two dominant modes of transportation in the country. Our approach borrows from the concept of gravity models [30], whereby both large populations and nearby populations may affect the risk of dissemination to a new locale. Specifically, we allow the force of infection in each state, λi, to be modified by the force of infection in neighboring states, λadj, and in two “hub” states (Mexico City and the Federal District), λhub. For each state, i, the λi at time t prior to mixing between states is defined as: λi(t)=Bi(t)Ii(t)N−1 where Bj is the transmission rate for state i. Bi(t) and R0,i(t) are related by: Bi(t)=R0,i(t)α−1 When both travel effects are included, the force of infection in state i is modified by the force of infection in all adjacent states and two hub states as follows: λ^i(t)=λi(t)+cadj∑j=1λadj(t)+chub∑k=1λhub(t) where cadj and chub are free parameters that are allowed to vary from 0–1. Overall, we compare the fit of 3 increasingly complex spatial models: (a) models fitted independently to each state (chub = cadj = 0), (b) meta-population models with nearest neighbors coupling (cadj > = 0 and chub = 0) and (c) meta-population models with nearest neighbor coupling and hub centered around the capital (chub > = 0 and cadj > = 0). All models incorporate estimates of state population size from the National Council of Population, Mexico for 2009 [31]. Populations in each state were assumed to be fully mixed. The initial proportion of susceptibles, μ(0), was set to 95% across all states based on the age structure of the population in Mexico and estimates of worldwide prevalence of age-specific pre-pandemic cross-reactive antibody responses to A/H1N1pdm [32]. As a sensitivity analysis, the initial proportion of susceptible population was allowed to vary from 0.75–0.95. Further, in a separate model we allowed for spatial variation of μ(0) between southeastern states and central and northern states (see S1 Text). In all models, initial incidence in the simulations, τ(0), was uniform across states and allowed to vary from 1–1,000 per 100,000 people (Table 2). The SEIR models were continuous and deterministic. The ordinary differential equations were solved numerically using Matlab version 8.2 (The Mathworks, Inc). We employed an adaptive Metropolis-Hastings algorithm to perform MCMC simulations and estimate parameter values [33]. We estimated model parameters (Table 2) by fitting the model-derived daily case proportion to empirical data for all states during the pandemic period. Using daily case proportion rather than case incidence allows standardization for potential reporting differences between states. We assumed uniform prior distributions for estimated parameters. We allowed the algorithm to run for 100,000 iterations following an initial burn-in of 250,000. The Geweke diagnostic method was employed to assess convergence of chains [34] with values close to 1 deemed satisfactory. We compared the fit across models with the Akaike information criterion (AIC) using the observed and simulated daily time series across all states. We also assessed model fit by examining whether peak pandemic incidence in each state occurred during the summer or fall in simulated data, and how this corresponded to the observed timing using the chi-square test. There were three pandemic waves in Mexico during the 2009 pandemic (Fig 1). The spring wave was relatively minor and concentrated in the greater Mexico City area. The summer wave was predominant in southeastern states, and the fall wave was concentrated in central and northern states. The cumulative case proportion during the summer wave was associated with higher mean specific humidity conditions (Pearson’s correlation, r = 0.74, 95% CI: 0.63, 0.85). Mean temperature (r = 0.42, 95% CI: 0.20, 0.63) and rainfall (r = 0.59, 95% CI: 0.20 0.78) were also significantly associated with summer cumulative case proportion, but were strongly influenced by outliers (Fig 3). The cumulative case proportion during the fall wave was negatively associated with specific humidity (r = -0.75, 95% CI: -0.86, -0.62), temperature (r = −0.48, 95% CI: -0.73, -0.37) and precipitation (r = -0.56, 95% CI: -0.70, 0.01). The significant association between precipitation and cumulative case proportion was primarily due to a single outlier (Fig 3). We fit mechanistic transmission models to the progression of the pandemic in 32 states to assess the dynamical consequences of school cycles, social distancing interventions (that include but are not limited to school closure, see Fig 1), and spatiotemporal variation in environmental drivers. Since the strongest statistical association between influenza activity and environmental forcing was seen with specific humidity in exploratory analyses (Fig 3), we allowed R0 to vary flexibly as a function of specific humidity only. This aligns with laboratory and epidemiological evidence linking influenza activity with variations of specific humidity [1,5,6,20–23]. We considered a series of increasingly complex models that included different levels of spatial mixing. Simple transmission models including humidity, school and intervention terms, but no spatial coupling accurately identified the season of the largest pandemic outbreak in 29 of 32 states (P < 0.001; Table 2, model 1). The model also accurately described observed pandemic activity in Colima, where a relatively balanced proportion of cases occurred in summer and fall. However, the model overestimated early summer transmission in several central and northern states, in particular, Veracruz, Guerrero and Morelos where cases primarily occurred during the fall (Fig 4). Comparison of AIC values across model structures indicated a significant improvement with the addition of a spatial mixing term allowing for interaction between adjacent states; however the proportion of peak incidence correctly predicted with respect to season decreased (27/32; Table 2, model 2). Inclusion of a term representing connectivity with Mexico City did not improve model fit nor peak prediction accuracy (Table 2, model 3). The estimated relationship between R0 and specific humidity, and the effect of school terms and interventions, were similar across all models. Since the model without spatial mixing best predicted pandemic activity, and estimates of the majority of other parameters did not change significantly with inclusion of spatial mixing terms, we focus on the model with no spatial mixing in subsequent sections. Our MCMC estimation algorithm achieved high convergence with the exception of two parameters, w1 and τ(0) which determine the value of specific humidity at which R0 becomes minimum and the initial incidence of infection, respectively (Table 2 and S1 Text). For the best-fit model, the simulated infection attack rate averaged across all states was 28% and ranged from 26–34%. Models where the mean latency period (θ), the mean recovery rate (α) and the initial level of susceptibility μ(0) were allowed to vary were not as robust as the primary models; however the models suggested a strong and consistent role for the effect of specific humidity and school terms on the pandemic waves (see S1 Text). The estimated relationship between specific humidity and R0 was J-shaped, with greatest R0 at high levels of specific humidity and minimal R0 at moderate levels of specific humidity (Fig 5; Table 3). Our estimated R0 values ranged between 1.14 and 1.26, depending on specific humidity conditions. Our model suggests that population level susceptibility was uniformly high and near initial levels at the beginning of the summer wave (Fig 5). In late May and June 2009, Re was estimated to be greatest in the southeastern states (mean = 1.26) due to more humid conditions; whereas in the central and northern states Re was slightly lower (mean = 1.21) due to moderate levels of specific humidity. The model implies that these regional differences in estimated Re, driven by differences in specific humidity, were critical for the formation of the heterogeneous pandemic wave pattern and fomented greater incidence during summer in the more humid southeastern states. During the summer vacation period, estimates of Re decreased below 1 in nearly all states owing to school closure. At the beginning of the school term in late August 2009, estimated Re averaged 1.11 for the majority of central and northern states, where susceptibility remained high (~85%). In contrast, susceptibility estimates in southeastern states were significantly lower (~75%), which reduced transmission rates (Fig 5E). We estimate that school vacation decreased influenza transmission by 14% (95% CI: 10%, 19%) on average during the spring and summer breaks. During the 14-day period in May 2009 when stringent social distancing interventions were put in place, R0 was reduced by 20% (95% CI: 11%, 40%); the large 95% credible interval likely has to do with a strong correlation between estimates of w2 (minimum of R0) and z (impact of interventions) in the model. Although we put constraints on the impact of intervention and school vacation periods based on past studies (Table 2 and [15,16]), relaxing these constraints did not change our estimates, attesting to the importance of reduced transmission during these periods. Indeed, simulations indicate that if the summer vacations were eliminated, peak pandemic activity would have occurred in June-July 2009 for all states (see S1 Text). Our results indicate that the spatiotemporal structure of the 2009 influenza pandemic in Mexico can be understood by the interplay between specific humidity, school cycles, and susceptibility. Our results suggest that high specific humidity in the southeastern states allowed for relatively high transmission rates in late May-June 2009 and favored a substantial outbreak in these states (Fig 5); the pandemic wave in the southeastern states subsided during summer school vacations as Re decreased. The results suggests that when school activities resumed in the fall, transmission increased due to increased contact rates and triggered outbreaks in some southeastern states, but these were relatively minor due to reduced levels of susceptibility following the summer wave. This is similar to the process that explains differences in the progression of the pandemic across regions in the US during the fall and winter of 2009–2010 [20]. Our model predictions indicate that central and northern states experienced more than 85% of their cases during the fall wave, consistent with observations. Our analysis suggests that the lack of substantial viral activity in this region in late May-June 2009 was due to slightly lower transmission rates associated with moderate levels of humidity, then compounded by school vacation, further reducing transmission rates. When summer vacation ended, the subsequent increase in contact rates made large outbreaks possible in this region. Our simulations suggest that a large summer wave would have occurred in the central and northern states if summer vacation had not reduced transmission (see S1 Text). This highlights that both environmental variability and school cycles were critical for generating the distinct summer and fall wave patterns observed in Mexico. We developed an SEIR model framework that allowed local levels of specific humidity to modulate transmission rates across a wide distribution of relationships, including a bimodal relationship indicating that influenza activity is enhanced by very low levels and very high levels of specific humidity [1]. Unlike previous studies that examined the effect of specific humidity (among other variables) on the progression of pandemic waves in the US [20], Canada [21,22], and Chile [23], it was necessary to use a bimodal relationship here because Mexico encompasses both temperate and tropical regions. Previous studies have suggested that population mixing across states may have been a factor in the formation of the spring, summer and fall waves [35]. We developed a model that allowed for mixing between adjacent states and a “hub” region representing the greater capital area (Mexico and the Federal District). However, we found no evidence that connectivity with Mexico City, and hence a hierarchical pattern of spread, could explain the spatiotemporal structure of the pandemic waves. In contrast, connectivity with adjacent states was important, which is reminiscent of the slow diffusive pattern of 2009 pandemic in US cities [19]. This suggests that only local environmental conditions, social mixing and susceptibility patterns shaped the trajectory of the outbreak once the virus became established in the population. We found no relationship between environmental conditions and the intensity of the spring wave of the 2009 pandemic in Mexico, which remained focused in central states. The spring wave may have only materialized in states highly connected to the geographic origin of the A/H1N1pdm virus prior to the initiation of intervention measures. We are unable to test this hypothesis further as the origins of the 2009 pandemic remain debated. We estimated a number of disease parameters that compare favorably with independent information, reinforcing the validity of our modeling approach. In particular, our estimates of pandemic infection attack rates ranged from 26–34% for the cumulative period April-December 2009, which is commensurate with estimates derived from global serological surveys (both symptomatic and asymptomatic infections) [29]. Our estimated R0 values, ranging between 1.14 and 1.26, align with estimates from previous studies in Mexico and elsewhere [9,21,28]. However, our estimated Re values in the southeastern (1.2, summer) and in central-northern states (1.1, fall) are smaller than previous estimates for these regions, which may stem from different modeling assumptions [16]. Finally, we estimated that drastic social distancing interventions reduced transmission by 20% (95% CI: 14%, 40%) in Spring 2009 in Mexico while summer school vacations reduced transmission by 14% (95% CI: 10%, 19%), which is broadly consistent with previous estimates [16]. Our model is prone to a number of limitations. In the absence of prior information on baseline pre-pandemic immunity in Mexico, we considered prior immunity in baseline models to be spatially homogenous at 5%, informed by global age-specific serosurveys [29]. We did not explicitly incorporate the role of heterosubtypic immunity in mitigating the spread of the virus [22,36,37]. Indeed, previous modeling studies suggest that prior immunity from seasonal strains can inhibit the transmission of pandemic strains, thereby either delaying pandemic waves until immunity wanes or creating multi-wave patterns in a single population [38,39]. It is possible that prior immunity in Mexico varied across regions as a result of differential phasing of seasonal influenza across the southeastern and central-northern regions. Specifically, given that seasonal influenza activity typically occurs during winter in central and northern Mexico [40], the pandemic virus started disseminating directly following the circulation of seasonal influenza strains, potentially stimulating cross-subtype immunity and reducing transmission early during the pandemic. In contrast, 6–9 months may have passed since the most recent seasonal influenza activity in the southeastern region where seasonal influenza activity occurred in the summer [41], potentially resulting in populations with relatively lower pre-pandemic immunity levels. In sensitivity analyses, we explored the possibility that pre-pandemic susceptibility varied independently in northern and southern states. A model allowing for regional differences in pre-pandemic immunity performed better than a model without, supporting higher prior immunity in the northern and central states than in the southeastern states; however we have no specific biological information to support this model. Further, it should be noted the model with regional differences in pre-pandemic immunity benefitted from prior information on the spatial structure of the pandemic wave in different Mexican regions, data that were not provided to the other models tested. Overall, the epidemiological consequences of prior immunity to pandemic influenza remains heavily debated and a subject worthy of further experimental and modeling work [36–39,42]. Although our model accurately captured the overall spatial structure of the pandemic, correctly predicting the season of peak pandemic activity in 29 of 32 states, some details of the pandemic were missed. Specifically, in some southeastern states, simulated summer outbreaks lagged 2–4 weeks behind the observed outbreaks. Further, the model did not accurately describe the intense growth of fall outbreaks in states such as Sonora, Tlaxcala and Hidalgo (Fig 4). The lack of age-structure in our model and finer details of the relationship between influenza and environmental forcing could explain these differences. As descriptive analyses revealed high correlation between regional pandemic patterns and specific humidity, we selected specific humidity as the most likely driver of transmission in our mechanistic approach. In line with previous work [20], we did not further consider the putative effect of other environmental covariates, particularly temperature, due to high collinearity with specific humidity. Another possible source of error is that we used weather data corresponding to the largest population in each state. Specific humidity and other weather variables can vary significantly within states. By using weather data at the largest population center in each state we mitigated some of the effects of spatial climate variability, but it should be noted that these data may not accurately represent conditions for people living outside the population center. Further, we did not consider stochastic effects and assumed that these effects were limited due to our focus on the large populations of Mexican states in a pandemic period where incidence is high, following earlier work [43]. An alternative approach would be to do seed infected hosts at critical times prior to each wave as in [44], which would be important to consider for small populations. Another issue in developing meta-population models for Mexico is the lack of detailed mobility data. Although the addition of a hub centered around the greater Mexico city area did not seem to affect the dynamics of the 2009 pandemic, details of population mobility could be more important in inter-pandemic seasons [30]. Future work could concentrate on calibrating more detailed meta-population model to incidence and mobility data in Mexico. Our results do not support a substantial increase in transmission at low levels of specific humidity, which has been observed in previous studies for both seasonal [1,5,6] and pandemic influenza [20–23]. However, pandemic activity was concentrated during April-October in Mexico when specific humidity levels were at moderate-to-high levels in a majority of states, making it difficult to assess the relationship between low levels of specific humidity and transmission. Another possibility is that—as discussed above—prior immunity in the spring may have been high in the northern (drier states) due to recent seasonal influenza transmission thereby inhibiting the spread of influenza during the period at the beginning of the pandemic when specific humidity was relatively low. Finally, although we accounted for institutional intervention measures, we did not account for changes in personal behavior (e.g., hand washing, avoiding public spaces, masks) that may have varied across time and space. Indeed, behavior change may have contributed to the formation of multiple waves in the UK during the 1918–1919 influenza pandemic [45]. Limited evidence indicates there were changes in travel behavior in Mexico in response to the pandemic [46]. Changes in social behavior, however, would not explain regional differences in pandemic patterns across Mexico unless these changes were regionally heterogeneous. Overall, our results indicate that the occurrence of spatially-heterogeneous waves of the A/H1N1 pandemic virus in Mexico can be understood through consideration of local specific humidity conditions, susceptibility, and school-driven mixing patterns. The effect of humidity on pandemic influenza transmission in Mexico is consistent with a recent global model of seasonal influenza activity that stipulates a bimodal relationship between influenza and specific humidity, where transmission is favored by very high and very low levels [1]. Broadly, these findings suggest that a greater understanding of the mechanisms that drive inter-pandemic influenza epidemics may increase our capacity to predict the timing of major outbreaks associated with novel pandemic influenza viruses.

10.1371/journal.pgen.1005346

TopBP1 Governs Hematopoietic Stem/Progenitor Cells Survival in Zebrafish Definitive Hematopoiesis

In vertebrate definitive hematopoiesis, nascent hematopoietic stem/progenitor cells (HSPCs) migrate to and reside in proliferative hematopoietic microenvironment for transitory expansion. In this process, well-established DNA damage response pathways are vital to resolve the replication stress, which is deleterious for genome stability and cell survival. However, the detailed mechanism on the response and repair of the replication stress-induced DNA damage during hematopoietic progenitor expansion remains elusive. Here we report that a novel zebrafish mutantcas003 with nonsense mutation in topbp1 gene encoding topoisomerase II β binding protein 1 (TopBP1) exhibits severe definitive hematopoiesis failure. Homozygous topbp1cas003 mutants manifest reduced number of HSPCs during definitive hematopoietic cell expansion, without affecting the formation and migration of HSPCs. Moreover, HSPCs in the caudal hematopoietic tissue (an equivalent of the fetal liver in mammals) in topbp1cas003 mutant embryos are more sensitive to hydroxyurea (HU) treatment. Mechanistically, subcellular mislocalization of TopBP1cas003 protein results in ATR/Chk1 activation failure and DNA damage accumulation in HSPCs, and eventually induces the p53-dependent apoptosis of HSPCs. Collectively, this study demonstrates a novel and vital role of TopBP1 in the maintenance of HSPCs genome integrity and survival during hematopoietic progenitor expansion.

The rapidly proliferating hematopoietic stem/progenitor cells (HSPCs) require well-established DNA damage response/repair pathways to resolve the DNA replication stress-induced DNA damage, which is deleterious for the genome stability and cell survival. Impairment of these pathways could lead to the progressive bone marrow failure (BMF) and hematopoietic malignancies. Here we reported a novel function of topoisomerase II β binding protein 1 (TopBP1) in definitive hematopoiesis through characterizing zebrafish mutantcas003 with a nonsense mutation in topbp1 gene encoding TopBP1. The homozygous topbp1 mutants manifested decreased HSPCs during their pool expansion in the caudal hematopoietic tissue (CHT, an equivalent of the fetal liver in mammals) due to the p53-dependent apoptosis. Further investigation revealed that the deficient TopBP1-ATR-Chk1 pathway upon DNA replication stress in topbp1 mutants led to accumulated DNA damage and further affected HSPCs survival. These studies therefore emphasized the importance of topbp1 function as well as DNA damage response pathways during the fetal HSPC rapid proliferation.

Hematopoietic stem/progenitor cells (HSPCs) possess the capabilities of self-renewal and differentiation into all lineages of mature blood cells [1]. Dysregulated self-renewal of HSPCs is tightly associated with the human blood diseases including leukemia and bone marrow failure (BMF) syndrome [2–4]. Previous studies have illustrated that the genes causative for adult hematopoietic diseases virtually play critical roles in the early hematopoiesis [5,6]. Therefore, exploring the unknown genetic regulators of HSPCs in the hematopoiesis would give us better understanding of the sophisticated mechanisms of hematopoietic diseases in adults. Recently, zebrafish has emerged as an excellent animal model to study the development of hematopoiesis [7–9]. With multiple unique advantages including external fertilization and development, optically transparent embryos, small size and high fecundity, zebrafish is extraordinarily suitable for the unbiased large scale forward genetics screening to identify novel genes regulating HSPCs self-renewal in the embryonic development [10]. More importantly, the hematopoietic anatomy and the critical transcriptional factors involved in the development of hematopoiesis are highly conserved between zebrafish and mammals [1,11]. Similar to mammals, zebrafish hematopoiesis consists of two waves of hematopoiesis, i.e. primitive hematopoiesis and definitive hematopoiesis. The primitive hematopoiesis takes place in the anterior lateral plate mesoderm (ALPM) and intermediate cell mass (ICM) at ~12–14 somites stage, producing primitive macrophages and erythrocytes, respectively [12]. In zebrafish definitive hematopoiesis, HSPCs originate in the ventral wall of dorsal aorta (an equivalent of the aorta-gonad-mesonephros [AGM] in mammals) through endothelium to hematopoietic transition (EHT) from 26 hours post fertilization (hpf) [13,14], and then colonize in caudal hematopoietic tissue (CHT, an equivalent to the fetal liver [FL] in mammal) (at 2 days post fertilization [dpf]), thymus (at 3dpf) and ultimately kidney marrow to support adult hematopoiesis (equivalent to bone marrow (BM) in mammal) (after 5dpf) [15,16]. During fetal hematopoiesis in CHT, the nascent HSPCs undergo extensive proliferation for the pool expansion to support the embryo development [15]. It has been reported that 95–100% of HSPCs are actively cycling in the mouse fetal liver, whereas most of adult HSPCs are in a quiescent state [17]. During DNA replication, the slowed or stalled DNA replication fork, which is termed as DNA replication stress, occurs frequently due to intracellular and extracellular sources including the by-products of cellular metabolism (e.g. dNTP misincorporation, reactive oxygen species etc.), ultraviolet light and chemical mutagens [18,19]. Because the stalled replication forks are vulnerable and the collapse of the forks can result in DNA double strand breaks (DSBs) that are deleterious for the genome stability and cell survival, the DNA replication stress-induced DNA damage needs to be efficiently resolved by DNA damage response (DDR) pathways [18]. The phosphoinositide kinase-related kinase ataxia telangiectasia mutated (ATM) and ATM and Rad3-related (ATR) are two important kinases involved in DDR. ATM mainly participates in the DSBs response, whereas ATR is activated by the single-stranded DNA (ssDNA) damage and DNA replication stress [20]. Recent studies have shed the light on the association between hematopoietic homeostasis and DDR. DDR impairment can lead to progressive BMF and hematopoietic malignancies [21–23]. Fanconi anemia (FA) pathway, which consists of 15 FA genes, mainly participates in repairing the DNA interstrand crosslinks (ICL). Most of the FA genes are associated with the replication fork protection and ATR activation pathway [24,25], and they are causally mutated in BMF or acute myelogenous leukemia [26]. Topoisomerase II β binding protein 1 (TopBP1) is a structurally and functionally conserved protein from yeast to human, which is essential as a scaffold protein in DNA replication initiation and DNA damage checkpoint activation [27–30]. TopBP1 plays a vital role in the DDR, it mainly protects against the ssDNA damage and DNA replication stress through the TopBP1-ATR-Chk1 axis [31–33]. In this process, the stalled replication forks will generate a typical double-stranded DNA-single-stranded DNA (dsDNA-ssDNA) structure. Following the replication protein A (RPA) coating, TopBP1-associated proteins including Rad9-Rad1-Hus1 (9-1-1 complex), ATR interaction protein (ATRIP) and ATR are recruited to the damage locus, then TopBP1 largely activates the ATR kinase activity through its ATR activation domain (AAD), which triggers the phosphorylation of Chk1 and stabilization of replication forks until the stress is resolved [34–38]. Other TopBP1 interacting components also facilitate the establishment of the TopBP1-ATR-Chk1 axis, including the mediator of DNA-damage checkpoint 1 (MDC1) and BRCA1 interacting protein C-terminal helicase (BRIP1, aka, FANCJ) [39–42]. Although the cellular function of TopBP1 has been established, its physiological role, especially the tissue specific requirement, is still largely unknown. TopBP1 null mice are embryonic lethal due to accumulated DNA damage and reduced cell proliferation, which is phenocopied by TopBP1 W1147R knock-in mice with abrogated AAD domain of TopBP1 [43,44]. Moreover, neuronal specific deletion of TopBP1 in mice demonstrates that TopBP1 is essential for neural progenitor cells to survive from the DNA replication stress [45]. Specific disruption of TopBP1 in the lymphoid cells blocks lymphocyte development due to aberrant V(D)J rearrangement [46]. However, whether TopBP1 participates in the HSPCs development is still unknown. Here we report a novel zebrafish mutantcas003, in which HSPCs can be generated normally, but fail thereafter in definitive hematopoiesis. Positional cloning and functional validation indicated that a nonsense mutation-caused C-terminal truncation of TopBP1 was responsible for its subcellular mislocalization and hematopoietic deficits. Disrupted TopBP1-ATR-Chk1 pathway and the accumulation of DNA damage were associated with the HSPCs defect and triggered apoptosis via a p53-dependent pathway. Our findings demonstrate that topbp1 is essential for the HSPCs survival under extensive DNA replication stress during the highly proliferative fetal definitive hematopoiesis. To explore new genes and regulatory mechanisms in vertebrate definitive hematopoiesis, we carried out a large-scale forward genetics screen on ENU-mutagenized F2 families in zebrafish by whole mount in situ hybridization (WISH) using c-myb probe (a key transcription factor and marker of HSPCs) [15,47]. In 5dpf wild-type zebrafish embryos, c-myb was expressed in all hematopoietic tissues including caudal hematopoietic tissue (CHT), thymus, and kidney (Fig 1); whereas homozygous mutantscas003 displayed normal morphogenesis (Fig 1A–A’), but dramatically decreased c-myb expression in CHT, kidney and thymus (Fig 1B–B’), suggesting the expansion of HSPCs was defective. To confirm the defective definitive hematopoiesis in mutantscas003, we further examined the expression of downstream hematopoietic lineage cell markers including ae1-globin (erythrocyte marker), mpx (granulocyte marker), lyz (macrophage marker) and rag1 (lymphocyte marker). The expression of all these markers was substantially decreased in the homozygous mutantcas003 embryos at 5dpf (Fig 1C–G’), which suggested hematopoiesis failure. Recent studies have demonstrated that vasculogenesis and blood flow are essential for HSPCs initiation and maintenance [48,49]. We examined the expression pattern of a pan-endothelial cell marker flk1 at 36hpf and an artery vessel marker ephrinB2 at 26hpf respectively, our results revealed that both of them were intact in mutantcas003 (S1A–S1D Fig). Consistently, heart beating rate and blood circulation were comparable between mutantcas003 and sibling control (S1 and S2 movie). In addition, live observation on mutantcas003, within Tg(fli1: EGFP) transgenic background [50], indicated that the vascular plexus in the CHT region was normal from 2dpf to 5dpf (S1E–S1L Fig). We further investigated the primitive hematopoiesis in mutantcas003. The WISH analysis data demonstrated that the expression of primitive hematopoietic cell markers were identical between siblings and mutantcas003 at 22hpf, including scl (hematopoietic progenitor marker), gata1 (erythrocyte progenitor marker), pu.1 (myeloid progenitor marker), lyz, l-plastin (myeloid cell marker) and mpx (S2A–S2L Fig, quantified in M). Taken together, we concluded that mutantcas003 displayed specific deficiency in definitive hematopoiesis during zebrafish circulation system development. HSPCs are generated from the ventral wall of dorsal aorta through the endothelia to hematopoietic transition (EHT) from 26hpf [13,14], and then migrate to the CHT, a proliferative hematopoietic microenvironment, for pool expansion at 2dpf [15,16]. To figure out when the HSPCs defect initiated in mutantcas003, we performed a time course analysis of c-myb expression from 36hpf to 5dpf. The WISH results demonstrated that the generation of HSC was intact in mutantcas003 as both c-myb and runx1 [51] expression were undisturbed at 36hpf (Fig 2A–B’ and 2F–G’), and the c-myb expression was still intact in the CHT at 2dpf in mutantcas003 (Fig 2C–C’ and 2H–H’). However, mutantcas003 displayed reduced c-myb expression in the CHT at 3dpf (Fig 2D–D’ and 2I–I’), and such defect was more profound at 4dpf (Fig 2E–E’ and 2J–J’), indicating that the HSPCs proliferation or maintenance was impaired in the CHT of mutantcas003. To consolidate this discovery, we carried out quantitative RT-PCR analysis on the c-myb mRNA level in zebrafish tails region including CHT from 2dpf to 5dpf. As expected, the c-myb expression level was attenuated from 3dpf to 5dpf (Fig 2K), which was consistent with the results of WISH analysis. To further confirm these findings, we crossed mutantcas003 with Tg(c-myb: EGFP), in which HSPCs could be visualized by EGFP [52]. Statistically significant reduction of EGPF+ cells was observed at 4dpf (Fig 2L, 2N and 2Q) and was more severe at 5dpf in mutantcas003 (Fig 2L, 2O and 2R) (Due to the long half-life of EGFP protein, the dynamics of c-myb expression indicated via Tg(c-myb: EGFP) was delayed, compared to WISH analysis via c-myb probe [5]). Collectively, our data revealed that, in mutantcas003, neither HSPCs specification in AGM nor their migration to CHT was affected, but their transitory expansion in the CHT was compromised. In order to elucidate the mechanism of hematopoietic failure in mutantcas003, we carried out positional cloning of the mutant. The mutation was first mapped to chromosome 24 by bulk segregation analysis (BSA). With a high resolution mapping approach, the mutation was revealed to be flanked by two closely linked SSLP markers, L0310_5 and R0310_4. The flanked region contained four candidate genes: topbp1 (topoisomerase II β binding protein 1), tmem108, cdv3 and vps41 (Fig 3A). After sequencing cDNA of all 4 genes, we identified a C to T nonsense mutation in topbp1 gene in mutantcas003 (Fig 3B), and confirmed this result through genomic sequencing. This mutation caused an earlier stop codon before the eighth BRCT (BRCA1 C-terminus) domain and a putative C-terminus nuclear localization signal (NLS) of TopBP1 protein (Fig 3C). This truncated form of endogenous TopBP1 (TopBP1cas003) protein was further confirmed by immunoblotting analysis of the CHT of heterozygote (Het cas003) and mutantcas003 embryos at 3dpf (Fig 3D). In order to examine whether the disruption of topbp1 was causative for phenotype of mutantcas003, we injected a validated topbp1 ATG morpholino oligo (MO) (S3A–S3B Fig) into one-cell stage wild-type embryos to block the translation of endogenous topbp1 mRNA (Fig 3E). Since topbp1 MO acted in a dose-dependent manner (S3C Fig), we applied morpholino microinjection causing no morphologic phenotype in the following studies. Topbp1 morphants manifested severe defective definitive hematopoiesis as that in mutantscas003 from 36hpf to 5dpf, while the primitive hematopoiesis at 22hpf, HSPC generation in AGM at 36hpf and vascular system in CHT at 3dpf were all intact in the morphants (Fig 3F–G’ and S4A–S4R Fig). To further consolidate our findings, we performed rescue assay by ectopic expression of wild-type topbp1 in mutantcas003. Consistent with previous report on the instability of topbp1 mRNA [53], ectopic expression of TopBP1 was barely detected at 3dpf after injection of in vitro synthesized topbp1 mRNA into 1-cell stage embryos. In order to overcome this obstacle, we employed a Tol2 transposase-mediated transgenic rescue approach [54]. The ubiquitin promoter (driving ubiquitous expression) and the coding sequence of topbp1WT or topbp1cas003 followed by P2A peptide-mCherry fusion protein (P2A peptide allows self-cleavage of transgenesis efficacy indicator-mCherry without affecting TopBP1 protein) were constructed into the plasmid containing Tol2 arms (hereinafter referred to as ubi: topbp1WT and ubi: topbp1cas003, Fig 3H) [55,56]. After co-injection with Tol2 transposase mRNA and ubi: topbp1WT or ubi: topbp1cas003 constructs into one cell stage mutantcas003 embryos, ubi: topbp1WT driven ectopic expression of wild-type topbp1 could rescue mutantcas003 phenotype at 5dpf (Fig 3I–3K), but not the ubi: topbp1cas003 construct (Fig 3I–3J and 3L). Taken together, the MO phenocopy assays and the wild-type topbp1 rescue assays revealed that the nonsense mutation in topbp1 was the causative mutation in mutantcas003. Meanwhile, we changed the name of mutantcas003 into topbp1cas003. To explore how topbp1 affected maintenance of HSPCs in CHT region, we first investigated the expression pattern of topbp1 during embryonic development. WISH analysis data indicated that topbp1 was a maternal mRNA, and was ubiquitously expressed during embryogenesis (S5A–S5J Fig). Previous reports had showed that topbp1 knock-out or knock-down could result in either cell proliferation blockage or cell apoptosis activation [44,45]. To investigate the cause of HSPCs abrogation, we conducted cell biology assessment of HSPCs in topbp1cas003 mutants in Tg(c-myb: EGFP) transgenic background. Double staining of c-myb and phospho-histone 3 (pH3) showed no significant difference in topbp1cas003 mutants, compared with siblings at 3.5dpf (Fig 4A–D’, quantified in Q), suggesting that the cell cycle of HSPCs was not affected in topbp1cas003 mutants. Furthermore, we performed 5-bromo-2-deoxyuridine (BrdU) incorporation assay on HSPCs, BrdU and EGFP double immunostaining results indicated that there was no significant difference in the percentage of BrdU+ HSPCs between siblings and topbp1cas003 mutants at 3.5dpf (Fig 4E–H’, quantified in R). However, TUNEL assay showed a significant increase of apoptotic EGFP+ HSPCs in CHT region of topbp1cas003 mutants, compared with that in wild-type siblings at 3.5dpf (Fig 4I–L’, quantified in S). At 4dpf, the percentage of apoptotic EGFP+ HSPCs was even more significantly increased in topbp1cas003 mutants in comparison with siblings (Fig 4M–P’, quantified in S), while the number of EGFP+ HSPCs were dramatically decreased (Fig 2N and 2Q). Notably, we could also detect the increased apoptosis in the cranial region and the neural tube in the topbp1cas003 mutants at 3.5dpf and 4dpf. Collectively, we concluded that the increased apoptosis in HSPCs was linked to the defective hematopoiesis in topbp1cas003 mutants. To determine how TopBP1 deficiency triggered apoptosis, we firstly checked the expression of several apoptosis-related genes in the CHT regions of topbp1cas003 mutants at 3dpf. The quantitative PCR results showed that the expression of p53, p21, cyclin G1 and mdm2 were upregulated in the CHT region of topbp1cas003 mutants, indicating the p53 signaling pathway was activated (Fig 5A). Furthermore, we employed ectopic expression of Bcl2 into topbp1cas003 mutants, which was known to inhibit p53 dependent apoptosis pathway [57]. WISH analysis on c-myb expression showed that bcl2 mRNA could partially restore the c-myb expression in the CHT regions of topbp1cas003 mutants (25 out of 43 embryos were partially rescued, S6B–S6D’ Fig, quantified in S6A Fig). To confirm the apoptosis in topbp1cas003 HSPCs mainly depended on the p53 pathway, we crossed topbp1cas003 mutant with the tp53M214K mutant (abbreviated as p53-/- below), which had been reported to abrogate p53 function in apoptosis [58]. Further investigation showed that the expression of c-myb in topbp1cas003 mutants was partially rescued in p53+/- heterozygous background at 4dpf (3/12 embryos were well rescued, 3/12 embryos were partially rescued, Fig 5B–F’), and the rescue effect was more obviously in p53-/- background at 4dpf (7/13 embryos were well rescued, 4/13 embryos were partially rescued, Fig 5B–F’). Taken together, we concluded that the apoptosis of HSPCs in topbp1cas003 mutants was p53-dependent. To further understand the molecular mechanism of HSPC apoptosis which was induced by this particular defective TopBP1 without its 8th BRCT domain and the putative NLS domain, we analyzed the subcellular localization of TopBP1cas003. Confocal imaging showed that flag-tagged TopBP1WT was predominantly localized in the nucleus of transfected HeLa cells (Fig 6A, left column). However, TopBP1cas003 was mistakenly localized in cytoplasm (Fig 6A, middle column), which was consistent with our previous sequence analysis on the lack of putative NLS in TopBP1cas003 (Fig 3C) and immunoblotting analysis on TopBP1WT/TopBP1cas003 protein in cytoplasmic and nucleus fractions of pooled embryos from heterozygote incrossing (S5L Fig). Moreover, addition of SV40 NLS at C terminus of TopBP1cas003 was sufficient to correct TopBP1cas003 subcellular localization defect (Fig 6A, right column). To test whether the hematopoietic deficiency in topbp1cas003 mutants could also be rescued by the nuclear localized TopBP1cas003, we carried out transient transgenesis of topbp1cas003-NLS or topbp1WT (as the positive control) in the topbp1cas003 mutants. WISH results of c-myb at 4dpf indicated that ectopic expression of topbp1cas003-NLS could rescue c-myb expression in topbp1cas003 mutants, as efficient as transgenesis with topbp1WT (Fig 6B–E’, quantified in F). Collectively, we concluded that the loss of NLS in TopBP1cas003 and the failure of nuclear localization directly caused HSPCs deficiency in topbp1cas003 mutants. Previous studies have demonstrated that TopBP1 plays conserved roles as a scaffold protein that is important for DNA replication and DNA damage response (DDR) [27,29,37]. Since the proliferation of HSPCs was not disrupted in topbp1cas003 mutants (Fig 4A–H’), it seemed that the function of TopBP1 in DDR instead of DNA replication was responsible for the HSPCs defect in the mutants. Firstly, we checked the activation of TopBP1-ATR-Chk1 pathway in topbp1cas003 mutants and siblings under the hydroxyurea (HU) treatment, which was extensively applied to mimic DNA replication stress and could activate ATR-Chk1 axis in mammalian cells and zebrafish embryos [30,59,60]. The phospho-Chk1 (pChk1) level in CHT region was significantly increased after 250mM HU treatment from 60hpf to 76hpf (Fig 6G, lane1 and 2). However, the activation of pChk1 was abrogated in topbp1 morphants (Fig 6G, lane 3). Consistently, we also observed dramatic ablation of pChk1 elevation in the CHT of topbp1cas003 mutants compared with wild-type siblings (Fig 6H). Furthermore, we analyzed protein-protein interaction sites in TopBP1 on the basis of previous biochemical and structural investigations [31,41–43,61,62]. The R122, R669 and W1156 sites in TopBP1 are involved in Rad9 or MDC1 interaction and ATR activation, respectively. All these sites are highly conserved among zebrafish, human and mouse (S7 Fig), and they are critical for TopBP1-ATR pathway [31,41,61,63]. Transient transgenesis of TopBP1ΔAAD, TopBP1W1156R, TopBP1R122E, TopBP1R669E and TopBP1WT (as positive control) in topbp1cas003 mutants was analyzed for hematopoiesis monitored by c-myb WISH. None of these mutated TopBP1 could rescue the hematopoietic failure in topbp1cas003 mutants, compared with TopBP1WT (Fig 6I), indicating that ATR activation function of TopBP1 was essential for HSPCs survival in topbp1cas003 mutants. Taken together, these data implied that the blockage of TopBP1-ATR-Chk1 pathway was correlated to the defective HSPCs in topbp1cas003 mutants. Since TopBP1-ATR-Chk1 axis was disrupted in topbp1cas003 mutants, the unresolved DNA replication stress would result in collapse of replication forks, which could introduce DNA double-stranded breaks ultimately [18]. To check whether the apoptosis of HSPCs was due to the accumulation of DNA damage in CHT region, we carried out fluorescent c-myb WISH analysis and immunostaining with phosphorylated histone H2AX (γH2AX) antibody, which was a typical DNA damage marker [64], from 39hpf to 3.5dpf. Interestingly, we couldn’t detect any γH2AX+ cells in AGM region at 39hpf in both topbp1cas003 mutants and siblings, but γH2AX+ HSPCs emerged in CHT region in topbp1cas003 mutants at the same stage (S8A–S8B Fig). Moreover, γH2AX+ HSPCs were accumulated in CHT region of topbp1cas003 mutants afterward (S8C Fig), and they were obviously increased at 3.5dpf, (Fig 7A–H’, S8C Fig) indicating the DNA damage was indeed accumulated in HSPCs in topbp1cas003 mutant. In addition, we could also observed several γH2AX+ cells in neuronal tissue (Fig 7G), which was consistent with previous investigation [45]. Furthermore, the immunoblotting of γH2AX within CHT regions of topbp1cas003 mutants at 3dpf also showed an increase of DNA damage (Fig 7I). Collectively, we found that DNA damage was accumulated in HSPCs in topbp1cas003 mutants. To further examine whether the hematopoietic failure was due to the defective DDR upon DNA replication stress in topbp1cas003 mutants, we challenged the embryos with HU. Indeed, high concentration treatment from 52hpf to 76hpf directly caused embryonic lethality in topbp1cas003 mutants (over 65%), however the effect on wild-type siblings was much milder (<15%) (S9A and S9B Fig) [60]. Furthermore, we carried out a recovery assay in the HU-treated zebrafish embryos (Fig 7J). Interestingly, despite of a suppression by HU treatment, the c-myb expression was recovered in wild-type sibling embryos after challenge removal (Fig 7K, 7L–L’, 7N–N’ and 7P–P’). In contrast, the c-myb expression level was not recovered, but decreased further in the HU-treated topbp1cas003 mutant embryos (Fig 7K, 7M–M’, 7O–O’ and 7Q–Q’). Taken together, all these observations suggested that the HSPCs in CHT of topbp1cas003 mutants were defective in replicative DNA damage response and they eventually underwent apoptosis through a p53-dependent signaling pathway. In this study, we reported a novel zebrafish mutant topbp1cas003, which manifested severe defect in definitive hematopoiesis. The reduction of HSPCs started from 3dpf, which was mainly due to the increased p53-dependent apoptosis, rather than proliferation deficiency. Genetic assessment revealed that a nonsense mutation in topbp1 gene was causative for the hematopoiesis failure. Further investigation revealed that the mutated TopBP1cas003 protein was decreased and mislocalized from nucleus to cytoplasm which compromised the DNA damage response. As a result, it led to accumulated DNA damage that triggered sequential apoptosis of HSPCs in topbp1cas003 mutants. In zebrafish definitive hematopoiesis, HSPCs undergo extensive proliferation in the CHT region around 3dpf, during which the replication stress, characterized by the stalled replication forks, can be induced by various endogenous and exogenous factors [18]. The stalled replication forks will generate typical dsDNA-ssDNA structure, followed with proper loading of RPA, ATR-ATRIP and 9-1-1 complex [37]. Sequential recruitment of TopBP1 can largely activate ATR kinase activity, and the latter will phosphorylate downstream molecules including Chk1. Activated Chk1 stabilizes the replication forks and arrests cell cycle in order to leave enough time for DNA damage repair machinery to work and to restart the replication fork, so that HSPCs can survive the stress and finish their pool expansion (Fig 8). The quantitative analysis and WISH results demonstrated that nonsense mutation in topbp1 might lead to nonsense mediated mRNA decay. The expression level of topbp1 was decreased over 80% in the whole embryo and about 50% in CHT of topbp1cas003 mutants (S5M–S5Q Fig). Although around 50% TopBP1cas003 protein remains in CHT, it was mistakenly localized in cytosol, while TopBP1WT was mainly in nucleus to play its role in DDR (S5L Fig). Our results suggest that TopBP1cas003 is decreased and fails in its nucleus entry due to the loss of its C-terminal NLS, abrogating the later ATR/Chk1 activation. In topbp1cas003 HSPCs, the unresolved stalled replication forks would collapse and generate multiple DNA fragile sites, which can induce dsDNA break [18]. As a result, p53-dependent apoptosis is elevated in topbp1cas003 HSPCs, impairing the HSPCs pool severely (Fig 8). Recently an improved clustered regularly interspaced short palindromic repeats (CRISPR)/ CRISPR-associated proteins (Cas9) system with custom guide RNAs (gRNAs) and a zebrafish codon-optimized Cas9 protein showed high mutagenesis rate in zebrafish, which could even generate biallelic mutations in the F0 generation [65,66]. In order to confirm that the deficiency of TopBP1 could disrupt the development of HSPCs, we adapted this optimized CRISPR/Cas9 system to obtain other topbp1 zebrafish mutants (S10A–S10B Fig). Some of the topbp1 Cas9 injected wild-type embryos displayed dramatically decreased c-myb expression as same as topbp1cas003 mutant at 4dpf (S10C–S10D’ Fig). And this phenotype could be reached in higher efficiency when the injected embryos were generated from the outcross between topbp1cas003 heterozygote and wild-type fish (S10E–S10F’ Fig). Conclusively, these data provided additional evidence that definitive HSPCs were defective in the TopBP1 loss-of-function embryos. It is an intriguing finding that topbp1 plays an essential role in proliferative tissues, especially in the definitive hematopoiesis without affecting the morphogenesis at the early stage, whereas its transcripts were ubiquitously distributed in the embryogenesis (S5 Fig), and TopBP1 knockout mice were reported to be lethal at the peri-implantation stage [44]. The WISH analysis showed maternal expression of topbp1 (S5A Fig), suggesting that homozygote topbp1 mutant embryos can inherit wild-type topbp1 mRNA from the female parents to support its early development until zygotic topbp1 expresses latter in the development. Nevertheless, we attempted to figure out whether topbp1 was expressed and functional in the HSPCs. Quantitative PCR analysis on the CD41+ cell population in the tail region of Tg(CD41: EGFP) embryos, which was reported to be an enriched population of HSPCs at 5dpf [67,68], showed that the level of topbp1 mRNA was 3-fold enriched in CD41+ cells, compared to cells in the whole tails, demonstrating its expression in HSPCs (S5K Fig) [5]. Furthermore, due to the lack of definitive hematopoiesis-specific promoter, we used hemangiogenic promoter lmo2, which was also expressed in definitive HSPCs, to drive the ectopic expression of wild-type topbp1 into topbp1cas003 mutants [52,69], we could indeed observe the expression of mCherry driven by lmo2 promoter in CHT region at 5dpf, and this construct could partially rescue the HSPCs deficits at 5dpf (S11 Fig). In addition, the vascular plexus in CHT region was normal in topbp1cas003 mutants or morphants from 2dpf to 5dpf (S1E–S1L Fig and S4O–S4R Fig), and low dose microinjection of topbp1 morpholino was sufficient to induce definitive hematopoiesis deficits in CHT without affecting the primitive hematopoiesis and vascular system in wild-type embryos (S3C–S3D Fig, S4 Fig). Taken all these data together, we concluded that TopBP1 played an essential and HSPC-intrinsic mechanism during definitive hematopoiesis. It is intriguing that whether the truncated TopBP1 can potentially function as a dominant negative protein. Ectopic expression of cas003 mutant form of TopBP1 (TopBP1cas003) driven by ubiquitin promoter was performed in wild-type fish, and it did not cause defective definitive hematopoiesis (S12 Fig). The possible reason for this phenomenon was that the mutated TopBP1 could not enter nucleus to compete with wild-type TopBP1. Meanwhile, the hematopoietic phenotype of topbp1cas003 heterozygotes was checked, and no HSPCs defect was observed. Taking these results together, we concluded that TopBP1cas003 could not function as a dominant negative form. In definitive hematopoiesis, nascent HSPCs seldom proliferate in AGM region, but they become active in cell cycle and undergo extensive proliferation in CHT region supported by niche cells, meanwhile, they have to overcome DNA replicative stress [13,15,18]. BrdU incorporation assays within Tg(c-myb: EGFP) embryos confirmed that HSPCs underwent high proliferation at a constant rate from 2dpf to 5dpf, although the expansion of neural tube cells was gradually attenuated (S13 Fig). As a result, the defect in HSPCs was more profound after 3dpf in the topbp1cas003 mutant. Consistently, we indeed found obvious accumulation of γH2AX positive cells (2.5dpf) and increased apoptotic cells (3.5dpf) in cranial and neuron tube tissues of topbp1cas003 mutant, which was in agreement with previous observations in neuron-specific TopBP1 knock-out mice [45]. Besides, some of homozygote topbp1 mutant embryos developed smaller head and eyes after 6dpf, and all of them eventually died around 10–20 dpf. Previous works within zebrafish mutants revealed several genes and pathways which were critical for the HSPCs development in CHT region, including genes involved in mitotic spindle assembly, maintenance of centrosome integrity and mitotic progression; pre-mRNA processing; sumoylation of genes participating in DNA replication or cell cycle regulation [5,70,71]. All these genes were indispensable for cell to complete proliferation or division. Because the HSPCs were highly proliferative in CHT, these data depict a picture that the HSPCs in fetal stage are extremely sensitive to the disruption of genes participating in various processing to complete cell division successfully and faithfully. As the DDR pathway is essential for genomic fidelity and stability during DNA replication, our work revealed that DDR pathway is also critical for HSPCs development in fetal stage. It has been reported that Fanconi anemia pathway is critical for the repair of DNA cross-link damage [26]. Biallelic mutations in any of 15 FANC genes will result in Fanconi anemia (FA), which can most frequently develop into inherited bone marrow failure (BMF) syndrome [72]. The work of Raphael Ceccaldi et al. revealed that the FA patients showed profound HSPCs defect before the onset of BMF [73]. The p53-p21 axis, triggered by replicative stress, was highly elevated in FA HSPCs, and the p53 silence can rescue hematopoietic deficits [73]. They also pointed out that p53 activation, caused by unresolved cellular abnormality, may be the signaling mechanism for inherited BMF, and the p53 activation was commonly found in other types of inherited BMF syndromes, such as Diamond Blackfan anemia (DBA) and dyskeratosis congenital (DC) [73]. HSPCs in topbp1cas003 mutants manifested similar features as that in FA (Fig 5), whether topbp1 could be a putative pathogenic gene in human BMF syndrome needs further investigation. Zebrafish fancd2 morphant exhibited developmental abnormalities and p53-dependent apoptosis, however its hematopoietic phenotype had not been extensively investigated [57]. The emi1 homozygous mutants showed disrupted genomic integrity and hematopoiesis failure [74]. Studies on topbp1cas003 mutants revealed that DNA damage and apoptosis signaling was accumulated in the HSPCs of topbp1cas003 homozygous embryos, which linked to the hyper-activated p53-p21 axis (Fig 5) and failed ATR/Chk1 activation (Fig 6). Furthermore, TopBP1-involved c-myb regulated DDR pathway was proposed by recent studies on castration-resistant prostate cancer [75]. HU treatment of the developing zebrafish further emphasized the importance of DNA damage response and repair pathway for HSPCs survival during high proliferation stage. Collectively, we demonstrated a novel and essential role of TopBP1 in HSPCs during their rapid proliferation in fetal hematopoiesis. Due to the dramatic definitive hematopoiesis phenotype in embryogenesis, topbp1cas003 mutants provide a unique model for the mechanism study and small molecular chemical screen on BMF-like hematopoiesis failure, which is caused by defective replicative DNA damage response. The zebrafish facility and study were approved by the Institutional Review Board of the Institute of Health Sciences, Shanghai Institutes of Biological Sciences, Chinese Academy of Sciences (Shanghai, China), and zebrafish were maintained according to the guidelines of the Institutional Animal Care and Use Committee. Wild-type (WT) zebrafish strains Tubingen (TU) and WIK, the transgenic zebrafish line Tg(c-myb: EGFP) [52], Tg(fli1: EGFP) [50], Tg(CD41: EGFP) [76], the mutant zebrafish line tp53M214K/M214K [58] were maintained as previously described [77]. For the forward genetics screen, WT TU zebrafish line was treated with ethylnitrosourea (ENU, Sigma) to generate mutants, the screen approach was performed as previously described [78,79]. The desired mutants within F3 generation were identified by the whole-mount in situ hybridization (WISH) using c-myb probe at 5dpf. For the chemical treatment, the hydroxyurea (HU, Sigma) was dissolved with distilled water into 1M and stored at -20℃. The embryos were treated with 250mM HU as the indicated procedures in the egg water at 28.5℃ [59,60]. To prevent the formation of melanin pigment, the embryos were incubated in egg water containing 0.045% 1-phenyl-2-thiourea (PTU, Sigma) after gastrulation stage. The embryos were collected at the desired stages [80]. Positional cloning was carried out with WIK line as previously described [81]. Firstly, the mutation was mapped to chromosome 24 by bulk segregation analysis (BSA) with simple sequence length polymorphism (SSLP) markers. Through high resolution mapping analysis on 1041 mutants, the mutation was finally flanked by two SSLP markers, L0310_5 and R0310_4. The candidate genes in this range were sequenced with wild type sibling and mutant cDNA, and the putative mutation was confirmed by genomic DNA sequencing. The primers used in the positional cloning were provided in supplemental S1 Table. Most experiments in this study were carried out with the embryos generated by the incross of mutantcas003 heterozygote pairs (TU/WIK background) used in the positional cloning if possible. The mutants can be identified by flanked SSLP markers, such as Z9852 and R0306_4. Alternatively, the mutants can be distinguished by restriction fragment length polymorphism (RFLP) using EcoP15I (NEB), the RFLP primers were provided in supplemental data (S1 Table). To construct Tol2 transgenesis vectors, the ubiquitin promoter [55] or lmo2 promoter [69] followed by P2A [56] and in-frame mCherry was cloned into modified Tol2 backbone [82]. The vectors were referred as pUbi-Tol2 or pLmo2-Tol2 below. The genes of interest can be inserted between the promoter and P2A. Zebrafish topbp1WT or topbp1cas003 were amplified and inserted into pUbi-Tol2 or pLmo2-Tol2 vectors. To generate the mutated forms of topbp1, the mutagenesis was carried out following QuikChange mutagenesis kit instruction using pUbi-topbp1WT-Tol2 plasmid as the template. The region of TopBP1 (984–1206) are the putative ATR activation domain (AAD) between BRCT6 and BRCT7. In TopBP1ΔAAD, the coding sequence of TopBP1 (1083–1159) containing conserved RQLQ and WDDP sequences are deleted [31]. The fragment of topbp1 (-9–692) was amplified and inserted into the pCS2+ vector for in situ probe preparation. To construct topbp1 MO effect evaluation plasmid, a DNA fragment containing topbp1 ATG MO targeting site was inserted into the upstream of EGFP coding region in pCS2+. Zebrafish topbp1WTand topbp1cas003 were cloned into pCMV4-FLAG-4 vector (Sigma). The SV40 NLS (nuclear localization signal) sequence (5’-CCAAAAAAGAAGAGAAAGGTA-3’) [83] was firstly cloned into pCMV4-FLAG-4 vector in the 3’ end of FLAG tag, and then the topbp1cas003 sequence was inserted into the pCMV4-FLAG-NLS plasmid. All of the primers used were listed in S1 Table. The mRNA was synthesized in vitro by SP6 mMessage mMachine Transcription Kit (Ambion). The topbp1 gRNA was synthesized as described [66]. The information of the topbp1 gRNA target site was shown in S1 Table. The zebrafish optimized Cas9 mRNA was synthesized in vitro from the pCS2-nCas9n plasmid (addgene, #47929) as described [65]. bcl2-egfp mRNA (~100pg) was injected into 1-cell stage embryos [54]. For the ectopic-expression, Tol2 transposon-mediated transient transgenesis was applied and performed as previously described [84]. A series of topbp1 transgene constructs within Tol2 vectors (~40 ng/μl) were mixed with transposase mRNA (~60 ng/μl) and 0.2 M KCl, and then injected into 1-cell stage embryos, respectively [85]. The volume of the mixture injected was about 0.5nL. The topbp1 ATG morpholino oligo (MO) (5’-CCTTGCTGGCTTTCGACATGGTGAC-3’) and control morpholino (5’- CCTCTTACCTCAGTTACAATTTATA-3’) were synthesized by Gene Tool company and was injected into 1-cell stage embryos. For Cas9 assay, topbp1 gRNA (50pg) and Cas9 mRNA (150pg) were co-injected into one-cell stage embryos. The T7EI assay was performed as described [65]. c-myb, runx1, ae1-globin, mpx, lyz, rag1 and topbp1 probes were transcribed in vitro by T3 or T7 polymerase (Ambion) with Digoxigenin RNA Labeling Mix (Roche). One color WISH was performed as described previously [54]. Images were photographed by the Nikon SMZ1500 microscope with Nikon DXM 1200F CCD or Olympus SZX16 microscope with Olympus DP80 CCD. c-myb RNA and immuno-fluorescence double staining was carried out as described previously [70]. For the immunostaining, rabbit anti-pH3 antibody (1:500, Santa Cruz) and rabbit anti-γH2AX antibody (1:500, gift from Dr. James Amatruda, University of Texas Southwestern) were used. The embryos were stained with goat-Alexa Fluore488-conjugated anti-rabbit secondary antibody (1:500, Invitrogen). DAPI (1:500, Beyotime) staining was carried out along with the secondary antibody incubation if necessary. The 3.5dpf topbp1cas003 mutant/Tg(c-myb:EGFP) or sibling embryos were soaked in egg water containing 10mM BrdU (Sigma)/15% DMSO for 30 minutes at 28.5℃ or injected with 1nL 30mM BrdU into the yolk sac. Then they were transferred into fresh egg water and incubated for 2 hours. After fixation in 4% paraformaldehyde (PFA, Sigma), the embryos were dehydrated with methanol and stored at -20℃ overnight. For BrdU immunostaining, the rehydrated embryos were digested with Proteinase K(12 μg/ml, Roche) at 30℃ for 28 minutes and treated with acetone at -20℃ for 30 minutes. After re-fixation with 4% PFA, the embryos were blocked with the block solution (PBS + 0.3% Triton-X -100 +1% DMSO+ 10 mg/ml BSA+10% normal goat serum) for 2 hours at RT. The embryos were then incubated with anti-GFP Rabbit Serum (1:500, Invitrogen) followed by goat-Alexa Fluore488-conjugated anti-rabbit secondary antibody (1:500, Invitrogen) incubation. 2N HCl was used to treat the embryos for 1 hour at room temperature (RT). After that, the embryos were stained with mouse anti-BrdU primary antibody (1:50, Roche) and goat-Alexa Fluore546-conjugated anti-mouse secondary antibody (1:500, Invitrogen). TUNEL assay was performed with the In Situ Cell Death Detection Kit TMR red (Roche). Similar to the BrdU immunostaining, 3.5dpf and 4dpf topbp1cas003 mutant/Tg(c-myb:EGFP) or sibling embryos were fixed with 4% PFA. After methanol dehydration, rehydration, Proteinase K digestion and acetone treatment, the embryos were permeated with permeabilisation solution (0.5% Triton X–100, 0.1% sodium citrate in PBS) at RT for 4 hours. Then the embryos were stained with the TUNEL Kit (100ul, enzyme: labeling solution = 1:9) at 37℃ for 2 hours. Finally, the EGFP immunostaining was carried out as described above. The CD41+ cells were sorted from the tails of Tg(CD41: EGFP) embryos including the CHT region at 5dpf as previously described [86,87]. The total RNA was extracted from TRIzol (invitrogen) dissolved zebrafish whole embryos or the tails including CHT region or the sorted cells, and then transcribed into cDNA by PrimerScript RT Master Mix (TaKaRa). The quantitative PCR was carried out with SYBR Green Real-time PCR Master Mix (TOYOBO) with ABI 7900HT real-time PCR machine, and analyzed with Graphpad 5.1 software. The primers used were listed in S1 Table. HeLa and HEK293T cells were maintained in DMEM with 10% Fetal Bovine Serum (FBS) and penicillin-streptomycin antibiotics (1:100). Plasmid transfection was carried out with Lipofectamine 2000 (Invitrogen) according to manufacturer’s instruction. The immunostaining was carried out in HeLa cells as previously described [70]. FLAG-topbp1WT, FLAG-topbp1cas003 and FLAG-topbp1cas003-NLS plasmids were transfected into HeLa cells. Mouse anti-FLAG primary antibody (1:1000; Genomics Technology) and goat-Alexa Fluore488-conjugated anti-mouse secondary antibody (1:500) were used for immunostaining. DAPI (1: 500, Beyotime) was applied for nucleus staining. To extract the protein from the cell line, the cells were homogenized directly with 2 X SDS sample buffer and boiled for 5 minutes at 95℃. To obtain fish protein from the CHT region, the tails of embryos including the CHT region were cut down, then ultrasonicated in RIPA lysis buffer (50mM Tris(pH7.4), 150mM NaCl, 1% NP-40, 0.5% sodium deoxycholate, 0.1% SDS). After centrifugation at 12000rpm for 15 minutes, the supernatant was mixed with 2XSDS sample buffer and boiled for 10 minutes. Cytoplasmic and nuclear extracts were prepared from the 3dpf embryos with Nuclear and Cytoplasmic Protein Extraction Kit (Beyotime) according to the manufacturer’s instruction. The immunoblotting was carried out as previously described [85], with rabbit anti-phospho-Chk1 (Ser345) (133D) antibody (Cell Signaling Technology), rabbit anti-γH2AX antibody, rabbit anti-zebrafish TopBP1 antibody (generated by 840–940 amino acid of zebrafish TopBP1 protein as antigen), mouse anti-GAPDH antibody (1D4) (Santa Cruz), mouse anti-alpha-tubulin antibody (Sigma) or rabbit anti-Histon3 (H3) antibody (Abcam). Images of zebrafish immunofluorescence staining or live transgenic embryos were taken by Olympus FV1000 scanning confocal microscope. The embryos were mounted in 1% low-melt agarose in a self-made 35mm coverslip-bottom dish. The confocal images were captured with an UPLSAPO 20X or 60X objective. To obtain images of HeLa cells immunostaining, the slides were directly immersed in the PBS solution in a 10cm dish. The images were captured with an UPLSAPO 40X objective. The transient transgenesis embryos and embryos for bright field imaging were anesthetized with 0.03% Tricaine (Sigma-Aldrich), mounted in 3% methylcellulose and imaged using a Zeiss Axio Zoom. V16 microscope equipped with a Zeiss AxioCam MRm digital camera. Data were analyzed with the Graphpad Prism 5 software using the two-tailed Student’s t-test. The plot error values were calculated by standard error of the mean (SEM). All data in this study were repeated for at least twice.

10.1371/journal.pgen.1004509

Wnt-Mediated Repression via Bipartite DNA Recognition by TCF in the Drosophila Hematopoietic System

The Wnt/β-catenin signaling pathway plays many important roles in animal development, tissue homeostasis and human disease. Transcription factors of the TCF family mediate many Wnt transcriptional responses, promoting signal-dependent activation or repression of target gene expression. The mechanism of this specificity is poorly understood. Previously, we demonstrated that for activated targets in Drosophila, TCF/Pangolin (the fly TCF) recognizes regulatory DNA through two DNA binding domains, with the High Mobility Group (HMG) domain binding HMG sites and the adjacent C-clamp domain binding Helper sites. Here, we report that TCF/Pangolin utilizes a similar bipartite mechanism to recognize and regulate several Wnt-repressed targets, but through HMG and Helper sites whose sequences are distinct from those found in activated targets. The type of HMG and Helper sites is sufficient to direct activation or repression of Wnt regulated cis-regulatory modules, and protease digestion studies suggest that TCF/Pangolin adopts distinct conformations when bound to either HMG-Helper site pair. This repressive mechanism occurs in the fly lymph gland, the larval hematopoietic organ, where Wnt/β-catenin signaling controls prohemocytic differentiation. Our study provides a paradigm for direct repression of target gene expression by Wnt/β-catenin signaling and allosteric regulation of a transcription factor by DNA.

During development and in adult tissues, cells communicate with each other through biochemical cascades known as signaling pathways. In this report, we study the Wnt signaling pathway, using the fruit fly Drosophila as a model system. This pathway is known to activate gene expression in cells receiving the Wnt signal, working through a transcription factor known as TCF. But sometimes Wnt signaling also instructs TCF to repress target gene expression. What determines whether TCF will positively or negatively regulate Wnt targets? We demonstrate that activated and repressed targets have distinct DNA sequences that dock TCF on their regulatory DNA. The type of site determines the output, i.e., activation or repression. We find that TCF adopts different conformations when bound to either DNA sequence, which most likely influences its regulatory activity. In addition, we demonstrate that Wnt-dependent repression occurs robustly in the fly larval lymph gland, the tissue responsible for generating macrophage-like cells known as hemocytes.

It is a common theme in gene regulation that the same transcription factor (TF) can directly activate or repress target gene expression, increasing the transcriptional complexity these TFs can achieve [1], [2]. There are several mechanisms by which TFs exhibit this dual regulation. These include TFs interfering with the binding of other TFs to DNA or co-activators [3]–[5] or signal-dependent changes of co-regulators bound to the TF [6]–[8]. In many cases, specific differences in the nucleotide sequence of the cis-regulatory modules (CRMs) targeted by these TFs influence the transcriptional outcome. The sequence specificity that determines the activation/repression choice of TFs can occur in the TF binding sites themselves, or the surrounding sequences. Several TFs that appear to be intrinsic transcriptional activators can also repress transcription when bound to CRMs in conjunction with other TFs [9]–[11]. In the case of the Drosophila NF-κB family member Dorsal, mutation of TF sites flanking Dorsal binding sites converts CRM reporters that are repressed by Dorsal into ones that are activated [12], [13]. For other CRMs regulated by nuclear receptors [14], [15], P53 [16], the POU TF Pit1 [17] and some Smads [18], [19], it is the type of the TF binding site itself that determines output. For the latter cases, it has been proposed that the DNA binding site allosterically regulates the TF, leading to differential recruitment of co-regulators [17], [20]. Dual regulation of transcription has also been seen in Wnt/β-cat (hereafter called Wnt) signaling, an important cell-cell communication pathway that plays various roles throughout animal development, stem cell biology and disease [21]–[23]. Wnt-induced nuclear accumulation of β-catenin (β-cat) is a key feature of this pathway. Once in the nucleus, β-cat is recruited to CRMs hereafter referred to as Wnt-dependent CRMs (W-CRMs), where it facilitates regulation of Wnt transcriptional targets [24], [25]. The best-characterized TFs that recruit β-cat to W-CRMs are members of the T-cell factor (TCF) family [26]. Studies with synthetic W-CRMs containing multiple copies of high affinity TCF binding sites and mutagenesis studies of binding sites in many endogenous W-CRMs support the view that TCF/β-cat complexes are powerful transcriptional activators [26]–[28]. In many cases, TCFs also mediate default repression by binding to W-CRMs in the absence of signaling [23], [28]. This regulation is commonly referred to as the TCF “transcriptional switch” [1], [28]. While vertebrate TCFs have become more specialized for either default repression or β-cat-dependent activation, invertebrate TCFs such as Drosophila TCF/Pangolin (TCF/Pan) mediate both sides of the transcriptional switch [26], [28]. All TCFs contain a sequence-specific DNA binding domain called the HMG domain, whose high affinity consensus is SSTTTGWW, (S = C/G, W = A/T) [29]–[31]. Invertebrate TCFs and some vertebrate TCF isoforms contain a second DNA binding domain, C-terminal to the HMG domain, known as the C-clamp [26], [32]. C-clamps recognize GC-rich motifs called Helper sites, and this interaction is essential for the activation of many W-CRMs [33], [34]. These data support a model where C-clamp containing TCFs recognize W-CRMs in a bipartite manner, via HMG domain-HMG site and C-clamp-Helper site interactions [26]. While TCF/β-cat complexes are commonly associated with transcriptional activation, there are a few cases where they appear to directly repress target gene expression [35]–[38]. The HMG sites in these repressed W-CRMs are very similar to those found in activated targets. In one case, TCF/β-cat may achieve repression by interfering with the binding of another activating TF [35]. For another target, TCF/β-cat may form a complex with the transcriptional repressor Brinker, and HMG and Brinker binding sites are both required for the repression [38]. In contrast to the aforementioned examples, we previously showed that TCF/Pan mediated Wnt-dependent repression of a W-CRM from the Ugt36Bc locus through HMG sites with a consensus that is distinct (WGAWAW) from classic ones [39]. In addition to mediating Wnt-induced repression, TCF/Pan is required for basal expression of Ugt36Bc in the absence of signaling [39]. This suggests a “reverse transcriptional switch” occurs at Ugt36Bc compared to the switch seen in activated targets. Instead of TCF/Pan default repression and Wnt-dependent activation, the reverse switch consists of TCF/Pan basal activation and Wnt-dependent repression. In this report, we have explored the mechanism of this reverse switch/direct repression mechanism by TCF/Pan and Wnt signaling in more detail. We identified another repressed W-CRM from the Tiggrin (Tig) gene, which contains functional WGAWAW sites bound by TCF/Pan. Regulation of the Ugt36Bc and Tig W-CRMs by TCF/Pan requires the C-clamp, which binds to Helper-like (r-Helper) sites adjacent to the WGAWAW sites. Swapping these sites in the Tig W-CRM to classic HMG and Helper sites converts the W-CRM into one that is activated by Wnt signaling. Conversely, an activated W-CRM from the naked cuticle (nkd) locus was converted to a repressed W-CRM by replacing its classic HMG-Helper pairs with pairs from the Tig W-CRM. Partial protease digestion indicates that TCF/Pan adopts a different conformation when bound to classic or repressive sites, supporting allosteric regulation of TCF/Pan by its binding sites. In addition, we have extended this work from cell culture to the fly, showing that WGAWAW and r-Helper sites mediate basal activation and Wnt-induced repression in the larval lymph gland (LG). Wnt signaling is known to play an important role in regulating hematopoiesis in the LG [40]. Thus, our work provides insight into how TCF/Pan can activate and repress Wnt transcriptional targets, and extends the TCF reverse transcriptional switch mechanism to a physiologically relevant context. Ugt36Bc was originally identified as a candidate for repression by Wnt signaling from a microarray screen performed in Kc167 (Kc) cells [39], a Drosophila cell line likely of hemocytic origin [41]. Several other repressed targets were also identified in this screen, including Tig [39], which encodes an extracellular matrix protein that serves as a PS2 integrin ligand [42], [43]. Tig expression was repressed by DisArmed, a mutated version of Armadillo (Arm, the fly β-catenin) which is defective in gene activation but is still competent for repression [39]. While these results are consistent with Tig being directly repressed by Wnt signaling, the cis-regulatory information responsible for Wnt regulation of Tig expression had not been identified. The Tig locus is compact, with a small (∼1 kb) intergenic region and six introns, only the first of which is larger than 500 bp (Figure 1A). The intergenic region possibly also contains elements driving the expression of the adjoining gene, Fic domain-containing protein (Fic), a gene involved in fly vision [44]. Fic was expressed in Kc cells, but was not regulated by Wnt signaling (Figure S1B). A 1.8 kb fragment containing the intergenic region between Fic and Tig, as well as the first exon and intron and part of the second exon of Tig was cloned upstream of a luciferase gene reporter (Figure 1C). This reporter (Tig1) was repressed 2–5 fold by Axin RNAi in Kc cells, similar to the fold regulation of endogenous Tig mRNA (Figure 1B and 1C). Expression of a stabilized form of Arm (Arm*) [45] also repressed the Tig1 reporter to a similar degree (Figure S1C). These results suggest that Tig1 contains most of the regulatory information required for Wnt regulation of the Tig gene. To better understand which regions were responsible for basal expression and Wnt-dependent repression of Tig, smaller fragments of the regulatory sequences in Tig1 were analyzed. In some cases (Tig2–Tig4), sequences were cloned upstream of the hsp70 core promoter, which is unregulated by Wnt signaling [33], [39], [45], while the Tig5 reporter used the endogenous Tig promoter. These reporters (Tig2–Tig5) all had basal expression higher than the hsp70 promoter control (Figure 1C). Much of the repressive activity appeared to be contained in a 578 bp fragment containing part of the first exon and most of the first intron (Tig3). However Tig1 was used for further functional experiments, to retain the endogenous promoter and additional cis-regulatory information of the Tig locus. TCF/Pan has previously been shown to activate Ugt36Bc and Tig in the absence of signaling, and to be required for Wnt-mediated repression [39]. To determine whether the C-clamp of TCF/Pan was required for these activities, RNAi rescue experiments were performed. Endogenous TCF/Pan was depleted from Kc cells using dsRNA corresponding to the 3′ UTR of TCF/Pan. Cells were then transfected with Ugt36Bc or Tig reporters, as well as expression plasmids for TCF/Pan, either wild-type control or a C-clamp mutant where five amino acids have been altered [33]. Wnt signaling was activated using Arm*. In control TCF/Pan depleted cells (transfected with empty vector), the Tig and Ugt36Bc reporters were not regulated by Arm* (Figure 2A and 2B). Wild-type TCF/Pan elevated basal expression and enabled significant repression by Arm*. In contrast, the C-clamp mutant neither activated nor repressed the reporters (Figure 2A and 2B). These data suggest that the C-clamp is required for TCF/Pan-dependent basal activity and Wnt-mediated repression of both reporters. To ensure that the C-clamp mutant TCF/Pan was functional, a synthetic reporter containing multimerized HMG sites and lacking Helper sites (6×TCF) was also examined (Figure 2C). As previously reported [33], the C-clamp mutant was able to rescue 6×TCF activation by Wnt signaling, albeit not completely under the conditions used (Figure 2C). Nonetheless, these data support an important role for the C-clamp in TCF/Pan regulation of the Ugt36Bc and Tig. A search through the Tig1 sequences using the open access program Target Explorer [46] failed to find classic HMG sites (SSTTTGWWS) [29], [31] or the Helper sites characterized in activated fly W-CRMs (GCCGCCR) [33]. However, the first intron of Tig contained several sequences that were similar to sites in the Ugt36Bc W-CRM that were footprinted by the HMG domain of TCF/Pan [39]. Therefore, similar footprinting of a 300 bp region of the Tig intron containing these putative sites (Figure 3A) was performed, comparing the footprint of GST and GST-HMG domain recombinant proteins (see Material and Methods for details). Several regions of this Tig regulatory DNA were protected by the HMG domain (Figure S2A), two of which are similar to the three WGAWAW sites previously found in the Ugt36Bc W-CRM [39]. Together, the five Tig and Ugt36Bc motifs defined a consensus of RNWGAWAW (Figure 3C). In addition, the regions of the Ugt36Bc and Tig loci containing the WGAWAW sites were footprinted with GST-HMG and GST-HMG-C-clamp, to identify C-clamp bound sequences. Three additional regions were protected only in the presence of the C-clamp (Figure 3B, S2A and S3). Alignment of these regions revealed a consensus of KCCSSNWW (K = G/T; Figure 3C), which was distinct from the classic Helper sites found in activated W-CRMs. These motifs are hereafter referred to as repressive-Helper (r-Helper) sites and the HMG bound sequences as WGAWAW sites. The r-Helper sites in the Ugt36Bc and Tig W-CRMs are adjacent to the WGAWAW sites (Figure 3A), similar to the HMG-Helper clustering in activated W-CRMs [33], [34]. To test whether these motifs act together to form a high affinity binding site for TCF/Pan, labeled probes containing a WGAWAW-r-Helper pair from Tig and Ugt36Bc were synthesized (Figure 3D) and analyzed for binding to recombinant GST-TCF/Pan fusion proteins using EMSA (Electrophoretic Mobility Shift Assay). Both probes were bound by GST-HMG-C-clamp, and mutation of the WGAWAW site abolished binding (Figure 3E). Mutation of the r-Helper site abolished binding in the case of the Ugt36Bc probe, and resulted in a small but reproducible reduction in binding of the Tig probe (Figure 3E). This difference was also seen with the footprinting data, where GST-HMG-C-clamp protection of the Ugt36Bc r-Helper site (Figure 3B) was more pronounced than the r-Helper sites in the Tig W-CRM (Figure S3). Consistent with being C-clamp binding sites, the r-Helper motifs were not required for binding by GST-HMG protein (Figure 3F). Taken together, these data support a model in which TCF/Pan binds to the Ugt36Bc and Tig W-CRMs through bipartite binding of HMG domain to WGAWAW sites and C-clamp binding to r-Helper sites. To determine whether the WGAWAW and r-Helper sites in the Tig W-CRM were functional, site-directed mutagenesis of the Tig1 reporter was performed. Altering either WGAWAW or r-Helper sites resulted in a strong reduction of basal expression and Wnt-dependent repression (Figure 4A). These data were similar to those obtained when the WGAWAW sites in the pHsp-178 Ugt36Bc reporter were altered [39]. When the r-Helper site in pHsp-178 was mutated, a similar defect was observed as when the adjacent WGAWAW site was destroyed (Figure 4B). These data demonstrate that the distinct bipartite TCF/Pan binding sites found in the Tig and Ugt36Bc W-CRMs are necessary for basal expression of the reporters. In the absence of these motifs, Wnt signaling causes little reduction in expression of these reporters, either due to loss of basal expression and/or loss of active repression by the pathway. In addition to the two WGAWAW sites in the Tig intronic W-CRM, five additional sequences were footprinted by the HMG domain, most of which were enriched with a TG-rich motif (Figure S2A). All five motifs were mutated, but the expression of these mutant reporters were not affected in a significant manner (Figure S2B). While it is possible that these motifs are functionally redundant, they were not analyzed further in this study. Since WGAWAW and r-Helper sites contribute to both basal activation and Wnt-mediated repression of Tig and Ugt36Bc W-CRMs (Figure 4) [39], these bipartite TCF binding sites could be sufficient for this regulation. To test this, a synthetic reporter containing two repeats of a small stretch (40 bp) from the Tig W-CRM (each repeat contains two pairs of WGAWAW and r-Helper sites) was constructed (Figure S4A). This reporter, termed “minR” for “minimal repressed W-CRM”, was repressed about two-fold by Axin RNAi or Arm* expression in Kc cells (Figure 5A; Figure S5A). Like the Tig and Ugt36Bc W-CRMs, the basal expression of the minR reporter is dependent on the WGAWAW and r-Helper sites (Figure S5B). These results demonstrate that these bipartite TCF sites are necessary and sufficient for the “reverse TCF/Pan transcriptional switch” that regulates targets repressed by Wnt signaling. The behavior of minR is the qualitative opposite of classic HMG-Helper site pairs, which are highly activated by Wnt signaling [33]. This suggests that the TCF/Pan sites themselves dictate whether a W-CRM is activated or repressed by the Wnt pathway. To test this, the HMG-Helper sites in the nkd-IntE W-CRM, which is activated by Wnt signaling in Kc cells and flies [33], [47], were replaced by WGAWAW-r-Helper sites (see Figure S4B for base pair changes). The basal activity of this “TCF sites swapped” nkd-IntE was significantly higher than either the original nkd-IntE or minR, suggesting a synergistic effect between the repressive TCF sites and the remaining sequences of nkd-IntE (Figure 5B). Strikingly, this W-CRM was repressed upon activation of Wnt signaling (Figure 5B). To determine whether the Tig1 W-CRM could be converted into an activated W-CRM, the functional WGAWAW and r-Helper sites identified in Figure 4 were converted into classic HMG and Helper sites (Figure S4C). This swapped Tig1 reporter was robustly activated by Wnt signaling (Figure 5C). To assess the individual contribution of each type of binding site to the switch in transcriptional output, r-Helper site only (H-only) and WGAWAW site only (W-only) swaps were constructed in the Tig1 reporter (Figure S4C). These “partial swap” W-CRMs lost the high basal expression of Tig1, and lacked the high activation seen when both motifs are swapped (Figure 5D). Taken together, these data argue that both the HMG domain and C-clamp binding domains are instructive in determining whether a W-CRM is activated or repressed by Wnt signaling. Our findings that the transcriptional output can be reprogrammed by altering the TCF binding sites suggests that DNA is allosterically regulating TCF/Pan. To test this, recombinant HMG-C-clamp protein was incubated with excess oligonucleotides containing activating or repressed TCF sites followed by partial digestion with two proteases, chymotrypsin or endoproteinase Glu-C. The digested product was then separated on SDS-PAGE gels. The digestion patterns between HMG-C-clamp bound with a classic HMG-Helper site pair (TH) and WGAWAW-r-Helper pair (WH) were distinct, with several proteolytic fragments observed with TH that were not detectable with WH (Figure 6A and 6B). Analyzing HMG-C-clamp mobility on a native gel indicates that the majority of the protein was complexed with either the TH [33] or WH probe (compare the shift with a control SS probe which does not bind TCF in Figure 6C). These data strongly suggest that the conformations of the HMG and/or C-clamp domains are distinct when bound to activating or repressing TCF sites. The HMG domain of LEF1 (a vertebrate TCF) is known to induce a sharp bend in DNA when bound to a classic HMG site [48]. Therefore, the possibility exists that differences in DNA bending could contribute to the transcriptional specificity of activated and repressed W-CRMs. To address this, probes where the position of the binding site was altered were tested via EMSA (Figure S6). If protein binding induced a bend in the DNA, mobility will be slowest when the binding site was present in the middle of the probe [49]. Consistent with the LEF1 data, the HMG domain of TCF/Pan exhibited bending when bound to a classic HMG site (Figure S6B). In addition, GST-HMG could bend a WGAWAW site probe, though the bend was slightly less than the classic HMG site (Figure S6B). The presence of a C-clamp in the protein and a Helper site in the probe did not alter the degree of bending (Figure S6C). Likewise the reduction of bending of the WGAWAW site was still observed when paired with an r-Helper site and bound by GST-HMG-C-clamp (Figure S6D). The data demonstrated a small difference in bending between the activated and repressed binding sites, which could contribute to the transcriptional specificity. To extend the analysis of Tig1 and minR reporters to the whole organism, these W-CRMs were cloned into P-element Pelican vectors [50], carrying the LacZ reporter gene plus insulators to minimize position effects, either using the endogenous Tig promoter (Tig1) or a heterologous one from hsp70 (minR). Transgenic lines were established and analyzed for LacZ expression in embryos and larva. Both reporters were active in embryonic hemocytes, as indicated by co-localization with MDP-1, a hemocyte marker (Figure 7A–7H) [51]. We also found staining of both reporters in the larval lymph gland (LG), fat body and circulating hemocytes (Figure 8; data not shown). These patterns are similar to that of endogenous Tig in the LG (Figure 8A–8C), as well as embryonic hemocytes and fat body [42]. These results indicate that both reporters can be used to study regulation by Wnt signaling in vivo. The Tig1 and minR reporters are both expressed at much higher levels in the cortical zone (CZ) of the LG, an irregularly shaped region containing mature hemocytes enriched in the periphery of the LG (Figure 8B, 8D, 8H). This pattern is largely non-overlapping with Wingless (Wg, a fly Wnt), which is enriched in the medullary zone (MZ) containing prohemocytes [40] (Figure 8E and 8I). The Wg pattern is more apparent in younger late 3rd instar larvae, i.e., ∼96–104 after egg laying (∼96–104 AEL; Figure 8D–8K), but the lacZ reporters expressed highest in older late 3rd instar larvae (∼104–112 AEL; Figure 8A–8C). The expression of the reporters did not overlap with Lozenge-Gal4≫UAS-GFP (Lz≫GFP), which marks crystal cells, a hemocyte lineage found in the CZ that often has high Wg expression [40] (Figure S7). While the presence of Wg in the MZ doesn't necessarily imply active Wnt signaling, these results support a model where Wnt signaling represses Tig and minR expression in this portion of the LG. To test whether the Tig1 and minR reporters were repressed by Wnt signaling in the LG, the Gal4 misexpression system [52] was used to modulate the Wnt pathway. Serpent-Gal4 (Srp-Gal4), which is active throughout the LG [53], was combined with UAS lines expressing Arm* or DisArmed in a background containing either reporter. Expression of either Arm* or DisArmed in the LG repressed the minR (Figure 9A, 9D and 9G) and Tig (Figure 9J, 9M and 9P) reporters with 100% penetrance. Under the conditions employed, no detectable change in expression of Cut, a CZ marker (Figure S8) [53], was observed (Figure 9B, 9E, 9H, 9K, 9N and 9Q), ruling out a gross change in cell fate in the LG being responsible for the loss of reporter expression. With stronger or longer expression of Arm*, we did observe a strong reduction of the CZ cell fate as previously reported (Figure S9) [40]. The results indicate that Wnt signaling can repress the Tig and minR reporters in the CZ without detectably altering cell fate. In addition, the finding that DisArmed can mediate this regulation suggests that the transcriptional activation activity of Arm is not required for this regulation. To test whether the Tig1 and minR reporters were repressed by Wnt signaling in embryonic hemocytes, we expressed Arm* or DisArmed under the control of two embryonic hemocyte drivers, Srp-Gal4 or Croquemort-Gal4 (Crq-Gal4). No detectable repression was observed (data not shown). To examine whether the negative results were due to perdurance of LacZ, we assayed circulating hemocytes from mid 3rd instar larvae (∼88–96 AEL). This is prior to release of LG hemocytes, so all circulating hemocytes are of embryonic lineage at this developmental stage [54]. Hemese-Gal4 (He-Gal4) [55], a circulating hemocyte driver, was used to drive the expression of UAS-Arm* or UAS-DisArmed. Expression of either transgene resulted in a significant repression of the minR reporter (Figure 10), demonstrating Wnt repression of this reporter in the embryonic hemocyte lineage. Our working model is that TCF/Pan activates Tig1 and minR expression in the CZ of the LG, while Wnt signaling represses these reporters in the MZ. To test this, we examined reporter expression when dominant-negative versions of Frizzled and Frizzled2 (FzDN and Fz2DN) [56], [57] were expressed via the MZ driver Dome-Gal4 [58]. We observed a strong expansion of minR in these LGs, but there was also a concomitant expansion of the CZ, indicated by a reduction of Dome≫GFP (Figure S10). This is consistent with a previous report demonstrating that Wnt signaling is required for maintenance of the MZ [40]. Depletion of TCF/Pan in the CZ using RNAi caused the predicted reduction in reporter gene expression, but there was also a reduction in the CZ (Figure S11). In both cases, the change in reporter expression was coupled with a change in cell fate, preventing a definitive demonstration that endogenous TCF/Pan and Wnt signaling regulates the minR and Tig reporters in the LG (see Discussion for further comment). To confirm that the Tig1 and minR reporters are directly regulated by TCF/Pan in vivo, the WGAWAW sites and r-Helpers in these elements were mutated. Mutation of either motif abolished expression of both reporters in the LG (Figure 11). In embryonic hemocytes, the WGAWAW site mutant of minR had no detectable expression (Figure 12 G–I), while there was some residual hemocytic expression in the r-Helper mutant (Figure 12 D–F). There was no obvious reduction in the Tig1 reporter in embryonic hemocytes when the two functional WGAWAW or two r-Helper sites identified in Kc cells were destroyed (data not shown). This caveat aside, the results indicate that the reverse transcriptional switch documented in Kc cells ([39] and this report) is also operational in the Drosophila hematopoietic system. This study extends our previous work characterizing WGAWAW sites in the Ugt36Bc W-CRM [39], identifying additional sites in another repressed target, Tig, and refining the consensus of these sites to RNWGAWAW (Figure 3C). These sites are distinct from traditional HMG sites (SSTTTGWWS) identified in earlier studies of TCF binding [29], [31]. These studies failed to identify WGAWAW sequences as TCF binding sites, perhaps because their experimental designs were biased for the highest affinity sites. However, Badis and coworkers used a microarray of randomized 8-mers to survey DNA binding domains of TFs found WGAWAW sites among the preferred binding sites for HMG domains derived from the four human TCFs [59]. To illustrate this point, we examined where eight functional classic HMG sites from activated W-CRMs and the five WGAWAW sites from the Tig and Ugt36Bc W-CRMs rank among the nearly 33,000 8-mers tested by Badis and coworkers (Table S1). Two classic sites from a Notum/wingful W-CRM [33] were the top-ranked site for all four HMG domains, while the third site from this W-CRM ranked 2–4th, depending on the protein. For classic sites in two nkd W-CRMs [33], [47], the rankings were lower, on average between 112th and 2833rd. The repressive WGAWAW sites we identified ranked between 98th and 4167th (Table S1). This work highlights the diversity of DNA recognition by HMG domains (which was also observed for half of the 104 TFs tested in this study) [59], and reveals that WGAWAW sites are a preferred class of HMG binding for TCF/Pan and vertebrate TCFs. In addition to HMG domain-WGAWAW site binding, we found that C-clamp interactions with r-Helper sites are required for TCF/Pan to regulate the Tig, Ugt36Bc and minR W-CRMs. The C-clamp is required for regulating the Ugt36Bc and Tig reporters (Figure 2), and WGAWAW and r-Helper sites in these W-CRMs are required for expression in Kc cells (Figure 4) as well as for the Tig1 W-CRM in the larval LG (Figure 11). Multimerized WGAWAW-r-Helper site pairs are sufficient for high basal expression and repression by Wnt signaling (Figure 5A, 9 and 11). The three characterized r-Helper sites share a loose consensus of KCCSSNWW and the spacing between adjacent WGAWAW and r-Helper sites is less than 7 bp among the sites we have examined (Figure 3B and S3). More functional WGAWAW, r-Helper site pairs need to be identified to better understand the sequence, spacing and orientation constraints on what constitutes this class of bipartite TCF binding site. In contrast to the Ugt36Bc and Tig W-CRMs, in several other cases traditional HMG sites have been found to mediate Wnt repression in Drosophila [35], [38] and mammalian cell culture [36], [37]. An examination of the sequences surrounding the functional HMG binding sites in the fly repressed W-CRMs did not reveal obvious candidates for r-Helper or Helper sites (C. Zhang and K. Cadigan, unpublished observations). In these cases, TCF/Pan is proposed to act with other TFs, either competing for binding with an activator [35] or acting in concert with the transcriptional repressor Brinker [38], [60]. We favor the view that the mechanism described in this report is distinct from these other examples of Wnt-mediated repression. The common models for signal-induced repression require the presence of a default activator bound to DNA near the repressive sites [2], [18], [36]. In contrast, in the TCF-mediated repression described in this report, both basal activation and repression occur through the same TCF binding sites (Figure 13). Mutagenesis of WGAWAW sites and r-Helper sites argue that they are both required for basal activation (Figure 4, 11 and 12), while repression of the minR and Tig reporters by Arm* and DisArmed argue that these sites are also responsible for Wnt-dependent repression (Figure 5, 9 and 10). Consistent with a dual role in regulating these W-CRMs, depletion of TCF/Pan via RNAi resulted in a reduction of basal activation and loss of Wnt-repression (Figure 2). Our data supports the model of a “reverse TCF transcriptional switch” that we have published previously [39], and this work extends this mechanism to the Tig W-CRM and highlights the importance of the C-clamp and r-Helper sites in this regulation (Figure 13). While we favor the model outlined in Figure 13, it is possible that it is an over-simplification and several things remain to be clarified. For example, mutation of the WGAWAW or r-Helper sites results in a dramatic loss of basal activation (Figure 4, 11 and 12) while depletion of TCF/Pan has a more modest reduction (Figure 2) [39]. This raises the possibility that other TFs could also act through the WGAWAW and r-Helper sites to achieve basal expression. For example, it is possible that TCF/Pan and Arm inhibit transcription by displacing other activating TFs from W-CRM chromatin. Another possibility is that Arm interaction with TCF/Pan disrupts its ability to bind to the bipartite site, though this model is not supported by ChIP data at the Ugt36Bc locus [39]. Further investigation is needed to determine whether additional regulators of these W-CRMs exist and if so, how do they functionally interact with TCF/Pan. Our report provides a dramatic example of how the DNA site can influence the transcriptional output of the TF binding to the site. Replacing classic HMG and Helper sites in a W-CRM (nkd-IntE) with low basal expression and a high degree of Wnt activation completely inverted the regulation: the altered W-CRM had high basal expression and was repressed by Wnt signaling (Figure 5B). Just as strikingly, changing 22 bps in the 1.8 kB Tig1 reporter, which converted two WGAWAW and two r-Helper sites into classic motifs, resulted in a W-CRM that behaves like a conventionally activated W-CRM (Figure 5C). Both the HMG and C-clamp binding sites needed to be swapped for this switch in regulation to occur (Figure 5D). These results clearly demonstrate that the type of bipartite TCF binding site to which TCF/Pan binds determines whether it acts as an activator or repressor upon Wnt stimulation. There are other examples of switching the transcriptional output of CRMs through altering the sequence of TF binding sites. Mutating sequences adjacent to Dorsal binding sites converts a repressed CRM into an activated one, suggesting that for Dorsal, transcriptional activation is the default state [12], [13]. Altering the binding site of Thyroid receptor or POU1 converted CRMs from repressed to activated elements [14], [16], [17]. In these cases, the conversion was only made in one direction, leaving open the possibility that the TF binding sites are not completely sufficient for determining the activation/repression decision. In our previous report on Wnt mediated TCF/Pan repression, the repressed Ugt36Bc W-CRM was converted to an activated one by changing three WGAWAW sites into classic HMG sites [39]. However, Wnt activation was only achieved when the Ugt36Bc W-CRM was placed adjacent to the metallothionein (MT) promoter and a small amount of Cu2+ was added [39]. When the hsp70 promoter was used, the altered Ugt36Bc W-CRM was not active, similar to the HMG site only swap in the Tig1 W-CRM (Figure 5D). Our new data strongly suggests that the complications in the prior report were due to our lack of knowledge of Helper sites, which we have now demonstrated to be essential for controlling the transcriptional output of W-CRMs. The conformation of the HMG and/or C-clamp domains of TCF/Pan is different when bound to a classic HMG-Helper pair compared to a WGAWAW-r-Helper pair, as judged by protease digestion patterns (Figure 6). In addition, the degree of bending of the DNA by the HMG domain is reduced when it is bound to a WGAWAW site (Figure S6). Presumably, these structural differences are transmitted to Arm protein bound to TCF/Pan, leading to differential recruitment of transcriptional co-regulators, as has been suggested for other TFs [20], [61]. Our results add to the growing recognition that TF binding sites are not just for recruiting TFs to regulatory DNA, but also have a profound influence on the TF's functional activity. Repressed W-CRM reporters, either natural (Tig1) or synthetic (minR), are active in embryonic and larval hematopoietic systems (Figure 7, 8 and 10), and are regulated by Wnt signaling (Figure 9 and 10). The data in the LG are especially interesting, given that Wnt signaling has been shown to control several cell fate decisions in this tissue. The Wnt pathway is required for maintenance and proliferation of the posterior signaling center (PSC), which functions as a hematopoietic niche in the LG [40], [62]. In addition, Wnt signaling promotes prohemocytic cell fate, blocking their differentiation in the MZ of the LG as well as promoting proliferation of crystal cells [40]. The Tig and minR reporters displayed minimal expression in the MZ and crystal cells (Figure S7 and S10), and their high expression in the CZ can be repressed by ectopic activation of Arm and DisArmed (Figure 9). Since DisArmed has little/no ability to activate transcription but retains repressive activity [39], these data suggest the existence of Arm-dependent repression of gene expression in the prohemocytes of the MZ. Wnt-mediated repression of the Tig and minR W-CRMs in the LG is likely direct, based on site-directed mutagenesis of the WGAWAW and r-Helper sites (Figure 11 and 12). However, we were unable to demonstrate that endogenous TCF/Pan and Wnt signaling regulates these reporters, because the genetic manipulations also altered the ratio of pro-hemocytes (MZ) and differentiated hemocytes (CZ; Figure S10 and S11). Thus, we could not uncouple cell fate change from regulation of the reporters in our loss of function experiments. It may be that the thresholds for maintaining the CZ and MZ cell fates and regulating the reporters are too similar. Another possibility is that Wnt signaling works redundantly with another factor to repress these reporters in the MZ. Having said this, it's interesting to note that the expression of Peroxidasin (Pxn), normally restricted to the CZ of the LG, expands into the MZ when Wnt signaling is inhibited [40]. Pxn has also been shown to be repressed by Wnt signaling and DisArmed in Kc cells and embryonic hemocytes [39], suggesting a similar relationship in the LG. The minR synthetic reporter is regulated by Wnt signaling in Kc cells, as well as hemocytes derived from embryos and the LG (Figure 5, 9 and 10). This regulation depends on the WGAWAW and r-Helper sites in all three contexts (Figure 11, 12 and S5). The Tig1 reporter is similarly regulated in Kc cells (Figure 1) and the LG (Figure 9 and 11). In contrast, we found no detectable regulation in embryonic hemocytes (data not shown), even though the reporter is expressed there (Figure 7) and Tig transcripts were repressed by Wnt signaling in these cells [39]. We suspect that the 1.8 kb Tig1 reporter may lack some cis-regulatory information required for Wnt regulation in embryonic hemocytes. Whether the repressive TCF sites can respond to Wnt signaling in other tissues remains unclear, since the minR and Tig reporters have no basal activity outside the hematopoietic system and fat body. To explore whether WGAWAW and r-Helper sites function outside of these tissues, we utilized a GFP reporter containing binding sites for Grainyhead (GRH), which provide basal activity in the imaginal discs [63]. Classic or repressive TCF sites were placed downstream of the GRH sites and transgenic flies generated and analyzed (Figure S12). While classic HMG-Helper site pairs (4TH) displayed strong expression consistent with activation by Wg signaling (Figure S12B, S12G and S12L), insertion of the minR sequences had no effect on the GRH-GFP reporter (Figure S12D, S12I and S12N). These results suggest that WGAWAW and r-Helper sites only respond to Wnt signaling in specific tissues (e.g. the LG). Conversely, 6TH and several other reporters that are activated by Wnt signaling in many tissues [33], [47] are not expressed in the LG (Figure S13). These data argue that the mechanism of Wnt gene regulation in the LG is different from other tissues such as imaginal discs, perhaps because the reverse transcriptional switch mechanism plays a greater role in this tissue. Further studies are needed to identify additional W-CRMs that are active in the LG, and to determine whether the regulatory mechanism uncovered in this report underlies Wnt control of PSC, pro-hemocyte and crystal cell fate in the fly LG. Kc cells were cultured and transient transfections were carried out as previously described [45]. For RNAi treatments, cells were seeded at 1×106 cells/ml in growth media supplemented with 10 µg/ml dsRNA for 4 days, diluted to 1×106 cells/ml without additional dsRNA, and grown for 3 more days for luciferase assay using Tropix Chemiluminescent Kits (Applied Biosystems) or 2 more days for mRNA preparation using Trizol Reagent (Invitrogen). dsRNAs targeting the 3′UTR of TCF/Pan [33] and the ORFs of Axin or a control gene (β-lactamase) were used [39]. qRT–PCR was performed as previously described [39]. Gene expression among different samples was normalized to tubulin56D levels. Each treatment in reporter assays was done in triplicate wells, each containing 2.5×105 cells. For standard reporter assays, 50 ng luciferase reporter and 6.25 ng LacZ per well were transfected with Axin RNAi or control RNAi. For TCF/Pan rescue assays, same amount of reporter and LacZ plus 50 ng TCF/Pan-expressing plasmid and 250 ng Arm* per well were transfected with TCF/Pan RNAi. pAc5.1-V5/His-A vector was used to equalize DNA content between samples and as a negative control for expression vectors. Luciferase activity was normalized to β-galactosidase activity from pArm-LacZ to control for differences in transfection efficiency among samples. In the figures, each bar represents the mean of biological triplicates and the data shown are representative of three independent experiments. All RLA units are arbitrary units unless otherwise specified. All luciferase reporter vectors are derivatives of pGL2 or pGL3 (Promega). pHsp-178, Tig1, Tig5 and all site mutants and swaps based on these W-CRMs were cloned into pGL2-basic. Tig2–4, minR, nkd-IntE and all site mutants and swaps based on these W-CRMs were cloned into pGL3-basic containing an Hsp70Bb minimal promoter. Vector with a Hsp70Bb promoter but containing no W-CRM was used to control for basal promoter activity. A MluI site was introduced into Tig1 upstream of the TCF sites for the ease of cloning of the swap constructs. Sequence changes were done using site-directed mutagenesis (QuickChange SDM kit, Stratagene) or recursive PCR [64]. Restriction sites and primer sequences are in Table S2 or as previously described [33], [39]. For expression plasmids, pAc-TCF (WT/C-mut), pAc-Arm*, pGEX-GST, pGEX-GST-HMG and pGEX-GST-HMG-C-clamp (WT/C-mut) have been described elsewhere [33], [39]. pArm-LacZ, a derivative of pAc-LacZ (Invitrogen) using the Arm promoter [65] was used as a transfection control. EMSAs were performed as previously described [39]. All GST-tagged proteins used in this study were purified from E. coli. 4 nM biotinylated probe (IDT, Coralville, IA) and 7–20 µM protein were used in each reaction. The conditions in the DNA bending assays were similar to the EMSA assays except for the following modifications: 4 nM biotinylated probe was incubated with 20 µM (for WH and WS), 200 nM (for TH) or 500 nM (for TS) protein before separating on 5% native PAGE gel. The probes for the DNA bending assays were generated according to a previously described strategy [49]. In short, the indicated TCF binding sites were cloned into pGL2-basic vector, and seven pairs of primers at varied positions on the vector were used. PCR products were digested at both ends by EcoRI, whose sites were introduced by the primers, and biotinylated through Klenow reaction using Biotin-16-dUTP (Roche). Probes containing a SS site (with both HMG site and Helper site mutated from TH) were generated to confirm that TCF/Pan has no detectable affinity to the surrounding sequences on the probes (Figure S6E). The sequences of TCF binding sites are summarized in Table S2. The WH and WS probes have the binding sites from Tig1 used in the EMSAs shown in Figure 3B. The sequences of the TH, TS and SS sites are previously described [33]. DNaseI fluorescent footprinting was performed as previously described [39]. 20 µM GST-HMG or GST-HMG-C-clamp was used in 50 ul reactions with 12 nM labeled probes. The probes were generated by PCR using one labeled primer and one unlabeled primer (IDT) (Table S2). For comparison between GST and GST-HMG, or GST and GST-HMG-C-clamp, or GST-HMG and GST-HMG-C-clamp, FAM and HEX labeled probes were used in two parallel reactions with different proteins, and combined after digestion. 303 bp in the middle of the Tig intronic W-CRM and the full length Ugt36Bc W-CRM (178 bp) were footprinted (see Table S2 for sequence information). 20 µl reactions containing 3–6 µM GST-HMG-C-clamp and 20× of the indicated DNA oligonucleotide were incubated for 5 min on ice and 15 min at room temperature. The buffer was the same as used for EMSA but without poly-dI*dC. Protease was then added (for partial proteolytic digestion) or not (for reverse EMSA) at a final concentration of 5–50 ng/µl for chymotrypsin (Roche) or 50–150 ng/µl for endoproteinase Glu-C (New England Biolabs). The mixture was incubated at 25°C for 2.5–3 hours. Then the digested product was loaded onto 16% tricine SDS-PAGE gel [66], and the undigested mixture was loaded onto 6% native PAGE-gel. After running, the gels were silver stained as previously described [67]. Tig (Tig1) and minR fly reporters were generated by cloning the corresponding sequences into pPelican and pHPelican vectors, respectively [50]. All 3×GRH-W-CRM fly reporters were generated by cloning the corresponding sequences into pDestination-eGFP vectors via pENTR/D-TOPO using the Gateway technique, then injecting into integration site 86Fb [68], [69]. Transgenic flies were generated by BestGene Inc. (Chino Hills, CA), Genetic Services Inc. (Cambridge, MA) and Rainbow Transgenic Flies Inc. (Thousand Oaks, California). All the Gal4 and UAS lines used in this study have been previously described: Srp-Gal4 [70], Dome-Gal4 [71], Lz-Gal4 [72], Cg-Gal4 [73], HmlΔ-Gal4 [74], UAS-Arm* and UAS-DisArmed [39], UAS-FzDN and UAS-Fz2DN [56], [75] and the DHH triple marker line containing Dome≫EBFP, Hml≫dsRed and Hh≫GFP [76]. The UAS-TCF/Pan-RNAi was a recombinant of two TCF/Pan RNAi lines, one from Vienna Drosophila Resource Center and the other from the Drosophila RNAi Screen Center. The Srp≫Arm* and DisArmed experiments were carried out in the presence of tub-Gal80ts. Crosses were set up at 18°C, and the larvae were transferred to 25°C for 2 days (Figure 9) or 3 days (Figure S9) before assaying. 3rd-instar larvae were dissected in ice cold PBS from the ventral midline in a similar manner as body wall muscle preparations [43]. For β-galactosidase stainings, exposed LG were fixed in 1% glutaraldehyde at room temperature for 15–20 min, then washed twice and stained in X-gal staining solution [77] with 1–2% X-gal for 10–60 min. Preparation of embryos, immunostaining and microscopy were as previously described, and methods for immunostaining of wing discs were adapted for LG [33]. At least 20 embryos or 12 LGs were analyzed for each condition, and the examples presented are representative. Primary antibodies were used at the following dilutions: mouse α-wg at 1∶150, mouse α-Cut at 1∶100 and rabbit α-Tig [78] at 1∶75 for LG staining; mouse α-MDP-1 [51] at 1∶100 for embryo staining, and rabbit α-LacZ (MP Biomedicals) at 1∶400 for embryo or 1∶600 for LG staining. Secondary antibodies were described previously [45]. Collection and processing of circulating hemocytes were as described previously [76]. Immunostained circulating hemocytes carrying the minR or Tig1 lacZ reporters were imaged using the Leica SP5 laser scanning confocal microscope with four channels representing LacZ, He≫GFP, P1 (a plasmatocyte marker) [79] and DAPI. Random hemocytes were circled as regions of interest (ROI) and quantified using the Leica LAS AF software. We observed little or no difference between control (He-Gal4≫+) and experimental groups (He≫Arm* or He≫DisArmed) for the DAPI and P1 and some fluctuation in the GFP channel, which could be due to Arm* or DisArmed affecting cell fate/identity. Therefore, we only used hemocytes whose He≫GFP signal intensity falls into the range of control hemocytes. For quantification, 10–15 hemocytes per larvae and 5 larvae per genotype were used.

10.1371/journal.pgen.1003254

An Evolutionarily Conserved Synthetic Lethal Interaction Network Identifies FEN1 as a Broad-Spectrum Target for Anticancer Therapeutic Development

Harnessing genetic differences between cancerous and noncancerous cells offers a strategy for the development of new therapies. Extrapolating from yeast genetic interaction data, we used cultured human cells and siRNA to construct and evaluate a synthetic lethal interaction network comprised of chromosome instability (CIN) genes that are frequently mutated in colorectal cancer. A small number of genes in this network were found to have synthetic lethal interactions with a large number of cancer CIN genes; these genes are thus attractive targets for anticancer therapeutic development. The protein product of one highly connected gene, the flap endonuclease FEN1, was used as a target for small-molecule inhibitor screening using a newly developed fluorescence-based assay for enzyme activity. Thirteen initial hits identified through in vitro biochemical screening were tested in cells, and it was found that two compounds could selectively inhibit the proliferation of cultured cancer cells carrying inactivating mutations in CDC4, a gene frequently mutated in a variety of cancers. Inhibition of flap endonuclease activity was also found to recapitulate a genetic interaction between FEN1 and MRE11A, another gene frequently mutated in colorectal cancers, and to lead to increased endogenous DNA damage. These chemical-genetic interactions in mammalian cells validate evolutionarily conserved synthetic lethal interactions and demonstrate that a cross-species candidate gene approach is successful in identifying small-molecule inhibitors that prove effective in a cell-based cancer model.

Anticancer therapeutic discovery is a major challenge in cancer research. Because cancer is a disease caused by somatic genetic mutations, the search for anticancer therapeutics is often driven by the ability to exploit genetic differences specific to tumor cells. Recently, cancer therapeutic development has sought to exploit synthetic lethality, a situation in which the combination of two independently viable mutations results in lethality. If a compound can be found to selectively kill a specific genotype via inhibition of a specific gene product, this is known as a chemical-genetic interaction, and it mimics a synthetic lethal genetic interaction. The ideal therapeutic would be broad spectrum, that is, active against multiple cancer genotypes within a tumor type and/or across a variety of cancers. We have developed an approach, taking advantage of the evolutionary conservation of synthetic lethal interactions, to identify “second-site” targets in cancer: genes whose chemical inhibition leads to selective killing of tumor cells across a broad spectrum of cancer genotypes. We identified small-molecule inhibitors of one such target, FEN1, and showed that these compounds were able to selectively kill human cells carrying cancer-relevant mutations. This approach will facilitate the development of anticancer therapeutics active against a variety of cancer genotypes.

Cancerous cells carry somatic mutations that genotypically distinguish them from surrounding noncancerous cells, and this provides an opportunity that can be exploited for therapeutic development. One strategy for the specific targeting of cancer genotypes relative to nonmutated somatic cells is to exploit synthetic lethal interactions [1]. For example, breast cancer cells with mutations in BRCA1 or BRCA2 are extremely susceptible to knockdown or chemical inhibition of PARP1, which encodes poly(ADP)ribose polymerase (PARP) [2], [3]. While exploiting synthetic lethality has the potential to be an effective approach to treating tumors, a major challenge is the identification of clinically relevant small-molecule inhibitors. One approach, pioneered by the National Cancer Institute, is to screen many thousands of unknown potential therapeutics on cancer cell lines [4]. Compounds generate a “fingerprint” of activity against certain cell lines, which can then be deconvolved, usually by mutation sequencing, to yield novel gene-drug interactions, in a so-called “bottom-up” approach. Alternatively, a “top-down” approach applies compounds of known target or mode of action to known genotypes, again to identify new gene-drug interactions. Recently, two groups used such an approach to screen more than 100 compounds against hundreds of cancer cell lines whose mutational status was known [5], [6], observing that gene-drug interactions tended to be more significant for targeted therapies, such as compounds targeting the BCR-ABL fusion protein, than for generally cytotoxic drugs, such as DNA damaging agents or antimitotics [6]. Thus, screening for compounds targeting a specific genetic lesion is preferable to developing new cytotoxic agents. Such targeted compounds can then be deployed as first-line anticancer therapeutics either singly or in a combination regime that would lessen the likelihood of drug-resistant clones developing within the tumor cell population [7], [8]. Many different cancer mutations lead to a limited repertoire of cancer phenotypes, such as chromosome instability, checkpoint dysfunction, and hyperplasia [9]. It is possible to identify a gene target that results in synthetic lethality with a large number of unlinked gene mutations by screening for targets that result in synthetic lethality with a common tumor phenotype. For example, chromosome instability (CIN), an increase in the rate of gain or loss of whole or parts of chromosomes, is observed in the form of aneuploidy in more than 90% of solid tumors and over 75% of blood cancers [10]. As the maintenance of genomic stability is an essential cellular process, CIN represents a phenotype that could potentially be leveraged towards selective killing of cancerous cells relative to normal cells. A gene that is synthetic lethal with a large number of cancer-related CIN genes would be an attractive therapeutic target in a large fraction of tumors. Genetically tractable model organisms, such as the budding yeast Saccharomyces cerevisiae, facilitate the identification of human CIN genes, via identification and sequencing of their human orthologs. For example, identification of yeast CIN genes [11]) led to the sequencing of the human homologs of 200 yeast CIN genes in human colorectal cancers, and it was discovered that human homologs of the yeast CIN genes SMC1, SCC2, BUB1, PDS1, MRE11, and CDC4 collectively account for approximately 25% of the mutational spectrum of colorectal cancer [12]–[15]. Thus, if a common synthetic lethal interacting partner could be identified for all of these genes, and a highly potent and specific inhibitor of its activity could be developed, inhibition of this target would offer a potentially broad means of targeting CIN cancers. In yeast, technologies exist to screen for genome-wide synthetic lethal interactions with relative ease [16], and identification of the synthetic lethal interaction network of the yeast orthologs of cancer-mutated genes has in previous cases revealed a small number of “hub” genes having synthetic lethal interactions with many yeast cancer-orthologs [17]. Previous studies have found a high degree of conservation between yeast and metazoan genetic interactions [18], [19], suggesting hub gene identification based on a yeast CIN gene synthetic lethal interaction network should yield broad-spectrum, second-site target genes applicable to human cancers. Here we present and validate a cross-species candidate-based approach to the identification of anticancer targets and the discovery of anticancer therapeutics. We show that a genetic interaction network comprised of colorectal cancer CIN genes is largely conserved between S. cerevisiae and a human cancer cell line. We develop an in vitro assay for the activity of the protein encoded by one such highly connected gene, FEN1, and use this assay to screen for small-molecule inhibitors. Finally, we show that flap endonuclease inhibitors recapitulate conserved genetic interactions. These data demonstrate the effectiveness of a cross-species synthetic lethal approach to the discovery of potential anticancer therapeutics. The human genes SMC1, SMC3, NIPBL, STAG3, RNF20, FBXW7/CDC4, MRE11A, RAD54B and BLM have been found to be mutated in colorectal cancer, and together account for approximately 25% of the CIN mutational spectrum of this disease [13]–[15], [20]–[22]. Protein BLAST was used to identify the budding yeast orthologs of these human genes (Table 1) and we constructed a synthetic lethal interaction network (Figure 1A), using literature and publicly available genetic interaction data (BioGrid and the Saccharomyces Genome Database) [18], [23]. To investigate the conservation of this network between yeast and a human cell line, we used siRNA-mediated knockdown of potential synthetic lethal gene pairs in the cell line HCT116. Knockdown efficiencies were evaluated by Western blots (Figure S1A). All pair-wise combinations between the three “central” synthetic lethal partner genes, WDHD1, FEN1, and CHTF8, and the ten outer cancer-mutated CIN genes were evaluated for synthetic lethality (Figure 1B, 1C, 1D). (CHTF8 was selected as a representative of the alternative RFCCHTF18, comprised of Dcc1, Ctf8, and Ctf18 in S. cerevisiae). Of the 30 possible synthetic lethal interactions among the genes tested, 22 have been reported in yeast [18], [23]. We found 16 of the predicted interactions (73%) were conserved between yeast and human cells, and 6 predicted interactions did not appear conserved in our assay (27%). Furthermore, one interaction, between FEN1 and STAG1, was not predicted based on yeast data; however, we detected a genetic interaction between these genes (Figure 1F and Table S1). No interactions were observed with STAG3, which functions primarily in human meiosis [24]. As in yeast, all three central genes – WDHD1, FEN1, and CHTF8 – were highly connected to sister chromatid cohesion genes (e.g. cohesin and/or cohesin loaders) (Figure 1F). As FEN1 encodes an enzyme, whereas WDHD1 and CHTF8 do not; it may be amenable to biochemical inhibitor screening. Thus, we sought to further validate genetic interactions between FEN1 and other genes in the network. To ensure that these observed interactions were not cell line-dependent, we attempted to recapitulate interactions between FEN1 and each of CDC4, RAD54B, and RNF20 in the karyotypically stable, immortalized fibroblast cell line hTERT. As in HCT116 cells, genetic interactions were observed following knockdown of all three gene pairs (Figure 1E, Table S2, Figure S1C). We found that individual siRNAs could recapitulate the genetic interactions observed with the pooled siRNAs (Table S3). These data validate a subset of genetic interactions identified in the HCT116 cells and thus confirm FEN1 as a strong candidate therapeutic target. FEN1 (Flap ENdonuclease 1) encodes an enzyme previously shown to be amenable to biochemical assay development in vitro [25] that has been implicated in almost all DNA transactions, including DNA repair and replication [26]). Adapting a previous radiolabel-based in vitro assay, we developed an in vitro assay for FEN1 activity based on fluorescence quenching [25]). In this assay, three oligonucleotides are annealed to generate the synthetic substrate, positioning a fluorophore and fluorescent quencher in close proximity. The flap endonuclease activity of FEN1 cleaves the 5′ flap to which the fluorophore is attached, allowing it to diffuse away from the quencher and fluoresce (Figure 2). Using a potent, previously described in vitro FEN1 inhibitor, compound 16 from Tumey, et al. [27], we observed significant inhibition of flap endonuclease activity (Figure 3, upper left panel). A screen of 30 000 compounds, from libraries containing known and FDA-approved drugs, and the Canadian Chemical Biology Network library, yielded approximately 90 hits, following a counterscreen using a quencherless substrate to eliminate false positives caused by fluorescent compounds and fluorescent quenchers. Ultimately, 13 compounds were selected for further investigation based on structural diversity and having drug-like properties (as described by Lipinski's “Rule of Five”; [28]). These compounds were found to have mid-nanomolar to low micromolar IC50s in vitro (Figure 3, remaining panels). We next sought to determine whether the flap endonuclease inhibitors we identified could recapitulate any of the genetic interactions found previously (Figure 1F). We first targeted the interaction between FEN1 and CDC4, owing to the fact that CDC4 has been shown to be a CIN gene mutated in many tumor types [11], [29]–[32]. We took advantage of a matched pair of cell lines in which both copies of CDC4 had been inactivated in HCT116 cells [13]. siRNA-mediated knockdown of FEN1 in this cell pair resulted in selective proliferation inhibition (Figure 4A). We applied the small-molecule hits from the screen to this matched pair of cell lines and found six compounds that selectively inhibited the proliferation of CDC4-knockout HCT116 cells relative to wild type cells (Figure 4B and Figure S2B). To ensure that these results were not cell line-specific, we utilized another matched pair of cell lines with inactivated CDC4, this time in DLD-1 cells. The six compounds showing selective proliferation inhibition of CDC4-knockout HCT116 cells were applied to CDC4-knockout and wild type DLD-1 cells [13], and RF00974 and NSC645851 were found to selectively inhibit the proliferation of CDC4-knockout DLD-1 cells relative to wild type (Figure 4C). To further test the idea that CDC4 activity is responsible for the observed effect, cells in which CDC4 had been inactivated in a heterozygous state were also treated with RF00974 and NSC645851. As with homozygous CDC4−/− cells, heterozygous CDC4+/− cells displayed a statistically significant decrease in proliferation relative to wild type CDC4+/+ cells, albeit lesser in magnitude (Figure 5). We next attempted to recapitulate the interaction between FEN1 and MRE11A, as MRE11A has been shown to be mutated at a frequency of 4% in colorectal cancers [15]. We treated cells in which MRE11A had been depleted via siRNA with the more potent of the two flap endonuclease inhibitors described above, RF00974, and found that MRE11A depletion sensitized cells to flap endonuclease inhibitor treatment (Figure 6A). We also found that treatment with a previously-described small-molecule inhibitor of MRE11A, mirin [33], was able to sensitize cells to treatment with RF00974 (Figure 6B). Taken together, these data suggest that inhibition of flap endonuclease activity is sufficient to recapitulate evolutionarily conserved, colorectal cancer-relevant synthetic lethal genetic interactions. Finally, we wished to characterize the mechanism by which inhibition of flap endonuclease activity may lead to cell death. Given the role of FEN1 in DNA replication and repair, we asked whether endogenous DNA damage increases as a result of FEN1 inhibition. We used HCT116 cells in which 53BP1 had been stably tagged with mCherry to ask whether 53BP1 focus formation, indicative of DNA repair centers [34], [35], increased. We found a statistically significant (p<0.05) increase in the frequency of cells with many 53BP1 foci following siRNA-mediated knockdown of FEN1. Furthermore, we observed a similar increase (p<0.05) following treatment with the flap endonuclease inhibitor RF00974 (Figure 6C). We next measured the level of H2AX phosphorylation (γ-H2AX), an independent indicator of DNA damage [36], in HCT116 CDC4+/+ and CDC4−/− cells in response to RF00974. We found that, similar to increasing 53BP1 focus formation, RF00974 treatment increased H2AX phosphorylation (Figure S3). H2AX phosphorylation was increased even in untreated HCT116 CDC4−/− cells, so no increase in phosphorylation was observed. In order to determine whether RF00974 leads to an increase in apoptosis in CDC4-deficient cells, we asked whether PARP cleavage, a marker of apoptosis [37], is increased following RF00974 treatment. We found that RF00974 treatment did not increase PARP cleavage in either wild type or CDC4-deficient cells. Taken together, these results suggest that loss of FEN1, or inhibition of flap endonuclease activity, lead to an increase in endogenous DNA damage that inhibits the proliferation of CDC4-deficient cells by non-apoptotic means. In this study, we used a cross-species candidate approach to identify new anticancer therapeutic targets for small-molecule inhibition having a potentially broad spectrum of applicability. We found that a yeast CIN synthetic lethal interaction network is largely conserved between S. cerevisiae and a human tumor cell line. Based on this network, we screened for in vitro inhibitors of the highly connected enzyme FEN1. Flap endonuclease inhibitors discovered in this screen recapitulated synthetic lethal interactions between FEN1 and each of CDC4 and MRE11A, demonstrating that evolutionarily conserved genetic interactions in a core cellular process, such as the maintenance of genomic stability, can be exploited as a means to inhibit the proliferation of tumor cells carrying specific and cancer-relevant mutations. The idea of using the unique genetic profile of tumor cells relative to somatic cells to selectively kill cancer has been applied by various groups, such as in the case of the chemical-genetic interaction between BRCA1/2 and PARP inhibitors [2], [3]. Several studies have focused on DNA damage, usually by identifying inhibitors of DNA damage response proteins that either directly kill tumor cells, or that potentiate the effects of DNA damaging agents [38]–[42]. Recently, two large-scale studies examining chemical-genetic interactions between new or established anti-cancer treatments and cancer cell lines of known genotype demonstrated the promise of such top-down approaches by identifying previously unknown sensitivities of many cancer genotypes, such as between Ewing's sarcomas and PARP inhibitors [5], [6]. An alternative means to construct genetic interaction networks for the discovery of therapeutic targets is to take a cross-species candidate approach in a genetically tractable model organism. In S. cerevisiae, defined genetic changes can be introduced and subsequently screened in a high-throughput manner [16], [43] (though mammalian genome editing technologies are advancing rapidly [44], [45]). The nearly 75% (16/22) conservation of synthetic lethal interactions we found between yeast and human cells is similar to the degree of conservation of genetic interactions between S. cerevisiae and the model metazoan Caenorhabditis elegans in a related network, identified by our group and others [18], [19], [46], and expands upon previous proof-of-principle work by our group [47]. Although we ultimately targeted the highly conserved flap endonuclease FEN1 in the current study, yeast genetic data has the potential to implicate biological processes, as opposed to specific proteins, as therapeutic targets; in this way, targets can be identified that are not conserved in S. cerevisiae. For example, we recently demonstrated that mutation of cohesin genes in yeast was synthetic lethal with mutation of proteins playing a role in replication fork stability. siRNA-mediated knockdown of cohesin genes was found to sensitize human cells to inhibition of PARP, a protein involved in replication fork progression, but without a known ortholog in yeast [19]. Thus, the versatility of yeast synthetic lethal networks to predict therapeutic targets makes our approach complementary to large-scale screening for gene-drug interactions [4]–[6]. Therapeutics that target a specific genotype, such as EGFR family inhibitors in the case of ERBB2 (also known as HER2) amplification, produce more significant gene-drug interactions than more general cytotoxic agents [6]; however, the indications for such agents are limited to a handful of genotypes. FEN1 plays a critical role in nearly all DNA transactions, including DNA replication via Okazaki fragment maturation [48], [49], long-patch base excision repair [50], [51], the prevention of trinucleotide repeat expansions [25], [52], and restart of stalled replication forks [53]. Yeast RAD27 is one of the most highly genetically connected genes in the yeast genome (Tables S5 and S6); many of these interactors are CIN genes [11], and many of the corresponding human orthologs may prove to be mutated and cause CIN in tumours. Given that the majority of the genetic interactions were conserved in the CIN synthetic lethal interaction network interrogated here, FEN1 may be a widely applicable target in cancers harboring mutations in a variety of CIN genes. More generally, DNA repair and replication protein inhibitors are being actively developed as anticancer therapeutics [2], [3], [41], [54] and the process of DNA replication forms a genetic hub in S. cerevisiae [16], [23], [43], [55]. The critical role of FEN1 in DNA transactions is analogous to that of PARP, a protein playing a role in DNA repair and the protection of stalled DNA replication forks [56], [57]. PARP is synthetic lethal with mutations in BRCA1/2 [2], [3], and its therapeutic range has been extended more recently to include cells with mutations in PTEN [38] and cohesins [19]. Thus, like PARP, FEN1 potentially represents a potent, broadly-applicable target for anticancer therapeutic development. In turn, the ideal anticancer therapeutic would have a broad spectrum, suggesting it would be more advantageous to target a phenotype common in cancer. CIN in the form of aneuploidy is seen in >90% of solid tumors [10] and represents a sub-lethal mutation in an otherwise essential process. Of relevance to the current work, moderate aneuploidy and CIN correlate with poor prognosis in cancer, but extreme aneuploidy correlates with improved patient outcomes [58], [59]. Yeast RAD27 is a CIN gene [17], and FEN1 mutation in various systems leads to CIN and has been associated with cancer [17], [60]; thus, inhibition of FEN1 in cancers that already exhibit CIN could lead to a level of CIN incompatible with viability. In the present study, flap endonuclease inhibitors were found to recapitulate the synthetic lethal interactions between FEN1 and each of CDC4 and MRE11A [18], [23]. We observed that both depletion and inhibition of flap endonuclease activity led to an increase in endogenous DNA damage. Recent reports have shown that γ-H2AX levels are not increased following FEN1 depletion [61]; however, we observed increases in DNA damage using two independent assays following two means of FEN1 inhibition, and attribute these results to cell background differences, such as the mismatch repair deficiency present in HCT116 cells. Furthermore, this increase in DNA damage led to a non-apoptotic inhibition of proliferation. Thus, one explanation for the lethality in combination with inactivation of CDC4 is that the cell is inappropriately driven through the cell cycle, owing to elevated levels of cyclin E [13], when otherwise it would arrest to try to repair DNA damage. Likewise, increased endogenous DNA damage combined with loss of MRE11A, a protein playing a critical role in the first steps of the DNA damage response [62], could lead to a level of DNA damage or mutation that is incompatible with proliferation. CDC4 has been reported to be mutated in a wide variety of tumor types, at frequencies ranging from 6% to >30%, depending on the tumor type [13], [21], [29], [63], [64]. Recently, it has been suggested that reduction of CDC4 activity to some level below that of wild type, but above complete abrogation of function, is optimal for tumor progression [63]. Thus, the fact that two flap endonuclease inhibitors described here were able to selectively inhibit the proliferation of both heterozygous and homozygous CDC4-knockout cell lines suggests that CDC4 loss, whether complete or partial, sensitizes cells to inhibition of flap endonuclease activity. As well, the fact that both genotypes were sensitive to inhibition of flap endonuclease activity adds weight to the suggestion that this response is specific to CDC4 activity, in the same way that changing response following alteration in dosage in biochemical screening is suggestive of target identity [65]. In summary, here we have presented a rational, cross-species approach to the identification of anticancer therapeutic targets by targeting CIN, a common cancer phenotype. The use of conserved synthetic lethal interaction networks to identify highly-connected second-site targets is an accessible alternative to large scale screens: it narrows down the number of synthetic lethal gene pairs to be directly retested from tens of thousands to dozens, and is based on strong synthetic lethal interactions discovered in yeast networks. We have demonstrated the potential of this approach to identify targets and therapeutics, such as FEN1 and the flap endonuclease inhibitors described here, having potentially broad applicability in the treatment of cancer. HCT116 cells were purchased from ATCC. HCT116 derivatives, DLD-1 and DLD-1 derivatives were gifts of Dr. Bert Vogelstein (Johns Hopkins University). (Importantly, we observed that the deleted exon in CDC4 in these cell lines is not exon 5, as previously reported [13], but exon 8. We attribute the difference to changing annotations in public sequence databases between 2004 and the present.) 53BP1-mCherry HCT116 cells were a gift of Dr. Sam Aparicio (UBC). These cells were grown in McCoy's 5A medium with 10% FBS. Immortalized (telomerase) BJ normal human skin fibroblasts, hTERT [66], were generously provided by Dr. C.P. Case (University of Bristol) and were grown in DMEM containing 10% FBS. Mirin was purchased from Sigma-Aldrich. RF00974 was purchased from Maybridge, Ltd. Western blots were performed as detailed elsewhere [47]. Antibodies used for Western blots are described in Table S4. Subconfluent and asynchronous cells were transiently transfected with siRNAs. HCT116 cells were transfected with ON-TARGETplus siRNA pools at a total siRNA concentration of 25 nM using DharmaFECT I (Dharmacon). In dual siRNA experiments, the total siRNA concentration was 50 nM. Cultures were replenished with fresh medium 11 hours after transfection. hTERT cells were transfected with ON-TARGETplus siRNA pools, or independent duplexes, at a total siRNA concentration of 100 nM using RNAiMax (Invitrogen). Cultures were replenished with fresh medium 24 hours after transfection. HCT116 cells were harvested 24 hours after siRNA transfection and re-plated in 96-well optical bottom plates. hTERT cells were transfected directly in 96-well plates. HCT116 cells were fixed four days after transfection, and hTERT cells were fixed seven days after transfection, in 4% paraformaldehyde/PBS. Nuclei were labelled with Hoechst 33342. Stained nuclei were counted using a Cellomics Arrayscan VTI fluorescence imager as described previously [47] or a Zeiss AxioObserver Z1 equipped with an LED Colibri light source, a 20× plan apochromat dry lens (numerical aperture = 0.8) and AxioVision v4.8 software. Images were analyzed using the Physiology Analyzer (Assaybuilder) option within the AxioVision software. Data were normalized to GAPDH-silenced controls and conventional statistics (e.g. column statistics and Student's t-tests) were performed. Experiments were performed twice; indicated numbers are averaged from at least 6 wells. To determine the presence of a synthetic lethal interaction, the proliferative defect was calculated, and is defined aswhere the predicted proliferation was the product of the proliferation of the two individual gene knockdowns, following a multiplicative model of genetic interactions [67]. Synthetic lethal interactions were scored as a proliferative defect of three times the average SEM of the experiment or greater. During compound incubation experiments, cells were incubated in compound of interest in 96-well optical bottom plates for approximately three days prior to fixation and analysis. Data (from six independent wells) were analyzed using a one-way ANOVA followed by a Tukey test. FEN1 was expressed in BL21 E. coli from pET28b(+) (a generous gift from R. Bambara, University of Rochester) using 1 mM IPTG. Bacteria were lysed in lysis buffer (50 mM NaH2PO4, 300 mM NaCl, 10 mM imidazole, pH 8.0 containing 2× protease inhibitor) via a French press at 10 000 psi. The lysate was clarified and passed through a 0.22 µM filter before being loaded onto a HisTrap FF column (1 mL, GE Healthcare) in an ÄKTAFPLC P-920 system (GE Healthcare). The column was washed in 10 volumes of wash buffer (lysis buffer+20 mM imidazole), and FEN1 was eluted with 5 volumes of elution buffer (lysis buffer+125 mM imidazole). The lysate was diluted with 9 volumes HI buffer (30 mM HEPES-KOH, 0.5% myo-inositol, pH 7.8) with 30 mM NaH2PO4 and concentrated in a protein concentrator (Amicon). It was then loaded onto a hydroxyapatite resin (HA Ultrogel, Pall Life Sciences). The hydroxyapatite resin was washed with 10 volumes of HI-30 mM PO4, and FEN1 was eluted with 5 volumes of HI-200 mM PO4. The eluate was diluted with 5 volumes HI-30 mM KCl prior to concentration, and then loaded onto a strong cation exchange column (1 mL HiTRAP SP FF FPLC, GE Healthcare Life Sciences). The column was washed with 10 volumes of HI-30 mM KCl, then 10 volumes of HI-200 mM KCl, and FEN1 was eluted with a gradient from HI-200 mM KCl to HI-500 mM KCl over 10 column volumes. Purified FEN1 was concentrated in FEN1 dilution buffer (30 mM HEPES-KOH, 5% glycerol, 0.1 mg/mL BSA, 0.01% NP-40), and aliquots of known concentration were frozen at −80°C. Oligonucleotides used were as follows: “template”, 5′-GGTGGACGGGTGGATTGAAATTTAGGCTGGCACGGTCG-3′, “upstream”, 5′-CGACCGTGCCAGCCTAAATTTCAATC-3′, “downstream”, 5′-6-FAM-CCAAGGCCACCCGTCCAC-BHQ-1-3′. (6-FAM is 6-carboxyfluorescein; BHQ-1 is black hole quencher 1.) The three oligonucleotides were annealed at equimolar amounts in annealing buffer (50 mM Tris, 50 mM NaCl, 1 mM DTT, pH 8.0) by heating to 94°C, cooling to 70°C, and gradually cooling to room temperature. FEN1 assays were carried out with 6 pmol FEN1 and 20 nM annealed substrate in FEN1 buffer (50 mM Tris pH 8.0, 30 mM NaCl, 8 mM MgCl2, 0.1 mg/mL BSA, 2 mM DTT). Assays were carried out at room temperature and kinetic reads were taken over approximately ten minutes in a Varioskan plate reader (Thermo Fisher Scientific), using excitation and emission wavelengths of 492 nm and 517 nm, respectively. 53BP1-mCherry cells were grown on cover slips. Following desired treatment (either two hours of bleomycin treatment at 5 µg/mL, four days following siRNA transfection, or after 24 hours of RF00974 treatment at 10 µM), cells were fixed for five minutes in 4% paraformaldehyde/PBS, mounted in Vectashield mounting medium containing DAPI (500 ng/mL), and imaged on a Zeiss Axioplan microscope with a Coolsnap HQ camera, using appropriate filters and controlled by Metamorph software. Cells were treated with RF00974 for 48 hours prior to harvesting of medium and cells in lysis buffer (50 mM Tris, 150 mM NaCl, 1% Triton-X-100, pH 7.5). Lysates were sonicated and clarified by centrifugation at 13 000 rpm for 15 minutes at 4°C. As a positive control, HCT116 cells were treated with 1 µM staurosporine prior to harvesting. Lysates were subjected to Western blotting as described above. Synthetic genetic array analysis of rad27Δ against a collection of yeast essential DAmP alleles [68] and temperature sensitive alleles [69] was carried out as described previously [11], [19].

10.1371/journal.pcbi.1004201

Bacterial Temporal Dynamics Enable Optimal Design of Antibiotic Treatment

There is a critical need to better use existing antibiotics due to the urgent threat of antibiotic resistant bacteria coupled with the reduced effort in developing new antibiotics. β-lactam antibiotics represent one of the most commonly used classes of antibiotics to treat a broad spectrum of Gram-positive and -negative bacterial pathogens. However, the rise of extended spectrum β-lactamase (ESBL) producing bacteria has limited the use of β-lactams. Due to the concern of complex drug responses, many β-lactams are typically ruled out if ESBL-producing pathogens are detected, even if these pathogens test as susceptible to some β-lactams. Using quantitative modeling, we show that β-lactams could still effectively treat pathogens producing low or moderate levels of ESBLs when administered properly. We further develop a metric to guide the design of a dosing protocol to optimize treatment efficiency for any antibiotic-pathogen combination. Ultimately, optimized dosing protocols could allow reintroduction of a repertoire of first-line antibiotics with improved treatment outcomes and preserve last-resort antibiotics.

Antibiotic resistance is a growing problem that the World Health Organization describes as “one of the top three threats to global health.” To date, bacteria have developed resistance to all antibiotics used in clinical settings. Unfortunately, the evolution of antibiotic resistant bacteria is accelerating, as antibiotics continue to be misused and overused. As the antibiotic pipeline is drying up, it becomes increasingly critical to utilize the antibiotics already on the market more effectively. The key to designing better regimens lies in the ability to predict how bacteria will respond to a particular antibiotic treatment. Because of this, we need a simple metric that characterizes this pathogen-antibiotic interaction that can be easily measured and used to design dosing protocols that will effectively clear an infection. To help guide the design of effective protocols, we use quantitative modeling to develop a metric that is easy to measure and quantifies the pathogen-antibiotic interaction. Through optimized antibiotic regimens, our strategy could extend the use of first-line antibiotics, improve treatment outcome, and preserve last-resort antibiotics.

Bacteria eventually develop resistance to all antibiotics they encounter [1–3]. Unfortunately, the evolution of antibiotic resistant bacteria is accelerating due to the widespread use of antibiotics [4,5]. As the antibiotic pipeline is drying up and the threat of antibiotic resistance is becoming more urgent [6,7], it is critical that we better utilize the antibiotics already on the market [8–10]. One of the largest and most commonly used classes of antibiotics for treating both Gram-positive and Gram-negative bacteria is the β-lactams [11–13]. Many β-lactams, such as penicillin V, amoxicillin, and first-generation cephalosporins, are first-line antibiotics; they are recommended for initial therapy because they are highly effective against non-resistant pathogens, have a lower risk of side effects, and are less expensive, relative to second-line antibiotics [14–16]. However, the rapid emergence of extended spectrum β-lactamase (ESBL) producing pathogens has greatly limited the use of β-lactam antibiotics [13,17]. ESBL-producing pathogens have significant adverse effects on clinical outcomes due to their ability to hydrolyze penicillins, broad-spectrum cephalosporins, and monobactams [6,18,19]. Patients infected with ESBL-producing pathogens have worse prognoses and, if given the incorrect treatment, mortality rates of 42–100% greater than patients receiving the correct treatment [18,20]. Additionally, β-lactams could promote horizontal gene transfer of virulence factors [21] and could be responsible for the spread of ESBL genes. As a precaution, most first-line β-lactams are ruled out if ESBL-producing pathogens are detected, even for ESBL-producing pathogens that appear to be sensitive to a particular β-lactam [22–25]. This is done largely out of concern for complicating drug responses that have been observed in vitro, such as the inoculum effect, a phenomenon in which the minimum inhibitory concentration (MIC) of an antibiotic increases as the bacterial density increases [24,26–30]. With first-line β-lactams ruled out, second-line antibiotics, such as carbapenems, fluoroquinolones, β-lactam/β-lactamase inhibitor combinations, glycopeptides, and cephamycins, are typically administered [31]. Although this practice is based on a valid concern, it has limitations. Specifically, second-line antibiotics are associated with higher costs and more adverse effects [32–37]. Additionally, the more frequently bacteria are exposed to second-line antibiotics, the faster the pathogens are likely to develop resistance to our last resort antibiotics [2,5]. Given the dearth of new antibiotics entering the market and the limited number of effective antibiotics already available, we cannot afford to disregard potentially effective antibiotics. First-line β-lactams could represent a missed opportunity for treating pathogens producing moderate levels of ESBLs. Individual bacteria that produce moderate levels of ESBL can remain sensitive to the antibiotic due to insufficient production or activity of ESBL; however, if enough bacteria are present, then the population’s collective ESBL concentration will be sufficient to render the population resistant to the antibiotic [38,39]. In other words, a low density population of moderate ESBL producers would lyse entirely because its collective ESBL concentration would be insufficient to inactivate the β-lactam, while a high density population would only experience partial lysis before its collective ESBL concentration can inactivate the β-lactam and promote the recovery of the surviving bacteria. This collective population recovery is time dependent [40]. Shortly after the antibiotic is first applied, the population will be reduced due to lysis and appear susceptible because it will not have yet benefited from the activity of ESBLs. Ideally, a treatment could pinpoint the time window when the most lysis has occurred and the least benefit has been experienced. Extensive studies have been carried out to devise methods to optimize treatment efficacy of antibiotics by changing the dosing period and amplitude. These studies typically examine which metric(s) can capture the pharmacokinetic/pharmacodynamics (PK/PD) of an antibiotic and be used to predict antibiotic efficacy [41–44]. Current metrics adopted in the clinical setting, such as the MIC, do not account for the time course of antimicrobial activity and are not sufficiently predictive of treatment efficacy [22,45–47]. Therefore, there is a need for a simple metric that characterizes this pathogen-antibiotic interaction that can be easily measured and used to design dosing protocols that will effectively clear an infection. Here, we use quantitative modeling to demonstrate a strategy for customizing regimens for a particular bacteria and antibiotic combination without needing to know the full mechanistic basis for the bacteria-antibiotic interaction. Specifically, we focus on optimizing a dosing protocol to enable β-lactams to effectively treat a moderate ESBL-producing pathogen. To help guide the design of effective protocols, we develop a metric, the recovery time, which is easy to measure and quantifies the pathogen-antibiotic interaction. Even though we assumed specific molecular mechanisms underlying this collective antibiotic response, our model illustrates that the predictive power of the recovery time is maintained for different specific molecular mechanisms and for different initial conditions. Through optimized antibiotic regimens, our strategy could extend the use of first-line antibiotics, improve treatment outcome, and preserve last-resort antibiotics. We developed a kinetic model comprising a set of ordinary differential equations (ODEs) to capture the population dynamics of collectively tolerant, ESBL-producing bacteria being treated by a β-lactam (S1 Text) [40]. We further nondimensionalized the model to facilitate analysis. In this model, introduction of the antibiotic inhibits bacterial growth and causes lysis. β-lactamase (Bla) is naturally found in the periplasm of Gram-negative bacteria, where it can benefit the host bacterium by hydrolyzing the β-lactams that diffuse into the periplasm [48]. However, moderate amounts of periplasmic Bla are insufficient to protect a bacterium from high concentrations of antibiotic [38,49]. Conversely, sufficient amounts of Bla can accumulate to protect a population if enough bacteria are initially present. With a dense enough population, the collective intracellular and extracellular Bla, due to lysis or leaky secretion [50], will be sufficient to degrade the antibiotic to a sublethal concentration before all cells are eliminated (Fig. 1A). Thus, the survival of the population depends on establishing a collective antibiotic tolerance (CAT)[30]. In general, Bla expression can be constitutive or inducible by the antibiotic [51–54]. Here, we focus on constitutive Bla expression, which is most clinically relevant to the pathogens that express plasmid-mediated ESBLs [39,55]. However, our conclusions also apply to the case where Bla expression is inducible. They will likely apply to bacterial responses to other antibiotics if the antibiotic causes an initial decline in the population density by killing a subpopulation of cells and the population can recover when the antibiotic is subsequently degraded by an enzyme produced by the cells (whether or not the enzyme is released into the culture). Using physiologically relevant parameters, our model generates PK/PD profiles that are characteristic of Bla-mediated CAT. Starting from a sufficiently high initial density, the population exhibits an initial decline upon antibiotic treatment, followed by eventual recovery due to intrinsic and Bla-mediated degradation of the antibiotic (Fig. 1B). Sufficient time is needed to observe this apparent drug tolerance. If examined shortly after antibiotic treatment, the population will have just experienced significant lysis and will appear susceptible because the effects of Bla have not yet been fully recognized. For a fixed initial antibiotic concentration, the model predicts a switch-like dependence of population survival over the initial population density: the population can only survive if starting at a sufficiently high density (Fig. 1C). If too few bacteria are present, the total expression of Bla from the entire population is insufficient to degrade the antibiotic fast enough to allow the population to recover. If enough bacteria are present, however, the population can endure the initial crash in density for a longer period. As such, some bacteria remain when the antibiotic concentration decreases sufficiently, due to Bla-mediated degradation, to allow the population to recover. The density-dependent survival of the population is the defining feature of the inoculum effect [28,56]. Our results illustrate the defining features of a CAT bacterial response involving antibiotic-triggered death. In particular, the population will appear resistant when its initial density is sufficiently high and it is given enough time to recover. These features form the basis for the preemptive practice of disregarding β-lactams when an ESBL-pathogen is identified. However, our model also indicates that the population is sensitive when its initial density is sufficiently low or when it is examined in a short time window. Given these properties, we reason that optimal antibiotic dosing may remain effective in eliminating bacteria. If so, an immediate next question is how to best design the treatment protocol. This task would be straightforward if we could determine the specific molecular mechanisms and defining parameters for each pathogen-antibiotic pair: under such a scenario, we could in theory use a model specific to the pair to examine efficacy of different dosing protocols. This is impractical, however, as many ESBL pathogens are poorly characterized at the molecular level and there are many different ESBL enzymes [57]. A more practical option would be to identify an easy-to-measure, lumped metric based on a bacterial population’s response to a single dose of antibiotic that will allow us to reliably predict its response to periodic antibiotic treatment without needing to know the underlying molecular-level parameters. A typical metric to quantify efficacy of an antibiotic is the minimum inhibitory concentration (MIC), which can be measured by disk diffusion and microbroth dilution methods after a certain duration of antibiotic treatment [58]. However, the MIC measured at a particular time point does not capture the rich temporal dynamics of bacterial responses due to antibiotic-triggered death. Instead, we propose to use another lumped metric: the recovery time; specifically, this defines the time it takes a population to return to its initial density after being exposed to a dose of antibiotic (Fig. 2). By definition, the recovery time captures the dominant dynamic features of bacterial temporal response. As such, it may be a more predictive metric for the long-term outcome of periodic antibiotic treatment. We first tested the predictive power of the recovery time in injection-based dosing protocols. With the base-parameter set, our model predicts a monotonic dependence of the recovery time on the antibiotic concentrations for single-dose treatment (Fig. 3A). Once the initial antibiotic concentration is high enough to cause cell lysis (a0 > 0.5), the recovery time increases exponentially with the initial antibiotic concentration until the antibiotic concentration is too high (a0 > 10) and the recovery time becomes infinite. This dependence is an intrinsic property of antibiotic-mediated lysis. Under low concentrations of antibiotic (0.5 < a0 < 10), the recovery time is primarily determined by how fast the antibiotic is degraded by Bla. Under increasing concentrations of antibiotic (a0 > 10), the rate of antibiotic degradation is essentially saturated (limited by the population size and the constant production rate of Bla) and the recovery time is primarily determined by the lysis rate. β-lactams’ killing rate is time-, not dose-, dependent and is reflected in the model’s lysis rate’s non-linear dependence on the antibiotic concentration (Hill coefficient = 3) [59]. Once the antibiotic concentration is high enough, further increasing the concentration does not increase the lysis rate. As noted above, the recovery time could represent a simple, yet reliable, metric in predicting outcomes from periodic treatment. To test this notion, we examined the consequence of periodic dosing of varying antibiotic concentrations. For each concentration, we varied the dosing periods from 0.1 to 2 times the corresponding recovery time, and obtained the final population density after 100 doses. Our modeling results confirmed the predictive power of the recovery time: as long as the initial antibiotic concentration is sufficiently high to cause significant initial lysis, the population will reach a high final density if the period is greater than the recovery time; the population goes extinct otherwise (Fig. 3B). Of the regimens leading to eventual population extinction (period < recovery time), different combinations of antibiotic concentrations and dosing periods eliminate a population with varying efficacy. To quantify this efficacy, we calculated the minimum number of doses necessary to reduce the population density to below 10-10 (Fig. 3C). The resulting landscape shows a strong dependence on antibiotic concentration and the corresponding recovery time. When the antibiotic concentration is too low and the recovery time is close to 0, the number of doses required to clear the infection is very large, regardless of the dosing frequency. When the antibiotic concentration is very high and the corresponding recovery times approach infinite, then the number of doses is very low. However, there is an intermediate range of antibiotics with intermediate recovery times that show variation in the number of doses necessary to clear the infection. Concentrations producing the longer recovery times require fewer doses because they can reduce the bacterial density more severely than concentrations with shorter recovery times. For intermediate antibiotic concentrations (1 < a0 < 10) to be most effective, the model suggests they should be delivered at low-to-intermediate period lengths (period = 20–50% recovery time) at which the population is most vulnerable. At the end of each period, the bacteria are still lysing, have almost reached minimum density, but have not yet experienced the benefits of Bla. At this point, the antibiotic has not been completely removed; thus the population will be subjected to a slightly higher concentration of antibiotic at each additional dose. If the antibiotic is delivered too frequently, the accumulated antibiotic increases the rate of lysis, thus causing higher amounts of Bla to be released, ultimately leading to the faster removal of the antibiotic. However, Bla cannot fully degrade the antibiotic before the next dose is added and the population quickly dies off. Although the population is cleared, a higher number of doses is necessary because the degree of lysis per dose is not maximized. In other words, subsequent doses are applied before the full extent of lysis from the previous dose is observed. However, if the antibiotic is delivered too infrequently, then the population will have the chance to recover between doses. Once again, these conditions do not maximize the degree of lysis per dose and more doses are necessary to achieve the same amount of population decrease associated with doses applied more frequently. A final aspect to consider when designing a regimen is the total amount of antibiotic delivered (Fig. 3D). Although some of the model’s regimens using higher concentrations of antibiotic (a0 > 10) are associated with fewer doses, they have the highest net antibiotic concentration. These concentrations may not be optimal, due to potential adverse effects associated with using excessive amounts of antibiotic, such as the destruction of the normal microbial flora, interference with the immune response, increased nephrotoxicity, and selection for antibiotic resistant mutants [32,60–63]. Also, efficient use of antibiotics can help reduce treatment cost [14,35]. Using dose number and total antibiotic delivered, an effective and realistic regimen can be designed by minimizing the number of doses, the delivery frequency, and the total antibiotic delivered. We note that the predictive power of the recovery time is maintained for low or moderate inoculum sizes. In particular, our modeling demonstrates that a multi-dose regimen will clear a population if the time between doses is less than one recovery time, regardless of effective antibiotic concentration and inoculum size (S1 Fig). Similar to the base case, the regimen can be optimized to have the fewest doses and the lowest net antibiotic concentration delivered by selecting the lowest concentration of antibiotic associated with the longest recovery time. Additionally, the predictive power of the recovery time is maintained for an antibiotic with dose-dependent killing (Hill coefficient = 1) or an antibiotic with time-dependent killing (Hill coefficient = 10): a multi-dose regimen will clear a population if the time between doses is less than one recovery time, regardless of effective antibiotic concentration and degree of antibiotic-mediated killing (S2 Fig). The predictive power of the recovery time can be applied to bacteria with varying rates of Bla synthesis and accumulation as long as the antibiotic concentration applied has an effective recovery time (S3 Fig A-T). When the bacteria are producing and accumulating Bla at a very fast rate (S3 Fig P-T), most individual bacteria can sufficiently protect themselves (CAT is no longer necessary) and the population experiences little or no decline in density. Consequentially, the model predicts that effective treatment protocols would shift to higher antibiotic concentrations capable of inducing significant lysis in more resistant bacteria. The predictive power is upheld as long as the recovery times associated with subsequent doses are sufficiently similar to the original recovery time measured from a single dose. Recovery times of subsequent doses depend on two main factors: the activity of Bla in the environment and the concentration of antibiotic. On one hand, if there is insufficient time for Bla to degrade between doses, then it will compound with each dose until the population is being protected by higher concentrations of Bla, relative to when the first dose was administered. As a result, the increasing pool of Bla will degrade the antibiotic faster, the recovery time of subsequent doses will decrease, and the population can recover when dosed at period lengths less than the original recovery time (S4 Fig A-B). This would happen in scenarios where the antibiotic concentration applied is insufficient to counterbalance the Bla that is either expressed at high levels or has an increased rate for hydrolyzing an antibiotic. The loss of predictive power in this case can be avoided by using a sufficiently strong antibiotic concentration. On the other hand, if there is insufficient Bla to degrade the antibiotic between doses, then the antibiotic will compound with each dose until the population is being exposed to higher concentrations of antibiotic, relative to when the first dose was administered. As a result, the increasing concentration of antibiotic will kill more cells, the recovery time of subsequent doses will increase, and the population will not be able to recover when dosed with period lengths equal to the original recovery time (S4 Fig C-D). Many antibiotics, such as β-lactams, are most effective when applied continuously for long periods of time [64,65]. Thus, we also modeled the predictive power of the recovery time in intravenous (IV)-drip based protocols, where a set concentration of antibiotic is delivered over a set duration during each dosing period. Here, we delivered the antibiotic dose over three time units and measured the corresponding recovery time (Fig. 4A). Similar to the injection recovery times, the IV-drip recovery times increase monotonically as the concentration of the dose increases, more Bla is required to remove the antibiotic, and more of the population lyses. In contrast, the IV-drip therapy has a narrower range of intermediate antibiotics with 0 < recovery time < 100. Some of the lower concentrations that are effective for injection treatment (0.5 < a0 < 1) are ineffective for IV drip treatment because the dose is too weak when delivered over a longer period of time. However, when the dose concentration is sufficiently high, the IV-drip recovery time is longer than the injection recovery time because the IV-drip is exposing the bacteria to a higher concentration for a longer period of time (Fig. 4B). Again, we use the recovery time from a single IV dose to establish the range of dosing frequencies able to eliminate the bacterial population. At each dosing concentration (for a fixed time duration of 3), we applied 100 doses of the antibiotic at periods ranging from the infusion duration (τ = 3) to 2 times the corresponding recovery time and calculated the resulting final bacterial density. The model shows that the predictive power of the recovery time is maintained when the antibiotic dosing concentration is sufficiently large with a long enough recovery time (a0 > 1.5): a multi-IV-dose regimen will eventually eliminate the population if the dosing period is less than one recovery time, regardless of effective antibiotic concentration (Fig. 4C). There is slight deviation from this for a0 < 1.5 due to the corresponding recovery times being too short for the Bla to be reduced to a baseline concentration before the next round of lysis and Bla release occurs. As a result, periods less than one recovery time could fail to eradicate the infection because the Bla concentration compounds with each subsequent dose, the antibiotic is degraded more quickly, fewer cells lyse, and the population can recover (S4 Fig). Similar to the injection based therapy, the IV-drip reduced a population constitutively producing high concentrations of Bla as long as the period was less than one recovery time and the initial antibiotic concentration was sufficiently high to cause significant initial decline (S3 Fig U-Y). However, the IV-drip protocols retained a larger range of effective antibiotic concentrations than the injection based protocols. This robustness is due to the antibiotic concentration continuously being replenished from the IV-drip. If a high enough concentration is maintained for sufficient time, the population’s Bla concentration will not be able to remove the antibiotic fast enough to prevent lysis and the population will decrease with each subsequent round of IV-drip infusion. Thus, these results suggest that IV-drip based regimens could serve as a platform for effectively applying first-class β-lactams to clear constitutive producers of high levels of ESBLs. We next evaluated the efficacy of each effective concentration-period combination by calculating the minimum number of doses necessary to reduce the population density to below 10-10 (Fig. 4D). Relative to the injection protocol, the IV-drip therapy has a narrower region of intermediate dose numbers, reflecting the narrow region of intermediate recovery times. Similarly to the injection based regimens, the intermediate antibiotic concentrations (1 < a0 < 5) require the least doses when delivered at low-to-intermediate period lengths (period = 20–60% recovery time) because that is when the population is most vulnerable. Again, the initial antibiotic concentrations too low to have a recovery time (a0 < 1) do not clear the infection, regardless of the dosing interval or number of doses applied. The concentrations with an infinite recovery time (a0 > 5) require only a single dose and thus the dosing frequency does not matter. Although the number of doses necessary to clear an infection might be the same for a range of antibiotic concentrations and periods, the least amount of total antibiotic is needed for intermediate antibiotic concentrations applied at 20–60% of the associated recovery time (Fig. 4E). A bacterial population often consists of phenotypically or genetically heterogeneous subpopulations[66,67]. For instance, different cells may express different levels of Bla, have different growth rates, or exhibit different sensitivities to the same antibiotic. This heterogeneity could compromise the predictive power of the recovery time. To examine this notion, we extended our injection-based model to account for two cases, each dealing with a mixture of two subpopulations (S1 Text). In one case, one subpopulation grows much more slowly and exhibits much greater tolerance to the antibiotic. In the other, two subpopulations display different degrees of collective antibiotic tolerance. Most antibiotic regimens are based on empirical observations of how bacterial infections responded to an antibiotic [32,78,79]. However, these regimens may be suboptimal both because they were not initially designed to handle resistant bacteria and because the current diagnostic assays cannot accurately predict how resistant pathogens will respond to them. It is critical that we develop a new strategy for using the existing antibiotics more effectively or our medical care will return to a state equivalent to that of a pre-antibiotic era. Ideally, the new strategies would be based on the molecular mechanisms underlying antibiotic resistance. However, this is impractical, given that many pathogens’ resistance mechanisms have not been characterized and they evolve rapidly. To this end, we propose using the recovery time as a lumped metric that can characterize a pathogen’s response to an antibiotic without requiring knowledge of the underlying mechanism. We used a kinetic model to test the ability of the recovery time to predict ESBL-producing pathogens’ responses to periodic dosing of β-lactams. Our simulation results suggest that the recovery time of a single dose can be used to design optimal multi-dose regimens for multiple delivery methods, including injections and continuous IV drip, various inoculum sizes, bacteria with a range of Bla production levels, and certain heterogeneous populations. Optimal dosing regimens for treating Bla-producing bacteria with a β-lactam would apply intermediate concentrations of antibiotic that have long recovery times at time intervals corresponding with when the bacterial density has been minimized. Furthermore, our modeling results suggest that regimens using lower, yet still lethal, concentrations of antibiotic can be as effective as regimens using higher concentrations. Reducing the amount of antibiotic the host is exposed to may be important to minimize the perturbation of the host’s microbiota and other defense mechanisms [32,60–62,80], which could have long-lasting detrimental effects. Also, under certain conditions, a higher concentration of antibiotic can lead to selection of more resistant subpopulation of bacterial pathogens [81]. Although this model considers the population level response to an antibiotic, there is a significant amount of gene-expression noise at the single cell level [66,67]. If an antibiotic were applied such that the population would have the chance to recover between doses, then the antibiotic would select for the bacteria expressing higher levels of resistance genes (as demonstrated in S6 Fig F). Ultimately, this would direct the evolution of the population towards an inherently more resistant infection than before the antibiotic treatment was applied. Our proposed method would minimize this problem by delivering subsequent doses of antibiotic before a more resistant population grew to a significant density. The recovery time of a pathogen under a single dose of antibiotic is a metric that is easy to measure and could guide the choice of an appropriate multi-dose antibiotic regimen for a wide range of infections. Measurements of the recovery time can be carried out in high resolution using commercially available microplate readers [82]. A critical step entails the construction of a comprehensive recovery time database for various pathogens under different antibiotics (Fig. 5). When a new bacterial pathogen is identified, its recovery times to a range of antibiotic concentrations will be recorded in vitro for different starting densities. Based on these measurements, regimens with varied concentrations and period lengths will be tested for different inoculum sizes. From these results, the period length, dose number, and antibiotic concentration can be optimized for a particular pathogen in vitro. Before entering this information into the database, the PK/PD of the particular antibiotic will be necessary to determine the concentration of antibiotic that should be delivered such that the concentration at the site of infection matches the concentration selected from the in vitro experiments. Given this database, a proper diagnosis of a pathogen, and an estimate of the severity of the infection (e.g. inoculum size), one can readily identify the scenarios in which first- and second-line antibiotics may still be applied and chose the most effective treatment protocol. Whenever a new pathogen arises, it can be evaluated and added to the library. The ability to predict the outcome of a multi-dose treatment without knowing the underlying resistant mechanism would remove the uncertainty that prevents clinicians from using first-line β-lactams when an ESBL-producing pathogen is detected. Given ESBL-producing bacteria’s prevalence [19,83–85], our proposed strategy could help to minimize the rate at which these bacteria develop resistance to more extreme antibiotics by ensuring that we do not overlook effective first-line antibiotics before moving on to more extreme antibiotics. The interaction between a β-lactam and a bacterial population expressing Bla can be simplified to the interactions between three main components: population density (n), antibiotic concentration (a), and Bla concentration (b). Our base model consists of the following ordinary differential equations: dndτ=(g−l)n (1) dboutdτ=lbin*−γ2bout−κIV(τ)bout (2) da dτ=κIV(τ)ainject−(bout+αbin*)(a1+a)−γ3a−κIV(τ)a (3) g=(1−n)(σ1σ1+a) (4) l=γ1(aHσ2H+aH)(σ4σ4+bin) (5) bin=κ(rg+γ4) (6) bin*=βnbin (7) r=aσ3+a (8) where g and l represent bacteria growth and lysis, respectively. Initial conditions of n(0) = 0.1, b(0) = 0, and a(0) = 0.01–100 were used for all simulation results, except for S2 Fig where n(0) = 0.01 or 0.001. The rest of the parameters are defined in a table in S1 Text. See S1 Text for further details of the model development and for the extended models that account for heterogeneous populations. Minor modifications are introduced to account for the IV drip protocol or dynamics of a mixture consisting of two subpopulations (S1 Text).

10.1371/journal.pntd.0005274

Whole Genome Sequence Analysis of Salmonella Typhi Isolated in Thailand before and after the Introduction of a National Immunization Program

Vaccines against Salmonella Typhi, the causative agent of typhoid fever, are commonly used by travellers, however, there are few examples of national immunization programs in endemic areas. There is therefore a paucity of data on the impact of typhoid immunization programs on localised populations of S. Typhi. Here we have used whole genome sequencing (WGS) to characterise 44 historical bacterial isolates collected before and after a national typhoid immunization program that was implemented in Thailand in 1977 in response to a large outbreak; the program was highly effective in reducing typhoid case numbers. Thai isolates were highly diverse, including 10 distinct phylogenetic lineages or genotypes. Novel prophage and plasmids were also detected, including examples that were previously only reported in Shigella sonnei and Escherichia coli. The majority of S. Typhi genotypes observed prior to the immunization program were not observed following it. Post-vaccine era isolates were more closely related to S. Typhi isolated from neighbouring countries than to earlier Thai isolates, providing no evidence for the local persistence of endemic S. Typhi following the national immunization program. Rather, later cases of typhoid appeared to be caused by the occasional importation of common genotypes from neighbouring Vietnam, Laos, and Cambodia. These data show the value of WGS in understanding the impacts of vaccination on pathogen populations and provide support for the proposal that large-scale typhoid immunization programs in endemic areas could result in lasting local disease elimination, although larger prospective studies are needed to test this directly.

Typhoid fever is a systemic infection caused by the bacterium Salmonella Typhi. Typhoid fever is associated with inadequate hygiene in low-income settings and a lack of sanitation infrastructure. A sustained outbreak of typhoid fever occurred in Thailand in the 1970s, which peaked in 1975–1976. In response to this typhoid fever outbreak the government of Thailand initiated an immunization program, which resulted in a dramatic reduction in the number of typhoid cases in Thailand. To better understand the population of S. Typhi circulating in Thailand at this time, as well as the impact of the immunization program on the pathogen population, we sequenced the genomes of 44 S. Typhi obtained from hospitals in Thailand before and after the immunization program. The genome sequences showed that isolates of S. Typhi bacteria isolated from post-immunization era typhoid cases were likely imported from neighbouring countries, rather than strains that have persisted in Thailand throughout the immunization period. Our work provides the first historical insights into S. Typhi in Thailand during the 1970s, and provides a model for the impact of immunization on S. Typhi populations.

Salmonella enterica subspecies enterica serovar Typhi (S. Typhi) is a human restricted bacterial pathogen and the etiological agent of typhoid fever. S. Typhi is transmitted faeco-orally and can establish asymptomatic carriage in a small subset of an exposed population [1]. Recent estimates [2–4] place the global burden of typhoid fever at 25–30 million cases annually, of which 200,000 are associated with deaths. Typhoid fever occurs most commonly in industrialising countries, specifically in locations with limited sanitation and related infrastructure [5]; children and young adults are among the most vulnerable populations in these settings [6–8]. Antimicrobial therapy together with water sanitation and hygiene (WASH) interventions are the major mechanisms by which typhoid fever is controlled [9, 10]. However, none of these approaches are optimal and resistance against antimicrobials has become increasingly common in S. Typhi since the 1970s [11–13]. A number of typhoid vaccines are licenced for use [14–18], however, they are not widely used as a public health tools in endemic areas, with the exception of controlling severe outbreaks such as those following natural disasters [19–22]. A sustained typhoid fever outbreak occurred in Thailand in the 1970s. A sharp increase in cases was observed in 1973–1974, which finally peaked in 1975–1976. In response, the government of Thailand established a national typhoid immunization program, which represented the first programmatic use of a typhoid vaccine in the country [14, 22, 23]. The immunization program targeted over 5 million school aged children (7–12 years) each year in Bangkok between 1977 and 1987 (80% of the eligible population). Thus, Thai school children were eligible to receive a single locally produced heat/phenol-inactivated subcutaneous dose of 2.5 x 108 S. Typhi organisms annually [14, 22, 23], before the program was halted in the early 1990s because of high rates of adverse reactions caused by the vaccine [22]. To our knowledge this is the only such programmatic use of a vaccine for controlling Typhoid fever in children in Thailand. Data from four teaching hospitals in Bangkok showed a 93% reduction in blood culture confirmed infections with S. Typhi between 1976 (n = 2,000) and 1985 (n = 132) [14, 23]. Notably, no significant decline was observed in isolation rates of Salmonella Paratyphi A (S. Paratyphi A), a Salmonella serovar distinct from S. Typhi that causes a clinical syndrome indistinguishable from typhoid fever, but for which S. Typhi vaccines provide little or no cross-protection [14]. This observation suggests that the reduction in S. Typhi infections was not attributable to improvements in infrastructure and hygiene practices only [5, 14, 20, 23]. While the inactivated S. Typhi vaccine was found to be highly efficacious [22, 23], it is no longer used as a consequence of being overly reactogenic [14, 16, 22, 23, 24]. A Vi capsular polysaccharide vaccine [15] and live-attenuated oral vaccine of strain Ty21a [16] have since replaced this vaccine for travellers to endemic locations [5, 21, 24]. The typhoid immunization program in Thailand provided a unique opportunity to investigate the impact of immunization on S. Typhi populations circulating within an endemic area. Here we present an analysis of a historical collection of 44 S. Typhi isolates obtained from patients in Thailand between 1973 and 1992 (before and during the immunization program). As S. Typhi populations demonstrate little genetic diversity, we used whole genome sequencing (WGS) to characterise these isolates, and core genome phylogenetic approaches to compare the historic isolates from Thailand to a recently published global S. Typhi genomic framework [4]. This is a retrospective study of bacterial isolates unlinked to patient information and was not subject to IRB approval. Forty-four S. Typhi isolated from patients with suspected typhoid fever attending hospitals in Bangkok, Nonthaburi, Loi, and Srakaew, in Thailand between 1973 and 1992 were available for genome sequencing in this study (Fig 1 and S1 Table). At the time of original isolation, bacterial cultures were transferred on nutrient agar slants to the department of Enteric Diseases, Armed Forces Research Institute of Medical Sciences (AFRIMS), Bangkok, Thailand for identification and antimicrobial susceptibility testing. At AFRIMS, bacterial isolates were subcultured on Hektoen Enteric agar (HE) and identification was performed by biochemical testing on Kligler iron agar slants, tryptone broth for indole, lysine decarboxylase medium, ornithine decarboxylase medium, urease test, mannitol and motility media (Becker Dickenson, Thailand). Serological agglutination was performed using Salmonella O antisera and Salmonella Vi antiserum (Difco, USA). Bacterial strains were stored frozen at -70°C in 10% skimmed milk or lyophilised in 10% skimmed milk; lyophilized ampoules were stored at 2–8°C. Prior to DNA extraction for sequencing, lyophilized bacteria were rehydrated with trypticase soy broth, inoculated on McConkey agar and incubated at 37°C for 18–24 hours. If bacteria were stored frozen in skimmed milk, organisms were inoculated directly onto McConkey agar after thawing and then incubated at 37°C for 18–24 hours. Antimicrobial susceptibility testing against ampicillin, chloramphenicol, cephalothin, gentamicin, kanamycin, neomycin, sulfisoxazole, trimethoprim/sulfamethoxazole, and tetracycline was performed by disk diffusion according to Clinical and Laboratory Standards Institute (CLSI) [25–28]. Genomic DNA from the 44 S. Typhi from Thailand was extracted using the Wizard Genomic DNA Extraction Kit (Promega, Wisconsin, USA). Two μg of genomic DNA was subjected to indexed WGS on an Illumina Hiseq 2000 platform at the Wellcome Trust Sanger Institute, to generate 100 bp paired-end reads. For analysis of SNPs, paired end Illumina reads were mapped to the reference sequence of S. Typhi CT18 (accession no: AL513382) [29] using the RedDog (v1.4) mapping pipeline, available at https://github.com/katholt/reddog. RedDog uses Bowtie (v2.2.3) [30] to map reads to the reference sequence, then high quality SNPs called with quality scores above 30 are extracted from the alignments using SAMtools (v0.1.19) [31]. SNPs were filtered to exclude those with less than 5 reads mapped or with greater than 2.5 times the average read depth (representing putative repeated sequences), or with ambiguous base calls. For each SNP that passed these criteria in any one isolate, consensus base calls for the SNP locus were extracted from all genomes (ambiguous base calls and those with phred quality scores less than 20 were treated as unknowns and represented with a gap character). SNPs with confident homozygous allele calls (i.e. phred score >20) in >95% of the S. Typhi genomes (representing a ‘soft’ core genome of common S. Typhi sequences) were concatenated to produce an alignment of alleles at 45,893 variant sites. The resultant allele calls for 68 of these SNPs were used to assign isolates to previously defined lineages according to an extended S. Typhi genotyping framework [32] code available at https://github.com/katholt/genotyphi). SNPs called in phage regions, repetitive sequences (354 kb; ~7.4% of bases in the CT18 reference chromosome, as defined previously [33] or recombinant regions (~180kb; <4% of the CT18 reference chromosome, identified using Gubbins (v1.4.4) [34]) were excluded, resulting in a final set of 1,850 SNPs identified in an alignment length of 4,275,037 bp for the 44 isolates. SNP alleles from Paratyphi A strain 12601 [35] were also included as an outgroup to root the tree. For global context, raw read data [4] were also subjected to genotyping analysis and those isolates sharing the genotypes that were observed in the Thai collection (n = 340; details in S2 Table) were subjected to the same SNP analyses, resulting in a final set of 9,700 SNPs for a total of 386 isolates. Maximum likelihood (ML) phylogenetic trees (Figs 1 and 2) were constructed using the 1,850 and 9,700 bp SNP alignments, respectively, using RAxML (v 8.1.23) [36] with a generalized time-reversible model and a gamma distribution to model site specific recombination (GTR+Γ substitution model; GTRGAMMA in RAxML), with Felsenstein correction for ascertainment bias. Support for ML phylogenies was assessed via 100 bootstrap pseudoanalyses of the alignments. For the larger tree containing global isolates, clades containing only isolates from only a single country were collapsed manually in R using the drop.tip() function in the ape package [37]. Subtrees were extracted for each subclade, which are therefore each rooted by the other subclades. Pairwise SNP distances between isolates were calculated from the SNP alignments using the dist.gene() function in the ape package for R [37]. Acquired antimicrobial resistance (AMR) genes were detected, and their precise alleles determined, by mapping to the ARG-Annot database [38] of known AMR genes using SRST2 v0.1.5 [39]. Plasmid replicon sequences were identified using SRST2 to screen reads for replicons in the PlasmidFinder database [40, 41]. Raw read data was assembled de novo with SPAdes (v 3.5.0) [42] and circular contigs were identified visually and extracted using the assembly graph viewer Bandage (v0.7.0) [43]. These putative plasmid sequences were annotated using Prokka (v1.10) [44] followed by manual curation. Where IncHI1 plasmid replicons were identified using SRST2, and their presence confirmed by visual inspection of the assembly graphs, IncHI1 plasmid MLST (pMLST) sequence types were determined using SRST2 [13, 39, 45, 46]. Where resistance genes were detected from short read data, Bandage was used to inspect their location in the corresponding de novo assembly graph in order to determine whether they were encoded in the bacterial chromosome or on a plasmid. Assembled contigs were concatenated and putative prophage genomes were identified with the PHAge Search Tool (PHAST) [47], and their novelty determined by BLASTN analysis against the GenBank database. Pairwise alignments between novel and known prophage sequences were visualised using the genoPlotR package for R [48]. Raw sequence data have been submitted to the European Nucleotide Archive (ENA) under project PRJEB5281; individual sample accession numbers are listed in S1 and S2 Tables. Assembled phage and protein sequences were deposited in GenBank, accession numbers are listed in Table 1. All 44 S. Typhi isolates collected between 1973 and 1992 were subjected to WGS and SNP analysis. Genome-wide SNPs were used to construct a ML phylogeny and isolates were assigned to previously defined genotypes [32] using a subset of SNPs (see Methods). These analyses subdivided the population into ten distinct genotypes, each corresponding to a specific lineage in the ML phylogeny (Fig 1). Genotype 3.2.1 (which includes the reference genome CT18, isolated from Vietnam in 1993 [29]) was the most common (n = 14, 32%), followed by genotype 2.1.7 (n = 10, 23%). Genotypes 2.0 (n = 1, 2%) and 4.1 (n = 3, 7%) were observed only in 1973 (pre-vaccine period). Genotypes 2.1.7 (n = 10, 23%), 2.3.4 (n = 1, 2%), 3.4.0 (n = 2, 5%), 3.0.0 (n = 3, 7%), 3.1.2 (n = 2, 5%), were observed only after 1981 (post-vaccine period). Each of these post-immunization genotypes was from a single location and time period (Fig 1), consistent with short-term localised transmission. The only exceptions were the two S. Typhi 3.1.2 isolates, that were from Srakaew in 1989 and Bangkok in 1992 and separated by just 4 SNPs. Genotypes 3.2.1 and 2.4.0 were observed amongst both pre- and post-vaccine isolates. Based on the Thai S. Typhi genotyping results we hypothesised that the post-immunization typhoid infections in Thailand resulted from occasional re-introduction of S. Typhi from outside the country, as opposed to long-term persistence of S. Typhi lineages within Thailand. To explore this possibility, and to provide a global context for our analysis, we examined 1,832 S. Typhi genomes from a recently published global collection that included isolates from 63 countries [4]. Genome-wide SNP-based ML trees for each of these genotypes, showing the relationships between Thai and global isolates, are shown in Fig 2. In general, post-vaccine Thai isolates were closely related to recent isolates sourced from neighbouring countries including Vietnam, Laos and Cambodia (Fig 2), consistent with regional endemic circulation. In contrast, most pre-vaccine isolates had no close neighbours in the global collection, particularly 2.0.0 strains (Fig 2A), suggesting they may have been Thailand-specific lineages that have died out following the vaccine program. The S. Typhi genomes in the global collection were mainly isolated 2–3 decades after the Thai isolates as we did not have access to contemporaneous isolates from these countries that could identify specific transfer events. However, all but three of the post-vaccine Thai isolates shared shorter SNP distances with isolates from neighbouring countries than they did with pre-vaccination Thai isolates (see Fig 3), consistent with these cases being caused by occasional re-introduction of genotypes circulating in the region. Notably, Thai S. Typhi 3.2.1 that were isolated in 1986–7 clustered separately from the 1973 pre-vaccine isolates (≥60 SNPs apart), but closely with isolates from Vietnam and Cambodia (differing by as few as 7 SNPs; Fig 2H). Post-vaccine Thai S. Typhi 2.4 formed two distinct groups that were not consistent with direct descendance from earlier isolates (Fig 2E). These data are therefore consistent with transfer of S. Typhi into Thailand from neighbouring countries during the post-immunization program era, although the long-term circulation of ancestral populations in Thailand remains an unlikely alternative explanation. We identified acquired AMR genes in the genomes of four S. Typhi genotype 3.2.1 that were isolated in Srakaew in 1986 (Fig 1, Table 1). These isolates shared the same four AMR genes: sul1 (sulphonamides), catA1 (chloramphenicol), tet(B) (tetracyclines), and aadA1 (aminoglycosides) which were carried on near-identical plasmids of IncHI1 plasmid sequence type 2 (PST2). Although the presence of insertion sequences (IS) in these plasmids prevented the complete sequences from being assembled, the regions of these plasmids encoding the AMR genes were identical in all assemblies. This commonality suggests they are a single plasmid (referred to as pTy036_01 in Fig 1 and Table 1) that was likely acquired in a common ancestor of this clade. The chromosomal and IncHI1 plasmid sequences for these four isolates were very closely related to those of a 1993 Vietnamese isolate (Viety1-60_1993) in the global S. Typhi collection [4, 45], consistent with regional transfer. We identified three non-AMR related plasmids amongst the Thai isolates (Fig 1, Table 1). Ty004 (genotype 2.2) carried two novel plasmids that assembled into circular sequences, pTy004_01 and pTy004_02. The largest, pTy004_01, was a novel variant of the cryptic plasmid pHCM2 [29, 49] (Fig 4). Ty004 was isolated in Bangkok in 1973, making pTy004_01 the earliest example of a pHCM2-like plasmid reported to date. pTy004_01 was distant from other pHCM2-like plasmids in the global S. Typhi genome collection, sharing 92% coverage and 99% nucleotide identity with the reference sequence pHCM2 of S. Typhi CT18 (genotype 3.2.1) which was isolated approximately 20 years later in Vietnam [29]. The pTy004_01 sequence (Fig 4) appears to be ~2 kbp larger than pHCM2, and encodes an additional tRNA-Lys as well as an insertion of a hypothetical protein (orf17) into a putative DNA polymerase gene (HCM2.0015c in pHCM2, divided into orf16 and orf18 in pTy004_01). Plasmid pTy004_02 was ~38 kbp in size and similar to E. coli plasmid pEQ2 (65% coverage, 98% nucleotide identity), encoding genes for conjugation, chromosomal partitioning, addiction systems and an abortive infection protein (orf44). Three isolates (Ty031, Ty042, and Ty049) all of genotype 3.0.0 and obtained from Srakaew in 1986, carried a ~40 kbp cryptic plasmid that we named pTy031_01. This plasmid was similar to that carried by Enterobacter hormaechei strain CAV1176 (83% coverage, 96% identity) and encoded genes for chromosomal partitioning, addiction systems, and a putative restriction modification system (orf33-orf34). PHAST analysis revealed the presence of novel intact prophages in three Thai S. Typhi isolates (Fig 1, Table 1). Two S. Typhi 3.1.2, isolated from Srakaew in 1989 and Bangkok in 1992, shared a novel phage STYP1 that was similar to fiAA91-ss infective for Shigella sonnei (Fig 5A). However, the S. Typhi phage lacked the cytolethal distending toxin cdt genes and the IS21 element found in phage fiAA91-ss [50]. This prophage sequence had a mosaic architecture, incorporating a number of putative insertions of phage tail fiber genes that were not present in the fiAA91-ss reference genome (Fig 5A). Additionally, a single isolate of genotype 4.1 obtained from Bangkok in 1973 contained a novel SfIV-like phage, here named STYP2, that lacked the serotype conversion gene Gtr cluster and IS1 element of phage SfIV [51]. Again, the novel Thai phage variant also encoded novel tail fiber genes not in the SfIV reference genome, as well as a Dam methylase gene (orf37) (Fig 5B). These data provide a historical insight into the population structure of S. Typhi in Thailand in 1973 (pre-immunization program, n = 11) and 1981–1992 (post- immunization program, n = 33). It has been reported that the national S. Typhi immunization program in Thailand, which commenced in 1977, was highly effective in reducing the burden of typhoid fever [14]. Our data are consistent with the hypothesis that the vaccine program successfully depleted the endemic S. Typhi population to the extent that most subsequent typhoid cases resulted from sporadic introduction of non-indigenous S. Typhi, rather than long-term persistence of the pre-vaccine era population. It is apparent that these introductions were sometimes accompanied by limited local transmissions, resulting in small, localized outbreaks, but we found no evidence to suggest that these result in the establishment of stable local source populations. Notably, the post-immunization S. Typhi isolates from Loi (in the north of Thailand near the border with Laos, from which it is separated by the Mekong river) were most closely related to Laos isolates, whilst those from the capital Bangkok and nearby Nonthaburi and Srakaew districts were closely related to other isolates from across Southeast Asia (Fig 2), suggesting there may have been multiple routes of import into Thailand. Our study is limited by the sample of isolates available for analysis, which was small and reflects opportunistic sampling of sporadic local cases in the four sites and historical storage. A larger collection of historical isolates from Thailand and neighboring countries in the 1970s and 1980s would help to further elucidate the epidemiological patterns of S. Typhi before and after the vaccination program. However, from our data, it is notable that the Thai isolates cluster according to site, consistent with limited local transmission rather than dissemination of lineages between locations. The only exception to this was two genotype 3.1.2 isolates, which were collected from Srakaew in 1989 and Bangkok in 1992 and differed by only 4 SNPs. This is consistent with either transfer between these cities in Thailand following an initial introduction into the country, or two independent transfers into Thailand from a common source. The phylogenetic structure is most suggestive of the latter, but denser samples from Thailand and/or potential source populations would be required to resolve this with confidence. While our sample is small, this study is nevertheless the largest to date exploring genetic diversity amongst S. Typhi from Thailand. An earlier global haplotyping study that included seven Thai isolates [52] identified five distinct haplotypes in Thailand (H3, 1989; H42, 1990; H50, 2002; Vi- H52, 1990; H79, 2002), three of which are related to genotypes that we identified amongst Thai strains in this study (H79, 2.3.4; H52, 3.4; H42, 3.1.2) [32]. Genotype 4.3.1 (H58) was not found amongst our historical Thai isolates. This is consistent with previously published spatiotemporal analyses of the global isolate collection, which showed this rapidly expanding clone only began spreading throughout Asia after 1990 [4]. To our knowledge the only evidence to date of the presence of 4.3.1 (H58) in Thailand comes from the global study [4], in which three isolates were identified from 2010–2011, most likely introduced from India. Therefore, our genomic snapshot of the Thai S. Typhi population is consistent with previous insights and is likely reasonably representative for the study period. In the years following the vaccination program the prevalence of Typhoid fever in Thailand has continued to decline [53, 54]. The vaccination program has been credited with reducing disease incidence in Thailand and was followed by increased economic development in the region as well as improvements to both water and sanitation systems that have likely improved the control of such outbreaks [53, 54]. Consequently, Typhoid fever is no longer considered a serious public health threat in Thailand [53]. The presence of novel plasmids and prophages in the Thai isolates is also noteworthy. While small plasmids of unknown function have been observed in S. Typhi previously [55], they are infrequent compared to the IncHI1 MDR plasmid and the cryptic plasmid pHCM2 [33]. Presumably, such plasmids are ephemeral; possibly because their maintenance imposes a fitness burden on the host cells so a strong selective advantage is required for retention [56, 57]. It is also possible that the lack of previous reports regarding the diversity of small plasmids in S. Typhi reflects a technological complexity, however, this is bypassed with high-throughput WGS and we detected negligible small plasmid content in the global collection of 1,832 genomes using the same screening approach [4, 32, 58]. Notably, few of the Thai plasmids share nucleotide sequence homology with those previously described in S. Typhi, but were closely related to those found in other Enterobacteriaceae. The novel pHCM2-like plasmid (pTy004_01) and two additional plasmids (pTy004_02 and pTy031_01) harbored genes associated with phage resistance, which could provide protection against phage predation [59–62]. We also observed two novel prophages integrated into Thai genomes, which both showed variation in their phage tail structural regions compared to close neighbors found in Shigella/E. coli. These regions are typically responsible for binding of phage to host receptors [63–65], thus the variation in these regions may be associated with recent adaptations to the S. Typhi host. While genomic data from more recent S. Typhi collections shows limited evidence for genetic exchange with other organisms [4], the detection amongst older Thai isolates of both phage and plasmids that have been previously associated with E. coli/Shigella suggests that genetic exchange may have been more common in the past or in certain localized populations. Overall, these data provide valuable historical insights into the S. Typhi populations circulating in Thailand during the 1970s and 1980s, and early examples of the two most common S. Typhi plasmids, as well as other mobile elements identified within the S. Typhi population. Importantly, while genomic epidemiology has been applied to study typhoid transmission, antimicrobial resistance evolution and antibiotic treatment failure in various settings [66–68], this study provides an important proof-of-principle demonstration that this approach can also provide useful insights into the impact of typhoid vaccines on circulating bacterial populations. This should motivate the adoption of WGS methods to monitor S. Typhi populations during future immunization programs and other large-scale interventions, which could potentially identify differential impacts on distinct genotypes.

10.1371/journal.pcbi.1005014

Modeling Electrophysiological Coupling and Fusion between Human Mesenchymal Stem Cells and Cardiomyocytes

Human mesenchymal stem cell (hMSC) delivery has demonstrated promise in preclinical and clinical trials for myocardial infarction therapy; however, broad acceptance is hindered by limited understanding of hMSC-human cardiomyocyte (hCM) interactions. To better understand the electrophysiological consequences of direct heterocellular connections between hMSCs and hCMs, three original mathematical models were developed, representing an experimentally verified triad of hMSC families with distinct functional ion channel currents. The arrhythmogenic risk of such direct electrical interactions in the setting of healthy adult myocardium was predicted by coupling and fusing these hMSC models to the published ten Tusscher midcardial hCM model. Substantial variations in action potential waveform—such as decreased action potential duration (APD) and plateau height—were found when hCMs were coupled to the two hMSC models expressing functional delayed rectifier-like human ether à-go-go K+ channel 1 (hEAG1); the effects were exacerbated for fused hMSC-hCM hybrid cells. The third family of hMSCs (Type C), absent of hEAG1 activity, led to smaller single-cell action potential alterations during coupling and fusion, translating to longer tissue-level mean action potential wavelength. In a simulated 2-D monolayer of cardiac tissue, re-entry vulnerability with low (5%) hMSC insertion was approximately eight-fold lower with Type C hMSCs compared to hEAG1-functional hMSCs. A 20% decrease in APD dispersion by Type C hMSCs compared to hEAG1-active hMSCs supports the claim of reduced arrhythmogenic potential of this cell type with low hMSC insertion. However, at moderate (15%) and high (25%) hMSC insertion, the vulnerable window increased independent of hMSC type. In summary, this study provides novel electrophysiological models of hMSCs, predicts possible arrhythmogenic effects of hMSCs when directly coupled to healthy hCMs, and proposes that isolating a subset of hMSCs absent of hEAG1 activity may offer increased safety as a cell delivery cardiotherapy at low levels of hMSC-hCM coupling.

Myocardial infarction—better known as a heart attack—strikes on average every 43 seconds in America. An emerging approach to treat myocardial infarction patients involves the delivery of human mesenchymal stem cells (hMSCs) to the damaged heart. While clinical trials of this therapeutic approach have yet to report adverse effects on heart electrical rhythm, such consequences have been implicated in simpler experimental systems and thus remain a concern. In this study, we utilized mathematical modeling to simulate electrical interactions arising from direct coupling between hMSCs and human heart cells to develop insight into the possible adverse effects of this therapeutic approach on human heart electrical activity, and to assess a novel strategy for reducing some potential risks of this therapy. We developed the first mathematical models of electrical activity of three families of hMSCs based on published experimental data, and integrated these with previously established mathematical models of human heart cell electrical activity. Our computer simulations demonstrated that one particular family of hMSCs minimized the disturbances in cardiac electrical activity both at the single-cell and tissue levels, suggesting that isolating this specific sub-population of hMSCs for myocardial delivery could potentially increase the safety of future hMSC-based heart therapies.

Ischemic heart disease, which results from reduced coronary flow of oxygenated blood, is a leading cause of myocardial infarction and heart failure. This insufficient oxygenation results in the death of cardiomyocytes, which are normally incapable of substantial regeneration. Therefore, despite tremendous advancements in pharmacological and interventional therapeutic approaches, ischemic heart disease continues to be responsible for nearly 1 out of 6 deaths in the United States [1, 2]. This has motivated novel cardiotherapeutic strategies to repair and regenerate heart muscle, including human mesenchymal stem cell (hMSC) therapy, the method of interest in this study. In clinical trials for treating myocardial infarction, the delivery of autologous bone marrow derived hMSCs has demonstrated improved ventricular ejection, enhanced angiogenesis, decreased fibrosis and scar size, and minimal immune response [3]. However, the benefits have often been modest and transient [4, 5], underscoring a need to better understand and exploit the underlying mechanisms by which hMSCs interact with human cardiomyocytes (hCMs) [6]. This limited mechanistic knowledge further makes it difficult to ensure long-term stability, with seamless structural and functional integration into the host tissue [7–9]. Therefore, deeper investigation into the mechanisms of how hMSCs impact cardiac function is necessary. Proposed hMSC-hCM interactions predominantly include: reprogramming of host hCMs, transdifferentiation of hMSCs into hCMs, paracrine signaling, electrophysiological coupling, and cellular fusion [6, 10]. Indirect paracrine signaling through the release of largely unidentified soluble factors is thought to play an important role [6, 11]; however, hMSCs have also exhibited functional direct electrical interactions with cardiomyocytes both in vitro and in vivo [10, 12–17], motivating ongoing investigations of the electrophysiological coupling and cellular fusion mechanisms. In particular, Valiunas et al. showed that hMSCs form connexin 43-mediated gap junctions between each other and with acutely isolated canine cardiomyocytes, suggesting the ability to form heterocellular electrical networks [15]. Later in vitro studies showed that such electrical connections can be functional and potentially arrhythmogenic, as co-culturing murine cardiomyocytes with greater than 10 percent of hMSCs decreased conduction velocity (CV) and predisposed re-entrant arrhythmias [16]. Pro-arrhythmic characteristics were also detected in vivo, where pigs receiving intravenous injections of mesenchymal stem cells had decreased effective refractory periods [17]. Moreover, Shadrin et al. recently reported a 25–40% incidence of hybrid cell formation of hMSCs and neonatal rat ventricular myocytes through cell fusion [10]. However, species-specific effects can limit the clinical relevance of such animal and in vitro studies, and similarly controlled experiments are difficult to perform in human patients. While hMSC therapy clinical trials are yet to report arrhythmogenicity [18], such adverse effects remain a concern. Therefore, in this study, it was of interest to assess the electrophysiological safety of various levels of direct hMSC-hCM electrical interactions under healthy conditions [18], and to predict methods of improving the safety of this therapy. Mathematical modeling is a powerful tool that can simulate direct intercellular electrical interactions between hMSCs and hCMs. Electrophysiological models have been established to describe hCMs [19–21], as well as their interactions with other resident heart cells [22–25], but never before with hMSCs. Therefore, in this study, the various types of currents experimentally characterized in hMSCs [26–29] were mathematically modeled to simulate an empirically classified triad of hMSC families distinguished by their respective functional ion channels: Type A) delayed rectifier-like hEAG1 and calcium activated potassium currents; Type B) delayed rectifier-like hEAG1, calcium activated potassium, tetrodotoxin (TTX)-sensitive sodium, and L-type calcium currents; and Type C) calcium activated potassium and transient outward currents [26, 28]. The empirical distinction of these three hMSC families was originally reported by Li et al. [26] based on patch clamp measurements of bone marrow-derived hMSCs obtained commercially and maintained in monolayer culture. We then simulated the electrical activity of hMSCs coupled to healthy hCMs, and interpreted the model findings within the context of prior in vitro and in vivo experiments to identify possible opportunities to minimize arrhythmic potential in future hMSC-based cell delivery cardiotherapies. The hMSC transmembrane voltage can be modeled as: d V d t = - 1 C m ( I stim + I tot , i ) (1) where V is voltage, t is time, Cm is the cell capacitance, Istim is a stimulus current, and Itot,i is the total transmembrane ionic current of Type i hMSCs (where i = A, B, or C). The total transmembrane ionic current for Types A, B, and C hMSCs are given by Eqs 2, 3 and 4, respectively: I tot , A = I KCa + I dr + I L , A (2) I tot , B = I KCa + I dr + I LCa + I Na + I L , B (3) I tot , C = I KCa + I to + I L , C (4) where IKCa is the calcium activated potassium current, Idr is the delayed rectifier-like hEAG1 current, IL,i is the leakage current for hMSC Type i (where i = A, B, or C), ILCa is the L-type calcium current, INa is the TTX-sensitive sodium current, and Ito is the transient outward current. To describe each type of hMSC ionic current, either Hodgkin-Huxley-like or Markovian-like approaches were taken. Parameters for these models were fit to published experimental hMSC data using numerical methods described in S1 Text. Parameters used in this study can be found in Tables A-G of S1 Text. To quantify the impact of each hMSC parameter on the hCM APD, an established multivariable regression analysis was performed [46, 47]. In 300 trials for each hMSC model, we randomly varied hMSC maximum conductance parameters and time constant parameters by a log-normally distributed pseudorandom scale factor with a standard deviation of 10%, as described elsewhere [48]. hMSCs were coupled to midcardial hCMs in a 1:1 ratio in this analysis. From the changes in the model APD outputs (Y) and parameters (X), a linear approximation can be made to find the normalized parameter sensitivity vector (B), such that Y ≈ XB. Therefore, a positive or negative sign of B (i.e., an element of B) indicates a positive or negative correlation between the parameter of interest and APD, respectively. Furthermore, the magnitude of B indicates the sensitivity of the APD to the parameter of interest. To better demonstrate the sensitivity of the APD output to each hMSC cell type, B was scaled by σAPD, the standard deviation of the APDs for each set of 300 trials for a respective hMSC cell type. In this study, three hMSC electrophysiology models were developed based on published experimental data. These three models were subsequently used to develop insight into hMSC-hCM electrical interactions. Three novel electrophysiological models were developed for the triad of hMSC families based on empirical data [26]. After successfully modeling each type of current expressed in hMSCs (Fig 1), it was necessary to validate the whole-cell models by simulating Itot,A, Itot,B, and Itot,C. Total current whole-cell voltage-clamp simulations of Types A, B, and C hMSCs are shown in Fig 3, along with schematics of functional currents for each cell type [26]. Like experimental recordings [26], our simulation had a conditioning potential of -80 mV, followed by 10 mV voltage steps for 300 ms between -60 mV and 60 mV, and a final holding potential of -30 mV. Overall, fitting individual currents (Fig 1) allowed for ample reconstruction of representative whole cell electrical activity. The simulations for Types A and B hMSCs, both of which possess delayed rectifier-like channel activity, generally agree with the magnitude and behavior of experimental total currents for a wide range of voltage contours [26]. As demonstrated by Li et al., Idr at a potential of 60 mV has a standard deviation of approximately 90 pA, and the activation time constant for Idr at a holding potential of -80 mV has a substantial standard deviation of approximately 25 ms [26]. Since these deviations affect the amplitude and activation kinetics of Types A and B hMSCs, we performed a sensitivity analysis to determine the impact of these parameters on hCM APD (see Sensitivity Analysis below for details). Type C hMSCs, absent of functional hEAG1 expression, also reproduced the magnitude and form of the experimental voltage-clamp experiments characterizing this hMSC family’s electrophysiological behavior [26]. Therefore, we used each of these hMSC models to predict the direct electrical interactions between hMSCs and hCMs. The three models developed in this study were each coupled and fused to hCMs to better understand direct cell-cell electrical interactions during hMSC cardiotherapies. To understand the arrhythmogenic effects of direct hMSC-hCM coupling at the tissue level, a VW analysis was performed on a single layer, anisotropic 5 cm × 5 cm 2-D midcardial tissue with 0% hMSCs (healthy control), 5%, 15%, or 25% randomly inserted hMSCs repeated for three different configurations per condition (see Fig 9A and S1–S4 Videos for sample re-entry simulations at selected S1–S2 intervals). As shown in Fig 9B, VWs lengthened with increasing percent of inserted hMSCs. Interestingly, at low (5%) insertion levels, VWs were dependent on the type of coupled hMSCs (Fig 9B); inserting hMSCs with delayed rectifier-like activity (i.e. Types A, B, and mixed populations of hMSCs) led to substantially larger VWs (approximately 15 to 20 ms) compared to Type C hMSCs (VW = 2.0 ± 0.5 ms, n = 3). At greater levels of hMSC insertion (i.e., 15% and 25%), VWs were nearly independent of the type of coupled hMSCs, and VWs exceeding 50 ms were observed. The S1–S2 intervals that led to re-entry for each hMSC type at low levels of insertion are shown in Fig 9C. As expected, the shifts in S1–S2 intervals leading to re-entry depended on the different mean tissue APDs (S12 Fig). Various modeling studies have demonstrated APD dispersion may influence re-entry [45, 49], while APD restitution slope, the range of DIs for APD restitution slopes greater than 1, and CV restitution slope are key factors in restitution-induced instability [20, 50–53]. Fig 10 illustrates the effects of the percentage and types of hMSCs on each of these arrhythmogenic factors. As expected, APD dispersion (ζ) increased with greater levels of hMSC insertion for all hMSC types (Fig 10A). However, APD dispersion was approximately 21%, 18%, and 17% lower for Type C hMSCs compared to hMSCs with delayed rectifier-like activity at 5%, 15%, and 25% hMSC insertion, respectively (Fig 10A). S12 Fig shows APD maps for cardiac tissues with 5% hMSC insertion. APD restitution slopes, as well as the range of DIs for slopes greater than 1, were slightly decreased following coupling with each hMSC type (Fig 10B; for raw APD restitution curves, see S13 Fig). Even at 25% hMSC insertion, the shift in maximum APD restitution slope was less than 10% (Fig 10B). CV restitution slopes markedly decreased following hMSC insertion, by as much as 71% at 25% hMSC insertion (Fig 10C). The coupling effects on CV restitution slopes were predominately dependent on percentage of hMSC inserted, rather than the type of hMSC (Fig 10C). Altogether, this dispersion of refractoriness and restitution analysis supports the claim that increased arrhythmogenic potential of inserted stem cells is minimized by Type C hMSCs at low levels of hMSC insertion, as VW and APD dispersion are lowest for this cell type, without adversely affecting APD and CV restitution slopes in comparison to delayed rectifier-like hMSCs. Our study provides insight into hMSC electrical activity, and the electrophysiological effects when directly coupling hMSCs to hCMs at both the single-cell and the tissue level. This computational analysis allowed us to hypothesize an electrophysiology-based approach for improved hMSC-based cardiotherapies, which has not been suggested elsewhere. We first developed three novel hMSC electrophysiology models based on a published empirical triad of hMSC families having distinct ion channel currents: Type A) Idr and IKCa; Type B) Idr, IKCa, INa, and ILCa; and Type C) IKCa and Ito [26, 28]. Subsequently, each hMSC model type was coupled to an adult ventricular hCM electrophysiology model to better understand the direct interactions of these cell types. The computational analysis led us to find that: 1) our model simulations are consistent with a range of empirical findings; 2) hMSC-hCM direct coupling can increase vulnerability to re-entry; and 3) vulnerability to re-entry can be minimized using Type C hMSCs at low levels of stem cell insertion. The ability of our computational models to reproduce empirical electrical hMSC and hMSC-hCM co-culture findings supports the validity of our results. As previously described, fitting individual currents (Fig 1) allowed reconstruction of representative whole cell voltage-clamp data by Li et al. (Fig 3) [26]. This enabled us to simulate hMSC-hCM coupling to develop insight into direct electrical effects of co-culturing these two cell types. The complex hMSC-hCM interactome, which also includes paracrine signaling [6, 11], makes it empirically infeasible to isolate direct electrophysiological coupling effects on APD. For example, Askar et al. have previously shown that the hMSC secretome alone significantly increases neonatal rat cardiomyocyte APD and significantly decreases Cav1.2 and Kv4.3 levels [54], while DeSantiago et al. demonstrated the hMSC paracrine factors stimulate the L-type calcium channel current and calcium transient activity in mouse ventricular myocytes [55]. Furthermore, Askar et al. found the paracrine effects on APD to be dose-dependent [54]. Several studies [13, 16, 54] demonstrate that hMSC co-culture does not lead to APD shortening in vitro, whereas our model studies suggest direct hMSC-hCM coupling alone would tend to shorten APD. Therefore, we hypothesize that in the experimental setting, hMSC-mediated paracrine effects may overshadow the model-predicted APD shortening effects of direct heterocellular coupling. Furthermore, the hMSC secretome reportedly alters atrial myocyte conduction [56], but does not significantly affect the conduction of ventricular myocytes [54], making it reasonable to compare our model results to empirical conduction and VW findings. Studies have shown that sufficient hMSC supplementation decreases CV and CV restitution slopes [16, 54], consistent with our simulations. Specifically, Chang et al. observed co-culturing cardiomyocytes with greater than 10 percent of hMSCs decreased CV and the CV restitution slope [16]. Moreover, sufficient hMSC supplementation increased inducibility of re-entry [16], which was also shown in our simulations (Fig 9). Based on our direct coupling-only simulations reproducing empirical co-culture conduction and VW findings, we hypothesize that in the experimental setting, hMSC-mediated paracrine effects on hCM conduction do not counteract the effects of direct heterocellular coupling demonstrated in this study, emphasizing the importance of understanding and minimizing the potential sources of hMSC-related arrhythmogenicity. Despite their non-excitable nature, hMSCs express gap junction proteins [15] and are therefore capable of influencing hCM action potentials. Furthermore, these effects cannot be presumed to be simply passive, as shown in Figs 5 and 6A. In the case of a passive cell, there is a consistent increase in hCM APD. The relatively large capacitance of hMSCs (approximately 60 pF [26], compared to 6.3 pF for cardiac fibroblasts [22]) makes this effect substantial, resulting in increases in APD of approximately 50 ms with a population of 80% passive hMSCs with midcardial hCMs. These passive effects were not duplicated once the cells expressed their respective ionic currents. Unlike passive hMSCs, Types A and B hMSCs decreased APD independent of hCM cell type. For example, the APDs of midcardial hCMs shortened by approximately 88 ms with a population of 80% Type A or B hMSCs. This effect was exacerbated in cellular fusion, where midcardial hCM APD was shortened by approximately 120 ms. During an hCM action potential, the peak hEAG1 current was two-fold greater than the maximum magnitude of Ito, and nearly twenty-fold greater than the maximum magnitude of IKCa. The larger outward current of Types A and B hMSCs resists hMSCs from approaching the transmembrane voltage of hCMs, resulting in an overall larger sinking effect that shortens phase 2 of the cardiac action potential, and initiates phases 3 and 4 of repolarization earlier. Such drastic changes in the action potential waveform could be possible in vivo if the delivered stem cells cluster in regions of the heart [57], such that hMSCs outnumber hCMs locally. This would be of even greater concern if the high incidence of cell fusion reported in vitro [10] were also found to occur in vivo as suggested by recent animal studies [58]. The implications of action potential variations include pathological electrical and mechanical states. Overall, our simulations suggest hMSC-hCM coupling: 1) alters action potential waveform at the single-cell and tissue level; 2) increases dispersion of APDs at the tissue levels; and 3) substantially decreases CV. Shortening of APDs by Types A and B hMSCs could have notable electrophysiological implications in the heart. Studies have shown that shortening of APDs could induce ventricular tachycardias, suggesting Types A and B hMSCs may be capable of pro-arrhythmic electrical remodeling [17, 59, 60]. Furthermore, one signature of ischemic patients is a loss of epicardial action potential dome, resulting in ST-segment elevation [61]. hMSC direct coupling to hCMs could exaggerate these effects by clustering in the epicardium and acting as an electrical sink, thus becoming pro-arrhythmic. Substantial decreases in APD due to Types A and B hMSCs could also portend altered Ca+2 transients in the hCM, resulting in decreased inotropy [62–67]. Such alterations could directly impact left-ventricular pressure development [22], which is of particular concern for myocardial infarction patients who already suffer decreases in ejection fraction, preload, stroke work, rate of pressure development, and overall mechanical efficiency [68]. The large variability in electrical activity of Types A and B hMSCs presents another potential source of arrhythmogenicity. hCM APD was negatively correlated and highly sensitive to Types A and B hMSC Gjunction (Fig 8A and 8B). This gap conductance has been shown empirically to be highly variable with a coefficient of variation of 87% [15]. The potentially irregular actions of Types A and B hMSCs are further amplified by the fact that hEAG1 activation kinetics are also highly variable, with a coefficient of variation of approximately 35% [26]. Since there is a highly negative correlation between hCM APD and numerous Idr components (e.g., its activating parameters and Gdr), and a highly positive correlation with its inactivating parameters, hMSCs with delayed rectifier-like currents are likely to be unpredictable in their direct effects on hCM APD. This is exacerbated by the fact that hMSC insertion leads to increased APD dispersion in a dose-dependent manner, which could unfavorably alter VWs and electrical stability [45, 49]. Decreased CV caused by hMSC supplementation (Fig 10C) is an established source of re-entrant loops [16], making hMSC-hCM direct coupling potentially arrhythmogenic. Chang et al. showed the potential of re-entrant arrhythmias in vitro was dependent on hMSC supplementation [16], which was confirmed in our simulations (Fig 9B). The decrease in CV is more drastic with increased hMSC supplementation (Fig 10C), which may occur if hMSCs cluster in a localized region, resulting in an increased probability for re-entry. Decreased CV also plays a significant role in ischemic patients. Specifically, ischemic patients also have signatures of transmural conduction slowing, resulting in ST-segment elevation and T-wave inversion [61]. These abnormalities may be exacerbated by the decreased CV effects of hMSC insertion. Current hMSC cardiotherapies involve implementation of electrically-unspecified hMSCs. As a result, Types A and B hMSCs, which reportedly account for a majority of hMSCs [26], will tend to dominate the electrical interactions with hCMs. This was seen in Figs 4 and 5, where the mixed population of hMSCs acted almost indistinguishably from Types A and B hMSCs. This model study suggests that the isolation of Type C hMSCs, absent of delayed rectifier-like currents, may offer superior effectiveness and safety as a cell-based cardiotherapy at low levels of hMSC insertion by minimizing VWs and action potential waveform perturbations compared to other hMSC types. Type C hMSCs exhibited unique electrical activity that was intermediate between the passive and delayed rectifier-functioning hMSCs, resulting in a favorable gap current. The equilibrating source-sink actions within the Type C hMSC gap currents resulted in smaller deviations in the APD (Figs 4 and 5), corresponding to longer action potential wavelengths at the tissue level following hMSC insertion (Table H of S1 Text), which we hypothesize contributed to this cell type having the smallest VW at low levels of hMSC insertion (Fig 9B and 9C). This suggests a decreased likelihood of the potential adverse electrical effects previously described. It is also important to note that overall at the tissue level, the VW increased at greater levels of hMSC-hCM direct coupling, and became independent of hMSC type at moderate and high levels of hMSC insertion. Previous findings suggest the hMSC-hCM interactome involves not only intrinsic, direct cell-cell coupling, but also indirect paracrine signaling through the release of largely unidentified soluble factors and exosome nanovesicles [6, 11]. Harnessing and delivering the key components of the hMSC secretome while circumventing the potentially pro-arrhythmic effects of direct cell-cell coupling may offer a superior cardiac therapy in the future. We note several limitations of the hMSC models developed. As previously discussed, the activation time constant for Idr at a holding potential of -80 mV has a coefficient of variation of approximately 35% [26]. This variability affects the output APD, as suggested by the sensitivity analysis, demonstrating the necessity for further empirical investigation into the kinetics of hMSC Idr. We also assumed that only a triad of families of hMSCs exist, but there may be more; for instance, it has been reported that ion channel expression varies with cell cycle progression [69–71], which may contribute to the variable electrical families and activities of hMSCs. However, the limited understanding of this behavior in the context of hMSCs motivated us to focus only on the three previously characterized hMSC families. We also assume constant ionic concentrations across the hMSC cell membrane. Currently, there is not enough experimental data to sufficiently model intracellular calcium levels in hMSCs. Our sensitivity analysis demonstrates that APD is not highly influenced by channels that are largely affected by these variations (e.g., IKCa), justifying this assumption. Collecting more electrophysiological data on carrier proteins within hMSCs [29] would encourage incorporating transient behavior of ionic concentrations into our models. A second limitation was that we assumed healthy hCMs in order to develop insight into the arrhythmogenic effects of hMSC insertion into healthy cardiac tissue, effectively performing an in silico Phase I clinical trial. However, we did not consider the effects of microfibrosis or random microscale obstacles [24, 72–77]. Each of these effects was purposely not considered in this study, as hMSC paracrine effects are expected to have a major impact on these changes [78–81]. The simulations performed in this study provide a framework for future investigation into each of these factors. Therapeutic hMSCs can disperse to both healthy and ischemic regions of the heart, motivating investigation of the effects of hMSC coupling with ischemic hCMs. This healthy hCM-only assumption made it appropriate to model local cardiac behavior (i.e., 5 cm × 5 cm heterogeneous anisotropic tissue) rather than whole heart behavior. Studying the effects of various spatial distributions of hMSCs using a fully three-dimensional anatomically detailed model of the heart could represent an area for future investigation building on the electrophysiology models developed herein. A fourth limitation was that we did not model other factors that may influence electrical instability, such as short-term cardiac memory and intracellular calcium dynamics [20, 82–84]. Instead, we prioritized other established factors of instability (e.g. APD dispersion, APD restitution slopes, CV restitution slopes), and found several advantages of Type C hMSCs compared to the other mesenchymal stem cell families. Finally, we assumed no interplay between paracrine signaling and electrophysiological coupling. However, it was recently shown that paracrine signaling can cause upregulation of Cx43 and increase intercellular conduction in atrial myocytes [56], as well as alter ion channel activity in ventricular myocytes [54]. We neglected paracrine mechanisms in our models, so investigating this time-dependent interaction would require further study. Based on these limitations, areas for future work include: 1) improving the models based on advancements in empirical data on hMSC electrophysiology; 2) considering the effects of microfibrosis or random microscale obstacles in combination with hMSC anti-fibrotic paracrine effects; 3) examining the electrical and electromechanical effects of hMSC models coupled with ischemic hCM models [85]; 4) modeling the interplay between electrophysiological effects and paracrine signaling in the hMSC-hCM interactome; and 5) empirically confirming our simulations, demonstrating that Type C hMSCs minimize the impact on APD and reduce the VW at low levels of hMSC insertion, offering a potential strategy for improving the safety of cardiac cell therapies. In conclusion, our study provides novel electrophysiological models of hMSCs that reproduce key experimental measurements from patch clamp studies, identifies mechanisms underlying the arrhythmogenic effects of hMSCs coupled to hCMs via gap junctions, underscores the electrical effects associated with hMSC-hCM fusion, and establishes the possibility of isolating a specific sub-population of hMSCs absent of hEAG1 delayed rectifier-like channel activity for minimizing the arrhythmogenic risk of future hMSC-based cell delivery cardiotherapies using low levels of hMSC coupling.