brickmanlab/human-scanvi

Description

Human preimplantation development model spanning early stages of development. The model was trained utilizing single‐cell ANnotation using Variational Inference (scANVI, Xu et al., 2021) implemented in scvi-tools. In short, scANVI raw single-cell RNA sequencing (scRNA-seq) count matrix - cell by gene, where values represent gene expression measured by counting number of transcribed RNA.

Model Training

• raw dataset
• notebook analysis

Metrics

Cell type (ct) prediction

Metric	Score
Accuracy score	0.7968144640551011
Balanced accuracy	0.8502734650790613
F1 (micro)	0.7968144640551011
F1 (macro)	0.8150578255414443

Model parameters

Below we provide settings for scANVI setup

lvae.init_params_["non_kwargs"]

{
    "n_hidden": 128, 
    "n_latent": 10, 
    "n_layers": 2, 
    "dropout_rate": 0.1, 
    "dispersion": "gene", 
    "gene_likelihood": "nb", 
    "linear_classifier": false
}

lvae.adata_manager.registry['setup_args']

{
    "labels_key": "ct",
    "unlabeled_category": "Unknown",
    "layer": "counts",
    "batch_key": "batch",
    "size_factor_key": null, 
    "categorical_covariate_keys": null, 
    "continuous_covariate_keys": null
}

References

Proks, M., Salehin, N. & Brickman, J.M. Deep learning-based models for preimplantation mouse and human embryos based on single-cell RNA sequencing. Nat Methods 22, 207–216 (2025). https://doi.org/10.1038/s41592-024-02511-3