scvi-tools

scvi-tools is a deep-learning skill for single-cell genomics analysis built on the scvi-tools framework. It helps researchers apply probabilistic, VAE-based models to common single-cell tasks—batch correction and integration, multi-modal analysis, deconvolution, label transfer, and RNA velocity—without having to assemble the modeling code from scratch. It solves the problem of choosing the right model for a given assay and running a correct, reproducible workflow around it.

The skill provides a model-selection guide mapping data types to models (scVI and scANVI for RNA integration and semi-supervised label transfer, PeakVI for scATAC-seq, totalVI for CITE-seq, MultiVI for multiome, DestVI for spatial deconvolution, veloVI for RNA velocity, sysVI for cross-technology batch correction). It ships numerous reference guides and modular Python CLI scripts—prepare_data, train_model, cluster_embed, differential_expression, transfer_labels, integrate_datasets, and validate_adata—that operate on AnnData (.h5ad) files and can be chained into pipelines. Reference material covers QC, binarization, feature selection, training with early stopping, latent-representation extraction, clustering/UMAP/Leiden, and differential accessibility/expression, plus environment and GPU setup and troubleshooting.

Its target users are computational biologists and single-cell researchers doing integration, multi-omics, or reference-mapping analyses, especially those wanting standardized scripts over ad-hoc notebooks. The operations are standard scientific computation on local data files (with optional GPU); there is no networked or destructive behavior beyond reading and writing analysis artifacts.