scvi-tools

Batch correction for scRNA-seq datasets is achieved by training probabilistic models like scVI to learn a shared latent space, effectively removing technical batch effects while preserving biological variance across your samples.

Multi-omics data integration is fully supported using specialized models like totalVI, PeakVI, and MultiVI, allowing you to jointly analyze scRNA-seq, ATAC, and CITE-seq datasets within a unified latent representation.

Single-cell batch correction requires preparing raw count data formatted as an AnnData or MuData object, performing highly variable gene selection, and configuring the setup_anndata call to register batch keys before model training.

Reference mapping is supported through the scANVI workflow, enabling you to map labels from a trained reference dataset onto new query single-cell datasets by leveraging the established shared latent space.

scVI performs unsupervised batch correction and data integration to generate a latent space, whereas scANVI extends this by incorporating semi-supervised cell type labeling using that same latent representation.

Highly variable gene selection is required before model training to reduce noise, lower computational overhead, and ensure the probabilistic model focuses on meaningful biological signals rather than technical background.