scvi-tools

Batch correction on scRNA-seq data uses probabilistic variational models to learn latent representations that reduce technical batch effects while preserving biological signal. You register raw counts and batch covariates using setup_anndata, then train the model to obtain a harmonized joint embedding.

Multimodal integration supports RNA, ATAC, and protein modalities using specialized models like totalVI, PeakVI, and MultiVI. These models learn joint probabilistic latent spaces from multi-omics inputs, enabling robust cross-modality integration and downstream analysis.

Uncertainty-aware differential expression uses Bayesian testing to quantify uncertainty while explicitly handling zero-inflation in raw count data. This probabilistic approach provides confidence intervals for expression differences across biological groups rather than deterministic p-values alone.

Probabilistic single-cell modeling requires AnnData objects containing raw count layers as input. You must correctly register batch covariates, sample metadata, and modality-specific keys using the setup_anndata function before training any models.

Query-to-reference transfer uses semi-supervised models like scANVI to annotate new single-cell datasets. By leveraging a pre-trained probabilistic reference model, you can predict cell types for unannotated query data while accounting for batch effects.

Spatial transcriptomics deconvolution is supported via specialized models like DestVI. It uses probabilistic modeling to estimate cell type abundances and spatially resolved gene expression profiles from mixed spatial spots, providing uncertainty estimates for the deconvolved signals.