scvi-tools

Batch correction for scRNA-seq data uses a variational autoencoder (scVI) to learn a shared latent space that separates biological signal from technical noise across batches. Register your AnnData with batch covariates and train the model to obtain a corrected latent representation.

Probabilistic differential expression in single-cell analysis uses Bayesian testing to quantify uncertainty in gene expression differences between cell populations. It yields posterior distributions rather than static values, allowing you to assess the statistical confidence of observed expression changes.

Multimodal integration of scRNA-seq and scATAC-seq datasets is supported for cross-modality analysis. Models like TOTALVI and MultiVI map these modalities into a shared latent space, enabling joint analysis and imputation across modalities even when profiles are incomplete.

scvi-tools integrates seamlessly with Scanpy for downstream analysis and supports GPU-accelerated training to handle large-scale single-cell omics datasets efficiently. GPU acceleration speeds up training but is not strictly required to run the variational autoencoder models.

Cell type annotation transfer to unlabelled single-cell data uses a semi-supervised approach (SCANVI) built on the scVI latent space. It leverages known labels in a subset of cells to predict and annotate cell types in the remaining unlabelled population.

Probabilistic modeling is preferred over deterministic methods when you need to quantify uncertainty in downstream analyses like differential expression or batch correction. It separates biological variation from technical noise, providing confidence estimates that deterministic methods cannot capture.