scvi-tools

scvi-tools provides deep probabilistic models for single-cell omics — principally for integrating datasets across batches, donors and technologies while modelling the noise structure of the data rather than assuming it away. Where Scanpy supplies the standard workflow, scvi-tools supplies the statistically serious end of it.

What it does

The library implements variational autoencoder models for integration, label transfer, differential expression and imputation, across RNA, protein, ATAC and spatial modalities. Because it models count noise and batch structure explicitly, integration tends to hold up better than heuristic correction on hard cases — many donors, several technologies, uneven composition.

Strengths

• Principled probabilistic modelling instead of ad-hoc batch correction.
• Strong performance on difficult multi-batch integration.
• Covers RNA, protein, ATAC and spatial modalities.
• Built on AnnData, so it drops into an existing Scanpy workflow.

Limitations

Training is GPU-work: slower and heavier than the standard Scanpy pipeline, and the results are stochastic, so seeds must be recorded for reproducibility. It demands more statistical understanding than clicking through a tutorial, and a model that converges is not automatically a model that is right for your design.