scvi-tools

Probabilistic batch correction for single-cell RNA-seq data uses a PyTorch-based variational autoencoder with amortized inference to learn robust latent representations. This Bayesian framework denoises data and integrates multiple studies while quantifying uncertainty across batches.

Yes, you can integrate multimodal datasets like ATAC-seq and RNA-seq together using probabilistic models designed for cross-domain integration. This framework supports RNA, protein, and ATAC modalities, enabling cross-modality imputation and combined analysis across multiple studies.

To perform semi-supervised cell-type annotation on scRNA-seq counts, prepare raw count data in AnnData format and train a probabilistic model. The framework learns latent representations that enable semi-supervised annotation alongside trajectory and velocity extensions.

The best way to run differential expression analysis after batch correction is to use the latent representations from the trained variational autoencoder. This probabilistic framework provides downstream statistical analyses, including differential expression, directly on the learned representations.

Yes, this single-cell integration framework works with AnnData and Scanpy workflows. It offers a consistent API and seamless interoperability with AnnData for scalable workflows, requiring raw count data to be prepared in AnnData format before model training.

You should use variational autoencoders for single-cell omics data integration when you need to denoise datasets, correct batch effects across multiple studies, or combine modalities like methylation and RNA-seq in a principled Bayesian framework with uncertainty quantification.