scvi-tools

Correct batch effects in scRNA-seq data by training a probabilistic deep generative model on raw count inputs. This approach learns a shared latent space that preserves biological signal while removing technical variation across batches.

Yes, variational inference supports multimodal integration for CITE-seq and ATAC-seq datasets. The model processes multimodal inputs to learn joint latent representations that enable cross-modality alignment and comprehensive cellular characterization.

Perform differential expression testing with uncertainty quantification by applying Bayesian testing to the trained latent space. This method accounts for model uncertainty and provides false discovery control, yielding more robust gene comparisons than ad hoc heuristics.

Yes, AnnData or MuData objects are required to train single-cell deep generative models. Inputs must contain raw count preprocessing and undergo model-specific setup before initiating training and extracting latent embeddings.

Yes, this approach supports spatial transcriptomics and methylation workflows for deconvolution. The probabilistic framework extends to these modalities, allowing you to estimate cell type proportions and perform annotation transfer across diverse genomic data.

Choose probabilistic deep learning over ad hoc heuristics to quantify uncertainty during single-cell integration. This method learns shared latent spaces directly from raw counts, providing mathematically grounded batch correction and uncertainty-aware downstream testing.