scvi-tools

To perform batch correction on single-cell RNA-seq data, this tool learns probabilistic generative models to produce latent, batch-corrected embeddings. You need to provide AnnData inputs with raw count layers, configure the model setup, and run variational inference training to get the integrated low-dimensional representation.

Yes, you can perform Bayesian differential expression for scRNA-seq with FDR control. This approach accounts for dropout and zero inflation while respecting batch structure, providing statistically principled differential expression results without relying on ad-hoc normalization or pseudocount heuristics.

The best way to integrate multimodal RNA and ATAC data is using modality-specific probabilistic models like MultiVI. By training variational inference models on AnnData objects with raw counts, you achieve a unified latent representation that captures both gene expression and chromatin accessibility.

Yes, probabilistic modeling works for spatial transcriptomics deconvolution using models like DestVI and Stereoscope. These approaches learn latent representations from raw count data to estimate cell type proportions and abundance within spatial spots.

Yes, you need raw counts in your AnnData object to use these probabilistic single-cell models. The models require modality-appropriate likelihood handling and use the setup_anndata function to configure raw count layers before initiating variational inference training.

You should not use deep generative models for single-cell integration if your dataset lacks raw count layers or if your environment cannot support variational inference training. Additionally, if your analysis does not require uncertainty-aware downstream tasks, simpler normalization methods may be more appropriate.