scVI

Summary

scVI (single-cell Variational Inference) is a deep generative model for single-cell transcriptomics that uses a variational autoencoder and a zero-inflated negative binomial distribution to model biological and technical variation in raw counts. It learns a low-dimensional representation of cell states for batch integration, clustering, and differential expression analysis. In pretraining evaluations as of 2026-06-09, scVI saturated in performance at 10% of the training corpus and outperformed more complex transformer architectures on several benchmarks.

Method

scVI uses a probabilistic neural network to model transcriptomic counts: - ZINB Likelihood: Takes raw counts directly and models the likelihood using a zero-inflated negative binomial distribution to account for dropout and overdispersion. - Latent Space: Projects high-dimensional cell expression profiles into a 10-dimensional latent embedding space. - Batch Correction: Integrates technical covariates directly into the decoder network, enabling joint representation learning and batch effect correction.

Benchmarking Results

In a systematic study of pretraining scaling for Single-Cell Foundation Models, scVI's latent embeddings were tested on cell-type classification and batch integration. It achieved its Learning Saturation Point at 1% to 10% of the training data (~200,000 to 2,000,000 cells) and consistently outperformed transformer-based methods like Geneformer on absolute classification accuracy and batch integration.

Citations

• Lopez, R. et al. Deep generative modeling for single-cell transcriptomics. Nat. Methods 15, 1053–1058 (2018). DOI: 10.1038/s41587-018-0229-2
• Key Evaluation: DenAdel, A. et al. (2026). Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance. Nature Methods, 23, 1447–1457. DOI: 10.1038/s41592-026-03120-y. Source paper: s41592-026-03120-y.pdf