The problem. Single-cell counts are sparse, noisy, and shot through with technical variation — dropout, library-size differences, batch effects. The usual pipeline patches each with a separate ad-hoc step. What you want is one model that treats the noise as noise and hands you a clean, low-dimensional representation.

The idea. scVI is a variational autoencoder for scRNA-seq. It models raw counts with a zero-inflated negative binomial likelihood, learns a latent embedding of each cell, and folds batch as a covariate so the latent space is corrected by construction. From that one fit you get denoised expression, batch-integrated coordinates for clustering, and a generative model you can sample from — all with calibrated uncertainty rather than point estimates.

Why it matters. This is the hierarchical/empirical-Bayes instinct again — the same thread running through DESeq2 and cell2location — now in deep-learning form, and it became the backbone of the whole scvi-tools ecosystem. For me it’s a satisfying convergence point: the probabilistic modeling I trust from bulk RNA-seq, scaled to single cell, and directly relevant to the spatial deconvolution methods that borrow its machinery.

Verdict. A genuine foundation, and the modeling is principled rather than a black box. Costs: it needs enough cells and compute, and the latent space is only as trustworthy as the model’s assumptions about your data. Read it back-to-back with scANVI, which adds labels to exactly this framework.