scVI (Single-cell Variational Inference)

The core deep learning architecture powering scVI. A VAE consists of an encoder network that compresses high-dimensional gene expression data into a probabilistic, low-dimensional latent space (the variational posterior), and a decoder network that reconstructs the original data from this latent representation. Unlike a standard autoencoder, the VAE enforces a prior distribution (typically a standard Gaussian) on the latent space, enabling generative sampling and robust imputation. scVI leverages this framework to separate biological signal from technical noise.

The statistical likelihood function used by scVI to model single-cell RNA-seq count data. This distribution accounts for two key characteristics of scRNA-seq:

• Negative Binomial component: Models the overdispersed nature of count data, where the variance exceeds the mean due to biological and technical variability.
• Zero-inflation component: Explicitly models the excess of zero counts caused by dropout events, where a gene is expressed but not captured during sequencing. By using the ZINB, scVI provides a more accurate probabilistic model than simpler Gaussian assumptions.

A defining feature of scVI is its explicit treatment of batch identity as a latent variable within the generative model. Instead of applying a pre-processing correction step, scVI learns a batch-specific shift in the latent space during training. This allows the model to:

• Disentangle biological variation from technical artifacts.
• Generate batch-corrected gene expression counts.
• Perform probabilistic normalization without discarding uncertainty. This approach is superior to hard-correcting the data matrix, as it preserves the full posterior distribution for downstream analysis.

The objective function maximized during scVI training. The ELBO is a tractable lower bound on the model's marginal log-likelihood and balances two competing terms:

• Reconstruction Loss: Penalizes the model for failing to accurately reconstruct the input gene expression counts from the latent representation.
• Kullback-Leibler (KL) Divergence: Acts as a regularizer, penalizing the learned latent distribution for deviating from the chosen prior (a standard Gaussian). Optimizing the ELBO ensures scVI learns a smooth, continuous latent space that captures meaningful biological structure while enabling generative capabilities.

The open-source Python framework that implements scVI and its model variants. Built on PyTorch and Pyro, it provides a standardized API for:

• scVI: The foundational model for scRNA-seq normalization and analysis.
• scANVI: A semi-supervised extension for cell-type annotation that harmonizes labeled and unlabeled datasets.
• totalVI: A model for joint analysis of paired gene expression and protein abundance data from CITE-seq.
• Solo: A doublet detection model built on the scVI latent space. The ecosystem enables seamless integration into standard single-cell analysis pipelines.

scVI enables principled Bayesian differential expression analysis directly from its generative model. Instead of using a point estimate of normalized counts, scVI computes the posterior distribution of expression for each gene in each cell. Differential expression is then assessed by sampling from these posteriors to estimate the log-fold change between conditions, providing a measure of uncertainty. This approach avoids the pitfalls of pseudobulk methods on sparse data and naturally accounts for batch effects, yielding more robust and calibrated results.