scVI

scVI is built on a variational autoencoder (VAE) framework that compresses high-dimensional gene expression data into a low-dimensional latent space. The encoder network maps raw count data to a probabilistic distribution, while the decoder reconstructs gene expression from latent variables. This generative formulation explicitly models the negative binomial distribution of single-cell RNA-seq data, naturally accounting for overdispersion and technical noise without heuristic preprocessing steps.

scVI explicitly models batch effects as a learnable parameter in the decoder, allowing the latent space to capture genuine biological variation while technical artifacts are absorbed by batch-specific variables. Unlike post-hoc correction methods such as Harmony or ComBat, scVI performs correction during representation learning itself. This enables seamless integration of datasets from different laboratories, sequencing platforms, or experimental protocols without requiring a separate data integration step.

Single-cell RNA-seq data exhibits an excess of zero counts due to both biological absence of expression and technical dropout events. scVI models this through a zero-inflated negative binomial (ZINB) distribution, which includes a separate parameter for dropout probability. This explicit treatment of zero-inflation eliminates the need for arbitrary imputation methods and preserves the uncertainty inherent in sparse single-cell measurements, leading to more statistically principled downstream analyses.

scVI leverages stochastic variational inference with minibatch training, enabling it to scale to datasets containing millions of cells. The model processes random subsets of cells during each training iteration rather than requiring the full dataset in memory. This design choice makes scVI suitable for large-scale cell atlas projects such as the Human Cell Atlas, where traditional dimensionality reduction methods like t-SNE or memory-intensive algorithms become computationally prohibitive.

Rather than producing a single point estimate for each cell's embedding, scVI outputs a mean and variance parameterizing a Gaussian distribution in latent space. This probabilistic formulation captures uncertainty in the representation—cells with low sequencing depth or ambiguous transcriptional profiles receive higher variance estimates. Downstream tasks such as differential expression testing and clustering can incorporate this uncertainty, producing more robust and reproducible biological conclusions.

scVI learns a cell-specific scaling factor that accounts for differences in total transcript capture across cells. This latent variable replaces manual normalization steps such as counts per million (CPM) or scran pooling, which can introduce biases. By jointly learning library size factors alongside the latent representation, scVI ensures that variation in sequencing depth does not contaminate the biological signal captured in the embedding space.

A computational process that removes technical variation introduced by different experimental batches, allowing genuine biological signals to be compared across separately processed single-cell datasets. scVI's latent variable model explicitly models batch effects as a learnable parameter, enabling harmonization without data distortion. Traditional methods like ComBat use linear adjustments, while scVI's deep generative approach captures non-linear batch effects.

The computational alignment of multiple single-cell datasets from different conditions, technologies, or donors into a shared latent space. scVI achieves this by learning a shared latent representation that captures biological variation while conditioning on batch-specific effects. This enables joint analysis of heterogeneous data without removing meaningful biological signals. Key competitors include Harmony and Seurat's CCA-based integration.

A probability distribution that models single-cell RNA-seq count data by combining a negative binomial component for expression counts with a zero-inflation component for technical dropout events. scVI uses the ZINB distribution as its generative output layer, allowing it to accurately capture both biological expression and technical zeros. This is critical because scRNA-seq data exhibits excess zeros beyond what standard count distributions predict.

A supervised machine learning approach that projects cell-type annotations from a well-characterized reference atlas onto a new query dataset. scVI's latent space can be used for label transfer by training a classifier on the reference embeddings and applying it to query cells. The probabilistic nature of scVI provides uncertainty estimates for each annotation, flagging ambiguous assignments that require manual review.

A statistical framework for identifying genes that are significantly up- or down-regulated between cell populations or conditions. scVI extends this by enabling Bayesian differential expression in its latent space, accounting for uncertainty in the generative process. Unlike pseudobulk methods, scVI's approach preserves single-cell resolution and can detect subtle expression shifts in rare populations.

A semi-supervised extension of scVI that incorporates partial cell-type labels during training. scANVI (single-cell ANnotation using Variational Inference) leverages both labeled and unlabeled cells to learn a more structured latent space that respects known biological hierarchies. This improves clustering accuracy and enables robust label transfer, especially when reference annotations are incomplete or noisy.