TotalVI

TotalVI uses a variational autoencoder (VAE) framework to learn a shared Joint Latent Space that simultaneously captures RNA transcript counts and surface protein levels. Unlike concatenation-based methods, it explicitly models the distinct statistical distributions of each modality—Negative Binomial for RNA and Mixture-of-Gaussians for proteins—within a single generative process. This unified posterior distribution enables the model to denoise both data types while preserving cross-modal correlations.

A defining capability of TotalVI is computationally predicting protein abundance from RNA expression alone. After training on paired CITE-seq data, the learned Cross-Modal Translation mapping in the latent space allows the decoder to generate realistic protein profiles for cells where only transcriptomic data is available. This is critical for integrating legacy scRNA-seq datasets with newer multi-omic atlases, effectively backfilling missing surface marker information without additional experimental cost.

TotalVI incorporates a dedicated latent variable that explicitly captures library size—the total number of sequenced reads per cell. By isolating this technical confounder from the biological latent factors, the model achieves robust normalization without the distortions introduced by heuristic scaling methods. This design ensures that downstream differential expression and clustering analyses reflect genuine biological variation rather than sequencing depth artifacts.

The model can condition its latent representation on categorical batch labels, functioning as a Batch Effect Correction Autoencoder. By learning a batch-invariant latent space, TotalVI harmonizes data collected across different laboratories, time points, or sequencing platforms. This enables the construction of large-scale harmonized cell atlases where technical variation is removed while preserving true biological signals such as disease state or cell type identity.

TotalVI employs amortized inference via neural network encoders, meaning that once trained, the model can infer latent representations for millions of new cells without re-optimization. This contrasts with non-amortized methods like Multi-Omic Factor Analysis (MOFA) that require iterative optimization per dataset. The encoder maps raw counts directly to the approximate posterior in a single forward pass, making TotalVI suitable for atlas-scale integration projects involving millions of single-cell profiles.

TotalVI includes a built-in Bayesian hypothesis testing framework for identifying differentially expressed genes and differentially abundant proteins between cell populations. By leveraging the full posterior distribution rather than point estimates, it quantifies uncertainty in effect sizes. Users can specify contrasts—such as treatment versus control—and obtain Bayes factors that measure the strength of evidence for differential expression, providing statistically rigorous results without external tools.

The shared, lower-dimensional mathematical representation where TotalVI embeds both RNA expression and surface protein abundance. This space captures the covariance between modalities, allowing the model to represent a cell's state holistically. Proximity in this space indicates functional similarity, enabling accurate cell-type clustering and visualization that reflects both transcriptomic and proteomic identity.

The computational task of converting one data modality into another. TotalVI performs protein-to-RNA and RNA-to-protein translation via its shared latent space. For example, the model can computationally predict surface marker abundance from a cell's transcriptome alone. This is invaluable for mapping legacy scRNA-seq datasets onto CITE-seq references without performing new experiments.

A core capability of TotalVI's generative design. The model explicitly accounts for technical confounders like donor variability or sequencing lane using a conditional architecture. It learns a latent representation of biological state that is invariant to batch effects, allowing for seamless integration of CITE-seq data collected across multiple experiments without distorting true biological signals.

TotalVI belongs to the broader class of computational methods designed to align and jointly analyze multiple data types measured from the same individual cells. Unlike horizontal integration (mapping scRNA-seq to scATAC-seq across different cells), TotalVI performs vertical integration on simultaneously measured modalities. This provides a direct, paired view of the central dogma at single-cell resolution.

A regularization technique used during TotalVI's training where one data modality (e.g., all protein data) is randomly zeroed out for a subset of cells. This forces the model to learn robust, cross-modal predictive distributions and prevents it from over-relying on a single data type. The result is a model that gracefully handles inference scenarios where one modality is partially or completely missing.