Total Variational Inference

Total Variational Inference models joint single-cell RNA expression and surface protein abundance from CITE-seq data to disentangle biological signals from technical artifacts for downstream single-cell analyses.

Key Features:

• End-to-end joint probabilistic modeling: Models RNA expression and surface protein abundance jointly using a probabilistic framework that accounts for protein background noise and batch effects.
• Dimensionality reduction: Produces low-dimensional latent representations that capture cellular heterogeneity from high-dimensional CITE-seq measurements.
• Data integration: Integrates datasets with varying measured proteins to enable comparison and combination across experiments and conditions.
• Correlation estimation: Estimates correlations between RNA transcripts and surface proteins to reveal potential regulatory or functional linkages.
• Differential expression testing: Supports differential expression analysis for genes and proteins across cell states or experimental conditions.

Scientific Applications:

• Immunology profiling: Jointly profiles gene expression and surface protein abundance to elucidate immune cell functions, demonstrated on murine spleen and lymph nodes.

Methodology:

Variational inference is used to probabilistically model observed CITE-seq RNA and protein counts, decomposing data into latent components that separate biological variation from technical artifacts and producing latent representations for downstream clustering, trajectory inference, and biomarker discovery.