zinbwave

zinbwave models zero-inflated negative binomial (ZINB-WaVE) distributions to produce robust low-dimensional representations of single-cell RNA-seq (scRNA-seq) count data while accounting for dropouts, over-dispersion, library size differences, and covariates.

Key Features:

• Zero-Inflated Negative Binomial Model: Implements the ZINB-WaVE zero-inflated negative binomial model to explicitly model zero inflation (dropouts) in scRNA-seq data.
• Over-Dispersion Handling: Models over-dispersion in gene expression counts beyond Poisson variability.
• Count Data Modeling: Operates directly on discrete scRNA-seq count data, respecting its count-based nature.
• Library Size Differences: Accounts for variations in library sizes across cells without requiring preliminary normalization.
• Batch Effects and Covariates: Optionally adjusts for batch effects and other sample- or cell-level covariates.
• Low-Dimensional Representations: Estimates robust low-dimensional latent factors suitable for downstream analysis.

Scientific Applications:

• Dimensionality Reduction for scRNA-seq: Produces low-dimensional representations used for visualization, clustering, and downstream analyses of single-cell transcriptomes.
• Comparative Performance: Demonstrated to outperform principal component analysis (PCA) and zero-inflated factor analysis (ZIFA) in stability and accuracy on simulated and real datasets.
• Data Integration Across Experiments: Enables integration of multiple experiments or conditions by adjusting for batch effects and covariates.
• Biological Studies: Applied in studies of cellular heterogeneity, developmental biology, and disease mechanisms at the single-cell level.

Methodology:

Applies the ZINB-WaVE zero-inflated negative binomial model to scRNA-seq count data to estimate low-dimensional latent factors while modeling zero inflation, over-dispersion, library size, and optional covariates without requiring preliminary normalization.