corral

corral implements correspondence analysis (CA) for dimensionality reduction, batch integration, and visualization of single-cell RNA-seq data, providing a count-based alternative to principal component analysis (PCA) that avoids distortive log-transformation.

Key Features:

• Single-table CA: Performs dimensionality reduction on a single matrix of single-cell data using decomposition of a chi-squared residual matrix.
• Multi-table CA (corralm): Leverages data-optimized scaling to align data generated from different sequencing platforms by projecting into a shared latent space.
• Overdispersion adaptations: Includes five adaptations of CA to address overdispersion and high sparsity in scRNAseq data, including Freeman-Tukey chi-squared residuals.
• Continuous data adaptation: Applies CA-style processing to continuous data such as proteomic TOF intensities using the Hellinger distance adaptation of CA.
• Computational efficiency: Utilizes sparse matrices and a fast implementation of singular value decomposition (SVD).
• Bioconductor integration: Can be called directly on Bioconductor objects such as SingleCellExperiment for pipeline integration.
• CA biplot visualization: Enables visualization of associations between genes and cell populations.

Scientific Applications:

• scRNA-seq dimension reduction: Computes cell embeddings with improved or comparable clustering accuracy relative to standard CA and glmPCA across diverse datasets.
• Cross-platform data integration: Aligns single-cell data from different sequencing platforms through multi-table correspondence analysis.
• Batch effect correction: Integrates scRNAseq data from multiple tables while preserving biological variation.
• Gene-cell association visualization: Generates CA biplots to explore relationships between genes and cell populations.

Methodology:

corral decomposes a chi-squared residual matrix via sparse matrix operations and fast SVD. Five adaptations address overdispersion in count data, including Freeman-Tukey residuals. Multi-table analysis (corralm) applies data-optimized scaling for cross-platform alignment into a shared latent space.