Correspondence analysis for dimension reduction, batch integration, and visualization of single-cell RNA-seq data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Principal Component Analysis (PCA) is standard for single-cell RNA sequencing (scRNAseq) dimension reduction.
• PCA requires continuous, normally-distributed data, often necessitating log-transformation which can distort scRNAseq data.
• Log-transformation can obscure meaningful biological variation in scRNAseq datasets.

Purpose of the Study:

• Introduce Correspondence Analysis (CA) as a count-based alternative to PCA for scRNAseq data.
• Propose and evaluate five adaptations of CA to handle scRNAseq data's overdispersion and sparsity.
• Develop an R/Bioconductor package, corral, for implementing CA in scRNAseq analysis.

Main Methods:

• Correspondence Analysis (CA) based on chi-squared residual matrix decomposition, avoiding log-transformation.
• Five novel adaptations of CA to address overdispersion and sparsity in scRNAseq data.
• Comparison with standard CA and generalized linear model PCA (glmPCA) on nine scRNAseq datasets.

Main Results:

• Adapted CA methods are fast, scalable, and outperform standard CA and glmPCA in clustering accuracy on 8 out of 9 datasets.
• CA with Freeman-Tukey residuals demonstrated particularly strong performance across diverse scRNAseq datasets.
• CA enables visualization of gene-cell population associations via CA biplots and supports multi-table analysis with corralm.

Conclusions:

• Correspondence Analysis (CA) provides an effective, count-based dimension reduction technique for scRNAseq data.
• Adapted CA methods enhance clustering accuracy and handle data sparsity and overdispersion better than PCA.
• The corral R package simplifies the adoption of CA for improved scRNAseq data analysis and integration.