DCATS: Differential composition analysis of single-cell data

Abstract

Single cell profiling technology such as single-cell RNA-seq (scRNA-seq) enables high throughput discovery and characterisation of diverse cell types or cell states in a population of cells. This ability has given rise to new statistical problems in robust quantification and comparison of cell-type proportion.. It remains challenging to effectively detect differential compositions of cell types when comparing samples coming from different conditions or along with continuous covariates, partly due to the small number of replicates and high uncertainty of cell clustering. Here, we introduce a new statistical model, DCATS, for differential composition analysis in single cells in a framework of beta-binomial regression. It leverages a confusion matrix to correct the bias of clustering and allows pre-estimated parameters across all cell types to account for its uncertainty. It also allows us to regress out the influence of confounding covariates except for the condition factor. Through multiple simulated and experimental data sets, we demonstrate the high effectiveness of DCATS in identifying variable cell types in various experiment designs. Combining the differential genes analysis, cell-cell interaction analysis, and other scRNA-seq analysis, DCATS deepens our understanding of cell types differential composition and gain biological insight.