Detection of differentially abundant cell subpopulations discriminates biological states in scRNA-seq data

Abstract Recently, single cell RNA sequencing (scRNA-Seq) has been applied to compare transcriptomic landscapes between biological conditions or states. This comparison allows the discovery of cell subpopulations that are differentially abundant between biological conditions or states. Typically, these subpopulations are determined by identification of clusters or cell types in which the proportion of cells from the two states deviates significantly from the corresponding proportion in the entire cell populations. This approach might be suboptimal as the most differentially abundant subpopulations may not perfectly overlap with these clusters or cell types. Here, we develop DA-seq, a multiscale algorithm that detects cell subpopulations with the highest disproportionate cell abundances between two samples. DA-seq does not require prior partitioning of the data based on clustering or cell subtypes. To evaluate DA-seq, we applied it to one simulated and two real scRNA-Seq datasets. The first scRNA-Seq dataset was obtained from a study on melanoma responders and non-responders to checkpoint immunotherapy. The second datasets is from a study that compares two time points of developing embryonic mouse skin.