scGSECA: identification of altered biological processes in single-cell RNA-sequencing data by discretization of expression profiles

Gene set analysis (GSA) aims at identifying gene sets whose cumulative expression is altered in the phenotype of interest. However, heterogeneous expression changes of single genes across samples can result in small pathway expression differences that are difficult to detect with state-of-art GSA algorithms. We recently showed that exploiting the bimodal behavior of RNA-sequencing gene expression profiles through data discretization allows for the identification of the altered biological processes, especially in presence of heterogeneity. Recently, this issue has begun to be exploited with single cell technologies. However, single-cell RNA-sequencing data are marked by technical noise and sparsity. To overcome these issues, we applied our “less is more” paradigm to single cell transcriptomes for the detection of altered gene sets. Here we present single-cell GSECA (scGSECA), a single cell extension of our Gene Set Enrichment Class Analysis. Applied to ~22.000 cells from 12 published studies from normal and tumor tissues, scGSECA highlighted the previously proposed bimodal behavior of single cell expression profiles. scGSECA exploits a cell-centric rather than gene-centric GSA. Non-zero expression levels are modeled as a superposition of gaussian distributions and discretized before statistical testing. We applied our method to synthetic data with controlled alterations and variable cohort sizes. We show that scGSECA provides higher performances than other available algorithms in detecting truly altered biological processes in large cohorts.