Assessing Statistical Methods for Clustering and Differential Gene Expression in scRNA-seq Data

Abstract

Single-cell RNA sequencing (scRNA-seq) is a field of transcriptomics that can provide information about cell heterogeneity, defined as the differences between cell states or types caused by changes during cell growth1. Differential gene expression (DGE) is crucial in evaluating the genomic changes that drive heterogeneity. As a result, improving confidence in estimates of highly  expressed genes becomes of high importance in scRNA-seq data analysis.

A common pitfall in determining these genes is the inflation of the type I error rate, where p-values are artificially low due to 'double-dipping', occurring when the same data are used for clustering and DGE. In the clustering stage, two methods - Countsplit and SVMs - are tested using a simulated scRNA-seq dataset with known parameters. Two DGE tests - MAST and Nebula - offer an alternative post-clustering approach to controlling type I error inflation.