An appropriate differential expression analysis in single-cell data depends on the study design

Abstract

Large-scale multi-subject single-cell data have become very common. However, these data have high heterogeneity in study designs, leading to confusion in differential gene expression (DGE) analysis. In this work, we show that pseudobulk produces a substantial amount of type 2 error when the group label being compared varies within a subject. When the comparison label is constant within a subject, type 1 error is inflated if pseudoreplication is not accounted for. As a general principle, we show that an appropriate method depends on the design of the study. We propose solutions to the inflated error rates and provide practical guidelines for researchers.