Motivation

Single-cell RNA-sequencing (scRNA-seq) provides more granular biological information than bulk RNA-sequencing; bulk RNA sequencing remains popular due to lower costs which allows processing more biological replicates and design more powerful studies. As scRNA-seq costs have decreased, collecting data from more than one biological replicate has become more feasible, but careful modeling of different layers of biological variation remains challenging for many users. Here, we propose a statistical model for scRNA-seq gene counts, describe a simple method for estimating model parameters and show that failing to account for additional biological variation in scRNA-seq studies can inflate false discovery rates (FDRs) of statistical tests.

Results

First, in a simulation study, we show that when the gene expression distribution of a population of cells varies between subjects, a naïve approach to differential expression analysis will inflate the FDR. We then compare multiple differential expression testing methods on scRNA-seq datasets from human samples and from animal models. These analyses suggest that a naïve approach to differential expression testing could lead to many false discoveries; in contrast, an approach based on pseudobulk counts has better FDR control.

Availability and implementation

A software package, aggregateBioVar, is freely available on Bioconductor (https://www.bioconductor.org/packages/release/bioc/html/aggregateBioVar.html) to accommodate compatibility with upstream and downstream methods in scRNA-seq data analysis pipelines.

Supplementary information

Raw gene-by-cell count matrices for pig scRNA-seq data are available as GEO accession GSE150211. Supplementary data are available at Bioinformatics online.

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.