Bias, robustness and scalability in differential expression analysis of single-cell RNA-seq data

Abstract

AbstractBackgroundAs single-cell RNA-seq (scRNA-seq) is becoming increasingly common, the amount of publicly available data grows rapidly, generating a useful resource for computational method development and extension of published results. Although processed data matrices are typically made available in public repositories, the procedure to obtain these varies widely between data sets, which may complicate reuse and cross-data set comparison. Moreover, while many statistical methods for performing differential expression analysis of scRNA-seq data are becoming available, their relative merits and the performance compared to methods developed for bulk RNA-seq data are not sufficiently well understood.ResultsWe present conquer, a collection of consistently processed, analysis-ready public single-cell RNA-seq data sets. Each data set has count and transcripts per million (TPM) estimates for genes and transcripts, as well as quality control and exploratory analysis reports. We use a subset of the data sets available in conquer to perform an extensive evaluation of the performance and characteristics of statistical methods for differential gene expression analysis, evaluating a total of 30 statistical approaches on both experimental and simulated scRNA-seq data.ConclusionsConsiderable differences are found between the methods in terms of the number and characteristics of the genes that are called differentially expressed. Pre-filtering of lowly expressed genes can have important effects on the results, particularly for some of the methods originally developed for analysis of bulk RNA-seq data. Generally, however, methods developed for bulk RNA-seq analysis do not perform notably worse than those developed specifically for scRNA-seq.