An evaluation of statistical differential analysis methods in single-cell RNA-seq data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables gene expression measurement at the individual cell level, revealing cell-to-cell variation.
• Differential gene expression analysis is a primary application of scRNA-seq, with numerous analytical methods developed.
• Evaluating these methods is crucial for accurate biological interpretation of scRNA-seq data.

Approach:

• Comparative performance evaluation of five popular open-source scRNA-seq differential expression analysis tools: DEsingle, Linnorm, monocle, MAST, and DESeq2.
• Assessment metrics included false discovery rate (FDR) control, sensitivity, specificity, accuracy, and area under the receiver operating characteristics (AUROC) curve.
• Analyses were conducted under varying simulation conditions, including sample sizes, data distribution assumptions, and proportions of zero counts.

Key Points:

• MAST demonstrated superior performance with the highest AUROC values across tested sample sizes and proportions of differentially expressed genes when data followed negative binomial distributions.
• With sample sizes of 100 per group, MAST consistently yielded the highest AUROC, irrespective of data distribution.
• Pre-filtering excess zeros improved the relative performance of DEsingle, Linnorm, and DESeq2, leading to higher AUROC values compared to MAST and monocle.

Conclusions:

• MAST is a highly effective method for differential gene expression analysis in scRNA-seq, particularly for negative binomial data.
• The choice of method can be influenced by data characteristics, such as the presence of excess zeros, and sample size.
• Data preprocessing steps, like zero-filtering, can significantly impact the performance of different differential expression analysis tools in scRNA-seq.