Evaluation of statistical methods for differential expression analysis of RNA-seq data with paired data design

Background: Recently, next generation sequencing has emerged as a powerful tool for gene expression profiling. Previous comparisons of statistical methods for RNA-seq data analysis were focused on data with two independent conditions. However, more complex experimental designs such as paired-data design are very common in RNA-seq experiments. Timely evaluations of innovative statistical methods for RNA-seq data with paired design are important and necessary in practical applications.

Methods: We assess the performance of six statistical methods including edgeR, DEseq, SAMseq, pairedBayes, baySeq, and TSPM for identifying differentially expressed (DE) genes from RNA-seq data with paired design. Via simulation studies with different sample sizes, competing methods are evaluated based on their classification accuracy via mean receiver operating characteristic (ROC) curves, predictive power via plot between positive predictive value (PPV) and 1- negative predictive value (1-NPV), and ability of controlling false discovery rate (FDR).

Results: Based on the ROC and predictive power curves, edgeR and DEseq perform well under all settings when compared to other methods. Additionally, Bayseq works well when sample size is small (n=3), and TSPM and SAMseq perform well under other settings with larger sample size. baySeq well controls the FDR under all settings while pairedBayes selects largest number of DE genes and has the worst FDR control. DESeq, SAMseq, and TSPM methods controls FDR better for studies with a large sample size (n=10).

Conclusions: Our results are useful in guiding the choice of statistical methods in detecting DE genes of RNA-seq data with a paired design.