BackgroundLong non-coding RNAs (lncRNAs) are typically expressed at low levels and are inherently highly variable. This is a fundamental challenge for differential expression (DE) analysis. In this study, the performance of 25 pipelines for testing DE in RNA-seq data is comprehensively evaluated, with a particular focus on lncRNAs and low-abundance mRNAs. Fifteen performance metrics are used to evaluate DE tools and normalization methods using simulations and analyses of six diverse RNA-seq datasets.ResultsGene expression data are simulated using non-parametric procedures in such a way that realistic levels of expression and variability are preserved in the simulated data. Throughout the assessment, results for mRNA and lncRNA were tracked separately. All the pipelines exhibit inferior performance for lncRNAs compared to mRNAs across all simulated scenarios and benchmark RNA-seq datasets. The substandard performance of DE tools for lncRNAs applies also to low-abundance mRNAs. No single tool uniformly outperformed the others. Variability, number of samples, and fraction of DE genes markedly influenced DE tool performance.ConclusionsOverall, linear modeling with empirical Bayes moderation (limma) and a non-parametric approach (SAMSeq) showed good control of the false discovery rate and reasonable sensitivity. Of note, for achieving a sensitivity of at least 50%, more than 80 samples are required when studying expression levels in realistic settings such as in clinical cancer research. About half of the methods showed a substantial excess of false discoveries, making these methods unreliable for DE analysis and jeopardizing reproducible science. The detailed results of our study can be consulted through a user-friendly web application, giving guidance on selection of the optimal DE tool (http://statapps.ugent.be/tools/AppDGE/).Electronic supplementary materialThe online version of this article (10.1186/s13059-018-1466-5) contains supplementary material, which is available to authorized users.

Background Protein-coding RNAs (mRNA) have been the primary target of most transcriptome studies in the past, but in recent years, attention has expanded to include long non-coding RNAs (lncRNA). lncRNAs are typically expressed at low levels, and are inherently highly variable. This is a fundamental challenge for differential expression (DE) analysis. In this study, the performance of 14 popular tools for testing DE in RNA-seq data along with their normalization methods is comprehensively evaluated, with a particular focus on lncRNAs and low abundant mRNAs. Results Thirteen performance metrics were used to evaluate DE tools and normalization methods using simulations and analyses of six diverse RNA-seq datasets. Non-parametric procedures are used to simulate gene expression data in such a way that realistic levels of expression and variability are preserved in the simulated data. Throughout the assessment, we kept track of the results for mRNA and lncRNA separately. All statistical models exhibited inferior performance for lncRNAs compared to mRNAs across all simulated scenarios and analysis of benchmark RNA-seq datasets. No single tool uniformly outperformed the others. Conclusion Overall, the linear modeling with empirical Bayes moderation (limma) and the nonparametric approach (SAMSeq) showed best performance: good control of the false discovery rate (FDR) and reasonable sensitivity. However, for achieving a sensitivity of at least 50%, more than 80 samples are required when studying expression levels in a realistic clinical settings such as in cancer research. About half of the methods showed severe excess of false discoveries, making these methods unreliable for differential expression analysis and jeopardizing reproducible science. The detailed results of our study can be consulted through a user-friendly web application, http://statapps.ugent.be/tools/AppDGE/