Detection of differentially expressed genes (DEGs) between different biological conditions is a key data analysis step of most RNA-sequencing studies. Conventionally, computational tools have used gene-level read counts as input to test for differential gene expression between sample condition groups. Recently, it has been suggested that statistical testing could be performed with increased power at a lower feature level prior to aggregating the results to the gene level. In this study, we systematically compared the performance of calling the DEGs when using read count data at different levels (gene, transcript, and exon) as input, in the context of two publicly available data sets. Additionally, we tested two different methods for aggregating the lower feature-level p-values to gene-level: Lancaster and empirical Brown's method. Our results show that detection of DEGs is improved compared to the conventional gene-level approach regardless of the lower feature-level used for statistical testing. The overall best balance between accuracy and false discovery rate was obtained using the exon-level approach with empirical Brown's aggregation method, which we provide as a freely available Bioconductor package EBSEA (https://bioconductor.org/packages/release/bioc/html/EBSEA.html).
Keywords: RNA-seq; differentially expressed genes; exon-level read counts; gene-level read counts; transcript compatibility counts; transcript-level read counts.