The aim of may RNA-seq studies is differential expression (DE) analysis between conditions of interest. The decreasing cost of sequencing has allowed for complex experimental designs, involving many samples over multiple conditions. Additionally, novel computational methods allow accurate estimation of transcript abundances, enabling analyses on the transcript level such as differential transcript expression (DTE) and differential transcript usage (DTU). Both types of experiments typically involve multiple hypotheses of interest per gene; i.e. testing DE between multiple conditions or for multiple transcripts. This leads to complex multiple testing problems. Indeed, a conventional data analysis strategy that assesses multiple hypotheses for every gene and controls the false discovery rate (FDR) on the hypothesis level fails to control the false discovery rate on a gene level, which will typically be inflated. This can lead to lower success rates of subsequent validation, since many genes without true treatment effects may be considered significant.
Researchers from Ghent University and the University of Zurich argue that multiple hypotheses at the gene level can be exploited in a two-stage testing procedure called stageR. In the screening stage, genes with effects of interest are prioritised using an omnibus test, e.g. a global F test, a global likelihood ratio test or by aggregating p values. Assessing the aggregated null hypothesis allows to control the gene-level false discovery rate and results in increased power in a DTU/DTE context. In addition, it enriches for genes with significant interaction effects in complex DE studies, thereby boosting power. In the confirmation stage, individual hypotheses are assessed for genes that pass the screening stage. Hence, it has the merit to combine the high power of aggregated hypothesis tests in stage I with the high resolution of individual hypothesis testing in stage II.
The authors show the benefits of adopting stage-wise testing in simulated and real complex DE experiments, where it results in a dramatic power increase for testing interaction effects and controls the gene-level false discovery rate. This is beneficial for downstream analyses of the results, for example gene set enrichment analyses. The power gain on DTE and DTU is shown using both simulations and real data, and the transcript-level DTU case study on a prostate cancer dataset highlights the advantages of combining gene-level power with transcript-level resolution of the results.