Improving the sensitivity of differential-expression analyses for under-powered RNA-seq experiments

High-throughput studies, in which thousands of hypothesis tests are conducted simultaneously, can be under-powered when effect sizes are small and there are few replicates. Here, I describe an approach to estimate the FDR for a given experiment such that the ground truth is known. A decision boundary between true and false positive calls can then be learned from the data itself along the axes of fold change and expression level. By excluding hits that fall into the false positive space, the FDR of any given method can be controlled providing a means to employ less conservative methods for detecting differential expression without incurring the usual loss of precision. I show that coupling this approach with a feature-selection method - an elastic-net logistic regression - can increase sensitivity 10-fold above what is achievable with the prevailing methods of the day. An R package implementing these methods is available at https://github.com/alextkalinka/delboy.