Tree-based differential testing using inferential uncertainty for RNA-seq

Identifying differentially expressed transcripts poses a crucial yet challenging problem in transcriptomics. Substantial uncertainty is associated with the abundance estimates of certain transcripts which, if ignored, can lead to the exaggeration of false positives and, if included, may lead to reduced power. Here, we introduce a data-driven differential testing method that maximizes biological resolution while retaining statistical power. Given a set of RNA-seq samples, TreeTerminus arranges transcripts in a hierarchical tree structure that encodes different layers of resolution for interpretation of the abundance of transcriptional groups, with uncertainty generally decreasing as one ascends the tree from the leaves. We introduce mehenDi, which utilizes the tree structure from TreeTerminus for differential testing. The nodes output by mehenDi, called the selected nodes, are determined in a data-driven manner to maximize the signal that can be extracted from the data while controlling for the uncertainty associated with estimating the transcript abundances. The identified selected nodes can include transcripts and inner nodes, with no two nodes having an ancestor/descendant relationship. We evaluate our method on both simulated and experimental data sets and compare its performance with other tree-based differential methods, as well as with uncertainty-aware differential transcript/gene expression methods. Our method detects inner nodes that show a strong signal for differential expression, which would have been overlooked when analyzing the transcripts alone.