Dividing out quantification uncertainty allows efficient assessment of differential transcript expression with edgeR

Differential expression analysis of RNA-seq is one of the most commonly performed bioinformatics analyses. Transcript-level quantifications are inherently more uncertain than gene-level read counts because of ambiguous assignment of sequence reads to transcripts. While sequence reads can usually be assigned unambiguously to a gene, reads are very often compatible with multiple transcripts for that gene, particularly for genes with many isoforms. Software tools designed for gene-level differential expression do not perform optimally on transcript counts because the read-to-transcript ambiguity (RTA) disrupts the mean-variance relationship normally observed for gene level RNA-seq data and interferes with the efficiency of the empirical Bayes dispersion estimation procedures. The pseudoaligners kallisto and Salmon provide bootstrap samples from which quantification uncertainty can be assessed. We show that the overdispersion arising from RTA can be elegantly estimated by fitting a quasi-Poisson model to the bootstrap counts for each transcript. The technical overdispersion arising from RTA can then be divided out of the transcript counts, leading to scaled counts that can be input for analysis by established gene-level software tools with full statistical efficiency. Comprehensive simulations and test data show that an edgeR analysis of the scaled counts is more powerful and efficient than previous differential transcript expression pipelines while providing correct control of the false discovery rate. Simulations explore a wide range of scenarios including the effects of paired vs single-end reads, different read lengths and different numbers of replicates.