Differential Expression Analysis of Complex RNA-seq Experiments Using edgeR

Abstract

This article reviews the statistical theory underlying the edgeR software package for differential expression of RNA-seq data. Negative binomial models are used to capture the quadratic mean-variance relationship that can be observed in RNA-seq data. Conditional likelihood methods are used to avoid bias when estimating the level of variation. Empirical Bayes methods are used to allow gene-specific variation estimates even when the number of replicate samples is very small. Generalized linear models are used to accommodate arbitrarily complex designs. A key feature of the edgeR package is the use of weighted likelihood methods to implement a flexible empirical Bayes approach in the absence of easily tractable sampling distributions. The methodology is implemented in flexible software that is easy to use even for users who are not professional statisticians or bioinformaticians. The software is part of the Bioconductor project. This article describes some recently implemented features. Loess-style weighting is used to improve the weighted likelihood approach, and an analogy with quasi- likelihood is used to estimate the optimal weight to be given to the empirical Bayes prior. The article includes a fully worked case study with complete code. 1