Detecting Differentially Expressed Genes with RNA-seq Data Using Backward Selection to Account for the Effects of Relevant Covariates

Abstract

A common challenge in analysis of transcriptomic data is to identify differentially expressed genes, i.e., genes whose mean transcript abundance levels differ across the levels of a factor of scientific interest. Transcript abundance levels can be measured simultaneously for thousands of genes in multiple biological samples using RNA sequencing (RNA-seq) technology. Part of the variation in RNA-seq measures of transcript abundance may be associated with variation in continuous and/or categorical covariates measured for each experimental unit or RNA sample. Ignoring relevant covariates or modeling the effects of irrelevant covariates can be detrimental to identifying differentially expressed genes. We propose a backward selection strategy for selecting a set of covariates whose effects are accounted for when searching for differentially expressed genes. We illustrate our approach through the analysis of an RNA-seq study intended to identify genes differentially expressed between two lines of pigs divergently selected for residual feed intake. We use simulation to show the advantages of our backward selection procedure over alternative strategies that either ignore or adjust for all measured covariates.

Multiple hypothesis testing and RNA-seq differential expression analysis accounting for dependence and relevant covariates

“…We illustrate the performance of our method using a backward selection procedure and the differential expression analysis method voom . If any of the covariates are uncorrelated or weakly correlated with the included variables, the performance of our method is similar to that of Nguyen et al (2015). On the other hand, if there are covariates strongly correlated with the included variables, we show that our method outperforms the method of Nguyen et al (2015).…”

“…There are many different ways to formally measure the relevance of covariate j using its p j|S . Nguyen et al (2015) considered two measurements: 1) the number of elements of p j|S less than 0.05 and 2) the Kolmogorov-Smirnov statistic (Kolmogorov, 1933;Smirnov, 1948) measuring the discrepancy between the uniform(0, 1) distribution and the Grenander estimate of a non-increasing distribution computed from the elements of p j|S . These two measurements are natural choices given the aforementioned behavior of the empirical distribution of p-values.…”

“…Now we apply the FSR variable selection method to the RFI RNA-seq dataset. This dataset has been used and described in Nguyen et al (2015). For completeness, we recall the details of the RFI data set in this article.…”

“…We use covariates to describe variables that are subject to variable selection. Nguyen et al (2015) addressed the covariate selection problem in the context of RNA-seq differential expression analysis by proposing a backward selection strategy to choose the most relevant covariates. They showed that their method works well when the included variables are uncorrelated or weakly correlated with one or more covariates.…”

“…We also compare the differential expression analysis produced by our method with several alternatives, such as the method of Nguyen et al (2015), methods including or excluding all covariates, and methods using surrogate variables Lee et al, 2017). As a result of accounting for truly relevant covariates, our approach outperforms others in terms of power, FDR control, and ability to distinguish the true and false signals with respect to the included variables.…”