Abstract
BACKGROUND: Measuring how gene expression changes in the course of an experiment assesses how an organism responds on a molecular level. Sequencing of RNA molecules, and their subsequent quantification, aims to assess global gene expression changes on the RNA level (transcriptome). While advances in high-throughput RNA-sequencing (RNA-seq) technologies allow for inexpensive data generation, accurate post-processing and normalization across samples is required to eliminate any systematic noise introduced by the biochemical and/or technical processes. Existing methods thus either normalize on selected known reference genes that are invariant in expression across the experiment, assume that the majority of genes are invariant, or that the effects of up- and down-regulated genes cancel each other out during the normalization. RESULTS: Here, we present a novel method, moose(2) , which predicts invariant genes in silico through a dynamic programming (DP) scheme and applies a quadratic normalization based on this subset. The method allows for specifying a set of known or experimentally validated invariant genes, which guides the DP. We experimentally verified the predictions of this method in the bacterium Escherichia coli, and show how moose(2) is able to (i) estimate the expression value distances between RNA-seq samples, (ii) reduce the variation of expression values across all samples, and (iii) to subsequently reveal new functional groups of genes during the late stages of DNA damage. We further applied the method to three eukaryotic data sets, on which its performance compares favourably to other methods. The software is implemented in C++ and is publicly available from http://grabherr.github.io/moose2/. CONCLUSIONS: The proposed RNA-seq normalization method, moose(2) , is a valuable alternative to existing methods, with two major advantages: (i) in silico prediction of invariant genes provides a list of potential reference genes for downstream analyses, and (ii) non-linear artefacts in RNA-seq data are handled adequately to minimize variations between replicates.