Differential expression analysis of complex RNA-seq experiments

© 2013 Dr. Yunshun ChenAs the cost of DNA sequencing decreases, sequencing technologies become more and more attractive to many researchers as platforms for studying gene expression. Although there exist many different combinations of technologies and protocols, we use the term `RNA-Seq' to denote the very broad class of experiments in which gene expression is studied by sequencing RNA. A very common goal for analysing RNA-Seq experiments is to identify genes that are differentially expressed across specified conditions in a designed experiment. It has proven to be very challenging for statisticians since it is a high-dimensional multiple testing problem in which one or more tests are performed for each of tens of thousands of genes. RNA-Seq data takes the form of integer counts, and there are a few issues that we need to address. Firstly, it is important to model the variance of the RNA-Seq count data correctly. Secondly, there is a need to estimate the variabilities of gene counts accounting for all sources of variation for any complex experimental design. Thirdly, the variation between biological replicates needs to be estimated as reliably as possible from a very small number of replicate libraries, and different genes should allow different degrees of biological variation. Finally, a powerful testing method is required for the purpose of detecting as many differentially expressed genes as possible while controlling the false discovery rate. These issues will be fully discussed in the thesis. In short, the mixture structure of biological variation and measurement error in the RNA-Seq counts implies a quadratic mean-variance relationship, hence negative binomial models for modelling the variance. Generalized linear models are used as a flexible statistical framework for the analysis of read counts from RNA-Seq gene expression studies, which provides the ability to analyse complex experiments involving multiple treatment conditions and blocking variables while still taking full account of biological variation. Biological variation between RNA samples is estimated separately from the technical variation associated with sequencing technologies using an adjusted profile likelihood approach. Also, the high-dimensional nature of sequencing data allows possibilities for borrowing information from the ensemble of genes which can assist in inference about each gene individually. A novel weighted likelihood empirical Bayes method is proposed to allow each gene to have its own specific variability while learning from the others. Finally, testing hypothesis while accounting for the variability of estimated parameters is discussed, and new methods for testing for differential expression are proposed. The analysis pipeline proposed in this thesis is implemented in the edgeR package of the Bioconductor project. The software implementation and computational algorithms designed for the purpose of computational efficiency are also discussed in this thesis. The methods developed here can be applied to count data arising from DNA-Seq applications, including ChIP-Seq for epigenetic marks and DNA methylation analyses