Statistical methods for microarrays and eQTL mapping.

摘要

Microarray technology can monitor expression levels of thousands of mRNA transcripts simultaneously. Its power has been demonstrated in thousands of studies. This thesis focuses on three topics arising from studies involving microarray data: (1) identifying differentially expressed (DE) genes among two or more conditions; (2) genetic linkage analysis for expression traits and (3) module construction to identify groups of co-regulated traits.;Kendziorski et al. (2003a) proposed an empirical Bayes approach for identifying DE genes among two or more conditions. An advantage of their approach is a hierarchical modeling framework that treats latent gene specific expression levels as arising from some population and thus allows a level of information sharing amongst genes. Although useful, the approach is limited by parametric assumptions prescribed by the models which assume a constant coefficient of variation (CCV) across all genes, which is not always true in real data. In this thesis, under the same hierarchical modeling framework, we propose the LNNMV model that relaxes the CCV assumption. This LNNMV model has been added to an R library called EBarrays and is now available in Bioconductor.;In the second topic, we consider quantitative trait loci (QTL) mapping experiments that treat microarray data as a phenotype for mapping. Recent so-called expression QTL (eQTL) experiments have shown that expression is a heritable trait that appears to be influenced by multiple QTL and sometimes interacting QTL. We propose a new automated and efficient multiple eQTL mapping approach that has the ability to identify both eQTL with large effect and eQTL with small or moderate effect as well as interacting eQTL. An application to a study of diabetes demonstrates many advantages of the approach.;Finally, we propose a method to identify groups of expression traits that are likely co-regulated, so-called co-expression co-regulation (CECR) modules. To do this, we build upon the popular work of Zhang and Horvath (2005) by extending their measure of adjacency from one based purely on correlation to one that accommodates both correlation and co-mapping. As in our eQTL mapping work, applications focus on a study of diabetes.