Mining large collections of gene expression data to elucidate transcriptional regulation of biological processes
Curry, Edward William James
A vast amount of gene expression data is available to biological researchers. As of October 2010, the GEO database has 45,777 chips of publicly available gene expression pro ling data from the Affymetrix (HGU133v2) GeneChip platform, representing 2.5 billion numerical measurements. Given this wealth of data, `meta-analysis' methods allowing inferences to be made from combinations of samples from different experiments are critically important. This thesis explores the application of localized pattern-mining approaches, as exemplified by biclustering, for large-scale gene expression analysis. Biclustering methods are particularly attractive for the analysis of large compendia of gene expression data as they allow the extraction of relationships that occur only across subsets of genes and samples. Standard correlation methods, however, assume a single correlation relationship between two genes occurs across all samples in the data. There are a number of existing biclustering methods, but as these did not prove suitable for large scale analysis, a novel method named `IslandCluster' was developed. This method provided a framework for investigating the results of different approaches to biclustering meta-analysis. The biclustering methods used in this work involve preprocessing of gene expression data into a unified scale in order to assess the significance of expression patterns. A novel discretisation approach is shown to identify distinct classes of genes' expression values more appropriately than approaches reported in the literature. A Gene Expression State Transformation (`GESTr') introduced as the first reported modelling of the biological state of expression on a unified scale and is shown to facilitate effective meta-analysis. Localised co-dependency analysis is introduced, a paradigm for identifying transcriptional relationships from gene expression data. Tools implementing this analysis were developed and used to analyse specificity of transcriptional relationships, to distinguish related subsets within a set of transcription factor (TF) targets and to tease apart combinatorial regulation of a set of targets by multiple TFs. The state of pluripotency, from which a mammalian cell has the potential to differentiate into any cell from any of the three adult germ layers, is maintained by forced expression of Nanog and may be induced from a non-pluripotent state by the expression of Oct4, Sox2, Klf4 and cMyc. Analysis of cMyc regulatory targets shed light on a recent proposition that cMyc induces an `embryonic stem cell like' transcriptional signature outside embryonic stem (ES) cells, revealing a cMyc-responsive subset of the signature and identifying ES cell expressed targets with evidence of broad cMyc-induction. Regulatory targets through which cMyc, Oct4, Sox2 and Nanog may maintain or induce pluripotency were identified, offering insight into transcriptional mechanisms involved in the control of pluripotency and demonstrating the utility of the novel analysis approaches presented in this work.