Methods for incorporating biological information into the statistical analysis of gene expression microarray data

Abstract

Microarray technology has made it possible for researchers to simultaneously measure the expression levels of tens of thousands of genes. It is believed that most human diseases and biological phenomena occur through the interaction of groups of genes that are functionally related. To investigate the feasibility of incorporating functional information and/or constraints (based on biological and technical needs) into the classification process two approaches were examined in this thesis. The first of these approaches investigated the effect of incorporating a pre-filter into the gene selection step of the classifier construction process. Both simulated and real microarray datasets were used to assess the utility of this approach. The pre-filter was based on an early method for determining if a gene had undergone a biologically relevant level of differential expression between two classes. The genes retained by the pre-filter were ranked using one of five standard statistical ranking methods and the most highly ranked were used to construct a predictive classifier. To generate the simulated data a selection of different parametric and non-parametric techniques were employed. The results from these analyses showed that when the constraints that the pre-filter contains were placed on the classification analysis, the predictive performance of the classifiers were similar to when the pre-filter was not used. The second approach explored the feasibility of incorporating sets of functionally related genes into the classification process. Three publicly available datasets obtained from studies into breast cancer were used to assess the utility of this approach. A summary of each gene-set was derived by reducing the dimensionality of each gene-set via the use of Principal Co-ordinates Analysis. The reduced gene-sets were then ranked based on their ability to distinguish between the two classes (via Hotelling’s T2) and those most highly ranked were used to construct a classifier via logistic regression. The results from the analyses undertaken for this approach showed that it was possible to incorporate function information into the classification process whilst maintaining an equivalent (if not higher) level of predictive performance, as well as improving the biological interpretability of the classifier.