A housekeeping gene-based procedure for the selection of differentially expressed genes for Affymetrix microarray experiments.

摘要

This thesis is concerned with modeling and analyzing experimental Affymetrix gene expression microarray data. For the analysis part, we propose a cross-gene-rank procedure for the detection of differentially expressed genes for Affymetrix microarray experiments. This procedure is based on a comprehensive exploration of housekeeping genes with the objective of classifying them (Eisenberg and Levanon (2003)) into "good" (non-differentially expressed) and "bad" (differentially expressed) genes using the non-parametric classification method based on cross-gene ranks developed in this work.;The advantages of the cross-gene rank procedure in the context of microarray gene expressions are as follows: Testing the null hypothesis that the expression measurements of an individual gene have identical means across treatments is reduced to test the null hypothesis that its cross-gene ranks of expression measurements are constant across treatments, given we know that most of the genes are non-differential genes; Instead of taking the genes independently from each other as in the model-based inferences and the traditional settings of rank statistcs, we put an individual gene in the context of other genes by considering it's position among them. In this manner, the cross-gene rank and its rank statistic are valuable expansions of currently available inferences. It puts an expression measurement at the cross point of a gene and a treatment in a grid composed of treatments as columns and genes as rows.;For the modeling part, models for expression measurements of non-differentially expressed "good" housekeeping genes are developed, where the expression measurements follow a truncated normal distribution with mean-dependent variances that follow "asymptotic" regression models. The variances depend on both the gene expression level and the experimental conditions. The variation in expression measurements of "good" housekeeping genes captured by the models includes both the observed variation within a single treatment and between a treatment and the control. Proper modeling of the dependence of the variance on the mean expression level help improve the estimation of the variation in gene expression when only a small number of replicates are available for the analysis.;Further, comparisons across a treatment and the control are analyzed using fold change as a statistical measure for all pairwise comparisons. Null confidence bounds of the fold change statistic are determined for all genes.;Our simulation study shows that our procedure has lower error rate than the constant confidence bounds procedure which is popularly used to provide comparative data by investigators using microarrays. Our procedure also greatly increases the yield of statistically significant genes delivered by our new analysis methods.;Key words: Affymetrix gene expression microarray, housekeeping genes, cross-gene rank, linear model, truncated normal distribution, asymptotic variance model, fold change (ratio), confidence bound.