| As the Human Genome Project comes to fruition, research on gene functions and genomic pathways (functional genomics) becomes more and more important. As one of the promising technologies to meet this challenge, cDNA microarray that can measure the gene expression levels of many genes simultaneously is becoming an important tool for biologists. As this new technology starts to produce large amounts of data, the need for various data analyses for this new kind of data arises. Hence, there has been early development of such analyses as unsupervised clustering analysis to associate unknown genes with the genes whose functions are known, classification of diseases with the help of various statistical tools, and genomic pathway analysis. However, the statistical considerations of those analyses, especially on the limitation of the number of samples available, have not been extensively addressed, even though most analyses are based on these statistical tools.; In this dissertation, two fundamental tools for the analysis of gene expression are addressed, with the constraints from the experiments in mind: multivariate gene expression analysis for genomic pathway network and finding strong feature (gene) sets to help molecular classification of diseases based on gene expression data. As we approach this problem, we need to keep in mind the very fundamental issue that should be addressed: a very limited number of samples (tissues) with many variables (genes). The statement of this problem is discussed and a few novel methods are proposed.; We also address the need for parallel implementations of the algorithms to accommodate the heavy computing necessary to handle the practical size of gene expression data and have implemented a parallel system, PAGE , Parallel Analysis of Gene Expression, on Beowulf clustered parallel computers. We also have developed visualization tools, VOGE, Visualization of Gene Expression, to help biologists exploit and interpret the huge amount of outputs from PAGE. |
