| Microarray analysis allows for genome-level exploration of gene expression by taking a snapshot of the cell at a specific point in time. Such datasets may provide insight into fundamental biological questions as well as address clinical issues such as diagnosis and therapy selection. The resulting data sets are very large and complex, and often suffer from sacrifice of specificity for scale. Sophisticated computational tools are needed for nontrivial, highly accurate, and consistent biological interpretation of microarray data.; This dissertation addresses the issue of improving the specificity of biological signal detection from microarray data. I address this problem on three levels. First, I developed two robust and accurate algorithms for missing value estimation for microarray data, KNNimpute and SVDimpute. The algorithms perform overwhelmingly better than row averaging or zero filling methods, and KNNimpute is robust to the choice of parameters used, percent of values missing, and type of data. Second, I created MAGIC, a flexible probabilistic framework for gene function prediction based on integrated analysis of high-throughput biological data, including gene expression data and protein-protein interactions data. I applied MAGIC to S. cerevisiae data and showed that it improves the specificity of gene grouping compared to its input microarray-based clustering methods. Finally, I suggested and evaluated methods for identification of differentially expressed genes and propose a general procedure for evaluation of other biomarker identification methods. |
