There has been a lot of work shows potential of cancer diagnosis based on gene expression profiles, but no standards have been established. One goal of this thesis is to develop analysis tools for microarray data that generate biologically meaningful outputs. In particular, focus is given to development of supervised sample classification methods. A binary sample classifier called Independently Consistent Expression Discriminator (ICED) is developed. ICED has a novel feature selection principle that searches for features whose values are consistent in one class and are not consistent at the same level in the other class. Statistical analysis and empirical evaluation of ICED show that ICED can provide accurate predictions and discover biologically relevant genes. ICED is intrinsically a binary classifier. A multi-class classification package ICED$^3$ is extended from ICED. Algorithms in ICED$^3$ achieved good prediction accuracies on tested datasets. A practical issue of ICED and ICED$^3$ is that the execution time of the algorithms increases rapidly as the number of genes, the size of sample, and the number of classes increase. For the purpose of time efficiency the code was enabled, optimized, and parallelized on IBM POWER4 systems. The execution time has been significantly reduced.
Although many experimental and computational approaches became available in recent years for inference of protein interaction in genome-scale, the study of these approaches shows that they are far from perfect. In this thesis a computational approach for protein-protein interaction inference is developed based on the similarity of the proteins' phylogenetic trees, which are inferred using Bayesian estimation techniques. Our approach not only can be applied to both fully and partially sequenced organisms, but shows a potential of better accuracy and coverage in genome-scale analysis.