Learning Pathway-based Decision Rules to Classify Microarray Cancer Samples

Despite recent advances in DNA chip technology current microarray gene expression studies are still affected by high noise levels, small sample sizes and large numbers of uninformative genes. Combining microarray data with cellular pathway data by using new integrative analysis methods could help to alleviate some of these problems and provide new biological insights. We present a method for learning simple decision rules for class prediction from pairwise comparisons of cellular pathways in terms of gene set expression levels representing the upand downregulation of pathway members. The procedure generates compact and comprehensible sets of rules, describing changes in the relative ranks of gene expression levels in pairs of pathways across different biological conditions. Results for two large-scale microarray studies, containing samples from prostate cancer and B-cell lymphoma patients, show that the method provides robust and accurate rule sets and new insights on differentially regulated pathway pairs. However, the main benefit of these predictive models in comparison to other classification methods like support vector machines lies not in the attained accuracy levels but in the ease of interpretation and the insights they provide on the relative regulation of cellular pathways in the biological conditions under consideration.