Motivation: Most approaches to gene expression analysis use real-valued expression data, produced by highthroughput screening technologies, such as microarrays. Often, some measure of similarity must be computed in order to extract meaningful information from the observed data. The choice of this similarity measure frequently has a profound effect on the results of the analysis, yet no standards exist to guide the researcher.

In the study of complex phenotypes, single gene markers can only provide limited insights into the manifestation of phenotype. To this end, protein-protein interaction (PPI) networks prove useful in the identification of multiple interacting markers. Recent studies show that, when considered together, many proteins that are connected via physical and functional interactions exhibit significant differential expression with respect to various complex phenotypes, including cancers. As compared to single gene markers, these “coordinately dysregulated subnetworks ” improve diagnosis and prognosis of cancer significantly and offer novel insights into the network dynamics of phenotype. However, the problem of identifying coordinately dysregulated subnetworks presents significant algorithmic challenges. Existing approaches utilize heuristics that aim to greedily maximize information-theoretic class separability measures, however, by definition of “coordinate ” dysregulation, such greedy algorithms do not suit well to this problem. In this paper, we formulate coordinate dysregulation in the context of the well-known set-cover problem, with a view to capturing the coordination between multiple genes at a sample-specific resolution. Based on this formulation, we adapt state-of-the-art approximation algorithms for set-cover to the identification of coordinately dysregulated subnetworks. Comprehensive experimental results on human colorectal cancer (CRC) show that, when compared to existing algorithms, the proposed algorithm, NetCover, improves diagnosis of cancer and prediction of metastasis significantly. Our results also demonstrate that subnetworks in the neighborhood of known CRC driver genes exhibit significant coordinate dysregulation, indicating that the notion of coordinate dysregulation may indeed be useful in understanding the network dynamics of complex phenotypes. 1.