Rod K. Nibbe, Mehmet Koyutürk, Mark R. Chance

Emerging evidence indicates that gene products implicated in human cancers often cluster together in "hot spots" in protein-protein interaction (PPI) networks. Additionally, small sub-networks within PPI networks that demonstrate synergistic differential expression with respect to tumorigenic phenotypes were recently shown to be more accurate classifiers of disease progression when compared to single targets identified by traditional approaches. However, many of these studies rely exclusively on mRNA expression data, a useful but limited measure of cellular activity. Proteomic profiling experiments provide information at the post-translational level, yet they generally screen only a limited fraction of the proteome. Here, we demonstrate that integration of these complementary data sources with a "proteomics-first" approach can enhance the discovery of candidate sub-networks in cancer that are well-suited for mechanistic validation in disease. We propose that small changes in the mRNA expression of multiple genes in the neighborhood of a protein-hub can be synergistically associated with significant changes in the activity of that protein and its network neighbors. Further, we hypothesize that proteomic targets with significant fold change between phenotype and control may be used to "seed" a search for small PPI sub-networks that are functionally associated with these targets. To test this hypothesis, we select proteomic targets having significant expression changes in human colorectal cancer (CRC) from two independent 2-D gel-based screens. Then, we use random walk based models of network crosstalk and develop novel reference models to identify sub-networks that are statistically significant in terms of their functional association with these proteomic targets. Subsequently, using an information-theoretic measure, we evaluate synergistic changes in the activity of identified sub-networks based on genome-wide screens of mRNA expression in CRC. Cross-classification experiments to predict disease class show excellent performance using only a few sub-networks, underwriting the strength of the proposed approach in discovering relevant and reproducible sub-networks.