Revealing protein networks and gene-drug connectivity in cancer from direct information.

The connection between genetic variation and drug response has long been explored to facilitate the optimization and personalization of cancer therapy. Crucial to the identification of drug response related genetic features is the ability to separate indirect correlations from direct correlations across abundant datasets with large number of variables. Here we analyzed proteomic and pharmacogenomic data in cancer tissues and cell lines using a global statistical model connecting protein pairs, genes and anti-cancer drugs. We estimated this model using direct coupling analysis (DCA), a powerful statistical inference method that has been successfully applied to protein sequence data to extract evolutionary signals that provide insights on protein structure, folding and interactions. We used Direct Information (DI) as a metric of connectivity between proteins as well as gene-drug pairs. We were able to infer important interactions observed in cancer-related pathways from proteomic data and predict potential connectivities in cancer networks. We also identified known and potential connections for anti-cancer drugs and gene mutations using DI in pharmacogenomic data. Our findings suggest that gene-drug connections predicted with direct couplings can be used as a reliable guide to cancer therapy and expand our understanding of the effects of gene alterations on drug efficacies.