UCLA

Drug combinations significantly expanded the opportunity space of druggable genome in cancer therapeutics, but the discovery of novel combinations is still limited by the capacity of our current drug screening technology. To address the challenge, we introduced a data-driven search method called the Predictive Optimization of Pharmaceutical Efficacy, or PROPHECY, for the selection of drugs in combinatorial cancer therapeutics. The user provides the genetic profile, of cancer cell lines or primary cells, and PROPHECY will select optimal drug combinations from a comprehensive list of drugs to meet clinical objectives. The decision making is accomplished by in silico drug screening in which the sensitivities of a cell on different drug combinations are ranked. The predictive model of sensitivity is trained to recognize signatures of information spread in the protein-protein interaction network. Once a comprehensive dataset of drug screening experiment is supplied, the computer could automatically learn interactions between drug targets and disease genes in the information signatures, and infer sensitivity for unseen drug and cell line pairs. We showed that the prediction have high correlation with experimental data by cross validation performed on a dataset of 40,000 entries, which represents 100 cancer drugs applied on 450 cell lines. We also verified the applicability of PROPHECY by performing an in vitro experiment with 36 two drug pairs suggested by the program and a panel of 6 cell lines. PROPHECY not only predicted the sensitivity with high accuracy, but also discovered novel high efficacy combinations and reproduced existing drug combinations. Unlike currently predominant approach of reductionistic drug development, the prediction of drug efficacy is based on network view of proteomic scale data, so can accurately reflect modular activity of the proteome and elucidate target gene interactions in de novo drug combinations.