Predicting Targeted Cancer Therapeutics

Abstract

Motivation: Cancer is a complex and evolving disease, making it difficult to discover effective treatments. Traditional drug discovery relies on high-throughput screening on reductionist models in order to enable the testing of 105 or 106 compounds. These assays lack the complexity of the human disease. Functional assays overcome this limitation by testing drugs on human tumors, however they can only test few drugs, and remain restricted to diagnostic use. An algorithm that identifies hits with fewer experiments could enable the use of functional assays for de novo drug discovery.\n\nResults: We developed a novel approach that we termed algorithmic ideation (AI) to select experiments, and demonstrated that this approach discovers hits 104 times more effectively than brute-force screening. The algorithm trains on known drug-target-disease associations assembled as a tensor, built from the (public) TCGA and STITCH databases and predicts novel associations. We evaluated our tensor completion approach using a temporal cutoff with data prior to 2012 used as training data, and data from 2012 to 2015 used as testing data. Our approach achieved 104-fold more efficient hit discovery than the traditional brute-force high-throughput screening. We further tested the method in a sparse, low data regime by removing up to 90% of the training data, and demonstrated the robustness of the approach. Finally we test predictive performance on drugs with no previously known interactions, and the algorithm demonstrates 103-fold improvement in this challenging problem. Thus algorithmic ideation can potentially enable targeted antineoplastic discovery on functional assays.\n\nAvailability: Freely accessible at https://bitbucket.org/aiinc/drugx.\n\nContact: quejebo@gmail.com, mcancobanoglug@gmail.com