Abstract B162: Small molecule target prediction and identification of novel anti-cancer compounds using a data-driven bayesian approach

Abstract

Abstract It typically takes 15 years and 2.6 billion dollars to go from a small molecule in the lab to a clinically useable drug. This cost has only increased over the past decade despite advancements in drug development, with one of the biggest bottlenecks being drug target identification. Current efforts are driven by case specific experimentation - a slow and failure-prone process - and each time a new drug is developed these experiments need to be redone from scratch. Recently there has been an explosion of genomic, chemical, clinical, and pharmacological data but no target prediction method has successfully integrated these multiple pieces of evidence into a cohesive approach. Here we propose BANDIT: a Bayesian ANalysis to find Drug Interaction Targets. BANDIT combines data on drug efficacies, post-treatment transcriptional responses, drug structures, side effects, bioassay results, and known targets within a Bayesian framework to predict the targets of small molecules. When applied to a test set of drugs with known targets, BANDIT achieved a predictive power of 91% at identifying drug pairs that share a target (AUROC = 0.91). The power of BANDIT's classifier steadily increased from 60% to 91% as we increased the number of included data types- demonstrating the strength of BANDIT's Big Data approach. Additionally, we found that BANDIT could reproduce the results of a published experimental kinase inhibitor screen with no additional data or training, indicating its potential to replace otherwise tenuous and costly experiments. We next used BANDIT to predict interaction targets for over 50,000 small molecules with no known targets and specifically identified a set of unstudied small molecules predicted to inhibit microtubule formation. We focused on microtubules because of their relevance as chemotherapy targets. Using immunohistochemical staining and tubulin binding assays, we were able to validate these predictions in vitro, identifying new potential chemotherapeutic options. We also found that by clustering drugs based on their BANDIT output, we were able to correctly predict the mechanism of action, as well as the target, for these new small molecule predictions. When this method was applied to the verified microtubule inhibitors, we were able to accurately separate small molecules based on their method of inhibition - polymerizing vs. depolymerizing. Using BANDIT's output we then created an interaction network of all known drugs characterized by their ATC type and intended use. We observed a definite structure to this “drug universe” network with certain drug classes, such as “Dermatologicals” and “Musculo-skeletal System,” clustering together. This information could help advance future drug development efforts by providing information how certain drug classes are similar and could potentially interact in a biological system. Altogether these results attest to BANDIT's ability to provide a novel, broadly applicable way to identify drug targets. By predicting targets for new and established drugs, without the need for numerous target-specific assays, BANDIT could greatly expedite future drug discovery efforts by reducing cost and time. Additionally, by rapidly screening compounds that have been developed but never fully studied we could efficiently increase our potential anti-cancer therapeutic options. Citation Format: Neel S. Madhukar, Linda Huang, Prashant Khade, Katie Gayvert, Paraskevi Giannakakou, Olivier Elemento. Small molecule target prediction and identification of novel anti-cancer compounds using a data-driven bayesian approach. [abstract]. In: Proceedings of the AACR-NCI-EORTC International Conference: Molecular Targets and Cancer Therapeutics; 2015 Nov 5-9; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2015;14(12 Suppl 2):Abstract nr B162.