The life cycle of drug discovery unsurprisingly starts with demonstrating the compound’s in vitro potency at a target of interest or through a phenotypic screen. However, the majority of drugs fail later on in clinical trials, illustrating that anticipating undesired activity early is arguably just as important as potency to maximize the investment of drug development. A preliminary solution is implementing machine learning methods to predict adverse effects, but toxicity testing is usually performed on a small subset of promising lead compounds, so data availability is limited. However, this is not the primary problem as there are many sources of public databases and literature, as well as government efforts for data transparency. Rather, the problem is that little has been done to transform this data into models or provide a simple, reliable means of identifying toxic features of novel compounds.<br/>We offer a solution to this problem with MegaTox, a toxicology-centric subset of Assay Central, a collection of Bayesian models of structure-activity relationship (SAR) data delivered in a self-contained executable program. SAR datasets are collated and stored using the code-management system Git, from which extended-connectivity fingerprints are calculated as contributing to activity the target of interest. Prediction scores of activity at undesired host targets (i.e. hERG and estrogen receptors) can be immediately generated by users, and visualization of active molecular features compliment human intuition to generate novel, nontoxic compounds. This technology ultimately aims to streamline drug design and reduce the superfluous synthesis of compounds that are more likely to fail due to common in vivo liabilities.<br/>We will describe the workflow of using publicly-available SAR datasets to create predictive Bayesian models to drive our own drug discovery efforts. Models of acute oral rat toxicity, cell cytotoxicity, hERG, estrogen receptors, and other targets at multiple thresholds will be discussed. Receiver Operator Characteristic scores range from 0.81 to 0.96 and Matthews Correlation Coefficients ranged from 0.34 to 0.84. Additionally, estrogen receptor models generated with alternative algorithms outside of Assay Central are presented for comparison.<br/><br/>

MegaTox for predicting compound liabilities

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.