[Part of this work was recently accepted to <i>ACS Central Science</i>, and has been given the following DOI until it is published: http://dx.doi.org/10.1021/acscentsci.8b00507 Preprint: https://arxiv.org/abs/1803.04465]<br/><br/>We report a step-change improvement in computational ADME-Tox prediction. Specifically, the PotentialNet family of graph convolutions achieved an 62% improvement over Random Forests by measure of R^2 across thirty-three large pharmaceutical datasets. The Absorption, Distribution, Metabolism, Elimination, and Toxicity (ADME-Tox) properties of drug candidates are estimated to account for up to 50% of all clinical trial failures. Predicting ADME-Tox properties has therefore been of great interest to the cheminformatics community in recent decades. Traditional cheminformatics approaches, whether the learner is a random forest or a deep neural network, leverage fixed fingerprint feature representations of molecules. In contrast, in this work, we "learn" the features most relevant to each chemical task at hand by representing each molecule explicitly as a graph, where each node is an atom and each edge is a bond. By applying graph convolutions to this explicity molecular representation, we achieve unprecedented accuracy in prediction of ADME-Tox properties. By challenging our methodology with difficult cross-validation procedures, we show that deep featurization better enables molecular predictors to not only interpolate, but to extrapolate to new regions of chemical space. Finally, we validate our model predictions with prospective experiments.

PotentialNet: Step-change improvement in ADME-Tox prediction with deep featurization

2019 Spring National Meeting

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