Accurate predictions of site selectivity for aromatic functionalization reactions is crucial for prioritizing target compounds in drug discovery and routes in process chemistry. Selectivity of aromatic C-H functionalization reactions is highly dependent on subtle electronic plus steric features of the substrate and can be controlled by reaction conditions such as catalysts. Prediction of reaction outcomes using machine learning is possible using both template and template free approaches and is generalizable to a large corpus of reactions but comes with a tradeoff of poor accuracy with site and regio-specificity. Current approaches to site selectivity prediction achieves good accuracy based on electronic calculations but does not address steric or reaction conditions and generally takes 1-10 min per molecule. We report a generalizable approach to prediction of site selectivity that is accomplished using a shared graph-convolutional neural network fed into a multitask predictor trained on ~58,000 examples from Reaxys. Overall mean reciprocal rank for all C-H functionalization tasks is 92% and predictions are made in approximately 200 ms. Likelihoods are given for each individual task allowing a chemist to quickly determine which C-H functionalization reactions might proceed with high selectivity and to prioritize those targets or routes.<br/>

Multitask prediction of site selectivity in aromatic C-H functionalization reactions

2019 Spring National Meeting

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