We describe how machine learning of historical data for organic chemistry informs synthesis planning algorithms that are able to propose fully-specified synthetic routes to target molecules. First, we present a supervised learning approach to predict the products of organic reactions given their reactants, reagents, and solvents. Analogous to human approaches, the conjugate neutral network method considers possible sites of reactivity and evaluates their relative likelihoods. The major product is predicted correctly over 85% of times in ~100 ms per example. As an extension, we report a generalizable approach to prediction of site selectivity in aromatic C-H functionalization reactions with an accuracy of 92%. Next, we summarize the results of a neural network model for identifying chemical reaction conditions, e.g., catalysts, solvents, reagents, and temperatures. Even though reaction context is often poorly represented in reaction databases, the neural network model trained on ~10 million example from Reaxys proposes conditions that closely match recorded chemical context data. The reaction and context prediction algorithms are integrated into a retrosynthesis workflow that, for a given molecular target, predicts a rank ordered list of reaction paths that connect the target to purchasable starting materials via a series of feasible reaction steps. Finally, incorporating reinforcement learning allows selection of pathways that optimize user objectives such as low toxicity or use of green solvents.

Machine learning for organic chemistry reaction prediction and retrosynthesis

2019 Fall National Meeting

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