Predicting Chemical Reaction Type and Reaction Products with Recurrent Neural Networks

Synthesizing chemical molecules is a necessary and often time-consuming process in fields like drug discovery and materials science. Here, we use recurrent neural networks to automate two tasks in chemical synthesis: 1) classifying the type of reaction two molecules will participate in, and 2) generating the correct products given the reactants and reaction type. We achieve 96% accuracy and 99% AUPRC in the task of reaction type classification, significantly improving on the baseline and previous approaches. In addition, our model for reaction-type generation stabilizes at 87.6% molecular similarity between the predicted and actual products, and reaches 96.6%. We demonstrate considerable improvement over rule-based prediction systems. These two completely data-driven systems can be useful to synthetic chemists seeking to predict the outcome of chemical reactions.