Designing chemist-like machine intelligence for retrosynthesis and reaction outcome prediction

Abstract

A reliable prediction of chemical reactivity remains in the realm of knowledgeable synthetic chemists. Automating this process by artificial intelligence can be useful to aid synthesis design in future digital laboratories. While several models have demonstrated promising results, the current best models are predicting the retrosynthesis and reaction outcome in a machinery fashion, such as language translation, which are not intuitive for scientists to understand and thus hard to trust the machine intelligence prediction. Here, we design two machine-learning based models, LocalRetro and LocalTransform, to predict the retrosynthesis pathway as well as the reaction outcome based on chemist intuition. For retrosynthesis, we design the model to perform the retrosynthesis by finding “strategy bonds” and “reactive atoms” and apply the possible reaction rule to decompose the reaction product to reactants. On the other hand, we design a deep learning model to find out all the possible reaction center and predict the possible electron rearrangements to predict the reaction outcome. Our retrosynthesis model shows 99.2% top-5 round-trip accuracy and the reaction outcome prediction model shows a 99.4% top-5 exact-match accuracy for a US Patent Reaction Dataset, containing 50 016 reactions. With the devise of novel reaction template extraction method, our proposed models are more interpretable and more accurate than all the existing machine intelligence methods. We further demonstrated practical application of our retrosynthesis model by correctly predicting the synthesis pathways of five drug candidate molecules from various literature and showed a comparable reactivity prediction accuracy compared with human experts by our reaction outcome prediction model.