A generalized-template-based graph neural network for accurate organic reactivity prediction

Abstract

The reliable prediction of chemical reactivity remains in the realm of knowledgeable synthetic chemists. Automating this process by using artificial intelligence could accelerate synthesis design in future digital laboratories. While several machine learning approaches have demonstrated promising results, most current models deviate from how human chemists analyse and predict reactions based on electronic changes. Here, we propose a chemistry-motivated graph neural network called LocalTransform, which learns organic reactivity based on generalized reaction templates to describe the net changes in electron configuration between the reactants and products. The proposed concept dramatically reduces the number of reaction rules and exhibits state-of-the-art product prediction accuracy. In addition to the built-in interpretability of the generalized reaction templates, the high score-accuracy correlation of the model allows users to assess the uncertainty of the machine predictions. To understand reactions in organic chemistry, ideally simple rules would help us predict the outcome of new reactions, but in reality such rules are not easily identified. Chen and Jung extract generalized reaction templates from data and show that they can be used in graph neural networks to predict the outcome of reactions and, despite simplification, still represent a high percentage of existing reactions.