(477c) A Formal Grammar-Based Machine Learning Approach for Predicting Reaction Outcomes

Prediction of reaction outcomes without performing time-consuming experiments offers several invaluable advantages -- reducing experimentation costs due to elimination of highly improbable reactions, lowering the time to hypothesis-validation-and-correction, enabling high-throughput experimentation for complex reactions, and allowing significantly more time for analyzing the results. In recent years, machine learning techniques have been successfully applied in this area that, either directly or indirectly, address one or more of the above issues. A few areas of application include discovery of novel molecules with desired properties, understanding structure-property relationships, identifying catalysts for better reaction catalysis, and data-driven modeling of chemical reaction kinetics. The success of machine learning methods could largely be attributed to their inherent proficiency in capturing complex non-linear dependencies between various factors that govern the reaction systems. Here, we present an approach for predicting the most likely product of a reaction based on a given set of reactants and agents (catalysts, reaction medium). The structural properties of reactants and the product molecules are encoded based on a formal grammar, akin to context-free grammars (CFG) in the area of natural language processing (NLP), that essentially gives rise to a parse-tree representation of the molecule. This representation is used in a convolutional neural network architecture to model the structural transformations from a set of reactants to the most likely product of the reaction. The proposed framework could be used for rapid experimentation and analysis of reaction outcomes, especially if the molecules involved are less-known and structurally novel. In addition, the latent space of the neural network architecture could also be used to discover insights on the transformations from reactants to products in a given reaction under different conditions.