Finding Reaction Mechanism Pathways with Deep Reinforcement Learning and Heuristic Search

Rojina Panta, Mohammadamin Tavakoli, Christian Geils, Pierre Baldi, Forest Agostinelli

Keywords: Reaction Mechanisms, Reinforcement Learning

Abstract: A chemical reaction, as the overall transformation of reactant molecules into product molecules, can be broken down into smaller reaction steps called elementary steps. An elementary step reaction involves only a single transition state, such as making bonds, breaking bonds, or donating electrons, that transforms the chemical reactants into chemical products. While Artificial Intelligence (AI) has been used to predict the outcomes of chemical reactions, most of these reaction predictors are designed to predict the major outcome of overall transformations, skipping the chemical reactions at a mechanistic level, making it difficult to identify intermediates and byproducts of the reaction. Information on the reaction mechanisms allows practitioners to identify intermediate molecules, predict the results of similar reactions under various conditions, and validate the feasibility of that reaction. Existing models for elementary steps, like OrbChain, can predict reaction mechanisms using a beam search algorithm, an incomplete search algorithm that does not explicitly consider path cost. To address this, we learn a heuristic function and perform a heuristic search that is complete and explicitly takes path cost into account.

Submission Number: 69