Improving Reaction Yield Prediction with Chemical Atom-Level Reaction Learning

Reaction yield, the percentage of reactants successfully converted into desired products relative to the theoretical maximum, is a critical metric for evaluating chemical reaction efficiency. Accurate prediction of reaction yields enables chemists to explore reaction space more effectively and design optimal synthetic pathways, which is essential for synthetic chemistry, drug development, and materials science. However, yield prediction remains challenging due to the complex influence of side reactions mediated by auxiliary (condition) molecules─such as catalysts, ligands, additives, and solvents─which typically facilitate the reaction without directly contributing atoms to the final product. Existing computational approaches often rely heavily on handcrafted features or overlook atom-level interactions that govern these effects, particularly the mechanisms by which auxiliary molecules modulate yield. In this work, we propose a chemical atom-level reaction learning (CARL) framework that leverages graph neural networks to explicitly model atom-level interactions between reactants and auxiliary molecules, thereby improving the fidelity of reaction modeling for yield prediction. Experimental evaluations on multiple benchmark data sets demonstrate that CARL achieves state-of-the-art (SOTA) performance without relying on handcrafted descriptors or domain-specific heuristics. Moreover, the framework highlights key substructures in both reactants and auxiliary molecules that critically influence reaction outcomes, underscoring its potential for systematically exploring larger reaction spaces.