Dynamic Reaction Path Descriptors: Integrating Mechanistic Insights for Enhanced AI-Driven Reaction Prediction

Efficient and accurate prediction of chemical reactions is crucial for advancing synthetic chemistry. Traditional AI models largely rely on initial molecular descriptors that capture only the initial structures of reactants, thereby overlooking the dynamic nature of chemical processes. Here, we introduce a path descriptor methodology that extracts features from mechanistic analysis by identifying key intermediates and transition states along the reaction pathway. Using the electrophilic cyclization of phosphines and alkynes as a model system, our approach achieves higher accuracy than the initial descriptor model (0.66 → 0.83) with a dataset of 65 samples, highlighting the benefit of incorporating mechanistic information into feature design. Importantly, this study serves as a proof-of-concept that validates the effectiveness of mechanism-informed descriptors for reactivity prediction, providing a conceptual foundation for future extensions to broader and more complex reaction systems.