Modeling chemical reaction with graph neural networks and reaction template

Abstract

This thesis presents the development of three graph-based machine learning (ML) models for single-step retrosynthesis, reaction outcome prediction, and atom-to-atom mapping (AAM) in the field of organic chemistry. The models, namely LocalRetro, LocalTransform, and LocalMapper, are designed following chemist-intuition to enhance their interpretability and performance. For each specific reaction task, we carefully design reaction templates with varying levels of simplifications to maximize their applicability while minimizing resolution requirements. Extensive experiments conducted on widely used USPTO reaction datasets demonstrate the superiority of our models over previous methods. These improvements can be attributed to the incorporation of chemist-intuitive design principles, which are suggested by our comprehensive analysis. Moreover, we showcase the potential of chemist-intuitive model design by introducing LocalRetro-mp, an advanced model that leverage both retro- and forward-synthesis predictions. This model enables more diverse and feasible synthesis planning, thereby enhancing the capabilities of retrosynthesis prediction. Overall, the results presented in this thesis highlight the benefits of incorporating chemist-intuition into ML models for reaction prediction in organic chemistry. The findings not only contribute to the advancement of the field but also pave the way for future research in leveraging chemist-intuitive design principles for further improvements in synthesis planning and other related areas.