(509ct) An Automated Workflow to Rapidly and Accurately Generate Transition State Structures Using Machine Learning

A powerful way to build predictive chemical models is by calculating reaction rates using transition state theory. Such methods, while offering both quantitative and qualitative estimates of preferred pathways in reaction networks, remain computationally expensive. A major bottleneck in this analysis is the determination of 3D transition state (TS) structures. Traditional quantum mechanical string-based methods to generate TSs are much too slow, while existing heuristic methods, such as synchronous transit approximations and KinBot, either have high failure rates or are parameterized for few specific reaction types.

In previous work, we developed a graph neural network to rapidly and accurately predict 3D TS geometries of isomerization reactions from the geometries of the reactant and product. Here, we extend our approach to reactions of type A + B → C. Our workflow consists of a novel alignment procedure where the A+B complex is first automatically aligned with C. Both the reactant complex and product are then fed through an updated graph neural network architecture and differential simulator to produce the final TS geometry. Importantly, we include additional constraints on the output geometry, which prevent critical failure modes from our previous approach.

We show that our method successfully generates transition state geometries for several A + B → C reaction types relevant to gas phase chemistry. We further integrate our workflow in a user-friendly package which allows users to automatically obtain TS geometries from atom-mapped reactant and product SMILES.