(621g) An End-to-End Workflow for Diverse Transition State Conformer Generation Using Machine Learning

Searching for transition state (TS) geometries is one of the key problems in chemical kinetics, where the obtained TSs are used to calculate kinetic parameters and generate reaction mechanisms. Moreover, researchers often require more than a single TS geometry in practice to obtain quantitatively accurate estimations for reaction barriers or partition functions. While efficient tools and workflows (e.g., ETKDG and ConfGen) have been developed and widely used for stable molecules, most researchers resort to laborious and expensive manual TS guess creation/DFT optimization procedures, since few studies have focused on TS conformer generation.

In this study, we developed an end-to-end workflow that converts reaction SMILES into a set of transition state conformers. It chains together five steps: reactant/product geometry generation, TS geometry guess generation, pre-screening, optimization, and verification.

The core of the workflow is an equivariant graph neural network (TS-EGNN) model that yields TS geometries equivariant to the input reactant with respect to translation and external and internal rotation. It allows generated TSs to inherit the conformer diversity from the input reactant and product whose conformer diversity can be more easily achieved. Further, we trained a machine learning model to pre-screen guesses not on the reaction path to reduce the computational burden of failed optimization and verification attempts and identified xTB as an affordable alternative to the density functional theory (DFT) for high-throughput TS optimization and verification. Finally, we developed a modularized, objective-oriented, and user-friendly package to integrate the workflow.

The proposed workflow allows the rapid generation of diverse TS conformers. It can contribute to creating large datasets for transition states geometries, beneficial for developing future data-driven approaches; it also acts as a handy tool liberating computational chemists from laboriously hand-crafting TS geometries.