Diffusion-based generative AI for exploring transition states from 2D molecular graphs

Exploring transition state (TS) geometries is crucial for understanding chemical reaction mechanisms and kinetics. Machine learning (ML) models have shown promise in TS geometry prediction, but they typically require 3D reactant and product conformations, demanding significant effort and computation. This paper introduces TSDiff, a generative model based on the stochastic diffusion method, that predicts TS geometries solely from 2D molecular graphs. TSDiff surpasses existing ML models in accuracy and efficiency, sampling various TS conformations and identifying more favorable reaction pathways with lower barrier heights than reference databases. This demonstrates its potential for efficient and reliable TS exploration.

Seonghwan Kim, Jeheon Woo, Woo Youn Kim