Reproduce “Generating transition states of chemical reactions via distance-geometry-based flow matching” on OpenResearch

About this paper

Yufei Luo, Xiang Gu, Jian Sun

Researchers from Xi’an Jiaotong University developed TS-DFM, a framework that predicts chemical reaction transition state geometries directly from reactant and product structures by operating in molecular distance geometry space. This approach achieves a mean RMSD of 0.1828 Å on the Transition1x dataset, representing an approximate 30% improvement over previous methods, and accelerates downstream transition state search algorithms.

Key insights

• Modeling TS prediction in distance geometry space, which directly captures bond breaking and formation, leads to higher accuracy and better generalization than Cartesian coordinate-based approaches.
• Optimal Transport–Conditional Flow Matching provides a robust and efficient framework for learning the velocity field that transforms an initial TS guess into the true TS structure.
• Perturbing reactant and product structures using normal mode sampling enables TS-DFM to discover diverse and potentially more energetically favorable alternative reaction pathways.

Summary by alphaXiv.