Researchers in China developed a unified machine learning framework to automate the calculation of chemical transition states, according to a study published Thursday in Nature Communications.
Computational chemists face persistent bottlenecks when calculating transition states, the fleeting molecular geometries that determine reaction kinetics. Earlier generative tools struggled with complex configurations such as transition metal catalysts, remaining largely restricted to simple reaction systems.
The system combines a curated structural database, UniTS-Lib, with a generative diffusion model named UniTS-Gen. UniTS-Lib contains 4,391 high-quality transition state structures spanning 42 chemical elements across diverse transformations. UniTS-Gen uses a custom higher-degree equivariant network to predict three-dimensional transition state configurations directly from two-dimensional reactant graphs.
Benchmarking tests confirmed that UniTS-Gen generalizes to unseen chemical systems and locates kinetically favored conformations. The authors demonstrated that the diffusion model delivers reliable initial geometric guesses, accelerating mechanistic discovery across organic synthesis.
Scientists from the Shanghai Academy of Artificial Intelligence for Science, Fudan University, ShanghaiTech University, and Golab Inc collaborated on the project. Density functional theory calculations ran on Fudan University’s CFFF platform and ShanghaiTech University’s high-performance computing system, while model training utilized CFFF alongside the Inspire platform at the Shanghai Academy of Artificial Intelligence for Science.
Individual contributors assisted with technical preparation: M.-J. Tang adapted the MolOP software to handle transition states, R.-K. Wu conducted quasi-classical molecular dynamics simulations of an ambimodal transition state, and X.-T. Yu alongside C.T. Ouyang manually inspected ambiguous entries in the UniTS-Lib dataset.
