Reactive machine learning potential for accelerating transition state search in organic synthesis

Understanding reaction kinetics is fundamental to organic synthesis, yet traditional quantum chemistry-based transition state searches are computationally expensive. Here we present DeePEST-OS, a reactive machine learning potential designed for rapid and accurate transition state optimization and energy barrier prediction spanning ten chemical elements. Trained on approximately 75,000 reactions generated by a low-cost data preparation strategy, this model integrates physical priors from semi-empirical quantum chemistry with equivariant message passing networks to predict potential energy surfaces nearly 10,000 times faster than quantum chemistry methods, while achieving high accuracy for transition state geometry (averaged root mean square deviation of 0.12 Å) and energy barriers (mean absolute error of 0.60 kcal/mol) on unseen reactions. DeePEST-OS enables practical applications including transition state conformer screening, barrier prediction for retrosynthesis of complex pharmaceuticals, and experimentally validated diastereoselectivity prediction in Diels-Alder reactions. Collectively, these results establish DeePEST-OS as a powerful tool for accelerating reaction kinetics studies in multi-element organic synthesis.