Structure-free generation of transition state conformations with Bayesian flow networks

Transition state (TS) conformation prediction is crucial for understanding reaction mechanisms and guiding catalyst or enzyme design yet remains challenging due to the intrinsic instability and structural sensitivity of TSs. Conventional determination relies heavily on expert intuition and computationally intensive quantum chemistry methods. Existing 3D-input search approaches are sensitive to initial conformations, while other methods tend to yield high-variance and imprecise results, limiting their practical application. Here, we introduce TSBFN, a structurefree transition state conformation generator based on Bayesian flow networks, directly from reaction SMILES and graphs. TSBFN achieves state-of-the-art performance, outperforming other machine learning methods with a median structural deviation of 0.064 Å and a 93.4% intrinsic reaction coordinate (IRC) validation rate. Notably, TSBFN generates inherently compact conformational distributions that naturally coalesce into a handful of high-quality candidates per reaction, substantially mitigating the challenge of TS selection. Furthermore, the structure-free design demonstrates enhanced robustness and efficacy in challenging multi-component systems. In an out-of-distribution Claisen rearrangement case, TSBFN efficiently generated two distinct classes of mechanistically relevant and valid TSs at millisecond speed, achieving 100% IRC validation and underscoring its potential for automated reaction exploration.