TransTS: AI Framework for Generalizable Transition State Generation

A novel machine-learning framework named TransTS has been developed to create transition-state (TS) structures for chemical reactions. Transition states are crucial as they represent the energetic barriers and pathways of fundamental chemical reactions, yet identifying them is computationally intensive due to the requirement for costly quantum-mechanical calculations. While traditional saddle-point searches are time-consuming, recent machine-learning techniques have improved TS generation by predicting structures based on reaction endpoint data. However, these approaches mainly focus on geometric relationships between endpoints and TSs, neglecting the underlying structural transformations. TransTS overcomes this by explicitly learning atom-level transformations and merging them with a cohesive atom-aligned geometric representation of reactants, TSs, and products. This allows for reaction-aware generalization in TS generation. The framework is applicable to various chemical reactions, utilizing atom-mapped reactant-product pairs. The research is documented on arXiv with the identifier 2608.14076 and an abstract type of 'cross'. Its significance lies in potentially lowering the computational costs associated with identifying transition states, thereby enhancing computational chemistry and drug discovery.