moTSart: accelerating automated transition state search with generative models in a low-data regime

Abstract

Transition states (TSs) are first-order saddle points on the potential energy surface, and thus the highest-energy structure on the (multi-step) minimum-energy reaction pathway that corresponds to a chemical reaction, determining the rate at which it proceeds. Locating TSs and thus elucidating reaction mechanisms is a computationally demanding and expertise-driven task. Recent automation strategies leverage heuristic rules or deep learning models to directly arrive at TS geometries from textual SMILES representations. However, standalone TS predictors lack robustness, whereas heuristic-based approaches lack scalability, limiting their applicability to reliable high-throughput reaction discovery under distribution shifts. Here, we present a generative machine learning framework embedded within a fully automated TS search pipeline and demonstrate accelerated high-throughput screening of bioorthogonal click reactions as a representative case study. The automated framework enables scalable exploration of large reaction spaces and progressively learns, reducing the number of required quantum mechanical evaluations as additional reactions are processed. In our case study, generative refinement decreased the required TS optimization cycles by nearly 40% after training on only 544 reactions. By coupling generative modeling with physics-based validation in a scalable workflow, our work establishes a framework for data-driven reaction mechanism exploration and advances the development of autonomous computational discovery of chemical reactivity.