2026/04/06 by Filipp Nikitin, Dylan M. Anstine, Olexandr Isayev · 1 voice
Chemistry · Materials Science · Physics and Astronomy · #Advanced Chemical Physics Studies #Cyclization and Aryne Chemistry #Machine Learning in Materials Science
paper · pdf · doi:10.26434/chemrxiv.15001681/v1
openalex publication_date 2026/04/06 · openalex created_date 2026/04/08 · openalex updated_date 2026/07/30
Determination of transition state structures is fundamental to the mechanistic understanding of chemical reactivity and selectivity, yet conventional approaches require extensive expert intervention for geometry preparation and saddle-point optimization, severely limiting the throughput of computational reaction analysis. We present RitS, a generative model trained on a curated dataset of approximately 2 million transition states computed at the GFN2-xTB level of theory. RitS generates three-dimensional transition-state geometries directly from reactant–product bond connectivity without requiring pre-optimized molecular structures as input. The model accommodates H, B, C, N, O, F, Si, P, S, Cl, Br, I elements, systems containing up to 51 total atoms, and both neutral and charged species, substantially extending the chemical scope of prior generative approaches. Benchmarking on the Transition1x dataset demonstrates that RitS achieves the lowest median root-mean-square deviation among current methods. Cross-dataset evaluation yields intrinsic reaction coordinate validation rates exceeding 90% from a single generated structure, rising to approximately 99% with three independent samples. A key capability of RitS is stereochemistry-aware generation: encoding product chirality in the input graph enables selective generation of competing diastereomeric transition states, as demonstrated for endo versus exo Diels–Alder cycloadditions and E versus Z elimination pathways. The model further generalizes to the multistep Hajos–Parrish–Eder–Sauer–Wiechert organocatalytic cycle, capturing each elementary step including the enantiodetermining intramolecular aldol transition state. These results establish RitS as a tool for automated, high-throughput transition-state generation applicable to mechanistic studies, selectivity analysis, and large-scale reaction network exploration.