Answers are grounded in this paper and cite their sources.
Machine-learning models for transition-state (TS) geometry prediction are almost universally benchmarked by the root-mean-square deviation (RMSD) of atomic positions, or by the mean absolute error of the interatomic-distance matrix (D-MAE), relative to a reference TS. We show that these Euclidean measures are unreliable proxies for the quantity a practitioner actually needs: whether a predicted geometry optimizes to a genuine TS at a target level of density functional theory. We optimized 520 interpolation-based guesses for 260 reactions from the Transition1x database [Schreiner M., Sci. Data 2022, 9, 779] at the r2SCAN-3c level. Every geometry-based quality measure–Kabsch-aligned RMSD, D-MAE, a reaction-coordinate projection, and cheap semi-empirical descriptors–saturates in a narrow band of the area under the receiver-operating-characteristic curve (AUC), 0.77–0.84, for predicting optimization success, and 18% of guesses within 0.28 Å of the reference fail nonetheless.
No takes yet. Share an insight, caveat, or question.
Kaplan et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: