Beyond the Training Domain: Robust Generative Transition State Models for Unseen Chemistry

TL;DR

This work exposes critical generalization gaps in generative TS prediction when extending chemistry beyond small organic molecules, revealing that unseen elements and transition metals induce unphysical geometries and large energy errors. It introduces Transition1x-2p3p4p and Transition1x-TMC benchmarks to stress-test OOD performance and demonstrates that self-supervised conformer-based pretraining substantially improves TS geometry accuracy and data efficiency, lowering median TS RMSD and reducing fine-tuning data needs by up to 75%. The authors further show that semi-empirical methods can bridge to DFT-level fidelity, with reasonable agreement between GFN2-xTB and DFT energetics and successful re-optimization of a substantial fraction of generated TS structures at the DFT level, especially when augmented with DFT-level conformer pseudo-reactions. Collectively, the paper defines a scalable, robust framework for TS prediction across chemically diverse landscapes, combining broad semi-empirical exposure with targeted high-accuracy refinement and setting the stage for exploring catalytically relevant reaction networks with unprecedented breadth.

Abstract

Transition states (TSs) govern the rates and outcomes of chemical reactions, making their accurate prediction a central challenge in computational chemistry. Although recent machine-learning models achieve near chemical accuracy in the prediction of TS structures and the associated reaction barriers for small organic reactions, their ability to generalize beyond the training domain remains largely unexplored. Here, we introduce targeted benchmarks to probe chemical and structural novelty in generative TS prediction. Building on Transition1x, a large-scale dataset of reactions involving small organic molecules, we construct curated extensions incorporating controlled elemental substitutions and diverse transition-metal complexes (TMC). These benchmarks reveal fundamental limitations of generative models in the generalization to previously unseen elements. As a result, they produce unphysical geometries and large energetic errors, even for reactions structurally similar to well-predicted organic systems. To address this challenge, we introduce a self-supervised pretraining strategy based on equilibrium conformers that exposes generative TS models to novel chemical environments prior to targeted fine-tuning. Across the newly proposed benchmarks, self-supervised pretraining substantially improves TS prediction for previously unseen systems, lowering the median RMSD of TS geometries on T1x-TMC reactions from 0.39 to 0.19 $\mathring{A}$ and reducing fine-tuning data requirements by up to 75%, enabling reliable performance even in low-data regimes. Overall, the integration of generative TS models with self-supervised pseudo-reaction pretraining provides an efficient, scalable, and chemically robust framework for elucidating TSs well beyond the small organic molecule domain, establishing a foundation for investigating complex and catalytically relevant reaction landscapes.

Figures (4)

• Figure 1: Overview of data generation workflow for Transition1x-2p3p4p and Transition1x-TMC.a A Transition1x TS structure serves as the starting point for generating both benchmarks. In Transition1x-2p3p4p, a single atom is replaced by an element from the same group up to the third period, producing systematically perturbed variants of the original reaction. In Transition1x-TMC, the source TS is incorporated as an organic ligand into one of ten catalytically relevant TMCs. All modified geometries are refined to valid TS structures using the P-RFO algorithm, and corresponding reactants and products are obtained via IRC calculations. All structures are computed at the GFN2-xTB level of theory. b Histogram of element frequencies, showing the number of reactions in which each element appears, colored according to the dataset (Transition1x-2p3p4p or Transition1x-TMC).
• Figure 2: Failure modes of the vanilla React-OT model on reactions involving novel chemistry.a Mean bond distance mean absolute error (DMAE) of generated TS structures as a function of the number of previously unseen atom types, showing a rapid increase in error with growing elemental novelty. Transition1x-2p3p4p introduces one novel atom type, while Transition1x-TMC introduces up to two. b Cumulative probability distributions of structural and energetic errors for the Transition1x-2p3p4p benchmark, illustrating the sharp performance degradation introduced by single-atom substitutions. c Ground-truth (solid) and predicted (dashed) bond-length distributions for representative C–C, C–Si, and C–Ge bonds, highlighting systematic geometric distortions in the model’s predictions.
• Figure 3: Conformer-based pseudo-reaction pretraining.a Construction of pseudo-reactions from sets of conformers. The conformer with the second-highest energy is assigned as the reactant, the highest-energy conformer as the TS, and the lowest-energy conformer as the product. b Effect of self-supervised pretraining using 2500 (Transition1x-2p3p4p and Transition1x-TMC) and 100 (Pt catalyzed) pseudo-reactions constructed from equilibrium minima of the corresponding datasets on the median performance of vanilla and fine-tuned React-OT models. The fine-tuned results are obtained using reduced training sets comprising 10% of Transition1x-2p3p4p and Transition1x-TMC and 50 reactions for the Pt-catalyzed benchmark. Transparent bars indicate models trained without pretraining, while solid bars denote models incorporating self-supervised pretraining.
• Figure 4: Validation at the DFT level of theory.a Absolute difference between energetic error evaluated at GFN2-xTB and DFT level of theory. Errors increase at the DFT level for all datasets, with the effect more pronounced for Transition1x-TMC. b Cumulative probability distributions of structural and energetic errors between the generated samples and their corresponding reference DFT TS. c Success rate of converging to the desired reference DFT TS structure after DFT optimization of the generated samples. A successful convergence is defined as achieving an RMSD of 0.05 Å or lower.