Chemical reactions are decided at a fleeting point that molecules occupy only for an instant: the transition state. It is the high-energy configuration through which reactants must pass before becoming products, and its structure helps determine not only how fast a reaction proceeds but also which chemical pathway wins when several outcomes are possible. Predicting this elusive arrangement has therefore become one of the central challenges in computational chemistry. A new study now suggests that artificial intelligence can be remarkably effective at generating transition-state structures—provided it has learned enough about the chemical world beyond the reactions it was originally shown.
The work, published in Nature Computational Science, examines a problem that has become increasingly important as machine-learning models move from laboratory demonstrations toward practical chemical discovery. Recent generative models can approach near-chemical accuracy when predicting transition states and reaction barriers for small organic reactions. Yet strong performance on familiar chemistry does not necessarily mean that a model understands chemistry in a transferable way. When an unfamiliar element, bonding environment, or transition-metal complex appears, a model may be forced to extrapolate far beyond its training experience, and its predictions can deteriorate sharply.
To investigate this boundary, the researchers developed targeted benchmarks based on Transition1x, a large-scale collection of reactions involving small organic molecules. Rather than measuring performance only on randomly selected test reactions, the new benchmarks were designed to expose specific kinds of novelty. They include controlled substitutions of chemical elements and expanded reaction systems containing diverse transition metal complexes, or TMCs. This strategy allows the researchers to separate ordinary prediction from genuine generalization: a model is not merely asked to recognize a familiar reaction pattern, but to construct a plausible transition state in a chemical environment it has not previously encountered.
The distinction matters because transition states are defined by delicate geometric and electronic relationships. At this point on a reaction pathway, bonds may be partly formed and partly broken, while atoms rearrange in ways that are often highly sensitive to their surroundings. Replacing one element with another can alter atomic size, electronegativity, preferred coordination number, and the distribution of electrons across the reacting system. Transition metals add another layer of complexity. Their partially filled d orbitals can support multiple oxidation states, coordination geometries, and spin configurations, making their catalytic behavior difficult to represent through simple structural analogy.
Generative transition-state models attempt to predict this complex arrangement directly from information about the reactants and products. In effect, they learn a distribution of chemically plausible structures rather than relying exclusively on a conventional step-by-step quantum-mechanical search. This can make them dramatically faster, a crucial advantage when screening many possible reactions. But speed does not remove the need for chemical knowledge. If a model has mostly seen main-group organic chemistry, it may generate geometries that look reasonable in a broad structural sense while missing the coordination patterns or bond distances required by an unfamiliar metal center.
The new benchmarks reveal precisely this weakness. According to the study, generative models trained without exposure to the relevant novel systems show fundamental limitations when asked to predict transition states involving previously unseen elements. Their difficulties become especially apparent for reactions containing transition metal complexes, where the chemical environment differs substantially from the small organic systems represented in much of the original training data. The result is a warning for the rapidly expanding field of AI-driven chemistry: a model can achieve excellent average accuracy while remaining unreliable at the frontier of chemical space, where the most valuable discoveries may occur.
The researchers address this challenge with a self-supervised pretraining strategy built around equilibrium conformers. These are stable or near-stable molecular geometries that can be generated or collected without requiring a labeled transition-state example for every chemical environment. During pretraining, the model is exposed to the shapes, local interactions, and coordination patterns found in equilibrium molecular structures. It is not directly told the answer to a particular reaction barrier or transition-state problem. Instead, it learns a broader representation of how atoms arrange themselves across diverse chemical settings before being fine-tuned for the more specialized task of transition-state generation.
This approach resembles giving a language model a broad vocabulary and knowledge of grammar before asking it to write about a specialized subject. For chemistry, the benefit is that the model can encounter unfamiliar elements and structural motifs before it must predict the high-energy configuration of a reaction. Equilibrium conformers do not contain all the information needed to determine a transition state, but they can provide a foundation for recognizing chemically meaningful distances, angles, coordination environments, and patterns of molecular flexibility. Fine-tuning can then adapt that foundation to the particular geometry and energetics of reactions.
Across the proposed tests, self-supervised pretraining substantially improved predictions for systems that were previously unseen. The study reports lower median root-mean-square deviation, or RMSD, for predicted transition-state geometries on the Transition1x-TMC benchmark. RMSD is a standard measure of structural error that compares corresponding atomic positions between a predicted geometry and a reference structure; lower values indicate closer agreement. The improvement is significant not merely because it raises an average score, but because it addresses a difficult form of extrapolation involving transition-metal chemistry. The researchers also find that pretraining reduces the amount of reaction-specific data needed during fine-tuning, allowing reliable performance in low-data regimes.
That finding could have broad implications for computational reaction design. High-quality transition-state calculations are often expensive, and experimental data for unusual catalytic systems can be limited. A model that can absorb general chemical knowledge from unlabeled molecular structures may reduce the burden of assembling enormous reaction datasets for every new element or catalytic family. It could help researchers prioritize candidate mechanisms, estimate which pathways deserve detailed quantum-chemical investigation, and explore areas of chemical space where conventional data-driven methods have little direct experience. The work does not eliminate the need for physics-based validation, but it points toward a more efficient partnership between machine learning and computational chemistry.
The study also highlights why future evaluations of scientific AI must be designed around meaningful novelty rather than random data splits alone. If nearly identical chemical systems appear in training and testing sets, a model may seem general while actually relying on memorized patterns or close structural neighbors. Controlled elemental substitutions and transition-metal benchmarks offer a more demanding test of whether a system has learned transferable chemical representations. By showing that equilibrium-based self-supervision can improve performance on these challenges, the researchers present a possible route toward models that are not only accurate on familiar reactions but more resilient when chemistry takes an unexpected turn. For AI-driven molecular discovery, that ability to confront the unknown may prove more important than another incremental gain on routine examples.
Subject of Research: Generative machine-learning models for predicting chemical transition states, with a focus on generalization to unseen elements and transition metal complexes.
Article Title: Robust generative transition-state models for unseen chemistry
Article References: Darouich, S., Toney, J.W., Luo, W. et al. Robust generative transition-state models for unseen chemistry. Nat Comput Sci (2026). https://doi.org/10.1038/s43588-026-01034-5
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s43588-026-01034-5
Keywords: transition states, generative models, machine learning, computational chemistry, chemical reaction prediction, Transition1x, transition metal complexes, self-supervised pretraining, equilibrium conformers, chemical generalization
Tags: artificial intelligence in computational chemistrybenchmarking AI models for chemical transition stateschallenges of extrapolation in chemical modelinggenerative models for high-energy molecular configurationsgenerative models for transition state predictionlimitations of machine learning in complex chemical environmentsmachine learning for chemical reaction pathwaysneural networks for predicting reaction outcomespredicting transition states in unseen reactionsreaction barrier prediction using deep learningrole of AI in understanding reaction mechanismstransferability of AI models in chemistry
