Every chemical reaction passes through a fleeting state that lasts only quadrillionths of a second. These ghostly molecular structures, called transition states, hold the key to understanding how reactions happen and designing better catalysts. Yet they are nearly impossible to capture in the lab. Now, researchers have developed a machine learning approach that can predict these elusive structures with remarkable accuracy in less than half a second.

The breakthrough, described in a study published in Nature Machine Intelligence, uses a mathematical framework called optimal transport to generate transition states from nothing more than the starting and ending molecules of a reaction. The method, called React-OT, is orders of magnitude faster than current computational chemistry approaches and more accurate than previous artificial intelligence attempts. The implications reach across drug discovery, materials science, and industrial chemistry, where the speed of identifying reaction pathways could accelerate everything from developing new medicines to designing more efficient catalysts.

The Problem That Slowed Chemistry Down

For decades, chemists have relied on quantum mechanical calculations to find transition states. These are configurations where bonds are stretched to the breaking point but haven't yet snapped, or where new bonds are forming but aren't yet complete. Transition states sit at the peak of an energy hill between reactants and products on what physicists call the potential energy surface.

Why do they matter so much? They govern reaction rates. The higher the energy barrier at the transition state, the slower the reaction. Understanding transition states is essential for predicting reaction mechanisms, designing catalysts that lower these barriers, and even controlling which products form from a given starting material.

The problem is that transition states are transient. They exist for perhaps a femtosecond, a millionth of a billionth of a second. They can't be isolated and measured in a test tube. Instead, chemists must calculate them using density functional theory, a quantum mechanical method that is computationally expensive. For complex reaction networks involving hundreds or thousands of reactions, finding all the transition states can require millions of individual quantum chemistry calculations. On a typical computer, each transition state search costs hours or even days.

This computational bottleneck has limited the scope of what chemists can explore. Building comprehensive reaction networks for complex chemical systems becomes impractical when each reaction requires so much computing power.

A New Approach: Learning From Data

Over the past few years, machine learning has begun to crack this problem. Neural networks trained on databases of known reactions can predict transition state structures directly, without the expensive quantum chemistry. A previous approach called OA-ReactDiff used a diffusion model, similar to technology behind image generation tools, to learn the joint distribution of reactants, transition states, and products.

But diffusion models have a weakness for this application. They generate structures through a stochastic process, meaning they include randomness. For a given pair of starting and ending molecules, multiple runs could yield different transition state predictions. That requires running the model many times and then ranking the results to pick the best one. This defeated the purpose of making the method fast.

The new React-OT approach takes a different path. Instead of diffusion, it uses optimal transport, a branch of mathematics developed centuries ago to solve the most efficient way to move material from one location to another. Modern versions of optimal transport theory have been adapted to machine learning problems, but applying it to chemical reactions required new insights.

The key innovation is reformulating the transition state search as a deterministic journey. Rather than starting from random noise and gradually denoising toward a transition state structure, React-OT begins with a reasonable initial guess: the simple linear interpolation between the starting and ending molecules. From there, an AI model learns to move atoms smoothly along an optimal path from this starting guess toward the true transition state.

The result is fully deterministic. Given the same starting and ending molecules, React-OT always produces the same transition state prediction. No need to run it dozens of times. One inference produces the final answer.

Accuracy and Speed

When tested on a dataset of over 1,000 organic reactions, React-OT achieved a median structural accuracy of 0.053 Angstroms, a measure of how close the predicted atoms are to their true positions. That's roughly half the error of previous AI approaches. For predicting barrier heights, the energy difference between reactants and the transition state, React-OT achieved a median error of 1.06 kilocalories per mole. Over 64% of predictions reached chemical accuracy, a threshold where errors are smaller than typical experimental variations.

And speed matters. React-OT generates these predictions in 0.4 seconds per reaction. That is roughly 20 times faster than the previous best AI method and nearly 1,000 times faster when accounting for the cost of running that method multiple times.

To put this in perspective: conventional quantum chemistry takes 12.8 hours on average to find a single transition state. React-OT completes 30,000 reactions in the time traditional methods take to handle one.

Making the Technology More Practical

The researchers didn't stop with lab benchmarks. They tested whether the method worked when fed lower quality input data. In real workflows, chemists often begin with reactant and product geometries optimized using cheaper, less accurate quantum chemistry methods rather than the high-precision calculations used for training.

React-OT remained robust. Even when the input structures were optimized with a semi-empirical method called GFN2-xTB, which is roughly 1,000 times cheaper than full quantum chemistry, the AI model still predicted transition states at comparable accuracy. This opens a path to truly practical acceleration: use cheap calculations to optimize reactants and products, feed them to React-OT, and skip the expensive quantum chemistry entirely for many reactions.

The researchers also trained the model on a much larger dataset of reactions calculated at the lower level of theory. This pretraining, followed by fine-tuning on higher accuracy data, improved performance by 25%. It suggests that the approach scales well as more reaction data becomes available.

Integration With Human Workflows

Perhaps most importantly, the team integrated React-OT into realistic computational workflows used by chemists. Rather than replacing quantum chemistry entirely, they developed a hybrid approach. React-OT generates a transition state prediction. A separate confidence model assesses how reliable that prediction is. If confidence is high, the structure is accepted. If confidence is low, a full quantum chemistry calculation takes over.

Using this workflow, 86% of transition states were predicted by React-OT to chemical accuracy while only 14% required expensive quantum calculations. The result was a sevenfold acceleration in overall speed compared to pure quantum chemistry, while maintaining accuracy. This balance of speed and rigor makes the method genuinely useful for chemists planning reactions, discovering new molecules, or optimizing catalysts.

Why This Matters

The implications ripple across chemistry and materials science. Researchers studying complex reaction networks, such as combustion chemistry, atmospheric chemistry, or the decomposition of pollutants, have always been constrained by computational cost. With React-OT, they can now explore these networks more comprehensively. Drug discovery could benefit from faster prediction of reaction pathways. Catalysis research could accelerate by enabling rapid screening of potential reaction mechanisms.

The method is limited to neutral organic molecules containing carbon, nitrogen, oxygen, and hydrogen. Reactions involving metals or charged species remain beyond current scope. But as datasets grow and the method is refined, these limitations will likely fade.

The underlying mathematical approach, optimal transport, is not specific to chemistry. The framework could eventually apply to protein-ligand binding, molecular adsorption on surfaces, or materials undergoing phase transitions. The authors hint at these possibilities, suggesting their work is opening a broader toolkit for modeling molecular transformations.

A Glimpse of Computational Chemistry's Future

React-OT represents a shift in how computational chemistry approaches difficult problems. Rather than building ever more powerful quantum computers or designing smarter optimization algorithms, this work trains AI on existing data to learn the patterns of chemical transformation. It combines machine learning efficiency with domain specific understanding of chemistry. The result is a tool that is both faster and more intelligent than what came before.

The researchers have released the code and trained models publicly, lowering the barrier for other scientists to use the method. As transition state prediction becomes faster and cheaper, chemists can ask bigger questions about reaction networks they've never had time to explore. In chemistry, as in so many fields, speed is opening new territory.

Credit & Disclaimer: This article is a popular science summary written to make peer-reviewed research accessible to a broad audience. All scientific facts, findings, and conclusions presented here are drawn directly and accurately from the original research paper. Readers are strongly encouraged to consult the full research article for complete data, methodologies, and scientific detail. The article can be accessed through https://doi.org/10.1038/s42256-025-01010-0