MIT Model Predicts Chemical Reaction Turning Points in Under a Second

Chemists designing new chemical reactions often seek to understand the reaction’s transition state—the critical juncture beyond which the reaction is destined to proceed. Knowing this allows them to optimize conditions for the desired outcome. Traditionally, predicting these transition states has been computationally intensive and complex.

Now, researchers at MIT have introduced a machine-learning model capable of predicting these crucial states in under a second, with impressive accuracy. This innovation promises to streamline the design of chemical reactions for creating a wide array of useful compounds, from pharmaceuticals to fuels.

“We’d like to ultimately design processes to take abundant natural resources and turn them into molecules that we need, such as materials and therapeutic drugs. Computational chemistry is really important for figuring out how to design more sustainable processes to get us from reactants to products,” explains Heather Kulik, the Lammot du Pont Professor of Chemical Engineering and a professor of chemistry at MIT, and the senior author of the study.

The research, led by Chenru Duan PhD ’22 (now at Deep Principle), Guan-Horng Liu (formerly of Georgia Tech, now at Meta), and Yuanqi Du (Cornell University), is detailed in a paper published in Nature Machine Intelligence.

A Leap in Estimation

Transition states occur at the energy threshold necessary for a reaction to proceed. They are so brief that direct experimental observation is virtually impossible. Computational techniques based on quantum chemistry offer an alternative, but these methods demand significant computing power and can take hours, or even days, to compute a single transition state.

“Ideally, we’d like to be able to use computational chemistry to design more sustainable processes, but this computation in itself is a huge use of energy and resources in finding these transition states,” Kulik notes.

The team’s earlier work in 2023 presented a machine-learning approach to predict reaction transition states. While faster than quantum chemistry techniques, it still required the model to generate approximately 40 structures and assess them via a “confidence model.”

The new model, React-OT, significantly improves upon this by starting with a linear interpolation estimate—positioning each atom halfway between its reactant and product locations. “A linear guess is a good starting point for approximating where that transition state will end up,” Kulik explains. “What the model’s doing is starting from a much better initial guess than just a completely random guess, as in the prior work.”

This enhanced starting point dramatically reduces the steps and time needed for prediction. The study demonstrates that React-OT achieves accurate predictions in about 0.4 seconds with only five steps. These predictions are also about 25 percent more accurate than those of the previous model and do not require a confidence model.

“That really makes React-OT a practical model that we can directly integrate to the existing computational workflow in high-throughput screening to generate optimal transition state structures,” Duan says.

Broad Applicability

React-OT was trained using a dataset of 9,000 chemical reactions with reactant, product, and transition state structures calculated via quantum chemistry. After training, it showed strong performance on held-out reactions from the same set and generalized well to other reaction types, including those with larger reactants.

“This is important because there are a lot of polymerization reactions where you have a big macromolecule, but the reaction is occurring in just one part. Having a model that generalizes across different system sizes means that it can tackle a wide array of chemistry,” Kulik says.

The researchers are now expanding the model’s capabilities to predict transition states for reactions involving elements like sulfur, phosphorus, chlorine, silicon, and lithium.

Markus Reiher, a professor of theoretical chemistry at ETH Zurich, commented, “To quickly predict transition state structures is key to all chemical understanding. The new approach presented in the paper could very much accelerate our search and optimization processes, bringing us faster to our final result. As a consequence, also less energy will be consumed in these high-performance computing campaigns. Any progress that accelerates this optimization benefits all sorts of computational chemical research.”

The MIT team has also created an app to facilitate the use of their model by other scientists in designing their own reactions. This is available here.

“Whenever you have a reactant and product, you can put them into the model and it will generate the transition state, from which you can estimate the energy barrier of your intended reaction, and see how likely it is to occur,” Duan says.

The research was supported by funding from the U.S. Army Research Office, the U.S. Department of Defense Basic Research Office, the U.S. Air Force Office of Scientific Research, the National Science Foundation, and the U.S. Office of Naval Research.