MIT Model Predicts Chemical Reaction’s Transition State in Under a Second

Chemists designing new chemical reactions often rely on understanding the reaction’s transition state—the critical juncture beyond which the reaction is destined to proceed. Identifying this point is vital for creating optimal conditions to facilitate the desired chemical transformation. However, conventional methods for predicting transition states and reaction pathways are computationally intensive and complex.

Now, researchers at MIT have introduced a machine-learning model capable of predicting these transition states with remarkable speed and accuracy—in under a second. This innovation promises to streamline the design of chemical reactions for producing a variety of valuable compounds, including pharmaceuticals and fuels.

“We aim to design processes that transform abundant natural resources into essential molecules like therapeutic drugs and materials. Computational chemistry plays a crucial role in devising more sustainable methods for converting reactants into products,” explains Heather Kulik, the Lammot du Pont Professor of Chemical Engineering and a professor of chemistry at MIT, who is also the senior author of the study.

The research, featured in Nature Machine Intelligence, was led by Chenru Duan PhD ’22 (now at Deep Principle), Guan-Horng Liu (formerly of Georgia Tech, now at Meta), and Yuanqi Du, a graduate student at Cornell University. [Source: Nature Machine Intelligence]

Enhanced Prediction Capabilities

Every chemical reaction must pass through a transition state, which occurs at the energy threshold required for the reaction to occur. These states are so transient that direct experimental observation is nearly impossible.

Computational techniques rooted in quantum chemistry offer an alternative for calculating transition state structures. However, these methods are resource-intensive and can take hours or even days to compute a single transition state.

“Ideally, we would leverage computational chemistry to design more sustainable processes. However, the computational cost of finding these transition states is itself a significant drain on energy and resources,” Kulik notes.

Building on earlier work from 2023, Kulik, Duan, and their colleagues developed a machine-learning strategy to predict reaction transition states. While faster than quantum chemistry techniques, this earlier model still required substantial computation, generating approximately 40 structures and evaluating them with a “confidence model.”

The need for multiple iterations arose from the model’s reliance on randomly generated starting points for the transition state structure. These random initial guesses often deviated significantly from the actual transition state, necessitating numerous calculations to refine the prediction.

The new model, React-OT, adopts a different approach, using linear interpolation to estimate the initial transition state. This technique positions each atom halfway between its location in the reactants and products.

“A linear guess offers a solid starting point for approximating the transition state,” Kulik explains. “The model benefits from a far better initial estimate than the completely random guesses used previously.”

React-OT requires fewer steps and less time to generate a prediction. The study demonstrates that the model can produce predictions in approximately 0.4 seconds with only five steps. Furthermore, these predictions are about 25% more accurate than those from the previous model and do not require a confidence model.

“React-OT is a practical model that can be seamlessly integrated into existing high-throughput computational workflows for generating optimal transition state structures,” Duan states.

Broad Applicability

The researchers trained React-OT on the same dataset used for their earlier model, comprising structures of reactants, products, and transition states for 9,000 chemical reactions. Once trained, the model demonstrated robust performance on withheld reactions and generalized effectively to new types of reactions, including those involving larger reactants with non-participating side chains.

“This capability is particularly important for polymerization reactions, where the reaction occurs at only one part of a large macromolecule. A model that generalizes across different system sizes can address a wide variety of chemistry,” Kulik emphasizes.

The team is now focused on expanding the model’s training to predict transition states for reactions involving elements such as sulfur, phosphorus, chlorine, silicon, and lithium.

Markus Reiher, a professor of theoretical chemistry at ETH Zurich, who was not involved in the study, commented, “Fast prediction of transition state structures is essential for chemical understanding. This new approach could greatly accelerate our search and optimization processes, leading us more quickly to our final results. This acceleration will also reduce the energy consumed in high-performance computing campaigns. Any progress that accelerates this optimization benefits all areas of computational chemical research.”

The MIT team encourages other scientists to use their approach in designing their own reactions and have made an app available for this purpose.

“By inputting a reactant and product, the model generates the transition state, which allows estimation of the energy barrier and likelihood of the reaction occurring,” Duan explains.

Funding for the research was provided by 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.