New model predicts a chemical reaction’s point of no return

Chemists striving to engineer novel chemical reactions often grapple with identifying the elusive transition state – the critical juncture from which a reaction is committed to proceeding. This vital piece of information is key to establishing the optimal conditions for desired chemical transformations. However, conventional methods for predicting these transition states and the subsequent reaction pathways are notoriously complex and demand immense computational resources, often taking days to yield a single result.

In a groundbreaking development, researchers at MIT have unveiled a new machine-learning model that promises to revolutionize this field. This innovative model can make these crucial predictions with high accuracy in less than a second. This exponential leap in speed and efficiency could significantly streamline the process for chemists to design a myriad of useful compounds, from life-saving pharmaceuticals to clean, sustainable fuels.

Heather Kulik, the Lammot du Pont Professor of Chemical Engineering and a professor of chemistry, and the senior author of the new study, emphasizes the broader impact: “We’d like to be able 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.”

The pioneering work was led by former MIT graduate student Chenru Duan PhD ’22 (now at Deep Principle), former Georgia Tech graduate student Guan-Horng Liu (now at Meta), and Cornell University graduate student Yuanqi Du. Their findings are detailed in a paper published recently in Nature Machine Intelligence.

Better Estimates with React-OT

Every chemical reaction must pass through a transition state – a fleeting, high-energy arrangement of atoms that is practically impossible to observe through experimentation. While quantum chemistry techniques can calculate these structures, the computational demands are prohibitive, often requiring hours or even days for a single transition state calculation.

“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, highlighting the paradox of resource-intensive computation for sustainable outcomes.

In 2023, Kulik and Duan, along with their colleagues, introduced an earlier machine-learning strategy for predicting transition states. While an improvement over quantum chemistry, that model still faced limitations due to its reliance on generating approximately 40 structures and then processing them through a “confidence model” to identify the most probable states. This iterative process was necessary because the model started with randomly generated guesses for the transition state structure, often far from the actual outcome, thus requiring numerous steps to converge.

The researchers’ new model, dubbed React-OT, represents a significant leap forward. Unlike its predecessor, React-OT employs a more refined strategy: it begins its predictions from an estimated transition state generated through linear interpolation. This technique shrewdly estimates each atom’s position by placing it halfway between its configuration in the reactants and in the products within a three-dimensional space.

“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 intelligent initialization dramatically reduces the computational burden. React-OT can now make highly accurate predictions in just about five steps, typically taking only 0.4 seconds, and crucially, it no longer requires an additional confidence model. These predictions also boast a remarkable 25 percent improvement in accuracy over the previous model.

Chenru Duan affirms the practical implications: “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.”

“A Wide Array of Chemistry”

React-OT was trained on the same extensive dataset used for the older model, comprising structures of reactants, products, and transition states for 9,000 diverse chemical reactions, primarily involving small organic or inorganic molecules. The model demonstrated robust performance on unseen reactions from this dataset and, impressively, generalized well to reaction types it had not been specifically trained on. This includes accurate predictions for reactions involving larger reactants, even those with side chains not directly participating in the core reaction.

“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 states, highlighting the model’s versatility across different chemical complexities.

The team is now working to further enhance React-OT, expanding its capabilities to predict transition states for molecules containing additional 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, commends the innovation: “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.”

In a move to foster broader scientific advancement, the MIT team has made their approach accessible through a dedicated app, enabling other researchers to leverage React-OT in their own reaction design efforts.

“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 explains, simplifying the powerful potential for real-world application.

This pioneering research received 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, underscoring its strategic importance.