When chemists design new chemical reactions, useful information implies the state of transition from the reaction – the point of no return from which a reaction must proceed.
This information allows chemists to try to produce the right conditions that will allow the desired reaction to occur. However, the current methods to predict the transition state and the path that a chemical reaction will take is complicated and requires a huge amount of computing power.
MIT researchers have now developed an automatic learning model that can make these predictions in less than a second, with great precision. Their model could allow chemists to more easily design chemical reactions that could generate a variety of useful compounds, such as pharmaceuticals or fuels.
“We would like to be able to design processes to take abundant natural resources and transform them into molecules we need, such as materials and therapeutic drugs. Computer chemistry is really important to understand how to design more sustainable processes to make us come from reagents to products ”
The former student graduate of MIT Chenru Duan Phd '22, who is now in depth; The former student graduate of Georgia Tech, Guan-Horng Liu, who is now in Meta; and the student graduated from Cornell Yuanqi University of are the main authors of the article, which appears today in Nature machine intelligence.
Best estimates
For a given chemical reaction to occur, it must go through a transitional state, which takes place when it reaches the energy threshold necessary for the continuous reaction. These transition states are so ephemeral that they are almost impossible to observe experimentally.
As an alternative, researchers can calculate the structures of transition states using techniques based on quantum chemistry. However, this process requires a lot of computing power and can take hours or days to calculate a single transition state.
“Ideally, we would like to be able to use computer chemistry to design more sustainable processes, but this calculation in itself is a huge use of energy and resources to find these transition states,” explains Kulik.
In 2025, Kulik, Duan and others reported On an automatic learning strategy they have developed to predict the transitional transition states. This strategy is faster than using quantum chemistry techniques, but always slower than what would be ideal because it requires that the model generates around 40 structures, then execute these predictions via a “model of confidence” to predict which states were most likely to occur.
One of the reasons why this model must be executed so many times is that it uses assumptions generated randomly for the starting point of the transition state structure, then performs dozens of calculations until it reaches its best estimate. These departure points generated at random can be very far from the real transition state, which is why so many steps are necessary.
The new model of researchers, React-Ot, described in the Nature machine intelligence paper, uses a different strategy. In this work, the researchers trained their model to start from an estimate of the transition state generated by linear interpolation – a technique that considers the position of each atom by moving it halfway between its position in the reagents and in the products, in a three -dimensional space.
“A linear supposition is a good starting point to get closer to the place where this transition state will end,” says Kulik. “What the model is doing is starting with a much better initial supposition than a simple random supposition, as in previous work.”
For this reason, less steps are needed and less time to generate a prediction. In the new study, researchers have shown that their model could make predictions with only about five stages, taking approximately 0.4 seconds. These predictions do not need to be fed by a model of confidence, and they are about 25% more precise than the predictions generated by the previous model.
“It really makes a react-ot a practical model that we can integrate directly into the workflow of calculation existing in high speed screening to generate optimal transition state structures,” explains Duan.
“A wide range of chemistry”
To create React-Ot, the researchers trained it on the same set of data they used to form their old model. These data contain structures of reagents, products and transition states, calculated using quantum chemistry methods, for 9,000 different chemical reactions, mainly involving small or inorganic inorganic or inorganic molecules.
Once formed, the model has performed well on other reactions from this set, which had been retained from training data. He also worked well on other types of reactions on which he had not been formed and could make specific predictions involving reactions with larger reagents, which often have side chains which are not directly involved in the reaction.
“This is important because there are a lot of polymerization reactions where you have a large macromolecule, but the reaction occurs in one part. Having a model that is generalized on different system sizes means that it can approach a wide range of chemistry, “explains Kulik.
Researchers are now working on the formation of the model so that it can predict transition states for reactions between molecules which include additional elements, including sulfur, phosphorus, chlorine, silicon and lithium.
“Quickly preaching the transition state structures is the key to any chemical understanding,” explains Markus Reiher, professor of theoretical chemistry at Eth Zurich, who was not involved in the study. “The new approach presented in the document could greatly accelerate our research and optimization processes, bringing us more quickly to our final result. Consequently, it will also be less consumed in these high performance computer campaigns. Any progress that accelerates this optimization benefits all kinds of computer chemical research. ”.
The MIT team hopes that other scientists will use their approach to design their own reactions and have created a Application for this purpose.
“Each time you have a reagent and a product, you can put them in the model and this will generate the transition state, from which you can estimate the energy barrier of your planned reaction, and see its probability of performing,” explains Duan.
The research was funded by the US Army Research Office, the US Department of Defense Basic Office, the US Air Force of Scientific Research, the National Science Foundation and the US Office of Naval Research.
