An artificial intelligence (AI) that can explain the principles beyond predicting the results of chemical reactions has emerged. This has been regarded as a notable achievement in both academia and the pharmaceutical industry.
Professor Joonyoung F. Joung of the Department of Applied Chemistry at Kookmin University and his research partner, Connor W. Coley, a professor at the Massachusetts Institute of Technology (MIT), announced in the international journal Nature on the 21st that they have developed a next-generation AI called "FlowER" that predicts the process by which chemical reactions occur.
Synthetic materials, ranging from pharmaceuticals to batteries and solar cells, are all created through complex chemical reactions. However, during the synthesis process, only the desired substances may not be produced, and unexpected byproducts or impurities may be generated. Due to these variables, developing new drugs or new materials takes a long time and is prone to failure. This is why developing AI that can predict reaction pathways has been a long-standing wish for researchers.
However, existing AIs focused on predicting products by inputting reactants as if translating a language. As a result, errors occurred where nonexistent atoms suddenly appeared or necessary atoms disappeared. This violates the "law of conservation of mass," which states that the types and numbers of atoms do not change before and after a reaction.
Professor Joung noted, "Existing AI models sometimes produced physically nonsensical results, making it difficult for chemists to trust them," adding, "I thought that if we reflected the basic laws of chemistry in AI, we could solve the problem."
The researchers developed the AI to follow the thought processes of chemists. Just as chemists indicate how electrons move with arrows to understand the reaction process, the AI also explains reactions centering on electron flow. It is designed to adhere strictly to the law of conservation of mass, and approximately 1 million cases of reaction data were used for training, greatly enhancing prediction reliability.
FlowER predicted not only the final products but also byproducts and new reaction combinations that frequently occur in experiments. In particular, while existing AIs needed thousands to tens of thousands of data points to learn new types of reactions, FlowER accurately predicted previously unseen reactions with over 65% accuracy with just 32 cases. Professor Joung explained, "This is similar to how a person quickly understands new rules based on a small amount of experience."
The research findings have great potential for use in the pharmaceutical industry. Predicting and determining the structure of impurities beforehand is closely related to the safety of new drug development. The researchers collaborated with global pharmaceutical companies like Denmark's Novo Nordisk, known for the obesity treatment Wegovy, Switzerland's Novartis, and the United States' Pfizer and Merck (MSD) through the MLPDS consortium.
Professor Joung stated, "The AI model we developed can predict the structure of impurities and inform what conditions they arise under," adding, "It can even suggest ways to eliminate impurities by changing the reaction conditions."
The idea for this research naturally emerged from Professor Joung's experiences. While pursuing his doctorate at Korea University, he self-studied AI in addition to conducting chemical experiments. The knowledge he studied through books and YouTube led to his current research on solving chemical reactions with AI.
Professor Joung mentioned, "At first, it was not easy to follow the coding, but during my postdoctoral research at MIT, I developed ideas while interacting with various majors like computer science, chemical engineering, and biotechnology," adding, "Many of the ideas that came up while chatting casually eventually led to papers."
Professor Joung believes this research will serve as a foundation for handling more chemical reactions in the future. He stated, "This is a case where simply reflecting the fundamental laws of chemistry in AI design significantly improved performance, demonstrating what possibilities can open up when basic science and AI meet," adding, "The research findings have already been disclosed on GitHub, and we plan to increase the data to train the AI to predict metal catalyst reactions in the future."
Professor Joung said, "Having seen my researcher mother closely, I naturally took the path of a scientist," expressing joy that both he and his mother set a family record of publishing a paper in the world-renowned journal Nature, with his mother, Hesson Chung, who is a senior researcher at the Korea Institute of Science and Technology (KIST), having published in Nature 31 years ago.
References
Nature (2025), DOI: https://doi.org/10.1038/s41586-025-09426-9