MutexaGPT Lets Chemists Turn Plain-English Intuition Into Enzyme Designs

Zhongyue (John) Yang, an assistant professor of chemistry, has built an AI tool he says is meant to lift the work of scientists rather than take it from them. Called MutexaGPT, it was developed with his former doctoral student Qianzhen Shao, now a postdoc at Stanford, and other team members. The study describing it appears in Nature Computational Science.

The tool targets enzymes, the specialized proteins that speed up chemical reactions and that Yang describes as nature's catalysts. Engineered versions are used to synthesize new medications, preserve nutrient-rich foods and help recycle polymers to cut plastic waste. Each enzyme is made of hundreds of amino acids, and swapping out even one, a change called a mutation, can substantially alter how the enzyme behaves.

That sensitivity is the problem. Yang says no one knows how to predict which mutations will do what without leaning heavily on trial and error in industry. For an enzyme engineering market he says generates hundreds of billions of dollars in revenue each year, that means large amounts of expensive testing and resource costs.

MutexaGPT takes requests written in plain English, driven by a researcher's intuition. Large language model agents then ask for any information that is missing, build physics-based models, and turn the results into actionable design proposals. Yang says the aim is to let researchers explore, predict and design enzyme mutations in natural language, while drawing on AI and fundamental principles of chemistry and physics.

He frames this against wider worry about automation. "I know there are increasing concerns about AI replacing humans in the workforce, and I share them," he said, adding that the point of MutexaGPT is not to get things done more efficiently without people. Researchers in enzyme engineering often build their own intuition about how things work, he said; they may hold a principle but not know how to quantify it. The tool is meant to translate those principles into actual designs.

The source lists several uses. One is cold-active enzyme technology: engineering enzymes to keep high catalytic competence at lower temperatures, which matters for processes such as making lactose-free milk and preserving food while retaining heat-sensitive nutrients. Another is laundry detergents that break down complex stains without hot water. A third is pharmaceutical manufacturing, where enzymes with greater catalytic efficiency and selectivity could reduce energy use, chemical waste and dependence on harsh reaction conditions.

The name is a contraction of "Alexa for Mutant." Yang says it reflects the team's original vision of an Alexa-like assistant for designing enzyme mutants, one that guides scientists through conversation. The "GPT" signals the use of large language models, which connect researchers' questions to computational tools for enzyme engineering.

Yang grew up in China and trained as a pure computational chemist, earning his Bachelor of Science in the Chang Po-ling Program of Chemistry at Nankai University. He enrolled at UCLA in 2013 and finished his doctorate in four and a half years, then held a postdoctoral fellowship at MIT before joining the Arts and Science faculty in 2020. Since then he has worked on developing Mutexa and running experiments on how well it works.

His stated goal is a future where scientists engineer enzymes for specific applications as readily as they consult an AI assistant. What to watch is whether outside groups validate the tool's predictions and whether pharmaceutical, food and detergent makers adopt it.