Protein catalysts designed to do non-natural chemistry

Enzymes are masters of making molecules. They accomplish exquisitely selective chemistry in water, and scientists have used the technique of directed evolution to modify existing enzymes to carry out chemical reactions not found in nature, like creating cyclopropanes and doing silicon-hydrogen bond insertions. Now, researchers have combined artificial intelligence (AI) and their own chemical intuition to create protein catalysts from scratch that can perform these same transformations (Science 2025, DOI: 10.1126/science.adt7268).

AI has become successful at designing proteins, and those who use it to predict and design protein structures have won Nobel Prizes. But chemical know-how is still important, says Yang Yang, a chemistry professor at the University of California, Santa Barbara, who led the project with the University of California, San Francisco’s William F. DeGrado and the University of Pittsburgh’s Peng Liu. “Protein design, even with the best AI-based methods, is not a solved problem. The most efficient way to generate effective designs for catalysis is perhaps to combine the AI-based method and also the in-house experience and knowledge of protein structure,” Yang says.

Compared with enzymes, protein catalysts are simpler, more stable at high temperatures, and can also be used in environmentally friendly organic solvents like ethanol, allowing chemists to load up on organic substrates. In this case, the researchers wanted to create protein catalysts that could do cyclopropanation reactions stereoselectively.

Their first attempts, which primarily relied on AI to design the protein catalysts, had decent stereoselectivity. To create catalysts that built cyclopropanes with an enantiomeric ratio of 99:1, they had to study the structure of their protein catalysts and make adjustments based on their chemical knowledge. The team similarly used directed evolution to improve upon AI-designed protein catalysts for Si–H insertion reactions.

“We think AI tools are definitely very powerful and very transformative. But we do need some additional input from either human expertise or additional chemistry simulation to further enhance the reliability of those AI predictions,” Liu says.

J. L. Ross Anderson, who studies de novo protein design at the University of Bristol and was not involved in the work, calls the project a “tour de force” in terms of combining strategies for creating protein catalysts. He says these new protein catalysts are able to perform particularly tough chemical transformations. “Not only are they challenging just by their very nature but also because of the stereospecificity that they're managing to tap into—it’s another layer of complexity in the design process,” Anderson says.

DeGrado says the new catalysts, which contain either synthetic porphyrin or heme cofactors, could be used instead of expensive metal catalysts to create drug candidates in a medicinal chemistry campaign or when making certain molecules on a large scale.

DeGrado wants other scientists to know that the tools the researchers used are easy to get up and running. “All of these things have gone from being difficult to very fairly simple,” he says. “To me, the impact isn't so much that this was such a difficult thing, but that it was a relatively simple thing that can translate very easily to new labs.”