AI designs new proteins

Proteins are tiny molecular machines, and researchers are now using artificial intelligence to draft some of them from scratch instead of copying versions found in nature. (nature.com) A protein is a chain of amino acids that folds into a 3D shape, like a string snapping into a specific knot. Tools such as RFdiffusion and AlphaFold help scientists propose those shapes and check whether the finished molecule is likely to hold together. (nature.com) (blog.google) That shift has moved the field from tweaking natural proteins to making “de novo” ones, meaning proteins with no direct natural ancestor. A 2025 review in *Nature Reviews Bioengineering* said AI systems now generate, evaluate, and optimize proteins far faster than older trial-and-error methods. (nature.com) The clearest recent example came on February 13, 2025, when a team led by University of Washington researchers reported AI-designed serine hydrolases in *Science*. The paper said the enzymes reached catalytic efficiencies up to 2.2 × 10^5 M^-1 s^-1 and used five folds not previously seen in natural serine hydrolases. (science.org) Serine hydrolases are enzymes that cut chemical bonds using a small reactive cluster of amino acids, like a lock pick arranged at atomic scale. Designing them is harder than designing a static protein because the active site has to stay precise through several steps of a reaction, not just one. (science.org) (cen.acs.org) Researchers are also aiming these methods at waste problems, including plastics. A February 2025 bioRxiv preprint from Universidad de Chile and collaborators said its protein-language-model system classified plastic-degrading enzymes with average accuracy of 89% and then generated candidate enzymes for polyethylene terephthalate, or PET, for future lab testing. (biorxiv.org) Companies are trying to turn that idea into industrial recycling. Protein Evolution, a Connecticut startup profiled by *Forbes* in May 2024, said it uses AI-designed enzymes to break down polyester textiles and post-industrial scrap into feedstocks for new material, but the company was still working to prove the process could scale. (forbes.com) Lab validation remains the bottleneck. The Baker Lab says its work still depends on repeated cycles between computation and experiments, and the 2025 plastic-enzyme preprint said its generated candidates had only been validated in silico, not yet confirmed in wet-lab tests. (bakerlab.org) (biorxiv.org) Safety researchers have started treating that gap as a governance problem as well as a technical one. *Nature* reported in 2025 that scientists were publishing safety guidelines for AI-designed proteins even as the same tools were opening new routes to make enzymes, antibodies, and other molecules not found in biology’s existing catalog. (nature.com) For now, the field is not replacing evolution so much as adding a new design workflow beside it. The next test is whether the growing number of computer-generated proteins can survive purification, measurement, and large-scale manufacturing as reliably as they survive on a screen. (nature.com)