David Baker's lab uses AI to build custom protein-cutting enzymes

The work, published in Nature, sits at the intersection of two trends: AI protein design tools maturing fast, and the long-standing challenge of engineering enzymes. The body notes that protease design is a tough test because peptide bonds have a half-life of several hundred years and require a precise arrangement of amino acids to break. Baker's lab started from known proteases — a zinc protease from crayfish and a cysteine protease from yeast — and kept only their catalytic cores, then used RFdiffusion2 and ProteinMPNN to build entirely new proteins around them. This mimics evolution but compresses it into a design cycle. The results show both promise and limitation: high structural success but low catalytic activity, which random mutation could partly rescue. The therapeutic angle is real but distant, since delivering proteins to the right place inside cells remains a major obstacle. The authors themselves frame the work as a demonstration of what the tools can and cannot yet do, with the gap between structure and activity as the central open problem.