Researchers use AI to design enzymes that cut other proteins

The study started from natural proteases: a zinc protease from crayfish digestive systems and a yeast cysteine protease. The design pipeline used RFdiffusion2 to generate backbones with constraints that gather key catalytic residues together and position the target cleavage sequence near the catalytic center; ProteinMPNN then filled in amino acids that maintain fold structure, form the catalytic site, and fit the target sequence. The cysteine protease project tested 12 backbones and 69 amino acid combinations with about 19% success, but activity was clearly lower than the yeast template enzyme. The zinc protease project was more problematic: without water molecules in the design constraints, only 14 of 89 designs showed weak activity, and modeling showed water often blocked the reaction site; with water positions constrained, 35 of 38 designs showed activity. Researchers then replaced amino acids near but not directly involved in catalysis one by one, screening versions with greatly improved activity: the best cysteine protease cut peptide bonds about 10^9 times faster, and a second zinc protease version had 30x higher activity than the original design with about 10^10 times faster cleavage. They also redesigned from scratch zinc proteases that cut proteins linked to Alzheimer's and frontotemporal dementia, validating cleavage in cells. For protein design developers, this shows the tools can build backbones not found in nature but also exposes the bottleneck: they excel at structure, not activity, and delivering enzymes into cells while avoiding immune reactions remains a hurdle before therapeutic use.