AI-designed proteases (Baker lab, Nature)

Your AI and this story

Key facts

• Paper 1: ‘De novo design of cysteine proteases’ (Nature, online Oct 14, 2026; first author Hojae Choi, senior authors Samuel J. Pellock and David Baker). Success rate 13/69 = 19%; kcat/kuncat up to 3x10^7; designed folds unlike natural proteases (TM score < 0.50); six crystal structures within 1.2 A Ca RMSD of the design models
• Paper 2: ‘Computational design of metalloproteases’ (Nature, online Oct 14, 2026; first author Anqi Chen, senior authors Donald Hilvert and David Baker). Of 135 designs, 36% cleave their intended peptide substrate at the targeted bond; the best accelerates hydrolysis more than 10^8-fold, and four substitutions take it above 10^10-fold
• Targets: designed metalloproteases selectively cleave the human disease-associated proteins TDP-43, amyloid-beta and serum amyloid A, with rate enhancements up to 10^9-fold; the paper also reports caged cytokines and receptor antagonists activated by the designed proteases (abstract)
• Pipeline (Ars Technica): RFdiffusion2 generates a backbone constrained to hold the catalytic residues and a target sequence; ProteinMPNN fills in the amino acids; AlphaFold3 predicts structures. Starting points were a crayfish digestive zinc protease and a yeast cysteine protease
• Ars Technica’s reading of the zinc work: designs that did not place a catalytic water often failed (14 of 89 weakly active), while adding the water as a design constraint raised this to 35 of 38
• Limitation stated by the authors (via Ars Technica): the tools can be told to satisfy structural constraints, not to achieve a specific catalytic activity, so designs are generally far less active than evolved enzymes; protein delivery and immune response are open obstacles to any therapeutic use