The peptide bond that links amino acids is chemically inert in aqueous solution, persisting for centuries without a catalyst. Conventional biology overcomes this barrier with a limited set of proteases that position reactive residues and water molecules to break the bond. By employing artificial-intelligence driven protein design, scientists have engineered entirely new proteases that retain the essential catalytic core while discarding the surrounding scaffold, enabling precise cleavage of chosen protein sequences.
Two classic mechanisms guided the engineering effort. In a zinc-dependent protease, the metal ion coordinates the target peptide bond and a water molecule, withdrawing a proton to generate a nucleophilic hydroxide that attacks the bond. A cysteine-based protease instead activates a thiol side chain to form a covalent intermediate with the substrate, which water later hydrolyzes to release the cleaved fragments. Both require a tightly arranged set of catalytic residues.
The design pipeline began with crystal structures of a crayfish zinc protease and a yeast cysteine protease. Using RFdiffusion2, the team generated backbone scaffolds that forced the known catalytic residues into proximity and positioned a user-specified target sequence near the active site. ProteinMPNN then populated each backbone with amino acids that satisfied folding stability, catalytic geometry, and target binding constraints. Hundreds of candidates were evaluated with AlphaFold3, confirming that the predicted folds were novel and not present in existing databases.
For the cysteine designs, twelve backbone models yielded sixty-nine amino-acid combinations, of which thirteen displayed measurable activity, a 19 percent success rate, though each was less potent than the native yeast enzyme. The zinc series explored one hundred-thirty-five designs, with fourteen of eighty-nine showing weak catalysis; many failures involved water molecules obstructing the active site. Subsequent mutagenesis of residues adjacent to the catalytic core produced variants that accelerated peptide-bond cleavage by up to ten-nine-fold for cysteine and ten-ten-fold for zinc proteases.
The refined zinc proteases were re-engineered to recognize sequences found in proteins linked to Alzheimer’s disease and frontotemporal dementia. After rebuilding the design from scratch with the new target, the researchers observed specific cleavage of the pathological proteins inside cultured cells. This proof-of-concept suggests that AI-generated proteases could one day be programmed to remove disease-causing protein fragments, offering a novel route to modulate neurodegenerative pathology.
Translating these enzymes into medicines faces substantial obstacles. Protein therapeutics must reach the intended tissue, avoid immune detection, and, when intracellular targets are required, deliver the encoding gene to patient cells. Moreover, the authors note that the design tools succeed only when supplied with explicit physical constraints; they do not guarantee functional performance beyond the modeled environment. Continued refinement of AI models and delivery platforms will be essential before clinical use becomes realistic.