Researchers have used AI protein-design tools to build new enzymes that cut selected protein sequences, including targets associated with Alzheimer’s disease and frontotemporal dementia. The work starts from the catalytic machinery of natural proteases but replaces the surrounding protein structure.
RFdiffusion2 generated backbones constrained to hold essential catalytic amino acids and a target sequence in the correct positions. ProteinMPNN then selected amino acids likely to fold into that backbone, preserve the active site and bind the intended target. AlphaFold3 checks suggested the results were not simple copies of known structures.
Initial performance was uneven. Thirteen of 69 cysteine-protease designs worked, while only 14 of 89 zinc designs made without an explicitly positioned water molecule showed weak activity. Adding that water as a design constraint raised the zinc success rate to 35 of 38, showing how a chemically important detail changed the outcome.
The first successful enzymes were still considerably less active than their natural starting points. Researchers improved them by mutating nearby amino acids: one version accelerated bond cleavage by a factor of one billion while remaining target-specific, and a zinc design became 30 times more active than its initial version. Cutting a disease-linked protein in cells demonstrates feasibility, not a therapy; delivery, safety and effects in organisms remain unresolved.