AI-Designed Proteases Cut Alzheimer’s-Linked Proteins in Cells

Summary

On October 14, 2026, Nobel winner David Baker’s University of Washington team used AI to replace nearly all of two natural proteases while preserving catalytic cores from a crayfish zinc enzyme and a yeast cysteine enzyme. Peptide bonds have estimated half-lives of several hundred years, so cleavage demands exact catalytic geometry. RFdiffusion2 generated backbones around catalytic residues and target sequences, ProteinMPNN selected compatible amino acids, and AlphaFold3 predicted structures unlike known database entries. The cysteine project combined 12 backbones with 69 amino acid sets, while the zinc project explored 135 designs.

Of 69 cysteine designs, 13 worked, a 19 percent success rate, although all were much less active than the yeast enzyme. Among zinc designs lacking a positioned water molecule, only 14 of 89 showed weak activity; making water an explicit constraint produced activity in 35 of 38. Most remained weaker than the starting enzyme, while one stronger design cut nonspecifically. Mutating nearby residues yielded a highly specific cysteine variant that accelerated cleavage by 10^9, while zinc v2 became 30 times more active than its original design and accelerated cleavage by 10^10. Fresh zinc designs cut cellular targets associated with Alzheimer’s and frontotemporal dementia after attempts to retarget an existing design performed poorly. Therapeutic use still faces protein delivery, immune response and intracellular gene-delivery barriers. The results also show current AI tools optimize structure better than catalytic activity, with random mutations still needed for major performance gains.