AI Designs Proteins That Break Stable Peptide Bonds

AI designs enzymes breaking stable peptide bonds

Researchers at the University of Washington have harnessed AI to build brand-new enzymes capable of slicing through peptide bonds-chemical links so stubborn they usually hang around for hundreds of years in water. The findings, published in mid-October 2026, come mainly from Nobel laureate David Baker’s lab and show that AI-driven protein design has moved way past just mimicking nature. Instead, the team constructed catalytic cores from scratch around two distinct protease mechanisms, then proved the resulting enzymes could be aimed at proteins linked to neurological disease.

Peptide bonds form the backbone of every protein, stitching amino acids together in a reaction that kicks out a water molecule. Reversing that process demands an enzyme because the bond itself is chemically inert; without a catalyst, a typical peptide bond survives for centuries. Nature has cooked up a few solutions, including zinc proteases that use a metal ion to wrangle the bond and a water molecule, and cysteine proteases that temporarily tether the target protein to the enzyme before water finishes the cut. Both mechanisms require a precise arrangement of just a handful of amino acids in the catalytic center, while the rest of the protein positions those residues and often decides which substrates get cleaved.

AI protein design engineering approach

The Washington crew treated that precision as an engineering problem. They started with known structures-a zinc protease from a crayfish digestive system and a cysteine protease from yeast-and pinpointed the essential catalytic amino acids in each. Then they fed those residues into RFdiffusion2, a generative model that dreams up protein backbones. The software was constrained to place the catalytic residues in the right geometry and to make room for a chosen target sequence near the active site. Once a backbone cleared those filters, ProteinMPNN-a sequence-design tool that also supports Google’s efforts to watermark AI-generated proteins-filled in the remaining spots with amino acids compatible with folding, catalysis, and substrate binding.

cysteine protease design results AI

For the cysteine protease project, the pipeline churned out 12 distinct backbones and 69 sequence variants. The zinc protease effort explored 135 complete designs. AlphaFold3 predictions confirmed that none of the resulting structures matched anything in existing databases, meaning the models hadn’t just retrieved known folds. When the researchers synthesized the candidates and tested them in a fused construct that released a fluorescent protein upon cleavage, 13 of the cysteine protease designs worked-a 19 percent hit rate-though all were slower than the natural yeast enzyme they were modeled on. The zinc protease results were still being analyzed at the time of publication.

AI protein design benchmark implications

The significance for anyone building with AI goes well beyond the raw success rate. The experiment proves current generative models can juggle multiple, stringent physical constraints at once: backbone geometry, catalytic residue placement, substrate recognition, and overall foldability. That matters because protease engineering has traditionally leaned on directed evolution, a slow, labor-intensive grind that only explores sequences next to a natural starting point. AI-driven design jumps straight to novel scaffolds, potentially unlocking catalytic solutions evolution never stumbled upon. For developers of protein-design platforms, the work validates that diffusion-based backbone generation coupled with sequence optimization can produce functional enzymes de novo, a milestone that could speed up the creation of therapeutic agents, industrial catalysts, or research tools.

Readers who maintain or evaluate AI protein-design pipelines should treat this as a benchmark. The 19 percent functional yield for cysteine proteases, while modest, sets a baseline for de novo enzyme design that future models have to beat. Teams deploying RFdiffusion2 or ProteinMPNN should verify their implementations reproduce the reported constraints, especially the simultaneous satisfaction of catalytic geometry and target-sequence accommodation, and should fold in AlphaFold3 or an equivalent structure prediction as a novelty filter. Those developing therapeutic programs should note that the designed proteases successfully cleaved disease-associated proteins in vitro, hinting at a path toward targeted protein degradation that sidesteps the need for natural protease specificity. However, the reduced activity relative to natural enzymes means any near-term application will need further optimization, likely through hybrid AI-evolution workflows. Keeping an eye on the zinc protease results when they drop will clarify whether the approach generalizes across catalytic mechanisms or stays constrained to the cysteine chemistry tested most thoroughly.

Source: Ars Technica