Scientists Designed Enzymes From Scratch to Cut Disease Proteins

For decades, one reaction has mostly beaten enzyme designers: the bond that holds a protein's own backbone together. Amide bonds between amino acids are so chemically stable that, left alone in neutral water, breaking one can take centuries. Two papers from the University of Washington's Institute for Protein Design, posted as preprints on bioRxiv and reported in secondary coverage to have appeared in Nature on October 14, 2026, describe enzymes built entirely from scratch — no natural scaffold borrowed or evolved — that cut that bond anyway, at a specific address chosen in advance, inside three proteins linked to Alzheimer's disease and ALS.

The work comes from the lab of David Baker, who shared the 2024 Nobel Prize in Chemistry for protein design, working with Donald Hilvert's group at ETH Zurich. The first preprint, "Computational design of metalloproteases," is credited to co-first authors Anqi Chen and Kejia Wu. The second, "Computational design of cysteine proteases," is led by Hojae Choi. Baker and Hilvert are corresponding authors on both.

Why cutting a peptide bond is so hard

Natural digestive enzymes and the proteases inside every cell do this constantly, but they're the product of billions of years of evolution refining a catalytic site down to the exact position of a few atoms. Building an equivalent from nothing is a different problem. As the metalloprotease preprint puts it, amide bonds are far more stable than the ester bonds enzyme designers have had more success with, the protein's natural leaving group isn't chemically activated the way it is in easier reactions, and the floppy, flexible peptide chains these enzymes need to grip don't hold still for a rigid active site the way small drug-like molecules do.

The two teams solved it with different chemistry. Chen and Wu's group used a fine-tuned version of RoseTTAFold Diffusion, an AI model that generates new protein shapes, to build a zinc ion into a pocket positioned next to a groove shaped to hold the target peptide — the same basic strategy many natural metalloproteases use, assembled without starting from one. Choi's group instead used a newer method called RFD2-MI to build proteases around an activated cysteine amino acid as the cutting tool, a chemistry found in a different family of natural enzymes.

What the designed enzymes actually did

In a single round of 135 designs, 36% of Chen and Wu's zinc-based enzymes were active and cut precisely at the intended bond, according to the bioRxiv abstract. The best of those designs accelerated the cutting reaction more than 100-million-fold (10⁸) over the uncatalyzed rate; later refinement pushed that past 10-billion-fold (10¹⁰), according to coverage of the published paper. The team then pointed the method at three disease-associated targets, by the same account: a design called TDPr3 that cuts within a disordered region of TDP-43 — a protein implicated in ALS and frontotemporal dementia — rather than destroying the whole molecule; AbetaF3, aimed at a specific sequence within amyloid-beta, the protein that aggregates in Alzheimer's disease; and SAAc8, aimed at serum amyloid A, which is linked to a form of inflammatory amyloidosis. TDPr3's existence and authorship are independently confirmed by its crystal structure entry in the Protein Data Bank.

Choi's cysteine-based enzymes were tougher to get working — 13 of 69 tested designs were active, a 19% hit rate — but reached rate enhancements up to 30-million-fold (3×10⁷) over background, per the structure's linked publication record. Their folds look nothing like natural proteases by a standard structural-similarity score, yet six crystal structures matched the computer-generated design models almost exactly. According to the secondary coverage, these enzymes also cut their targets inside living human cells in lab tests, not only in a test tube.

What's established, and what isn't

Both papers report laboratory biochemistry: designed proteins that cleave an isolated peptide or protein target at a chosen site, with one team adding a cultured-cell test. Neither paper, as described in the available coverage, reports results in animals, and nothing here is evidence that cutting TDP-43, amyloid-beta or serum amyloid A changes the course of any disease in a living organism. The success-rate figures — 36% and 19% — describe how many computer-generated designs turned out to be active and precise enough to validate in the lab in a first attempt, which is itself a marked jump from prior de novo enzyme design efforts, nearly all of which were limited to simpler, already-activated model reactions.

The research was funded by academic and nonprofit sources, including the Howard Hughes Medical Institute, the Gates Foundation, and U.S. federal science agencies, rather than a drug company. Provisional patent applications covering the design methods have already been filed, which is routine for Institute for Protein Design work headed toward eventual licensing, not evidence that any product is imminent.

What this means for researchers

This is a methods paper dressed as a medical one, and the distinction matters. What's been demonstrated is that computational design can now reliably produce an enzyme that performs a chemistry nature's own proteases handle easily and designers have struggled with for years, and point it at a sequence chosen by the researcher rather than one evolution happened to select. That's a different kind of tool than a drug candidate: it's closer to a programmable pair of molecular scissors that, in principle, could be aimed at almost any protein whose sequence is known, including ones that current antibody- or small-molecule-based strategies can't easily reach.

It's also part of a broader pattern this journal has tracked before — AI-driven peptide and protein design tools increasingly being turned on problems nature hasn't already solved, as in an ancient antimicrobial peptide a separate team resurrected last month to kill drug-resistant bacteria. Readers curious about how peptide bonds relate to the broader molecules this journal covers can find background in our compound notes.

What to watch next: whether either lab, or a licensee, moves these designs into an animal model; whether the enzymes stay specific to their intended targets when tested against the full complexity of a cell or organism rather than a purified protein; and whether the next round of design work closes the remaining gap to natural enzymes, which even the optimized designs described here still don't fully match. A bond stable enough to outlast fossils finally has a lab-made enzyme that can find it on command — what happens once that enzyme has to find it inside a living body is still an open question.