---
title: "Washington Scientists Use AI to Design Enzymes That Cut Other Proteins"
description: "Of 135 designed zinc proteases tested, 36% cleaved their intended target, including three proteins linked to human disease"
author: "rews desk"
published: 2026-10-14T17:10:15Z
modified: 2026-10-15T03:17:45Z
url: https://rews.cc/a/washington-scientists-use-ai-to-design-enzymes-that-cut-othe-f72f1b
language: en
tags: ["ai", "biology", "proteins", "enzymes", "rfdiffusion", "science"]
publisher: "Rews (https://rews.cc)"
---

# Washington Scientists Use AI to Design Enzymes That Cut Other Proteins

*Of 135 designed zinc proteases tested, 36% cleaved their intended target, including three proteins linked to human disease*

By rews desk · October 14, 2026 · https://rews.cc/a/washington-scientists-use-ai-to-design-enzymes-that-cut-othe-f72f1b

## In brief

- Two Nature papers published Oct. 14 report proteases designed from scratch using AI tools from David Baker’s lab at the University of Washington
- Of 135 zinc protease designs tested, 36% cleaved their target peptide at the intended bond; the cysteine designs had a 19% hit rate (13 of 69)
- The best zinc design, after four amino acid substitutions, sped up peptide-bond cleavage more than 10 billion-fold over the uncatalyzed reaction
- Designed proteases cut TDP-43, amyloid-beta and serum amyloid A, with rate enhancements up to a billion-fold
- Crystal structures match the design models to within 1.2 angstroms and score below 0.50 on TM score against natural protease folds

Researchers at the University of Washington reported on Tuesday that they had used artificial intelligence design software to build protein-cutting enzymes from scratch, among them enzymes that slice apart three proteins tied to human disease.

The work appeared in two papers in Nature from the lab of David Baker, who shared the 2024 Nobel Prize in Chemistry, with Donald Hilvert of ETH Zurich a corresponding author on both. In the [larger of the two studies](https://www.nature.com/articles/s41586-026-11167-2), 135 designs for zinc-based proteases were tested in the lab and 36 percent cut their matching peptide at the bond the designers had picked out.

That is the hard part. A protein is a chain of amino acids held together by peptide bonds, and left in water those bonds are stubborn, with a half-life estimated in the hundreds of years. Nature gets around this with proteases, enzymes that hold the bond, a water molecule and a few of their own amino acids in exactly the right geometry. Miss by an angstrom and nothing happens.

As [John Timmer wrote in Ars Technica](https://arstechnica.com/science/2026/10/researchers-use-ai-to-design-proteins-that-cut-up-other-proteins/), the teams short-circuited evolution: they kept the catalytic core of two natural proteases, one a zinc enzyme from a crayfish’s digestive tract and the other a cysteine enzyme from yeast, and threw out everything else. [RFdiffusion2](https://www.ipd.uw.edu/2025/04/introducing-rfdiffusion2) drew new protein backbones around those catalytic amino acids, with the target sequence held nearby as a constraint. ProteinMPNN, the same tool that [featured in Google DeepMind’s protein watermarking work](https://rews.cc/a/google-deepmind-finds-a-way-to-watermark-proteins-designed-b-c67845), filled in the sequence. AlphaFold3 predicted what the results would look like, and the researchers found nothing in the structural databases that resembled them.

The cysteine project was leaner. Twelve backbones, 69 sequences, 13 that worked, a hit rate of 19 percent, [according to that paper](https://www.nature.com/articles/s41586-026-11143-w). Every one was weaker than the yeast enzyme it came from.

The zinc designs taught the lab something about its own software. Of 89 designs made without explicitly placing the water molecule that does the chemistry, only 14 showed any weak activity, and models of the failures suggested a stray water molecule was sitting in the spot where the reaction needed to happen. When the researchers added that water to the design constraints, 35 of 38 designs worked.

Then they tuned them. Swapping amino acids near the active site that take no part in the catalysis produced a version more than 10 billion times faster than the uncatalyzed reaction, up from more than 100 million times for the best original design; four substitutions accounted for the jump, which the paper attributes to better preorganization of the active site. The redesigned cysteine proteases topped out at about 30 million times background.

For the disease work, the group built zinc proteases aimed at TDP-43, which aggregates in frontotemporal dementia and ALS, amyloid-beta, and serum amyloid A, with rate enhancements of up to a billion-fold. Retrofitting an existing design to a new target sequence went badly, so they ran the pipeline again from the beginning. One design cut TDP-43 in its disordered C-terminal tail inside cells.

The teams also made cytokines and receptor antagonists caged so that only a designed protease switches them on, which would let a researcher flip a cell’s state with an enzyme that has no natural counterpart to interfere with.

That absence is the point. Crystal structures of the designed cysteine proteases match their computed models to within 1.2 angstroms at the backbone, the papers report, while scoring below 0.50 on TM score against every known protease fold.