AI-Designed Proteases Cut Proteins That Resist for Centuries

Imagine a pair of molecular scissors that you design on a computer, with no natural template, and that cuts one precise spot in one precise protein. It sounds like a wish. On October 14, 2026, two teams at the University of Washington's Institute for Protein Design published papers in Nature that bring it closer to reality. According to Unite.AI's report on the studies, the teams used computational, AI-driven design to build proteases from scratch. Proteases are enzymes that cut other proteins. These were built to work at chosen sequences, including sites in the human disease-associated proteins TDP-43, amyloid-β and serum amyloid A.

The detail that makes this remarkable is how stubborn the target is. The bonds these enzymes cut are stable for hundreds of years under ordinary conditions. Let me unpack why that matters, what was actually shown, and where we should stay cautious.

Why cutting a protein is so hard

Proteins are long chains of amino acids, and the links between the amino acids are called peptide (or amide) bonds. According to the report on the cysteine protease paper, those bonds are stable for hundreds of years in neutral water because the energy barrier to breaking them is so high.

That is good news for life, since proteins don't fall apart on their own. It is bad news for anyone who wants to design an enzyme to break them. The paper says that success in designing enzymes from scratch ("de novo") had been largely limited to model reactions with low energy barriers. Cutting a peptide bond is a much taller hill to climb.

Think of it as the difference between designing a ramp and designing a rocket. Both are engineering, but the margin for error is nothing alike. An enzyme has to hold its target in just the right position and drive the chemistry with atomic-scale precision, and it has to be a stable protein that folds correctly in the first place.

Two teams, two chemical strategies

The work comes as two separate papers. Both list David Baker as a corresponding author, and Donald Hilvert of ETH Zurich is a corresponding author on both. They differ in the chemistry they use: one produced zinc-based metalloproteases, the other proteases that work through an activated cysteine.

The first, Computational design of metalloproteases, builds enzymes around a zinc atom. The zinc activates a water molecule, and the water does the cutting. Co-first authors Anqi Chen and Kejia Wu and colleagues used RFdiffusion methods to design proteins with zinc-based active sites sitting next to a pocket that grips the target peptide.

The second, De novo design of cysteine proteases, with Hojae Choi as first author, takes another route. It uses an activated cysteine, an amino acid with a sulfur atom, as the nucleophile, the chemical group that attacks the bond. The team used a new deep learning method called RFD2-MI, developed by team members Brian Coventry and Magnus S. Bauer.

Two independent chemical routes matter. In my reading, they suggest this is not a one-off trick but a more general capability: if you can describe the chemistry you want, AI-guided design can start to build a protein that performs it.

The numbers behind the headline

Numbers in science news can be numbing, so here is what I find most telling. In the metalloprotease paper, 135 designs were tested, and 36% of them catalysed cleavage of their intended peptide substrate at the targeted bond, as the paper reports. Whatever your baseline for enzyme design, a hit rate above one in three is strong for a reaction this difficult.

The best initial design sped up peptide-bond hydrolysis by more than 108-fold compared with the uncatalysed reaction. Four amino-acid substitutions improved how well the active site was pre-organised, and the acceleration then reached more than 1010-fold, according to the paper. For the cysteine proteases, 13 of 69 designs were active, which is 19%, with rate enhancements of up to 3 × 107.

There is also a gap worth stating plainly. The authors' own experimental and computational analyses attribute the remaining activity gap relative to the most active natural cysteine proteases to imperfect active-site pre-organisation and substrate positioning. In other words, the best natural enzymes are still better. These designs look like impressive first-generation tools, not replacements for what evolution has produced.

The AI toolbox behind the design

The pipelines were not a single model producing an answer. They were loops. In the metalloprotease work, one round of generating a protein backbone around the desired active-site motif was followed by sequence design and filtering with AlphaFold 3. Two rounds of refinement via partial diffusion came next, then a final back-and-forth between sequence design and AlphaFold 3 filtering. The cysteine work similarly cycled RFD2-MI backbone generation with ProteinMPNN sequence design and AlphaFold 3 filtering.

In plain terms: one AI proposes a shape, another proposes an amino-acid sequence that should fold into it, and a structure-prediction model filters the candidates. Weak candidates are screened out computationally, and the lab then tests what remains.

Refinement continued after the first hits. To optimise a metalloprotease called Zn45, the team screened 378 single-site variants and 148 combinatorial variants. For a cysteine protease named Dokki, 24 of 89 generated variants outperformed the parent design, with improvements from 1.2- to 31-fold.

The researchers also solved crystal structures of several designs and deposited them in the Protein Data Bank. For the cysteine proteases, six crystal structures matched the design models to within less than 1.2 ångströms. That agreement between prediction and reality is the kind of evidence that makes me trust the claims more. The designed proteases also adopt folds very different from natural proteases, with a TM score below 0.50. So the AI is not just copying existing blueprints. It found architectures that work and look different.

Cutting disease-linked proteins

The most eye-catching part is the targets. The metalloprotease designs selectively cleave the human disease-associated proteins TDP-43, amyloid-β and serum amyloid A, with rate enhancements of up to 109-fold. Named examples include TDPr3 for TDP-43, SAAc8 for serum amyloid A, and AbetaF3 for amyloid-β.

The TDP-43 case is subtle and clever. Rather than degrading the whole protein, the authors redesigned an existing scaffold to cut within its disordered C-terminal domain. That region governs liquid-liquid phase separation and oligomerisation. The paper presents this as a way to modulate that behaviour without wiping out the essential nuclear functions of the protein's structured domains. It is a bit like trimming a loose thread instead of unravelling the sweater.

The teams also designed "caged" cytokines and receptor antagonists that the designed proteases switch on selectively. The paper describes this as bioorthogonal control of cell state, meaning a way to flip a biological switch with a tool that does not interfere with the cell's own machinery. The designed cysteine proteases also efficiently cut their targets inside mammalian HeLa cells in a reporter assay, with catalytically inactive mutants as controls. That is a notable step beyond test-tube chemistry.

What this is not (yet)

Let me add some honest brakes. These are laboratory results: enzymes acting on peptides, in test tubes and in cultured cells. Nothing in the report shows a treatment in animals or people. Getting a protease into the right tissue, keeping it from cutting the wrong things, avoiding immune reactions and proving a benefit are separate, long roads. That is my own reading of the gap between a lab demonstration and a medicine, not a claim from the papers.

The commercial side is normal but relevant. A competing-interests disclosure in the metalloprotease paper states that provisional patent applications have been filed related to aspects of the methods and designs, listing several people including Baker as inventors or contributors. The cysteine paper's authors declared no competing interests. Funding statements list a broad group of funders, including the Howard Hughes Medical Institute, the Gates Foundation, the Defense Threat Reduction Agency, the National Science Foundation and parts of the National Institutes of Health, among others. Design and analysis code for the cysteine study is public on GitHub, which helps other labs check and build on it.

One more detail: Nature received both papers on October 14, 2025. They were accepted in September 2026, and preprints had appeared on bioRxiv in November 2025. That is nearly a year of review and revision, a healthy reminder that careful science moves slower than the news cycle. Also worth knowing: the Unite.AI report itself is labelled as written by an AI agent and reviewed by the site's editorial team, so I would read the original Nature papers for any detail that matters to you.

Why it matters beyond one lab

The authors conclude that de novo design can yield highly active, programmable proteases with potential applications in therapeutics, biotechnology and bioremediation. "Programmable" is the word to remember. Today, if you want to cut a particular protein at a particular place, you look among existing enzymes and hope. A programmable protease would mean you specify the target and the computer helps you build the tool.

That shift echoes what we see in other fields where AI moves from analysing to creating. I explored a similar feeling in AI's growing role in mathematics: the interesting part is not that the machine is fast, but that it starts to propose things we did not explicitly ask for. And much as Ithaca restores damaged ancient texts by filling gaps with informed predictions, protein design tools fill a huge search space with candidates that experts then verify.

There is a dual-use question lurking here too, as with any powerful biological design tool. Programmable molecular cutters look very promising for medicine, but the same openness that lets labs reproduce results also means the community needs thoughtful norms. That is my opinion rather than something the papers address, and I would like to see the field discuss it in public rather than only after something goes wrong.

A question to carry with you

Most of us will never touch a protease in our lives, yet proteins quietly run our health, our food and our medicines. If AI can now design enzymes for jobs nature never evolved to do, what should we ask of it first: cures for stubborn diseases, cleanup of pollution, or cheaper medicines? And who gets to decide, and who benefits?

I find this story hopeful, precisely because it is so grounded: real crystal structures, real hit rates, honest gaps. When AI touches your life in the coming years, it may not be through a chatbot but through a medicine whose key ingredient was designed on a screen. How comfortable would you be with that, and what would you want to know before trusting it?