← All articles

Mikhail Polyakov

AI has learned to design proteins that cut other proteins: what two papers showed and why this is not a medicine yet

Level: IntermediateReading time: 5 min

Topics: Science, AI, Medicine

A long chain of coloured beads in a tangle being cut by steel scissors, against a blurred laboratory rack with empty test tubes

The short answer: in two papers (an article in Nature and a preprint; the preprint's authors from the Institute for Protein Design at the University of Washington include David Baker) AI tools were used to create, from scratch, enzymes that cut other proteins at chosen sites. Creating such enzymes from scratch had remained a difficult task, and the preprint's authors consider the resistance of the bonds between protein building blocks a likely reason. According to the Nature abstract, some of these enzymes selectively cut proteins associated with disease. Judging by the abstracts, there is no mention of trials in humans. Below: what was done, what the figures mean and what still lies ahead.

Where the news comes from and what I read

The news appeared in Ars Technica on 14 October 2026. I was unable to open the full Ars Technica text and saw only its beginning. So I rely on what I read myself: the abstract of the Nature article "Computational design of metalloproteases", published on 14 October 2026, and the abstract of the bioRxiv preprint "De novo design of cysteine proteases" (November 2025). I did not read the papers in full beyond their open parts.

What these proteins are and why they are hard to design

A protein is a chain of building blocks, amino acids, held together by peptide bonds. According to Ars Technica, these bonds on their own are extremely stable, with their half-life estimated at several hundred years. But in nature proteins last far less, because organisms have evolved enzymes that break these bonds. They are called proteases.

The bioRxiv preprint says the same: proteases are widespread in nature and in biotechnology, and the preprint's authors call creating such enzymes from scratch a difficult task, probably because amide bonds resist breaking.

What was done

  • The first paper (Nature). The authors used methods from the RFdiffusion family, that is, AI methods for protein design. According to the abstract, they designed proteins with zinc-based active sites that activate water, and with a pocket next to them where the segment of chain to be cut (a peptide) sits. Of the 135 designs tested, 36% cut their peptides at the intended point.
  • The second paper (preprint). A new deep learning method, RFD2-MI, was used to design enzymes in which the bond is broken by an activated amino acid, cysteine. Such enzymes cut a protein chain depending on which building blocks are in it.
  • How Ars Technica describes it. The researchers took the "catalytic core" of two types of proteases and replaced the whole rest of the protein, obtaining enzymes that can be tuned to cut specific proteins.

Which figures are given

They come from the abstracts and should be read with care.

  1. Speed-up. The best zinc design speeds up the breaking of a peptide bond by more than 10 to the power of eight times, that is, more than a hundred million times, compared with the reaction without an enzyme. Four substitutions of building blocks brought the speed-up to more than 10 to the power of ten times. For the cysteine enzymes the preprint gives a speed-up of up to 3 times 10 to the power of seven.
  2. Disease-associated proteins. The zinc enzymes selectively cut the proteins TDP-43, amyloid beta and serum amyloid A, with a speed-up of up to 10 to the power of nine times. The abstract calls them "disease-associated human proteins"; the abstract does not claim that these particular proteins cause disease.
  3. Unlike nature. For the cysteine enzymes, according to the preprint, the shape of the protein differs strongly from natural proteases, and the crystal structures match the design with a deviation of less than 1.2 angstroms.
  4. An additional application. The Nature authors also created "caged" cytokines (signalling proteins) and receptor antagonists (substances that block a cell from receiving a signal), which such proteases selectively switch on. The authors call this bioorthogonal control of cell state.

What is reported here and what is not

This is my own breakdown, not words from the sources.

  1. Reported in the abstracts. Enzymes created from scratch cut peptides and some target proteins, and the crystal structures of the cysteine enzymes are close to the design (this comes from the preprint abstract). I did not check the measurement methods themselves.
  2. Stated by the authors as possible. Applications in therapy, biotechnology and environmental clean-up. In the Nature abstract this is worded as "applications", without promises of results.
  3. From Ars Technica. The beginning of the article says the researchers demonstrated the enzymes' effectiveness by cutting proteins associated with neurological diseases. Under what conditions and with what result is not visible in the part of the text available to me.
  4. What is missing from the parts I read. Data on treating people, safety, doses and comparison with existing drugs. There is also no comparison of activity with natural enzymes in the abstracts I read.

How to read news like this. Find where in the text the experiment was carried out: in a test tube, in cells, or in animals and humans. The abstracts I read say nothing about trials in humans, and the conditions of the experiments are not given. If the abstract does not name the experimental conditions, look for them in the full text of the paper, and until then do not draw conclusions about whether the result is applicable. The path from such a result to a medicine is not described in these texts.

Summary

According to the abstracts, the authors showed that AI tools can create, from scratch, enzymes that cut proteins: 36% of the 135 zinc designs worked, and the best one sped up the reaction more than a hundred million times. Some of the designs selectively cut disease-associated proteins. This is a step in enzyme design, but not a treatment: the parts of the sources I read say nothing about trials in humans.

See also another piece in this section: The virtual cell: what it is.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources