SEATTLE — Researchers at the University of Washington have used deep-learning software to design enzymes that find a chosen sequence inside another protein and cut it, including sequences taken from the proteins that pile up in the brains of people with Alzheimer's disease and amyotrophic lateral sclerosis. Two papers describing the enzymes were published in the journal Nature on Oct. 14.

The enzymes are proteases, the molecular machines that sever the bonds holding a protein chain together. Natural proteases digest food, clot blood and trigger cell death, and drugmakers have long wanted versions that attack one target and nothing else. The Nature papers report proteases that were not copied or tuned from anything in nature. They were drawn up on a computer, built in bacteria and then measured in the laboratory.

That matters because of how most drugs work. A conventional drug sticks to its target and blocks it, so one molecule of drug neutralises one molecule of disease protein. An enzyme works catalytically, cutting one target and moving to the next, which in principle allows far smaller doses. The University of Washington's Institute for Protein Design, which produced both papers, said designed proteases could also reach proteins considered undruggable because they have no pocket for a drug to grip.

"This is a major step forward for protein design," said David Baker, the institute's director and a professor of biochemistry at the UW School of Medicine, in the university's announcement. Baker shared the 2024 Nobel Prize in Chemistry for computational protein design.

The first paper, "De novo design of cysteine proteases," describes enzymes built around an activated cysteine that acts as the cutting blade. Of 69 designs tested, 13 cut their intended target, a success rate of 19%, according to the university. The most active sped up the cutting reaction by up to 30 million times compared with the same reaction with no enzyme present. The designed proteins fold into shapes unlike those of any known natural protease.

"This is a step toward designing tailor-made molecular scissors," said Hojae Choi, a doctoral student in biochemistry at the UW School of Medicine and the paper's lead author.

The second paper, "Computational design of metalloproteases," uses a bound zinc ion rather than an amino acid as the blade. Thirty-six per cent of 135 designs were active and cut at the intended bond, the institute said. The best initial design accelerated the reaction by more than 100 million times; four changes to its amino acids raised that above 10 billion times.

Those enzymes were then pointed at disease proteins. Designs cut selected sequences from TDP-43, which forms abnormal clumps in ALS and some dementias, from amyloid-beta, the fragment that accumulates in Alzheimer's disease, and from serum amyloid A, which damages organs in systemic amyloidosis. The amyloid-beta cut ran about 100 million times faster than the uncatalysed reaction, according to the university. The cysteine designs also cut their targets inside living human cells grown in the laboratory, not only in a test tube, the university said.

The designs came out of a pipeline of deep-learning models. RFdiffusion2 sketched the three-dimensional shapes, the ProteinMPNN family chose the amino acid sequences, and AlphaFold 3 predicted the resulting structures so that candidates straying from the blueprint could be discarded before anything was made, according to Phys.org's account of the two papers. Only after that filtering did the surviving designs go into bacteria for testing.

"Our results show that de novo enzyme design has advanced well beyond model reactions on activated substrates," the authors wrote, in a passage quoted by Phys.org.

The limits are plain in the numbers. Most designs failed: 56 of the 69 cysteine candidates did not cut their targets, and roughly two-thirds of the zinc designs missed. The cysteine enzymes remain weaker than the most active natural proteases, a gap the authors attribute to imperfect arrangement of the active site around the substrate. Neither paper tested the enzymes in an animal or a person, and the results say nothing about whether such an enzyme could be delivered safely to the brain or to any other tissue where a target protein collects. The institute described the medical uses as possibilities rather than findings.

Work on the enzymes was funded by the Howard Hughes Medical Institute, the Audacious Project, Open Philanthropy, the Schmidt Family Foundation and others, the university said.

The teams have deposited crystal structures of several designs in the Protein Data Bank, and the metalloprotease paper discloses provisional patent applications covering the work. The metalloprotease group also reported caged cytokines and receptor antagonists that its proteases switch on, a route to drugs that activate only where the enzyme has been delivered. "Proteases are just the beginning," said Anqi Chen, a postdoctoral scholar at the UW School of Medicine and a co-lead author of that paper.