Deep-learning-designed proteases cut chosen protein targets

Researchers at the University of Washington’s Institute for Protein Design report that they have computationally designed synthetic proteases, enzymes that cut other proteins at chosen sites. According to UW Medicine, the teams used deep-learning-based design methods to specify a target sequence, build enzymes to cleave it, and then verify activity in several experimental settings. Ars Technica, drawing on the underlying studies, describes the work as a way to “short-circuit” the long natural diversification of proteases by preserving only the catalytic core of known enzyme classes and redesigning the rest.

The work matters because peptide bonds, which link amino acids into proteins, are chemically very stable on their own. Ars Technica says a typical peptide bond in water would otherwise persist for hundreds of years without enzymatic help. Proteases solve that problem by arranging a small set of catalytic components with extreme precision so that the bond can be broken. That precision is also what has made protease design difficult: the catalytic “blade” must be placed with near-atomic accuracy, and natural proteases often rely on subtle flexibility to grip their targets, according to UW Medicine.

Two protease mechanisms, redesigned

The researchers worked with two distinct types of protease chemistry. One project focused on cysteine proteases, in which an activated cysteine in the enzyme attacks the target peptide bond. The other focused on metalloproteases, described by Ars Technica as zinc proteases, where a metal ion helps position and activate the chemistry needed for cleavage. In both cases, the starting point was not a blank slate: the teams began from known proteases, including a yeast cysteine protease and a digestive zinc protease from crayfish, and retained the key catalytic residues while redesigning the surrounding protein structure.

For the design workflow, Ars Technica says the researchers first used RFdiffusion2 to generate enzyme backbones under several constraints. The designed backbone had to bring the catalytic amino acids together, and it also had to accommodate a chosen target sequence near the active site. They then used ProteinMPNN to assign amino acid identities along that backbone, balancing structural compatibility, catalytic-site formation and target binding. UW Medicine summarizes the same process more broadly as using deep-learning-based design methods to develop structures, screen them for cutting a specific sequence, and confirm that experimental structures matched the computational blueprints.

The teams also checked whether these proteins merely reproduced existing natural structures. Ars Technica reports that AlphaFold3 predictions for many designs did not resemble entries in current databases, suggesting the models were not simply reproducing known protease folds. UW Medicine goes further for the cysteine work, stating that the resulting proteins adopt folds unlike any naturally occurring protease.

How the experiments were run

The initial tests were built around direct cleavage assays. Ars Technica reports that the researchers created a long fusion protein containing the designed protease, the target sequence it was meant to cut, and a fluorescent protein. They then measured whether cleavage released the fluorescent portion. UW Medicine says the teams tested the enzymes on purified proteins, on sequences drawn from proteins implicated in Alzheimer’s disease and amyotrophic lateral sclerosis, and, for one study, inside living human cells grown in the lab.

• Cysteine proteases were designed around an activated cysteine catalytic center.
• Metalloproteases were designed around a bound metal ion used as the catalytic element.
• Designs were screened for activity against specified target sequences.
• Some candidates were tested on disease-linked sequences such as amyloid-beta.
• At least one study included experiments inside cultured human cells.

The cysteine protease project produced a 19% success rate, according to both sources. Ars Technica states that 13 designs worked, while noting that all were less active than the natural yeast enzyme used as a model. UW Medicine says these enzymes cut their target sequence up to 30 million times faster than the uncatalyzed reaction. Those two descriptions are compatible: one compares the designs to a natural template, while the other compares them to no enzyme at all.

The metalloprotease results are presented somewhat differently in the two sources, so they should be read carefully. UW Medicine says 36% of the designs showed activity and precisely cut the intended target, amyloid-beta, 100 million times faster than the uncatalyzed reaction. Ars Technica provides a more detailed breakdown: among early designs that did not explicitly include the catalytic water molecule as a design constraint, 14 out of 89 showed weak activity. When the researchers instead constrained the design to position a water molecule correctly, 35 out of 38 showed some activity. Ars Technica also says one stronger design lacked specificity about where it cut. These figures appear to refer to different subsets or stages of the work rather than a direct contradiction, but the sources do not fully reconcile them.

Iteration improved activity and specificity

A notable part of the studies was what happened after the first designs worked only modestly well. Ars Technica says the teams identified residues near the active site that were not directly responsible for catalysis and substituted alternative amino acids unlikely to disrupt the structure. That local optimization improved performance substantially. For the cysteine system, Ars Technica reports that the best version accelerated peptide-bond cleavage by a factor of 10^9 while remaining highly specific for its intended target sequence. For the zinc-based system, a second version was said to be 30 times more active than the original design and to accelerate cleavage by a factor of 10^10.

UW Medicine presents lower top-line acceleration figures for the named Nature papers: up to 30 million times faster than the uncatalyzed reaction for the cysteine proteases, and 100 million times faster for the metalloproteases. Because these numbers differ from some of the later optimized results cited by Ars Technica, it is prudent to treat them as describing different reported endpoints, likely from different designs or summary levels, rather than assume a single headline number for each platform.

Disease targets and therapeutic promise

As a proof of relevance beyond model substrates, the teams aimed some designs at proteins associated with neurological disease. UW Medicine says they tested sequences drawn from proteins implicated in Alzheimer’s disease and amyotrophic lateral sclerosis. Ars Technica specifically says the metalloprotease work targeted proteins associated with Alzheimer’s disease and frontotemporal dementia, and that the researchers ultimately obtained a design able to cut the intended target in cells. It also notes an important lesson: simply retargeting an existing design did not work well, so the team restarted the design process from scratch for the new sequence.

The therapeutic rationale is straightforward. Unlike a conventional drug or antibody that binds one target molecule at a time, an enzyme can act catalytically and then move on. UW Medicine says a single designed protease could cut thousands to millions of specified targets over its lifetime, potentially allowing a lower dose than a binding-based medicine. The same source also argues that proteases could address proteins often described as “undruggable,” because cleavage does not require a conventional binding pocket. Potential uses mentioned include destroying a viral protein, eliminating a bacterial toxin, clearing a misfolded protein in neurodegenerative disease, or activating a drug only in a chosen tissue by cutting off a protective cap.

What remains unresolved

The studies are encouraging, but both the reported limitations and the experimental scope suggest a long path before any medical use. Ars Technica emphasizes that the first successful designs were generally less active than the natural enzymes that inspired them, even if subsequent optimization improved that. It also points to practical delivery hurdles: proteins can be hard to get to the relevant tissue, may provoke immune responses, and become even harder to deploy when the intended target is inside cells rather than outside them.

That intracellular challenge could require cells to produce the protease themselves from a delivered gene, which adds another layer of complexity. None of the source texts claim a therapy is near. Instead, they support a more measured conclusion: the work shows that deep-learning-guided protein design can now produce synthetic proteases with chosen specificity, measurable catalytic activity, and at least some function in cells. Whether those capabilities can be translated into safe, effective and deliverable medicines remains to be demonstrated.

For anglophone companies, the near-term change is not a new approved treatment but a new design capability to watch. Biotech and pharma teams may see these results as an early sign that custom catalytic proteins could expand target selection beyond binders, especially for disease proteins lacking classic drug pockets. In practical terms, that could influence R&D strategy, partner scouting and platform evaluation around protein engineering, while leaving major execution questions—manufacturing, delivery, specificity in complex tissues and regulatory proof—still unresolved.