Researchers use AI to design proteins that cut up other proteins

AI design tools make proteins that break bonds that are stable for hundreds of years.

A protein is a chain of amino acids held together by what are called peptide bonds. On their own, those bonds are extremely stable; they’re estimated to have a half-life of several hundred years. But in Earth’s environment, most proteins don’t last nearly that long. That’s because organisms have evolved a handful of ways to catalyze the break-up of peptide bonds, which have diversified into countless specialized protein-digesting enzymes. Proteases are everywhere and generally make short work of any proteins that find their way into the environment.

Now, researchers have basically short-circuited the diversification process, using AI-based protein design tools to take the catalytic core of two types of protease and replace the entire rest of the protein. The resulting enzymes could be tuned to cut specific proteins in half, and the researchers demonstrated their effectiveness by using them to cut proteins associated with neurological diseases.

Cutting proteins

The basic peptide bond isn’t especially interesting-looking—it’s what you get when you link an amino group to an acid, releasing a water molecule in the process. As their name implies, amino acids have both an amino and an acid group, allowing them to form peptide bonds on two sides of the amino acid. This is critical for building up the long chains found in proteins. (The amino acids also have side chains with distinctive chemistry, which gives a backbone specific properties.)

Breaking these bonds apart simply reverses the process by breaking apart the peptide bond and a water molecule, and linking their parts. Because both components are chemically stable, the reaction requires an enzyme to occur with any frequency. Without an enzyme, a typical peptide bond would be stable for hundreds of years in water.

Life has evolved a handful of enzyme mechanisms that manage this catalysis. The research at issue here relied on two different ones. One is a zinc protease, where the positive charge of the zinc atom coordinates the peptide bond that’s being broken, key parts of the protease, and a water molecule. The protease partially pulls a proton off the water molecule, activating it to attack and break up the peptide bond.

The second protease the researchers use is called a cysteine protease. In this case, a cysteine (one of the amino acids) in the enzyme is activated to attack the peptide bond, creating a temporary intermediate where part of the protein being cut is chemically linked to the enzyme. A water molecule eventually breaks this intermediate apart, releasing the second part of the cut protein.

Both of these mechanisms require a handful of specific amino acids from the protease to be located in close proximity within the catalytic center of the protease. Other amino acids in the protease help ensure these key catalytic amino acids are brought together in the right place with the right orientation. The rest of the protease may also enforce a degree of specificity on what proteins get cut. Some proteases, like the ones active in our digestive system, can make cuts next to any amino acid. Others are slightly more selective, preferring to cut near positively or negatively charged amino acids. And a few are very highly specific, such as the protease that activates insulin, which only cuts near a very specific series of amino acids.

Designer enzymes

The people behind the new work, primarily from Nobel-winner David Baker’s lab at the University of Washington, viewed designing a protease as a challenge. For one, the stability of the peptide bonds means that reactions wouldn’t take place without an effective catalyst. And to get that catalyst requires an extremely precise arrangement of specific amino acids. So it wasn’t obvious that the AI protein design tools the lab developed were up to the task.

Both projects took a similar approach. Just like evolution, they started with an existing protease. One was a zinc protease used in a crayfish’s digestive system. The other was a cysteine protease from yeast. In both cases, the protein structure was known, as were the key amino acids that catalyze cleaving other proteins. The design process started with these amino acids.

The design process started with software called RFdiffusion2, which creates a backbone for the enzyme (the backbone consists of the protein’s repeated peptide bonds). The researchers gave RFdiffusion2 a number of constraints. To start, any backbone it designed had to bring together the key catalytic amino acids needed for a functional protease. They also gave it a target amino acid sequence where the protease would eventually cut. The backbone had to accommodate that target sequence in close proximity to the catalytic center.

With a backbone laid out, the process next moved to ProteinMPNN, a tool that works along the backbone, populating it with specific amino acids. (That tool featured in Google’s efforts to watermark AI-designed proteins.) ProteinMPNN selects each amino acid based on a series of constraints. For starters, the amino acid must be compatible with the backbone folding up into its planned structure. Next, in critical areas, each amino acid chosen needs to be compatible with the formation of the enzyme’s catalytic site. And finally, in nearby areas, it has to be compatible with latching on to the intended target.

The researchers repeated this process dozens of times for both the cysteine and zinc proteases. For the cysteine protease, they made 12 different backbones and populated them with 69 different sets of amino acids. For the zinc project, they tried 135 different protein designs. They used AlphaFold3 to predict the structures of many of these designs and found nothing in the databases that looked much like them, confirming their tools weren’t simply stealing previously identified structures.

Do they work?

The researchers started by making a single, long protein that contained the protease, the sequence it was designed to target, and a fluorescent protein. They then tested to see if the fluorescent protein was cut loose. In the case of the cysteine protease, they got 13 designs that worked, a success rate of 19 percent, although all of them were considerably less active than the yeast enzyme that was used as a model.

The zinc project was a bit more complicated. They tested many designs that didn’t include the water molecule that normally takes part in that reaction. These had a relatively low success rate, with only 14 out of 89 showing weak catalytic activity. When they modeled the ones that didn’t work, they found that the designs often let a water molecule block the site where reactions should take place.

So, they also tried designs where they placed a water molecule in the right location as an additional design constraint. These had a much higher success rate, with 35 out of 38 showing some catalytic activity. Again, however, their activity was considerably lower than the enzyme that was used as the starting point for the designs. (They did get one design with a better level of activity, but it wasn’t specific about where it cut other proteins.)

To find out whether this activity could be improved, the researchers identified amino acids that were in close proximity to the active site of the protease, but not actively involved with the catalysis. And they swapped in every alternative amino acid that wasn’t predicted to disrupt the structure of the design. They found some versions that were considerably more active; the best version they identified accelerated the cleavage of the peptide bond by a factor of 10⁹ while remaining highly specific for the intended target protein sequence.

Similar things happened with the zinc-based protease. Their v2 version of their protease, which incorporated some of the random mutations, was 30 times more active than the original design it was based on. It also accelerated peptide-bond cleavage by a factor of 10¹⁰.

As a final demonstration of this approach’s promise, the researchers designed zinc proteases that could target proteins associated with neurological disorders, such as Alzheimer’s and frontotemporal dementia. Attempting to modify their existing design to target a new sequence didn’t work well, so they started the process over from the start and managed to find something that could cut the intended target in cells. For the researchers, this suggests that we might ultimately be able to influence the course of some genetic diseases by targeting the malformed proteins that cause them.

What do the results tell us?

The potential therapeutic use of this is complicated. Proteins are often difficult to deliver to the site where they’re needed, and can often trigger an immune response. Things get even more challenging if the protease target is inside a cell rather than on the exterior. That would likely require the target cells to make the protease, which would involve delivering the gene encoding it to the cells where it’s needed. So, there are considerable hurdles that would need to be cleared before we could potentially start developing therapies using this technology.

The researchers also highlight how the work shows some limitations of the tools. As used, they successfully designed proteins that matched all the physical constraints they were given: They folded up into structures that brought the amino acids needed to catalyze a reaction into roughly the right location, and latched on to specific target amino acid sequences.

But we don’t, at the moment, have the ability to tell these tools we want any specific catalytic activity. So many designs lacked some of the subtle details that make evolved enzymes such great catalysts. In other words, the tools are great at structure, but they’re not necessarily great at activity. And, while structure is essential for function, it isn’t everything. As the researchers saw, introducing random mutation could still make subtle structural tweaks that produced enzymes that were considerably more active.

Of course, this protein design software is only a few years old, and both developers and users might have ideas for incorporating constraints beyond physical location.

Nature, 2026. DOI: 10.1038/s41586-026-11143-w, 10.1038/s41586-026-11167-2  (About DOIs).