Researchers Use AI to Design Proteins That Cut Other Proteins
AI Design Tools Create Proteins That Break Bonds That Have Been Stable for Hundreds of Years
AI design tools create proteins that break bonds that are stable for hundreds of years.
Protein is a chain of amino acids connected by peptide bonds. These bonds themselves are very stable. It is estimated that their half-life is several hundred years. But in Earth's environment, most proteins don't last nearly as long. This is because organisms have evolved a handful of methods to catalyze the breaking of peptide bonds that have diversified into countless specialized protein-digesting enzymes. Proteases are ubiquitous and generally truncate proteins that enter the environment.
Now, researchers have essentially shortened the diversification process by using AI-based protein design tools to remove the catalytic core of two types of protease and replace the entire rest of the protein. The resulting enzymes can be tuned to cut specific proteins in half, and researchers have demonstrated their effectiveness by using them to degrade proteins associated with neurological diseases.
A basic peptide bond doesn't look particularly cool - it's what you get when you attach an amino group to an acid and release a molecule of water in the process. As their name suggests, amino acids have both amino and acidic groups that allow them to form peptide bonds on either side of the amino acid. It is vital for building the long chains found in proteins. (Amino acids also have side chains with distinct chemistry that give the backbone its special properties.)
Breaking these bonds simply reverses the process by breaking the peptide bond and a water molecule and linking their fragments together. Because both components are chemically stable, the reaction requires an enzyme to occur with any frequency. Without enzymes, a typical peptide bond would be stable in water for hundreds of years.
Life has evolved several enzymatic mechanisms that manage this catalysis. The research discussed here relied on two different cases. One of them is a zinc protease, in which the positive charge of the zinc atom in the broken peptide bond coordinates the key parts of the protease and a water molecule. The protease partially pulls a proton from the water molecule, activating it to attack and break the peptide bond.
The second protease used by researchers is called 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 cleaved protein becomes chemically linked to the enzyme. A water molecule eventually cleaves this intermediate, releasing the second part of the cleaved protein.
Both of these mechanisms require a handful of specific amino acids from the protease to be located near the catalytic center of the protease. The other amino acids in the protease help ensure that these key catalytic amino acids are assembled in the right place in the right orientation. The remainder of the protease may also exert some degree of specificity toward the cleaved proteins. Some proteases, such as those active in our digestive tract, can make cuts next to any amino acid. Others are a bit more selective, preferring to reduce positively or negatively charged amino acids. And a few are very specific, like the protease that activates insulin, which only degrades near a very specific set of amino acids.
The people behind the new work, mostly from the lab of Nobel laureate David Baker at the University of Washington, saw designing a protease as a challenge. For example, the stability of peptide bonds means that reactions cannot proceed without an effective catalyst. And to get that catalyst, a very precise arrangement of specific amino acids is needed. So it wasn't clear that the AI protein design tools the lab had developed would be up to the task.
Both projects had a similar approach. Just like evolution, they started with an existing protease. One of them was zinc protease, which was used in the digestive system of crabs. The other was a cysteine protease from yeast. In both cases, the structure of the protein was known, as well as the key amino acids that catalyze the breakdown of other proteins. The design process started with these amino acids.
The design process started with a software called RFdiffusion2, which creates a backbone for the enzyme (the backbone is made up of repeated peptide bonds of the protein). The researchers gave RFdiffusion2 a number of limitations. To begin with, any designed backbone must assemble the key catalytic amino acids required for a functional protease. They also gave it a target amino acid sequence where the protease would eventually be cut. The backbone had to place that target sequence near the catalytic center.
With the backbone in place, the process was next transferred to ProteinMPNN, a tool that works along the backbone and populates it with specific amino acids. (This tool is featured in Google's efforts to annotate AI-designed proteins.) ProteinMPNN selects each amino acid based on a set of constraints. To begin with, the amino acid must be compatible with the backbone folding into its programmed structure. Next, in the critical regions, each selected amino acid must be compatible with the formation of the catalytic site of the enzyme. And finally, in close quarters, it should be compatible with sticking to the intended target.
The researchers repeated this process dozens of times for both cysteine and zinc proteases. For cysteine protease, they made 12 different backbones and filled them with 69 different sets of amino acids. For the zinc project, they tried 135 different protein designs. They used AlphaFold3 to predict the structure of many of these designs and found nothing similar in the databases, confirming that their tools were not simply stealing previously identified structures.
The researchers started by making a single, long protein that contained a protease, a sequence designed to target it, and a fluorescent protein. They then tested to see if the fluorescent protein was loosened. In the case of cysteine protease, they obtained 13 designs that worked, a success rate of 19%, although all were significantly less active than the yeast enzyme used as a model.
The zinc project was a bit more complicated. They tested many designs that did not include the water molecule that normally participates in that reaction. These had a relatively low success rate, with only 14 of 89 showing poor catalytic activity. When they modeled the ones that didn't work, they found that these designs often allowed a water molecule to block the site where the reactions needed to take place.
Therefore, they also tried designs in which they placed a water molecule in the appropriate location as an additional design constraint. These had a much higher success rate, with 35 out of 38 catalytic activities. However, again, their activity was significantly lower than that of the enzyme used as a starting point for the designs. (They got a design with a better level of activity, but it wasn't clear where it cut other proteins.)
To find out if this activity could be improved, the researchers identified amino acids that were adjacent to the protease's active site, but not actively involved in catalysis. And they swapped in every substitution an amino acid that wasn't predicted to disrupt the structure of the design. They found versions that were significantly more active. The best version they identified accelerated peptide bond breakage by a factor of 109 while remaining highly specific for the desired target protein sequence.
Similar things happened with zinc-based protease. The v2 version of their protease, which incorporated some random mutations, was 30 times more active than the original design on which it was based. It also accelerated peptide bond cleavage by a factor of 1010.
As the latest show of promise for this approach, researchers have designed zinc proteases that can target proteins associated with neurological disorders, such as Alzheimer's and frontotemporal dementia. Trying to modify their existing design to target a new sequence didn't work well, so they started the process from scratch and managed to find something that could cut the target in cells. For researchers, this suggests that we may eventually be able to influence the course of some genetic diseases by targeting the defective proteins that cause them.
The potential therapeutic use of this is complex. Proteins often have a hard time getting to where they are needed and can often trigger an immune response. If the protease's target is inside the cell instead of outside, things get even more challenging. which presumably requires the target cells to make the protease, which involves delivering the gene encoding it to the cells where it is needed. Therefore, there are significant hurdles to overcome before we can potentially begin developing therapies using this technology.
The researchers also show how the work reveals some of the tool's limitations. As used, they successfully designed proteins that conformed to all the physical constraints they were given: they assembled into structures that brought the amino acids needed to catalyze the reaction into roughly the right place and attached to specific target amino acid sequences.
But we don't currently have the ability to tell these tools that we want a specific catalytic activity. Many of the designs lacked some of the finer details that make evolved enzymes great catalysts. In other words, tools are great in structure, but not necessarily great in function. And while structure is essential to performance, it's not everything. As the researchers observed, the introduction of random mutations can still produce subtle structural changes that produce enzymes that are significantly more active.
Of course, this protein design software is only a few years old, and both developers and users may have ideas to include constraints beyond physical location.
Nature, 2026. DOI: 10.1038/s41586-026-11143-w , 10.1038/s41586-026-11167-2 (About DOIs).
John is the science editor of Ars Technica. He holds a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in molecular and cellular biology from the University of California, Berkeley. When physically separated from his keyboard, he tends to look for a bike or a scenic spot to bond with his walking shoes.