Researchers, most of them from Nobel laureate David Baker’s laboratory at the University of Washington, developed AI-designed enzymatic proteins capable of cleaving peptide bonds in other proteins. The work relies on tools such as RFdiffusion2 and ProteinMPNN, and showed that these models can build structures that place catalytic amino acids near a target sequence, even when the design does not resemble known proteins in databases.
Peptide bonds have considerable chemical stability; their spontaneous breakdown in water may take hundreds of years. Living organisms use protease enzymes to accelerate this process. The researchers focused on two mechanisms: a zinc-dependent protease and a cysteine-dependent protease. In both cases, the challenge was to precisely arrange a small number of amino acids within a catalytic center capable of attacking the target bond.
How Were the Designed Enzymes Built?
The team began with two known structures: a zinc protease from the digestive system of a lobster and a cysteine protease from yeast. RFdiffusion2 was used to create the protein backbone while imposing constraints that ensured the catalytic amino acids were brought close to one another and to the sequence intended for cleavage. ProteinMPNN then filled the backbone with amino acids compatible with the structural fold, the catalytic site, and interaction with the target.
The researchers designed 12 backbones and 69 amino-acid sets for the cysteine-protease project, while 135 designs were tested in the zinc-protease project. They used AlphaFold3 to predict the structures of a number of designs and found no proteins in databases with a degree of similarity indicating that the tools had simply reused known structures.
Measurable Activity with a Gap from Natural Enzymes
The team tested the designs inside a long protein that combined the protease, the target sequence, and a fluorescent protein, so that cleavage would reveal the success of the reaction. Thirteen cysteine-protease designs succeeded, representing 19% of the tests, but their activity was far lower than that of the reference yeast enzyme.
Among the zinc-protease designs that did not explicitly constrain the water molecule necessary for the reaction, 14 of 89 showed weak catalytic activity. Analysis of the failed designs indicated that a water molecule sometimes occupied the position required for the reaction. When the position of the water molecule was included in the design, the success rate rose to 35 of 38 designs, while activity remained below that of the natural enzyme in most cases.
The researchers improved some designs by switching amino acids near the active site that did not participate directly in catalysis. The best modification in one project accelerated peptide-bond cleavage by 109-fold while maintaining high specificity for the target sequence. The second version of the zinc protease was also 30 times more active than the original design and accelerated the reaction by 1010-fold.
What Does This Mean for Medical Applications?
In a final experiment, the researchers designed zinc proteases targeting proteins associated with Alzheimer’s disease and frontotemporal dementia, and one of them succeeded in cleaving the intended target inside cells. However, modifying an existing design to target a new sequence did not work well, prompting the team to restart the design process from the beginning.
The results demonstrate the potential of this methodology for targeting abnormal proteins associated with certain diseases, but they do not represent a ready treatment. Delivering the proteins to the appropriate location is difficult, and they may trigger an immune response. If the target is inside the cell, it may be necessary to deliver the gene encoding the protease to the relevant cells, creating an additional challenge.
Limitations of AI Protein Design
The experiments show that current design tools are good at satisfying engineering and structural constraints: folding the protein into the required shape, bringing catalytic amino acids close together, and binding to a specific sequence. But they do not by themselves guarantee the precise details necessary for high enzymatic activity. Some designs were approximately correct structurally but weakly reactive, and small mutations subsequently improved their activity substantially.
The most important takeaway is that AI protein design has moved beyond proposing theoretical structures to producing enzymes that actually work on specific targets, but the gap between structure and activity, as well as delivery inside the body, still limits the transition to therapeutic applications. The results were published in the journal Nature under the digital object identifiers 10.1038/s41586-026-11143-w and 10.1038/s41586-026-11167-2.