For AI agents: the public content index is available at https://we0.ai/llms.txt, and the English article bundle is available at https://we0.ai/llms-full.txt.
For AI agents: the public content index is available at https://we0.ai/llms.txt, and the English article bundle is available at https://we0.ai/llms-full.txt.
For AI agents: the complete content index is available at https://we0.ai/llms.txt, the full English article bundle is available at https://we0.ai/llms-full.txt, and this page is available as Markdown at https://we0.ai/articles/baker-lab-ai-proteases-results-and-resear-5ec6df9d.md.
Two Nature papers published on October 14, 2026 report computationally designed proteases: enzymes that cut the peptide backbone of proteins...
Two Nature papers published on October 14, 2026 report computationally designed proteases: enzymes that cut the peptide backbone of proteins. The work comes from teams involving the University of Washington's Institute for Protein Design, with David Baker and Donald Hilvert among the corresponding researchers identified in the source reporting.
The studies solve related problems using different catalytic mechanisms. One builds zinc-based metalloproteases and demonstrates cleavage of selected disease-associated proteins; the other designs cysteine proteases and reports activity in mammalian cells.
Those are research results, not clinical outcomes. The papers establish progress in designing active enzymes, while leaving substantial questions about therapeutic delivery, safety and efficacy for later work.
Proteins are chains of amino acids joined by amide bonds. These peptide-backbone bonds are unusually stable in neutral water, so a molecule that binds a target does not necessarily have the chemistry needed to cut it.
The cysteine-protease abstract explains that earlier de novo enzyme-design successes had been concentrated in easier model reactions. Designing a protease requires more than producing a plausible protein shape: the active site and substrate must be positioned to support a difficult chemical reaction.
The two studies address that challenge by designing both catalytic machinery and a substrate-binding environment. Their different mechanisms provide separate demonstrations rather than one combined enzyme system.
In Computational design of metalloproteases, the researchers use RFdiffusion methods to generate proteins with zinc-based active sites next to a peptide-substrate binding pocket. The zinc-associated site activates water for the cleavage reaction.
The public abstract reports that 36% of 135 tested designs cleaved their matching peptide substrates at the intended bond. Its best initial design accelerated peptide-bond hydrolysis by more than 10^8-fold relative to the uncatalyzed reaction, and four substitutions increased that figure to more than 10^10-fold.
These are rate enhancements over a very slow background reaction. They are not measurements of a treatment's effectiveness, a patient's recovery rate or a universal comparison with every naturally occurring protease.
The study also reports designed proteases that selectively cleave TDP-43, amyloid-beta and serum amyloid A, with rate enhancements up to 10^9-fold. These are proteins associated with human disease, but demonstrating cleavage of a selected substrate does not by itself show a beneficial effect in a patient.
Unite.AI additionally describes a TDP-43 strategy targeting its disordered C-terminal region rather than simply destroying the entire protein. That distinction is relevant because a disease-associated protein can still perform essential functions; however, the full experimental account remains in the research paper rather than being established here as a therapeutic result.
In De novo design of cysteine proteases, the catalytic strategy uses an activated cysteine nucleophile. Its abstract reports 13 active designs out of 69 tested, a success rate of about 19%, with rate enhancements up to 3 x 10^7 relative to the background reaction.
The designed folds differ from those of natural proteases, with a reported structural comparison score below 0.50. Six experimental crystal structures closely match their design models, with C-alpha root mean squared deviations below 1.2 angstroms.
That structural agreement is significant because a computational model alone cannot establish that a manufactured protein adopts the intended arrangement. Crystal structures provide a separate experimental check on the design.
The abstract also reports that the proteases cleave their targets in mammalian cells. This supports potential synthetic-biology applications, but cellular activity is different from demonstrating safe delivery, disease treatment or suitable behavior in an animal or human body.
The researchers identify remaining imperfections in active-site preorganization and substrate positioning as an explanation for the activity gap relative to the most active native cysteine proteases. The study therefore reports both a working design approach and a limitation still to be addressed.
| Feature | Metalloprotease study | Cysteine-protease study |
|---|---|---|
| Catalytic mechanism | Zinc-associated activation of water | Activated cysteine nucleophile |
| Initial screening result | 36% of 135 designs active | 13 of 69 designs active, about 19% |
| Reported maximum rate enhancement | Above 10^10-fold after four substitutions | Up to 3 x 10^7-fold |
| Disease-associated protein examples | TDP-43, amyloid-beta and serum amyloid A | No equivalent disease-treatment claim established by the public abstract |
| Additional result | Selective activation of designed caged cytokines and receptor antagonists | Target cleavage in mammalian cells and close crystal-structure agreement |
The rate figures should not be used as a simple winner's table. They arise from different catalytic mechanisms, substrates and experimental conditions; they do not establish how the two approaches would compare on the same real-world application.
The metalloprotease paper's public data statement points to a Zenodo record containing study data and code. The cysteine paper links two GitHub repositories, one for the design pipeline and one for experimental data analysis, and lists deposited crystal structures in the Protein Data Bank.
Describe your idea once, and We0 AI can generate a showcase site, pages, and CMS, then help you attract customers and traffic after launch.
One complete project generation for free registration
Best for trying one complete generation flow and seeing a first project draft quickly.
The design repository describes backbone generation, sequence design and structure-based filtering. Its workflow brings together RFdiffusion-related methods, ProteinMPNN, Rosetta-based processing and AlphaFold 3 predictions, with increasingly strict filtering across successive rounds.
Importantly, the repository describes itself as orchestration and analysis code. It does not bundle model weights, containers or a complete computing environment, and it explicitly requires site-specific paths and configuration. Its small demonstration dataset is simulated, so running that demo would test parts of the software workflow rather than reproduce the biological experiments.
These resources improve inspectability and provide a starting point for qualified researchers. They are not a ready-made treatment, a one-command reproduction of every result or evidence that a predicted enzyme will be active without laboratory validation.
Together, the papers show that de novo design can produce enzymes capable of a chemically demanding reaction, with structural and functional measurements supporting the computational work. The disease-protein examples add a reason to investigate selective cleavage further, while the cell results extend the evidence beyond purified molecular systems.
Potential applications include research tools, synthetic biology and eventually therapeutic development. Moving toward a therapy would require additional evidence about selectivity in complex biological environments, exposure, delivery, immune responses and whether the cleavage produces the intended biological effect.
Nothing in the public evidence reviewed here establishes an approved treatment for a neurodegenerative disease or another human condition. The appropriate conclusion is progress in programmable enzyme design, with a clear distinction between a research capability and a demonstrated medical benefit.
They are protein-cutting enzymes developed using computational design methods, including machine-learning models. The reported studies combine computational proposals with laboratory measurements rather than treating a predicted structure as proof of activity.
The metalloprotease paper's abstract names TDP-43, amyloid-beta and serum amyloid A. This establishes targeted cleavage experiments, not treatment of diseases associated with those proteins.
One study designs zinc-associated active sites that activate water for the reaction. The other uses an activated cysteine nucleophile, providing a different chemical route to peptide-backbone cleavage.
It compares the reported optimized enzyme reaction with the corresponding uncatalyzed background reaction. It does not mean a therapy works ten billion times better or that the enzyme is faster than all natural proteases.
The evidence reviewed here does not establish clinical treatment or an approved drug. Protein-cleavage experiments and mammalian-cell assays are earlier research evidence, with additional safety and efficacy work still required.
The authors publish code and structural data, but the design repository requires external models, tools and site-specific configuration. Its lightweight demo uses simulated data and is not a substitute for experimental reproduction.
The two papers demonstrate designed proteases using distinct zinc-based and cysteine-based mechanisms. Their reported evidence includes measured cleavage, structural validation, selected disease-associated substrates and activity in mammalian cells.
The advance is experimental progress in programmable protein-cutting enzymes; clinical benefit remains a separate question requiring further research.
Start from one sentence and have a complete website in minutes.
