Baker lab's from-scratch zinc proteases cut their chosen target in 36% of 135 designs
David Baker's University of Washington group tested 135 AI-designed zinc proteases, and 36% cut their target peptides. Cysteine designs fared worse, with 13 of 69 cutting theirs, and both rates come from a bacterial screen.
The Scientist · Science desk
Figures for two designed protease types. Cysteine: 13 of 69 designs cut their target. Zinc: 36% of 135 designs tested cut their target.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Cysteine designs cut target | 13 of 69 cut | 5 | |
| Zinc designs cut target | 36% | 6 | |
| Zinc designs tested | 135 designs | 4 |
What happened
- RFD2-MI, a version of RFdiffusion2, designed the protein backbones, and ProteinMPNN with EnhancedMPNN chose the amino acid sequences.
- The best zinc design sped peptide-bond hydrolysis more than 10^8-fold over the uncatalysed reaction, and four substitutions pushed that past 10^10-fold.
- One designed protease cut full-length TDP-43, and others cut fragments of amyloid-beta and serum amyloid A.
- AI-guided refinement also produced a cysteine protease that cut its target at precisely the intended site.
Why it matters
- What's now possible capability A protease built for a chosen sequence can act as a switch for other designed proteins: the abstract describes caged cytokines and receptor antagonists that the designed enzymes selectively activate.
- What it limits constraint Medical use is a distance off. The enzymes have not been in human trials, and the rate figures compare each design with the uncatalysed reaction, so they say little about how a design stacks up against a natural protease.
- What doesn't add up contradiction Phys.org puts the gain from refining the top zinc protease at 30 times, while the abstract states its four-substitution gain as floors over the uncatalysed reaction. As worded, the two cannot be compared.
Phys.org frames the result as a leap past AI-generated protein shapes, because a protease has to recognize a chosen sequence and then cut it [12]. The authors place the advance somewhere narrower. "Our results show that de novo enzyme design has advanced well beyond model reactions on activated substrates," the team wrote [10]. The Nature abstract calls de novo design of enzymes that cleave amide bonds a major challenge [11], and the paper's reference list includes an earlier Nature paper titled "Computational design of metallohydrolases" [13]. In our view the new part is the pairing: a zinc active site built next to a binding pocket for a peptide substrate chosen in advance [14].
The zinc numbers come from the Nature abstract and the cysteine numbers from Phys.org's account of a second paper [16]. Of 69 cysteine candidates, 13 cut their targets [5], or 18.8% [23]. The zinc rate of 36% over 135 designs [6] works out to about 49 enzymes [25] and roughly twice the cysteine rate [26]. That makes 204 designs tested in all [24]. They came out of a pipeline that Phys.org says was repeated dozens of times for each protease type [15], with AlphaFold 3 predicting final shapes to check against the intended designs [3].
In the screen, bacteria made one long chain joining the candidate enzyme, its target sequence and a glowing marker protein, and a design counted as workable if it cut the target and released the marker [18]. That keeps enzyme and target in one molecule. The thing this doesn't tell you is how many of the hits would cut the same sequence on a separate molecule. The TDP-43 result comes closer to that test, and the abstract puts rate enhancements for the three disease proteins at up to 10^9-fold [7].
The data availability statement lists crystal structures deposited in the PDB under seven accession codes [17].
What to watch
- Hit rates for target sequences beyond TDP-43, amyloid-beta and serum amyloid A, ideally from groups outside Baker's lab.
- Tests of the designed proteases against free-standing target molecules, not only the single-chain fusion screen.
- Any step toward human trials; Phys.org reports that none of the enzymes has been tested in people.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence72
- Adoption
- Not scored
- Hype gap+15
- Incentives
- Not scored
- Confidence68
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
David Baker at the University of Washington and colleagues combined several AI tools to create both cysteine and zinc proteases from scratch.
- [2]
A version of RFdiffusion2 called RFD2-MI was used to design the 3D shape and core backbones of the proteins, and ProteinMPNN and EnhancedMPNN were applied to select the amino acid sequences.
- [3]
AlphaFold 3 was used to predict the final shape of the designed proteins and check how closely they matched the intended designs.
- [4]
The team tested 69 versions of the cysteine protease and 135 of the zinc protease.
- [5]
Out of the 69 cysteine protease candidates, 13 successfully cut their targets.
- [6]
Of 135 zinc protease designs tested, 36% catalyse cleavage of their cognate peptide substrates at the targeted scissile bond.
- [7]
The designed metalloproteases selectively cleave the human disease-associated proteins TDP-43, amyloid-beta and serum amyloid A, with rate enhancements of up to 10^9-fold (printed as '109-fold' in the page text).
- [8]
One AI-designed protease cut full-length TDP-43, a protein that builds up abnormally in ALS and some dementias; other customized versions cut targeted fragments of amyloid-beta and serum amyloid A.
- [9]
The enzymes have not yet been tested in human clinical trials.
- [10]
"Our results show that de novo enzyme design has advanced well beyond model reactions on activated substrates," the team wrote.
- [11]
The abstract says de novo design of enzymes that cleave amide bonds represents a major challenge.
- [12]
Phys.org says AI tools have already generated new protein shapes and structures, and that using AI to create functional proteases is a significant leap forward because the enzyme must both recognize a chosen protein sequence and cut it efficiently.
- [13]
The paper's reference list includes an earlier Nature paper, Kim, D. et al., 'Computational design of metallohydrolases', Nature 649, 246-253 (2026).
- [14]
The abstract describes using RFdiffusion methods to design proteins with water-activating zinc-based active sites adjacent to a peptide substrate binding pocket.
- [15]
The researchers repeated the design process dozens of times for both types of protease to find options that worked.
- [16]
The findings are detailed in two Nature papers: Anqi Chen et al., Computational design of metalloproteases, and Hojae Choi et al., De novo design of cysteine proteases.
- [17]
Crystal structures have been deposited in the PDB under accession codes 11CE, 11CM, 11CP, 11CR, 11CU, 11CY and 11DU.
- [18]
Bacteria produced a single long protein chain linking the candidate enzyme to its intended target sequence and a glowing marker protein; a design was deemed workable if it chopped the target sequence and set the marker free.
- [19]
The best zinc design accelerates peptide-bond hydrolysis by more than 10^8-fold relative to the uncatalysed reaction (printed as '108-fold' in the page text); four substitutions increase this acceleration to more than 10^10-fold by improving active-site preorganization.
- [20]
Using AI to refine the most promising options boosted the cutting speed of the top zinc protease by 30 times.
- [21]
The team also designed caged cytokines and receptor antagonists that are selectively activated by the designed proteases, enabling bioorthogonal control of cell state.
- [22]
The AI-guided refinement also led to the design of a cysteine protease that cut its target at precisely the intended site.
- [23]
13 of 69 cysteine candidates is about 18.8%.
- [24]
The two papers tested 204 designs in total.
- [25]
36% of 135 zinc designs is about 49 designs.
- [26]
The zinc hit rate is roughly twice the cysteine hit rate.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
2 independent publishers whose own reporting we read for this story.
- nature.comComputational design of metalloproteases
1 article · October 13, 2026
- phys.orgAI designs working molecular scissors that recognize and cut selected protein targets
1 article · October 16, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- De novo protein designFollow
- AI for BiologyFollow
- Enzyme EngineeringFollow
Entities
- ProteinMPNNFollow
- RFdiffusionFollow
- AlphaFold 3Follow
- NatureFollow
- David BakerFollow
- University of WashingtonFollow