Peptide bonds are among the most stable linkages in biology: left in neutral aqueous solution, they can persist for centuries. They form the load-bearing seam of every protein backbone, and they have long been considered the hardest target in de novo enzyme design—previous successes were largely limited to low-energy-barrier model reactions, and no designed enzyme had ever cleaved a peptide bond within a protein. In mid-October, the David Baker group at the University of Washington's Institute for Protein Design closed this gap with two papers published online in Nature on the same day.
A Scissor with a Zinc Blade
The metalloprotease paper takes the following approach: generative methods such as RFdiffusion build protein backbones, a zinc-based active site activated by water is placed at a designated position on the backbone, and a peptide substrate-binding pocket is attached alongside. The zinc ion polarizes the bound water molecule so it can attack the cleavable amide bond on the target peptide; the coordination geometry of the active site borrows from natural metalloproteases such as astacin and thermolysin. The full pipeline is motif-scaffold generation for one round of backbones, sequence design with AlphaFold 3 screening, then two rounds of partial-diffusion refinement, each interleaved with sequence design and screening, with a final iteration between the two.
The experimental numbers are solid: of 135 designs tested, 36% cleaved their corresponding substrates at the intended positions; the best initial design accelerated peptide hydrolysis to more than 100 million times the uncatalyzed rate. The team further optimized the active site of a design called Zn45, screening 378 single-site saturation mutants across 21 residue positions and 148 combinatorial mutants across 6 positions, ultimately using 4 substitutions to raise the acceleration beyond 10 billion-fold. Mechanistically, these substitutions improved preorganization: positioning the catalytic residues and the zinc-bound water closer to the transition-state geometry before the reaction occurs. Molecular dynamics simulations and machine-learning force fields provided supporting evidence: the way the designed active site lowers the activation energy of the rate-limiting step follows the same physics as the natural astacin family.
> Note: Preorganization is a core metric of catalytic proficiency in enzymology. Natural enzymes achieved near-perfect active sites through billions of years of evolution; the fact that computational design reached this regime through a few rounds of screening suggests the pipeline captured most of the reaction physics. Crystal structures matched too: several designs' experimental backbones align closely with the computational blueprints, and coordinates have been deposited in the PDB.
Cutting Disease Proteins: TDP-43, Aβ, and SAA
What really pushes this work beyond a demonstration is the target selection. The team built three scissors aimed at misfolded-protein diseases: TDPr3 targets TDP-43, SAAc8 targets serum amyloid A, and AbetaF3 targets beta-amyloid, with acceleration on disease-relevant substrates up to 1 billion-fold, outcompeting other peptides in specificity panels. The TDP-43 design is particularly thoughtful: degrading the full-length protein would damage its essential nuclear functions, so the scissor targets only the C-terminal disordered region that governs phase separation and aggregation—cutting here can modulate aggregation behavior without destroying the protein.
The Other Scissor: A Cysteine Route
The companion paper, published the same day, uses a different catalytic mechanism: an activated cysteine nucleophile. Functional elements such as the catalytic triad and oxyanion hole are retained, but the backbone is entirely new, modeled with RFD2-MI, sequences generated with ProteinMPNN, and filtered through multiple rounds of AlphaFold 3. Of 69 computational designs, 13 showed cleavage activity—a 19% success rate—with some variants accelerating the background reaction up to roughly 3 × 10^7-fold. After combinatorial mutation optimization, the best Dokki-series variant, Dokki-v2, reached a catalytic efficiency (kcat/KM) of 413 per M per second, an 1.8 × 10^9-fold acceleration of amide hydrolysis, a melting temperature above 80 °C, and retained activity at 70 °C. The enzyme was also validated inside mammalian HeLa cells: the protease was fused to a nuclear localization signal, and with a fluorescent reporter system, substrate cleavage redistributed the fluorescent protein within the cell—while a catalytically dead mutant did not. Crystal structures of six designs matched the computational models closely, with Cα RMSD below 1.2 Å, and their folds differ clearly from any known natural protease.
| Dimension | Metalloprotease | Cysteine Protease | |---|---|---| | Catalytic mechanism | Zinc-activated water | Cysteine nucleophile | | Design tools | RFdiffusion family | RFD2-MI family | | Tested / hits | 135 / 36% | 69 / 13 | | Best initial acceleration | >100 million-fold | ~30 million-fold | | Optimized acceleration | >10 billion-fold | 1.8 billion-fold | | Disease targets | TDP-43, Aβ, SAA | Intracellular cleavage validated | | Crystal structures | Four sets deposited in PDB | Six sets, RMSD < 1.2 Å |
Where the Gap Remains
The papers themselves name two limitations: the designed enzymes still trail the most active natural cysteine proteases, due to insufficient active-site preorganization and limited precision in substrate positioning within the binding pocket. Another more candid signal is that random mutagenesis can still substantially improve these enzymes—meaning the structural design tools are excellent, but catalysis design is not fully solved. Tools that excel at structure do not necessarily excel at catalysis; this boundary has never been drawn more clearly than here.
The path to therapeutics also has hurdles: proteins are hard to deliver to their targets and prone to triggering immune responses; for intracellular targets, the encoding gene must first be delivered. Between demonstration and application, the obstacle is not the algorithm—it's delivery.
Sources and caveats: Both papers went online at Nature in mid-October (October 14 in English sources; October 15 in some Chinese coverage), with corresponding authors including David Baker and Donald Hilvert, and co-first authors Anqi Chen and Kejia Wu on the metalloprotease paper. Figures here are compiled from Nature abstract pages, PDB entries, and multiple secondary reports; not every number has been cross-checked against the full-paper tables. Data and code for the metalloprotease paper are deposited in Zenodo; sequences, crystal coordinates, and analysis code for the cysteine protease paper have been made public.