2026-09-26
RFD2-MI designed cysteine proteases from scratch: 13 of 69 cut the target peptide. Dokki-v2 hits kcat 0.064 s-1 (1.8e9-fold over background) and works in mammalian cells.
Amide bonds hold proteins together. In neutral water they last for hundreds of years. Natural cysteine proteases clear that barrier with a Cys-His-Asp catalytic triad that activates the cysteine nucleophile, an oxyanion hole that stabilizes high-energy intermediates, and a peptide cleft that parks the scissile bond next to the thiol.
Computational enzyme design has mostly worked on low-barrier model reactions. Serine hydrolases that cut activated esters arrived recently; protein backbone amides are far less reactive and much harder to bind. Scaffolding polar catalytic geometry and a peptide-binding cleft at the same time has been stuck for decades. The Baker lab at the University of Washington now reports cysteine proteases built from scratch, on folds that do not exist in nature.
The active-site motif comes from yeast SUMO protease Ulp1 (PDB 1EUV): the Cys580-His514-Asp531 triad, an oxyanion hole from Gln574 plus the backbone of residues 578-579, and four substrate backbone residues. The cleavage sequence was left unspecified. Peptide conformation and sequence were sampled together with the enzyme.
Backbones came from RFD2-MI, a version of RFdiffusion2 tuned for enzymes and protein-protein interfaces. It can start from Gaussian noise and condition only on active-site atom coordinates, without pre-assigning catalytic residues to sequence positions. Cysteine was chosen over serine because it is easier to activate. Beta-strand pairing was chosen to seat the polar peptide backbone and offset the desolvation cost.
The pipeline ran in three rounds. 50,000 initial backbones were filtered by DSSP (drop long loops, low strand content, high coil) down to 3,232. ProteinMPNN designed sequences with catalytic residues frozen and 0.1 Å distance restraints, then AlphaFold 3 checked apo-state self-consistency. Two rounds of partial diffusion (T=40, then T=10) tightened the geometry. Finalists needed AF3 complex pLDDT above 90, apo-versus-complex Cα RMSD below 0.8 Å, and correct catalytic distances.
Screening used a cis fusion: the cognate peptide, then mScarlet, then a His tag, grafted onto the enzyme C-terminus. An active design cuts during E. coli expression and releases mScarlet, which shows up as a smaller band after nickel purification.
13 of 69 designs cleaved the fusion, a 19% hit rate spanning 3 of 12 unique backbones. The three families are named Gawi, Kal, and Dokki. Dokki and Kal went to completion; Gawi reached about 95%. LC-MS confirmed all 13 cut at the intended bond. Cys-to-Ala or His-to-Ala killed activity. Mutating the oxyanion Gln/Asn did not uniformly abolish it.
On trans FRET octapeptides, Gawi was too slow for kinetics and Kal substrates were poorly soluble. Only the Dokki family gave a usable signal. Parent designs sat near kcat 10⁻⁶ s⁻¹ (Gawi) and 10⁻⁴ s⁻¹ (Dokki). ProteinMPNN redesign produced 7 of 21 Gawi and 24 of 89 Dokki variants that were 1.2- to 31-fold faster. Steady-state numbers: Gawi-09, kcat (0.8±0.1)×10⁻⁵ s⁻¹, KM 34±4 μM, kcat/KM 0.24±0.03 M⁻¹ s⁻¹; Dokki-15, kcat (1.1±0.1)×10⁻³ s⁻¹, KM 89±10 μM, kcat/KM 12±2 M⁻¹ s⁻¹. Dokki-15 has a bell-shaped pH profile peaking near pH 7.0 (apparent pKa 6.0 and 8.0), a melting temperature above 80 °C, activity out to about 70 °C, and tight specificity for VPTLTGG/G over seven lookalike peptides.
A 185-mutant scan around the Dokki-15 active site mostly hurt. Four positions (Tyr41, Ser44, Pro67, Thr99) helped, all with low ProteinMPNN probabilities in the original design. Combining them gave Dokki-v2 (Y41T:S44V:P67V:T99F): kcat 0.064±0.005 s⁻¹, KM 154±14 μM, kcat/KM 413±50 M⁻¹ s⁻¹, more than 130 turnovers per active site. The gain versus Dokki-15 is a 58-fold jump in kcat. Against uncatalyzed peptide hydrolysis (kuncat = 3.6×10⁻¹¹ s⁻¹), that is a 1.8×10⁹-fold rate enhancement.
| enzyme | kcat (s⁻¹) | kcat/KM (M⁻¹ s⁻¹) |
| Gawi-09 | 8×10⁻⁶ | 0.24 |
| Dokki-15 | 1.1×10⁻³ | 12 |
| Dokki-v2 | 0.064 | 413 |
| TEV (literature) | 0.16 | 2620 |
| Ulp1 on SUMO | 6.1 | not reported on the same peptide |
Dokki-v2 sits at the low end of native sequence-specific cysteine proteases: about 2.5-fold below TEV in kcat and 6-fold below in kcat/KM. Ulp1 cuts SUMO-containing substrates at kcat 6.1 s⁻¹ and shows no detectable activity on peptide fusions that lack the SUMO domain.
Six crystal structures cover all three scaffolds, resolutions 1.77-3.38 Å, all within 1.2 Å Cα RMSD of the design models. Foldseek TM-scores against PDB and AlphaFold databases stay below 0.50 (lowest 0.46 for Gawi-09). In HeLa cells, Dokki-family enzymes, including the original computer output, cleave a nuclear H2B-sfGFP-substrate-mScarlet reporter and release mScarlet from chromatin. Cys-to-Ala controls keep both fluorophores nuclear.
Generative design can now attack amide hydrolysis, one of the harder chemical steps in biology, and land on new folds rather than redecorating a native protease. The practical recipe is reusable: RFD2-MI for the scaffold, ProteinMPNN for sequence, AF3 for filtering, a fluorescent fusion for wet screening. For synthetic biology, sequence-programmable proteases that work in mammalian cells mean protein circuits no longer have to be built around TEV or HRV3C.
This is a first and an increment stacked together. The first is de novo construction, a new fold, true amide cleavage, and live-cell activity. The increment is efficiency: Dokki-v2 is within a small factor of TEV and still about two orders of magnitude below Ulp1 on SUMO.
The remaining gap is pinned on two things: incomplete active-site preorganization, and sloppy substrate positioning. In PLACER ensembles, Dokki His45 samples a non-productive rotamer propped up by Ser44; Y41T and S44V in Dokki-v2 suppress that state. The oxyanion Gln is still farther than in Ulp1, so productive TS1 geometries are 36% versus 100% in Ulp1, which matches the modest 1.9- to 3.8-fold drop when that Gln is knocked out. MD of the Michaelis complex shows a smaller, more polar peptide interface than native proteases, with poorer alignment of the scissile carbonyl to the catalytic cysteine. The Kal family never yielded reliable kinetics because of substrate solubility. Some optimized variants show substrate inhibition.
The four mutations that actually raised kcat had been assigned low ProteinMPNN probabilities. Fixed-backbone sequence design missed them; the current objective does not score catalytic preorganization well. Apo crystal structures of Dokki show the catalytic cysteine oxidized to sulfonic acid and a 1.3 Å shift of the His-bearing hairpin, so the designed geometry and the resting functional state are not the same object. Claims about endogenous disease substrates sit in the discussion only; there is no disease-model data. This is a bioRxiv preprint and has not been peer-reviewed.