The Pulse
AI Designs Cysteine Proteases to Cut Stable Peptide Bonds
A team led by Hojae Choi and David Baker designed cysteine proteases that cut peptide bonds, which can persist for hundreds of years in neutral water. In tests, 13 of 69 designs worked, and the enzymes also cleaved targets in mammalian cell
AI.info Team · Written with AI ·
“Here we describe the de novo design of enzymes that use an activated cysteine nucleophile to hydrolyse the polypeptide backbone in a sequence-dependent manner.”
Hojae Choi et al., Nature
The enzymes designed by Hojae Choi, David Baker and their colleagues do something that has challenged computational enzyme design: they cut peptide bonds, the links joining amino acids into proteins. In neutral water, those bonds can persist for hundreds of years. The researchers report that their designed proteases cleave specific sequences and work inside mammalian cells.
From designed structure to working enzyme
The team’s paper, published in Nature on October 14, describes enzymes built around an activated cysteine, an amino acid whose sulfur atom can attack a peptide bond. The researchers designed the proteins to position that reactive group alongside other catalytic residues and a target peptide. The resulting enzymes hydrolyse the bond—breaking it through a reaction with water—rather than merely binding to the target.
The computational workflow combines RFD2-MI with ProteinMPNN and AlphaFold 3, according to the paper’s design overview. The team tested 69 designs and found 13 that cleaved their targets, a success rate of 19%. For the active designs, the authors report catalytic rate enhancements of up to 30 million times the uncatalyzed background reaction.
Six crystal structures matched the designs
Structural tests offered a check on whether the proteins took the shapes the software proposed. Six crystal structures closely matched their design models, with reported Cα root mean squared deviations below 1.2 ångströms. The designed proteins also had folds unlike those of natural proteases, with structural similarity scores below 0.50.
The result matters because enzyme design has often been limited to model reactions with lower energy barriers. Peptide bonds are harder targets: their stability makes them useful for building proteins that hold together, but also makes them difficult to cleave without a biological catalyst. These designs show that computational methods can produce new protein scaffolds capable of catalysing that reaction, rather than relying only on natural enzyme structures.
Cell tests show activity, not a treatment
The designed proteases cleaved their intended targets in mammalian cells, the authors report. That moves the work beyond purified-enzyme tests and demonstrates activity in a more complex biological setting. The paper presents this as a possible starting point for synthetic biology, not evidence of a therapy or of safety in people.
The team also identifies a limit: the designed enzymes remain less active than the most effective natural cysteine proteases. Its analyses point to imperfect preorganization of the active site and the positioning of the substrate as sources of that gap. Improving those features will be necessary to raise activity while retaining the desired sequence selectivity.
The study’s protein structures are available through the Protein Data Bank under accessions 9YNL, 9YNM, 9YOX, 9YOY, 9YOZ and 9YP0. Those deposited structures provide a concrete record of six designed enzymes whose experimentally observed forms closely match their computational models.