AI-Designed Enzymes Cut Peptide Bonds That Last Centuries, Preprint Reports

Researchers report designing enzymes from scratch that cut proteins at bonds meant to last for centuries. Ars Technica covered the work, and its summary is blunt: "AI design tools make proteins that break bonds that are stable for hundreds of years."

The numbers come from a bioRxiv preprint. Of 69 designs, 13 worked, a 19% success rate. The best enzymes reached rate enhancements over the background reaction (kcat/kuncat) of up to 3 × 10^7. Their folds differ sharply from natural proteases (TM score under 0.50). Six crystal structures landed close to the design models, with Cα RMSDs under 1.2 Å. The authors also say the proteases efficiently cleave their targets in mammalian cells.

The context: amide bonds along the peptide backbone are stable for hundreds of years in neutral water because of a high energy barrier to hydrolysis. The preprint says de novo enzyme design had been largely limited to low energy barrier model reactions. Cutting a backbone is a much harder test.

The gap to nature remains. The authors say the most active native cysteine proteases still win, because of imperfect active site preorganization and substrate positioning.

My take: this is the right kind of result, with a hit rate, a rate enhancement and crystal structures in the same abstract. What the material here does not say is which AI models generated the designs, or which proteins were cut in the cells. Those are the footnotes I want.