Deep learning-guided de novo design using RFDiffusion2-MI generates functional cysteine proteases with novel folds (TM-score < 0.50) that achieve rate enhancements up to 3 × 10⁷-fold over uncatalyzed peptide bond hydrolysis through computationally specified catalytic triads and substrate binding sites, with crystal structures confirming atomic-level accuracy (Cα RMSD < 1.2 Å) of the design methodology.

Despite advances in de novo enzyme design, success has been largely limited to low energy barrier model reactions. Amide bonds such as those linking amino acids along the peptide backbone are stable for hundreds of years in neutral aqueous solution because of the high energy barrier to hydrolysis. Here we describe the use of a new deep learning method, RFD2-MI, to de novo design enzymes which utilize an activated cysteine nucleophile to hydrolyze the polypeptide backbone in a sequence-dependent manner, achieving rate enhancements over the background reaction (kcat/kuncat) of up to 3 × 107. The generated designs have folds very different from the proteases in nature (TM score < 0.50), and crystal structures are very close to the design models (Cα RMSDs < 1.2 Å), highlighting the accuracy of the design methodology. Our approach has broad utility for advancing the design of novel proteases for both biotechnical and medical applications.