Computational design of enzyme repertoires

Antibodies are produced to target any antigen using a finite set of gene fragments, generating a huge diversity (>10¹⁰) distinct structures. In contrast, we are unaware of a system that can produce analogous diversity in enzymes. Inspired by antibody repertoires, I have developed the first strategy to design, synthesise, and experimentally test repertoires comprising millions of enzymes. Using evolution-guided atomistic design simulations, I designed thousands of protein fragments that exhibited high structure and sequence diversity, including within the active-site pocket, which can be genetically assembled into full-length enzymes. I also developed an ML-based algorithm to select a subset of the designed fragments that would give rise to stable and active proteins. Applied to a family of xylanases (sophisticated enzymes which are critical in biomass degradation) I designed a repertoire comprising a million enzymes at a cost of 0.3¢ per enzyme. Screening with an activity-based probe revealed thousands of functional xylanases based on nearly 1,000 unique backbones. Advanced machine-learning methods uncovered important elements that discriminate active from inactive designs, enabling us to design even more effective enzyme repertoires targeting, in principle, any desired substrate. Thus, enzyme repertoire design will enable a new generation of highly efficient and selective enzymes, while teaching us essential rules in biomolecular design.