New research represents a notable step forward in designing enzymes from scratch. With their approach, the authors designed an enzyme that uses a covalent intermediate to catalyze a two-step reaction, analogous to what many proteases do when breaking apart proteins. This establishes a framework for engineering enzymes capable of facilitating complex, multistep reactions. The rational design of enzymes capable of catalyzing arbitrary chemical reactions remains a formidable challenge in computational protein engineering. Traditional approaches rely on placing active sites within pre-existing protein scaffolds, which often limits catalytic efficiency due to constraints in structural flexibility and active site preorganization. Although chemical interventions have circumvented these challenges, the efficiency of the initial computational designs remains far below the range observed for natural enzymes. However, recent advances in deep learning have provided new opportunities for designing proteins from the ground up, specifically to scaffold complex active sites like those found in serine hydrolases – the largest known class of enzymes. Here, Anna Lauko and colleagues present PLACER (Protein-Ligand Atomistic Conformational Ensemble Reproduction) – a novel machine learning network that predicts the precise atomic structures of enzyme active sites by considering the protein backbone, amino acid identities, and the chemical structures of bound molecules. By employing RFdiffusion to create proteins with complex catalytic sites and PLACER to evaluate these proteins’ active site organization, Lauko et al. successfully designed functional serine hydrolase enzymes capable of efficiently catalyzing ester hydrolysis from minimal active site specification. Moreover, they used their approach to uncover new catalysts in low-throughput screens, yielding five distinct enzyme folds unlike any found in natural serine hydrolases.
