This study presents a computational approach for designing serine hydrolases with high catalytic efficiency. The researchers used RFdiffusion to generate proteins with complex catalytic sites and ChemNet to assess active site geometry and preorganization across the reaction coordinate. Experimental characterization revealed novel serine hydrolases that catalyze ester hydrolysis with catalytic efficiencies up to 3.8 × 10³ M⁻¹ s⁻¹, closely matching design models and having distinct folds from natural serine hydrolases. In silico selection based on active site preorganization significantly increased success rates, enabling identification of new catalysts in screens of as few as 20 designs. The de novo design approach provides insights into geometric determinants of catalysis, complementing structural and mutational studies of native enzymes. The study also demonstrates the ability of generative deep learning methods to find completely new solutions to design challenges. The results show that the designed catalysts have catalytic efficiencies up to 10³ M⁻¹ s⁻¹, which is significantly higher than previous designs but still orders of magnitude slower than native enzymes. The study highlights the importance of preorganization and catalytic geometry in enzyme design and shows that the combination of RFdiffusion and ChemNet can enable the design of complex enzymes that catalyze multi-step transformations. The findings suggest that future design efforts should focus on optimizing the oxyanion hole residues to more preferentially stabilize the transition state over the sp² ground state. The study also demonstrates the potential of ChemNet to rapidly generate ensembles for reaction intermediates, providing structural insights that would otherwise require labor-intensive studies. The results indicate that the designed catalysts are more active than previous de novo designed serine hydrolases but still less efficient than native enzymes, highlighting the need for further optimization. The study provides a roadmap for the design of industrially relevant serine hydrolases and complex enzymes that catalyze multi-step transformations.This study presents a computational approach for designing serine hydrolases with high catalytic efficiency. The researchers used RFdiffusion to generate proteins with complex catalytic sites and ChemNet to assess active site geometry and preorganization across the reaction coordinate. Experimental characterization revealed novel serine hydrolases that catalyze ester hydrolysis with catalytic efficiencies up to 3.8 × 10³ M⁻¹ s⁻¹, closely matching design models and having distinct folds from natural serine hydrolases. In silico selection based on active site preorganization significantly increased success rates, enabling identification of new catalysts in screens of as few as 20 designs. The de novo design approach provides insights into geometric determinants of catalysis, complementing structural and mutational studies of native enzymes. The study also demonstrates the ability of generative deep learning methods to find completely new solutions to design challenges. The results show that the designed catalysts have catalytic efficiencies up to 10³ M⁻¹ s⁻¹, which is significantly higher than previous designs but still orders of magnitude slower than native enzymes. The study highlights the importance of preorganization and catalytic geometry in enzyme design and shows that the combination of RFdiffusion and ChemNet can enable the design of complex enzymes that catalyze multi-step transformations. The findings suggest that future design efforts should focus on optimizing the oxyanion hole residues to more preferentially stabilize the transition state over the sp² ground state. The study also demonstrates the potential of ChemNet to rapidly generate ensembles for reaction intermediates, providing structural insights that would otherwise require labor-intensive studies. The results indicate that the designed catalysts are more active than previous de novo designed serine hydrolases but still less efficient than native enzymes, highlighting the need for further optimization. The study provides a roadmap for the design of industrially relevant serine hydrolases and complex enzymes that catalyze multi-step transformations.