Combining Machine Learning and Directed Evolution for Optimization of a Monooxygenase

L-3,4-dihydroxyphenylalanine, or L-Dopa, serves as a crucial pharmaceutical for treating Parkinson's disease and as a precursor for various catechol-containing compounds. The flavin-dependent monooxygenase HpaBC has shown potential in producing microbial L-Dopa, yet it lacks substantial native activity towards L-tyrosine. Researchers have turned to computational models and iterative engineering techniques to enhance the enzyme's performance.

To evaluate the effectiveness of machine learning (ML) models in predicting activity-enhancing mutations in HpaBC, the study benchmarked multiple ML algorithms. These models were compared based on their predictive capabilities for improving enzyme activity. Experimentally validated single mutants served as seeds to generate new variants using EVOLVEpro, a directed evolution software.

The team then assessed the impact of incorporating directed evolution-derived variants into the combinatorial predictions made by EVOLVEpro. This approach allowed for the exploration of an expanded sequence space, combining both ML-derived and directed evolution-derived mutations. The resulting HpaBC variants exhibited significantly improved activity levels.

Interestingly, while the integration of directed evolution data altered the predicted mutation trajectories of EVOLVEpro, both ML and directed evolution training strategies ultimately converged on enzyme variants with comparable levels of activity. This finding demonstrates that different regions of the sequence space can yield equally optimized enzymes.

In summary, this study provides a comprehensive comparison of zero-shot ML models and establishes an iterative framework for integrating machine learning techniques with directed evolution approaches. By combining these powerful tools, researchers can more efficiently accelerate enzyme engineering efforts, ultimately leading to the development of improved biocatalysts for industrial applications.