Functional Enhancement of Flavin-Containing Monooxygenase through Machine Learning Methodology

Directed evolution of enzymes often fails to obtain desirable variants because of the difficulty in exploring a huge sequence space. To obtain active variants from a very limited number of variants available at the laboratory scale, machine learning (ML)-guided engineering of enzymes is becoming an attractive methodology. However, as far as we know, there is no example of an ML-guided functional modification of flavin-containing monooxygenase (FMO). FMOs are known to catalyze a variety of oxidative reactions and are involved in the biosynthesis of many natural products (NPs). Therefore, it is expected that the ML-guided functional enhancement of FMO can contribute to the efficient development of NP derivatives. In this research, we focused on p-hydroxybenzoate hydroxylase (PHBH), a model FMO, and altered only four amino acid residues around the substrate binding site. ML models were trained with a small initial library covering only approximately 0.1% of the whole sequence space, and the ML-predicted second library was enriched with active variants. The variant with the highest activity in the second library was PHBH-MWNL (V47M, W185, L199N, and L210), whose activity was more than 100 times that of the wild-type PHBH. For elucidation of the mechanism of the observed activity enhancement, the crystal structure of PHBH-MWNL in complex with 4-hydroxy-3-methyl benzoic acid was determined. In the PHBH-MWNL crystal structure, the missing water molecule WAT2 was observed due to N199 hydrogen-bonding to WAT2, indicating that the L199N mutation contributed to the observed functional improvement by stabilizing the proton relay network proposed to be important in catalysis.