Deep Learning-Driven Enzyme Engineering in Pichia pastoris: A Sustainable Platform for High-Yield Nootkatone Biosynthesis

Nootkatone, a valuable sesquiterpene with broad bioactivities and application potential, faces yield limitations in microbial synthesis due to metabolic and enzymatic inefficiencies. In this study, we present an advanced strategy combining metabolic engineering and deep learning-guided enzyme design to optimize nootkatone production in Pichia pastoris. By systematically modifying the mevalonate pathway, optimizing cofactor supply, and minimizing competing metabolic pathways, a robust yeast strain producing 702.15 mg/L valencene was developed. To facilitate the efficient conversion of valencene to nootkatone, we applied ancestral sequence reconstruction (ASR) to identify hotspot amino acid residues, guiding the design of a variant library. The deep learning model DLKcat was then used to conduct virtual saturation mutagenesis screening on library sites, predicting their enzyme turnover number (k_cat). The engineered cytochrome P450 (HPO) variant H54A exhibited the highest activity, with catalytic performance 2.3 times that of the initial. Furthermore, the implementation of intermittent feeding fermentation significantly elevated the final nootkatone yield to 3365.36 mg/L, the highest reported to date. This study provided a green platform for an alternative sustainable access of high-value nootkatone, and exemplifies the potential of machine learning in optimizing metabolic pathway enzymes for efficient biosynthesis of other bioactive terpenoids in microbial systems.