Machine Learning-Guided Selection of Cyclodextrins for Enhanced Biosynthesis and Capture of Volatile Terpenes

Abstract

Nootkatone and limonene are valuable volatile organic compounds (VOCs), but their biosynthetic production is hindered by volatility. This study employed machine learning to guide cyclodextrin (CD) selection for encapsulating these VOCs, with a focus on nootkatone capture during fermentation to prevent losses and potentially replace dodecane as an organic solvent extractant. A LightGBM model accurately predicted complexation free energies (ΔG) between CDs and guest molecules (R² = 0.80 on a 10% test set, with a mean absolute error of 1.31 kJ/mol and a root-mean-squared error of 1.90 kJ/mol). Experimental ranking of 7 CD types validated the model’s ΔG predictions and encapsulation performance rankings. Nootkatone showed high encapsulation efficiencies ranging from 21.29% (α-CD) to 88.41% (Me-β-CD), capturing 22.61–116.71 mg/g CD. Notably, Hp-γ-CD, which is the least studied or used CD in research, performed well with nootkatone (63.64%, 84.01 mg/g CD) despite model discrepancies. For limonene, encapsulation efficiencies spanned from 0.62% (Hp-γ-CD) to 55.45% (β-CD), with 0.61–84.28 mg/g CD encapsulated. Constructed engineered Saccharomyces cerevisiae strains produced nootkatone (up to 97.30 mg/L captured by 10 mM Me-β-CD) from de novo fermentation using glucose as a carbon source. This approach demonstrated the potential of CDs to replace dodecane as an organic solvent for terpene extraction during fermentation. The study highlights machine learning’s potential for guiding CD selection to enhance volatile terpene biosynthesis, capture, and utilization during fermentation, offering a more environmentally friendly alternative to traditional organic solvent-based extraction methods.