Abstract
<jats:p>Biocatalytic depolymerization of polyethylene terephthalate (PET) to maximize bis(2-hydroxyethyl) terephthalate (BHET) is a novel strategy to enable efficient PET resynthesis in closed-loop plastic recycling. However, realizing this route is challenging as BHET production during enzyme-catalyzed degradation is intrinsically limited by a yield-selectivity trade-off as formation of the mono (2-hydroxyethyl) terephthalate (MHET), and the monomers terephthalic acid (TPA) and ethylene glycol (EG) needs to be suppressed. To overcome this challenge, we devised a multi-objective Bayesian optimization (BO) strategy that iteratively learns from biocatalytic data obtained for the gold standard LCCICCG enzyme and predicts reaction conditions to simultaneously maximize the BHET yield (YBHET) and selectivity (SBHET). From a full grid-based design of 3,780 possible experiments, the framework identified two Pareto-optimal solutions: YBHET = 11.7 mgBHET gPET−1, SBHET = 56% and YBHET = 8.1 mgBHET gPET−1, SBHET = 68% in five optimization cycles comprising only 25 experiments. With a high predictive accuracy (R2 > 0.96), this approach doubled the yield, increased the selectivity by approximately 20%, and exhibited 9-fold experimental efficiency compared to manual screening. Feature-importance analysis identified EG concentration as the key factor governing the yield-selectivity trade-off, and was complemented by Michaelis-Menten-based kinetic model for mechanistic interpretation. Finally, the framework was validated on the PES-H1/PHL7 enzyme for predicting YBHET, demonstrating transferability across PET hydrolases. Overall, this study establishes a generalizable and data-driven multi-objective optimization workflow for accelerating reaction optimization for selective enzymatic plastic recycling.</jats:p>