A responsible and sustainable circular economy of polymers requires efficient recycling processes with a low CO₂ footprint. Enzymatic depolymerization of polyethylene terephthalate (PET) is a first step to make PET polymers a part of a circular economy of polymers. In this study, a structure-based deep learning model was utilized to identify residues in TfCut2 that are responsible for improved hydrolytic activity and enhanced stability. Machine learning guided design identified novel beneficial positions (L32E, S35E, H77Y, R110L, S113E, T237Q, R245Q, and E253H), which were evaluated and stepwise recombined yielding finally the beneficial variant L32E/S113E/T237Q. The latter TfCut2 variant exhibited improved PET depolymerization when compared with TfCut2 WT (amorphous PET film, 2.9-fold improvement // crystalline PET powder (crystallinity > 40%), 5.3-fold improvement). In terms of thermal resistance the variant L32E/S113E/T237Q showed a 5.7 °C increased half-inactivation temperature (T₅₀⁶⁰). The PET-hydrolysis process was monitored via a quartz crystal microbalance with dissipation monitoring (QCM-D) in real-time to determine depolymerization kinetics of PET coated onto the gold sensor. Finally, conformational dynamics analysis revealed that the substitutions induced a conformational change in the variant L32E/S113E/T237Q, in which the dominant conformation enabled a closer contact between the catalytic site and PET resulting in increased PET-hydrolysis. Overall, this study demonstrates the potential of deep learning models in protein engineering for identifying and designing efficient PET depolymerization enzymes.

负责任且可持续的聚合物循环经济需要低 CO ₂足迹的高效回收工艺。聚对苯二甲酸乙二醇酯 (PET) 的酶解聚是使 PET 聚合物成为聚合物循环经济一部分的第一步。在这项研究中，利用基于结构的深度学习模型来识别TfCut2中负责改善水解活性和增强稳定性的残基。机器学习引导设计确定了新的有益位置（L32E、S35E、H77Y、R110L、S113E、T237Q、R245Q 和 E253H），对这些位置进行评估并逐步重组，最终产生有益变体 L32E/S113E/T237Q。与相比，后者的TfCut2变体表现出改进的 PET 解聚TfCut2 WT（非晶PET薄膜，提高2.9倍//结晶PET粉末（结晶度> 40%），提高5.3倍）。就热阻而言，变体 L32E/S113E/T237Q 的半失活温度增加了 5.7 °C ( T ₅₀ ⁶⁰）。通过具有耗散监测功能的石英晶体微天平 (QCM-D) 实时监测 PET 水解过程，以确定涂在金传感器上的 PET 的解聚动力学。最后，构象动力学分析表明，取代引起了变体 L32E/S113E/T237Q 的构象变化，其中主导构象使催化位点和 PET 之间的接触更紧密，从而导致 PET 水解增加。总的来说，这项研究证明了蛋白质工程中深度学习模型在识别和设计高效 PET 解聚酶方面的潜力。