Integrating graph learning and meta-learning to enhance PET hydrolase activity at elevated temperatures

The identification and design of efficient biocatalysts to address plastic waste accumulation are pressing global challenges. Leveraging recent advancements in artificial intelligence (AI), we developed an integrated machine-learning framework incorporating the geometric vector perceptron-graph neural network and learn2learn models to yield robust and thermostable variants based on three representative polyethylene terephthalate hydrolases (i.e., FAST-PETase, DepoPETase, and BHETase). The engineered hydrolase variants demonstrated up to a 5.5-fold improvement in hydrolytic activity across a temperature range of 40°C–70°C compared to the wild-type enzymes. Molecular dynamics simulations revealed key microenvironmental interactions sustaining hydrolytic activity at high temperatures—enhanced hydrogen bonding and hydrophobic interactions around active sites—and highlighted the unbound dynamics of bis(2-hydroxyethyl) terephthalate and mono-(2-hydroxyethyl) terephthalate near catalytic sites, motivating a two-enzyme system that achieved a 1.5-fold increase in terephthalic acid production. This study demonstrates a scalable AI-guided strategy for evolving robust polyethylene terephthalate hydrolases for industrial biodegradation. • Two-round machine learning discovers thermostable PET hydrolase variants • Meta-learning improves mutation compatibility and catalytic efficiency • MD-driven two-enzyme system achieves 1.5-fold increase in TPA yield Li et al. present a two-round machine-learning framework integrating GVP-GNN and meta-learning to evolve robust PET hydrolases. The engineered variants exhibit enhanced thermostability and catalytic efficiency, while a two-enzyme system achieves a 1.5-fold increase in terephthalic acid production from PET.