Chen_2025_Science_390_503

Recycling thermoset polyurethanes is hindered by their cross-linked structures and chemically stable urethane bonds. Although chemo-enzymatic approaches offer promise, known urethanases remain inefficient under industrial glycolysis conditions. Here, we present GRASE [graph neural network (GNN)-based recommendation of active and stable enzymes], a GNN-based framework that integrates self-supervised and supervised learning to identify efficient, glycolysis-compatible urethanases. Among these, AbPURase exhibited two orders of magnitude greater activity than previously known enzymes in 6 molar diethylene glycol, enabling near-complete depolymerization of commercial polyurethane at kilogram scale within 8 hours. Structural analysis revealed that a tightly packed hydrophobic core and proline-stabilized lid loop may confer AbPURase's stability and efficiency in harsh solvents. This work highlights how deep learning accelerates the discovery of biocatalysts with industrial potential and addresses a critical barrier in polyurethane recycling.