GNN-based Recommendation of Active and Stable Enzymes (GRASE)

Enzymatic recycling has emerged as a promising strategy for sustainable plastics, yet thermoset plastics like polyurethane remain challenging due to their resistance to mechanical amorphization. Although chemo-enzymatic degradation provides valuable insights, inefficient biocatalysts restrict their compatibility within industrial glycolysis. Here, we developed GRASE, a framework integrating self-supervised and supervised graph neural networks to identify a class of urethanases. Among these, AsPURase exhibited catalytic activity two orders of magnitude higher than the most active known urethanases in 6 M diethylene glycol (DEG), facilitating near-complete kilogram polyurethane depolymerization within 12 hours. Structural analysis further revealed the molecular adaptations underlying AsPURase's remarkable performance, unveiling an unrecognized multifunctionality. This study demonstrates how deep learning algorithms integrate chemical and enzymatic methods to overcome long-standing barriers in PUR recycling.