Scientists degrade PU with AI-designed enzyme

A paper published in the journal Science has explained how a team of researchers used neural networks to help them develop an enzyme that break down PU foam in as little as eight hours. The team was led by Yanchun Chen at the College of Life Sciences and Technology, Beijing University of Chemical Technology.

Conventional recycling methods have long struggled with PU’s chemical stability and cross-linked structure, which make depolymerisation difficult. To overcome this, the researchers created GRASE (Graph Neural Network–based Recommendation of Active and Stable Enzymes), a machine-learning system that combines self-supervised and supervised learning to predict which enzymes can remain active under harsh glycolysis conditions.

According to a report in Ars Technica  the scientists built GRASE using Pithia-Pocket an artificial intelligence (AI) model used in enzyme design and protein engineering. It was developed to predict and optimise the “active pockets”—the regions of enzymes that bind to and act on specific molecules (substrates).

Using this approach, the team discovered AbPURase, an enzyme that demonstrated activity levels two orders of magnitude higher than previously known urethanases when exposed to 6 mol diethylene glycol – a solvent typical of industrial polyurethane recycling. The new enzyme enabled near-complete depolymerisation of commercial polyurethane at kilogram scale within eight hours.

Structural analysis revealed that AbPURase’s stability and efficiency stem from a tightly packed hydrophobic core and a proline-stabilised lid loop – features that help it maintain activity in aggressive solvents. The researchers say these findings could pave the way for enzyme-assisted chemical recycling processes capable of handling the rigid, cross-linked PU materials used in insulation, coatings and elastomers.

The authors note that GRASE represents an important step toward integrating deep learning and biocatalysis for polymer recycling. By screening vast enzyme libraries and predicting stability under industrial conditions, the AI-driven system dramatically reduces the time and experimental cost required to find suitable biocatalysts.

As an illustration of AI's use in the creation of this enzyme, the above image of polyurethane foam was created using generative AI.

Science Magazine