Beijing University of Chemical Technology breaks through polyurethane recycling challenge in latest Science study

Deep learning is driving the discovery of novel enzyme catalysts to overcome challenges in polyurethane recycling. Globally, approximately 22 million tons of polyurethane are consumed annually, with its production being energy-intensive and emitting substantial greenhouse gases. Recycling efforts face significant hurdles due to the cross-linked structure and chemically stable urethane bonds. Particularly, polyether-type polyurethane, which accounts for 75% of the foam market, is difficult to biodegrade due to the lack of effective degrading enzymes. Currently, glycolysis stands as the most promising chemical recycling method in industry, yet its byproduct—a mixture rich in aromatic carbamate—often gets incinerated as hazardous waste, causing environmental and economic burdens. Although chemical enzymatic recycling shows potential, existing urethane enzymes exhibit low efficiency under industrial glycolysis conditions and poor solvent compatibility, limiting their practical application.
Recently, a research team led by Wu Bian and Cui Yinglu at Beijing University of Chemical Technology has developed a Graph Neural Network (GNN) framework named GRASE. By integrating self-supervised and supervised learning, the team successfully identified highly efficient glycolysis-compatible urethaneases. Notably, ADPURase demonstrated activity two orders of magnitude higher than known enzymes in 6 mol of diethylene glycol, achieving near-complete depolymerization of kilogram-scale commercial polyurethane within 8 hours. Structural analysis revealed that its tightly packed hydrophobic core and proline-stabilized "cap-like structure" may be key factors maintaining high stability and efficiency in harsh solvents. This study highlights the powerful capabilities of deep learning in accelerating the discovery of industrial-potential biocatalysts. The related paper, titled "Glycolysis-compatible Urethanases for Polyurethane Recycling," was published in Science. The first author is Special Research Assistant Chen Yanchun, with co-first authors being doctoral students Sun Jinyuan and Shi Kelun.

The research team first systematically evaluated the hydrolytic activity of 14 reported amylases toward aromatic carbamates. Results showed that only Aes72, GatA, and SP1 exhibited activity. Aes72 demonstrated the highest hydrolytic activity toward 2,4-TDA-DEG at 0.06 U/mg, preferentially hydrolyzing the ortho-acyl bond. Although SP1 exhibited broader regional selectivity, its catalytic efficiency was lower. Given that 2,4-TDA-DEG dominated the substrate, Aes72 was identified as the ideal template for further development.

Figure 1. Hydrolysis pathway of TDA-DEG and reported amylase activities (A) 2,4-TDA-DEG hydrolyzes to intermediates 2-TAC and 4-TAC, which further hydrolyze to 2,4-TDA. (B) 2,6-TDA-DEG hydrolyzes to intermediate 6-TAC, which further hydrolyzes to 2,6-TDA. (C, D) Published amylase activities for 2,4-TDA-DEG (C) and 2,6-TDA-DEG (D). Reactions were conducted in 100 mM sodium phosphate buffer (pH 7.5), with all enzymes incubated at 50°C except SP1 which was kept at 37°C to prevent thermal inactivation. Data are presented as mean ± standard deviation (n=3). To overcome limitations of traditional screening methods, the research team developed the GRASE framework integrating two neural network models: Pythia-Pocket for structural embedding of active site residues (assessing functional similarity via cosine similarity) and Pythia for protein stability prediction. GRASE automatically screened 24 candidate enzymes from the database, with 21 successfully expressed and demonstrating amylase activity. Eight enzymes outperformed Aes72, including ADPURase that showed 32-fold and 62-fold activity enhancement for 2,4-TDA-DEG and 2,6-TDA-DEG respectively, with particularly significant activity improvement in 6 M diethylene glycol.

Figure 2. Model Architecture and Functional Validation of GRASE-Identified Enzymes (A) Architectures of Pythia-Pocket and Pythia. Both tools abstract protein structures into graphical representations: Pythia-Pocket processes full-protein graphs to predict ligand-binding residues, while Pythia predicts amino acid distribution based on local structural context. (B) GRASE workflow: Pythia-Pocket calculates cosine similarity between candidate enzymes and the active site residues of query enzymes (e.g., Aes72), followed by Pythia's protein stability assessment to recommend enzymes with balanced activity-stability profiles. (C) Enzymatic activity evaluation for 2,4-TDA-DEG and 2,6-TDA-DEG. The analysis includes sequence search, structural retrieval, sequence clustering, and GRASE-filtered candidate enzymes. Undeveloped enzymes are underlined. (D) Heatmap of GRASE-filtered enzymes hydrolyzing TDA-DEG products at 30–60°C. Green indicates total product (TAC intermediates + TDA monomers), purple represents TDA monomers.
In the chemical enzymatic depolymerization experiment, ADPURase achieved 98% substrate conversion within 12 hours at 50°C with an enzyme loading of 8 mg/g, completely hydrolyzing the TAC intermediate into TDA monomer. In contrast, SP1 and Aes72 demonstrated conversion rates of only 15% and 26%, respectively. At the kilogram scale, ADPURase achieved 95% conversion within 8 hours and 98% within 12 hours at an 8 mg/g loading, with TDA and diethylene glycol recovery rates reaching 94.7% and 98.5%, respectively, demonstrating excellent industrial feasibility.

Figure 3. Chemical Enzymatic Depolymerization of BASF Polyurethane Foam (A) Kilogram-scale polyurethane foam glycolysis process: Glycolysis produces an upper phase rich in polyether polyols, a lower phase containing DEG and aromatic compounds, and insoluble solids. The pie chart shows the composition of the lower phase. ADPURase hydrolyzes TDA diaminoacetate into TDA monomers, with TDA and DEG recovered through distillation. (B) Hydrolysis kinetics of 30g foam glycolysis lower phase under ADPURase (50°C), Aes72 (37°C), and SP1 (30°C). The pie chart displays the relative proportions of TDA-DEG, TAC, and TDA at each time point. (C) Left: Hydrolysis curves of kilogram-scale foam lower phase under 8 mg/g and 16 mg/g ADPURase; Right: Time-dependent changes in TDA-DEG, TAC, and TDA composition at 8 mg/g enzyme loading. Structural comparisons reveal that ADPURase shares an open "cap-like structure" with other high-activity enzymes, forming a V-shaped active pocket conducive to binding large substrates. Low-activity enzymes, however, exhibit narrow active sites or lack cap-like structures. Kinetic analysis indicates that ADPURase's amidease activity far exceeds its esterase activity, which is the primary reason for its efficient hydrolysis of amide bonds. In 6M diethylene glycol, ADPURase shows minimal melting temperature changes and structural stability, while Aes72 demonstrates significant destabilization. Molecular dynamics simulations further confirmed that the tight hydrophobic core and proline-stabilized ring structure of ADPurase effectively resist solvent penetration and maintain structural integrity.

Figure 4. Enzyme Structure Comparison and Kinetic Parameters (A) Surface structure and cap-like region of Aes72. (B) Local structural differences in active sites of different enzymes, highlighting the critical role of cap-like structures in catalytic performance. Highly active enzymes (>1 U/mg) exhibit open cap-like structures, while low-active enzymes (<0.1 U/mg) show restricted or absent cap-like structures. (C) Comparison of Michaelis-Maxwell kinetic parameters for 2,4-TDA-DEG, NAMA, and MAMB between ADPURase and Aes72. Data are presented as mean ± standard deviation (n=3).
This study not only successfully identified highly efficient amylase with industrial application potential, but also highlighted the advantages of deep learning in enzyme function discovery. The GRASE framework achieves precise functional prediction of low-sequence-homologous enzymes through local structural embedding, providing a new paradigm for large-scale discovery of functionally unknown enzymes. In the future, further capturing and amplifying enzymes' "hidden functions" will deepen our understanding of enzyme diversity and promote the application of biocatalysts in broader industrial scenarios.