DELM: Predicting Enzymatic Degradation of Plastics with Language Models

DELM[1] predicts whether an enzyme degrades a given plastic, directly from the enzyme's amino-acid sequence — no structure required. A protein language model (ESM2) encodes the enzyme and a polymer language model (polyBERT) encodes the plastic; a cross-attention head fuses them to output a degradation probability. Paste an enzyme and DELM ranks which plastics it is most likely to degrade.

GitHub | Contact

Run a Prediction

Paste one or more enzyme sequences in FASTA format or upload a FASTA file, choose a plastic (or rank all of them), and submit. Each run opens a dedicated, bookmarkable result page.

Up to 1,500 residues per job. For longer sequences, please contact us.
or
Accepts .fasta, .fa, .txt, or .seq.

Predictions currently run on a CPU server. The enzyme embedding is computed once and scored against each plastic, so ranking all plastics is nearly as fast as scoring one. The result page shows live progress, can be bookmarked or shared, and is retained for one week.

How DELM Works

Protein sequences and polymer PSMILES are encoded by two pretrained language models — ESM2[2] for the enzyme and polyBERT[3] for the plastic. A cross-attention module lets protein residues and polymer atoms attend to one another; the fused representation is pooled and passed to a classifier that outputs the degradation probability. Predictions are averaged over five cross-validation folds.

3D structure of a PET hydrolase with high cross-attention residues highlighted in red.
Figure 1: High cross-attention residues mapped onto a PET hydrolase structure (UniProt A0A0K8P6T7). DELM highlights catalytic residues (e.g. Ser160, Asp206) from sequence alone, indicating it captures structural and functional signal.

Performance

Across nine protein language models, medium-sized models give the best test AUC; the largest (3B/15B) do not improve accuracy on this dataset. DELM therefore uses ESM2-t33-650M, which is both accurate and light enough to serve on CPU.

Bar chart of 5-fold validation and test AUC for nine protein language models; ESM2_T33 has the best test AUC.
Figure 2: Five-fold cross-validation AUC for different protein language models. ESM2-t33-650M (ESM2_T33) achieves the best test AUC.

Supported Plastics

DELM scores enzymes against the plastics it was trained on:

PETPCL PHBPEF PBSPLA PESPBSA PHVPE
Repeating-unit structures (PSMILES) of the polymers used to train DELM.
Figure 3: Repeating units (PSMILES) of the polymers in the training set.

References

Please contact shijie.xu@ees.hokudai.ac.jp for any questions.

Changelogs