PANTS: a triaged catalogue of candidate PET-degrading enzymes for therapeutic use

PANTS mines metagenomic sequence space for polyester hydrolases and triages them for therapeutic use: degrading PET at 37 °C and neutral pH, in serum, rather than in an industrial reactor above PET's glass transition. Version 0.2.0 completes the structure set and runs the full evaluation protocol.

14,804,920 predicted proteins scanned across five environments, 439 candidates retained (0.003%), 416 with a predicted structure and 402 with measured active-site geometry. The reference set holds 1140 characterised enzymes: 855 positives of which 342 carry a published measurement, 131 hard negatives and 153 near misses, with 75 activity measurements (73 carrying a DOI).

New in 0.2.0. The reference structures are complete at 354, of which 60 are experimental crystal structures rather than 12: the builder had always preferred a deposit over a model, but the PAZy import left the PDB identifier field empty, so that preference had never once fired. Linking UniProt's cross-references and ranking them by exact sequence match, then divergence, then resolution, recovered them.

The main result is negative, and it is the same result three independent ways. Sequence embeddings separate polyesterases from other folds at AUC 0.975 and PET-active from PET-inactive polyesterases at 0.493. Active-site geometry looks convincing raw (cleft depth AUC 0.808) and falls to 0.534 under cluster-grouped splitting — reproducing, on 342 enzymes, the 0.533 measured on 131. And the learned head, trained on labels somebody actually measured, scores 0.921 against a nearest-known-PETase retrieval baseline of 0.931: it does not beat looking the answer up. Leave-one-family-out is not evaluable at all, because every ESTHER family here is wholly positive or wholly negative. This dataset is label-limited, not method-limited. The negative class — polyesterases measured on a different plastic — numbers 26.

A methodological finding worth reusing. Predicted and experimental coordinates are not interchangeable. Paired within the same protein (n=49), the oxyanion hole differs systematically: the second donor's angle is 23.6° in the crystal against 15.6° in the model (p 1e-06). Predictions build a tighter, more idealised active site than the protein has. Cleft depth is source-invariant. Pooling sources without accounting for this lowered the activity AUC from 0.749 to 0.553; the source column exists so it can be accounted for.

Every number here was recomputed from the deposited CSVs by scripts/build_release.py, never copied from prose. Data CC BY 4.0; source code MIT.