InterPET: A Curated Benchmark of Sequence Embeddings and Graph Architectures with Interpretability and Biological Validation for PETase Activity Prediction

Motivation: Machine learning has emerged as a powerful accelerator for identifying PET-hydrolyzing enzymes (PETases). Yet, published models are often evaluated on benchmark performance alone, leaving their biological validity unexamined. Here we present InterPET, a curated benchmark and ablation study addressing both issues. Results: We aggregated sequences from four datasets (PlasticDB, PAZy, PlasticEnz, PEZY-miner), removing duplicate sequences, and filter data leakage, yielding a training set of 937 sequences and a benchmark of 139 sequences. Eight model configurations were trained and evaluated, spanning three embeddings (ESM-2, ProtT5, classical AAC/CTD descriptors), two tree-based classifiers (XGBoost, Random Forest), and two GraphSAGE variants differing in sequence-only and sequene plus 3D structure data. ESM-2 + XGBoost achieved the best performance (F1 = 0.91, AUC = 0.99, MCC = 0.90). SHAP-based feature attribution linked top-ranked AAC/CTD features (proline content, solvent accessibility, hydrophobicity) to known determinants of PETase activity, and cross-representation correlation showed that embedding-based models implicitly re-encode much of the same biophysical signal. However, in-silico mutagenesis revealed that the top-ranked M1 recovered only 0.5/3 catalytic-triad residues. These findings demonstrate that representation choice, classifier architecture, and evaluation criteria interact in ways a single leaderboard metric cannot capture. Availability and implementation: InterPET datasets and code are available at https://github.com/indiraprakoso/interpet/.