PAM: features, fine-tuned ESM-2 fold models and predictions for cryptic binding site prediction

Data, fine-tuned models and predictions accompanying the manuscript Tabular In-Context Learning for Cryptic Site Prediction (submitted to Bioinformatics). PAM predicts cryptic binding site residues on CryptoBench with TabICLv2 over 53 structural features and 64 principal components of a cross-fitted fine-tuned ESM-2 3B embedding. pam_features.zip: per-residue feature tables (381,704 residues, 1,136 apo chains), labels, folds and the LIGYSIS exclusion lists. pam_predictions.zip: per-residue predictions of PAM, of each fold model and of the released checkpoint of Škrhák et al.; cross-validation thresholds, bootstrap outputs and comparison tables. model_cleantest_{0..3}.pt: PyTorch state dicts of the four cross-fitted fine-tuned ESM-2 3B fold models. Code: https://github.com/Peter650-bot/tabicl_protein_cryptic_sites (MIT). README.md describes every file; bundles_manifest.json lists SHA-256 checksums.