µORScreen: a lightweight consensus modeling framework for µ-opioid receptor ligand prediction and virtual screening

Opioid use disorder (OUD) remains a major public health challenge, and the human µ-opioid receptor (µOR) is a central target in opioid pharmacology. Here, we report a reproducible ligand-based workflow for µOR antagonist classification named µORScreen which integrates rigorous split design, systematic model benchmarking, interpretation, virtual screening and lightweight local deployment. A curated set of 982 human µOR ligands was partitioned under three complementary strategies (similarity-based, scaffold-based, and random-based), each with a held-out test set and five train/validation folds. On the test evaluation, LightGBM (ECFP4 with RDKit 2D descriptors) generalized best (AUROC 0.714), closely followed by TabPFN (0.705) and Random Forest (0.696). The three top-ranked models were combined into a consensus classifier that prioritized unanimously predicted compounds as high-confidence antagonist-like candidates. Applied to GPCRdb, ZINC, REINVENT, and OUROBOROS, µORScreen revealed pronounced source dependence, with the strongest enrichment of antagonist-like candidates in GPCRdb. On an independent set of 17 non-overlapping, literature-derived ligands (10 antagonists, 7 non-antagonists), the consensus achieved a balanced accuracy of 0.68. SHAP analysis attributed predictions to a concentrated subset of fingerprint features, and the workflow was deployed as a web server supporting SMILES-based prediction and RF-based SHAP analysis. µORScreen thus provides a computationally efficient, openly accessible framework for early-stage µOR ligand prioritization and external-library triage.