μOR-ligand: target-aware view-based hybrid feature selection for μ-opioid receptor ligand functional classification.

Understanding active functional class (agonist vs antagonist) at the human μ-opioid receptor (μOR) is critical for drug discovery and safety assessment. While recent machine learning models such as ExtraTrees (ET) and message-passing neural networks (MPNNs) achieved ROC AUC scores of 0.915 ± 0.039 and 0.918 ± 0.044, respectively, it remains unclear how target-conditioned interaction features influence functional class detection and how resampling choices (e.g., SMOTE) impact robustness when evaluated under identical, fixed splits. Therefore, we introduce the μOR-Ligand framework-a target-aware view-based hybrid feature selection to improve performance in identifying whether an active ligand is an agonist or antagonist. To realize μOR-Ligand, three views have been constructed: (1) fingerprint, (2) ligand descriptors, and (3) molecular interaction features, yielding a comprehensive feature space of 114,552 variables (1190 fingerprints, 1618 ligand descriptors, 111,741 interaction descriptors). Feature selection is performed per view to obtain three view-specific subsets; each trains a base learner, and their out-of-fold predictions are fused via a linearly weighted multimodel feature selection stage. In parallel, the three selected feature sets are merged and trained with a stacking model (ensemble feature selection). Finally, μOR-Ligand forms a view-based hybrid feature selection by linearly combining the multimodel and ensemble outputs. Such a target-aware view-based hybrid feature selection for the stacked ensembles framework achieved an improved ROC AUC of 0.930 ± 0.026, supported by a promising significant p-value of 0.046 and a t-statistic of 1.707 (> t-critical=1.663) against the recent model, MPNNs. Also, μOR-Ligand further increased ROC AUC to 0.977 on internal cross-validation, as the highest ROC AUC score. In addition, μOR-Ligand is evaluated under a resampling-controlled μOR evaluation protocol that pairs ± SMOTE on identical, fixed splits. Overall, the study (1) demonstrates that target-aware interaction features, though weak alone, contribute a complementary signal in multimodel fusion, improving performance in functional classification and stability, and (2) establishes a resampling-controlled evaluation protocol for μOR modeling, and (3) identifies correlations between top features and μOR pocket chemistry/residues, and (4) case study to show effectiness on unseen external data, as a real-world application. Overall, the study demonstrates that hybridizing ligand-based and target-conditioned views-via target-aware view-based hybrid feature selection for stacked ensembles-adds complementary signal beyond ligand-only baselines, particularly for functional class (agonist vs antagonist).