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

Area of Science:

• Computational Chemistry and Cheminformatics
• Pharmacology and Drug Discovery
• Machine Learning in Bioinformatics

Background:

• Accurate classification of μ-opioid receptor (μOR) ligands as agonists or antagonists is crucial for drug discovery and safety.
• Existing machine learning models like ExtraTrees and MPNNs show promise but have limitations in understanding feature influence and evaluation robustness.
• The impact of target-conditioned interaction features and resampling methods (e.g., SMOTE) on model performance requires further investigation.

Purpose of the Study:

• To introduce the μOR-Ligand framework, a target-aware, view-based hybrid feature selection method for improved μOR functional class prediction.
• To enhance the identification of whether an active ligand acts as an agonist or antagonist.
• To establish a robust evaluation protocol for μOR modeling, controlling for resampling effects.

Main Methods:

• Developed the μOR-Ligand framework utilizing three views: ligand fingerprints, ligand descriptors, and molecular interaction features.
• Employed a hybrid feature selection strategy combining multimodel fusion and ensemble feature selection (stacking) for stacked ensembles.
• Implemented a resampling-controlled evaluation protocol using identical, fixed splits with and without SMOTE for robustness assessment.

Main Results:

• The μOR-Ligand framework achieved a superior ROC AUC of 0.930 ± 0.026, outperforming recent models like MPNNs (p-value=0.046).
• Achieved a high ROC AUC of 0.977 on internal cross-validation, demonstrating strong predictive performance.
• Demonstrated that target-aware interaction features provide complementary signals, improving classification performance and stability.

Conclusions:

• Hybridizing ligand-based and target-conditioned views via the μOR-Ligand framework significantly enhances functional classification accuracy for μOR ligands.
• The study establishes a standardized, resampling-controlled evaluation protocol for μOR modeling.
• Identified correlations between top predictive features and μOR pocket chemistry, offering insights for future drug design.