An interpretable consensus pipeline for candidate dopamine D2 receptor ligands

Abstract Machine learning accelerates GPCR ligand discovery but often lacks interpretability and struggles to generalise to structurally novel chemical space. We present an integrated framework combining ensemble learning, SHAP interpretability, and complementary structure-based analyses to prioritise putative candidates for the dopamine D₂ receptor (DRD2). Models were trained on 1247 curated ChEMBL compounds using Bemis–Murcko scaffold-split validation. Feature ablation showed that Morgan fingerprints alone were sufficient for classification (AUC = 0.910), whereas combined physicochemical descriptors improved regression (R² = 0.606). Random Forest achieved scaffold-split classification AUC = 0.906 ± 0.016, and external PubChem BioAssay validation on non-overlapping compounds achieved AUC = 0.954. SHAP analysis identified nitrogen-containing substructural motifs consistent with Asp3.32 recognition, supporting the canonical DRD2 pharmacophore hypothesis. The pipeline recovered nine known DRD2-active compounds among the top 100 docked DrugBank candidates, corresponding to ≈ 14.6-fold enrichment over random selection. Virtual screening of DrugBank prioritised Roginolisib (DB18088), DB08010, and Bitopertin (DB12426) as putative DRD2 candidates. The near-zero correlation between ML predictions and docking scores ( r = 0.090, p = 0.37, n = 100) supports their use as complementary evidence streams. Molecular dynamics simulations (200 ns) and MM-GBSA calculations indicated stable binding modes with geometries distinct from the reference ligand 8NU. By requiring agreement across complementary computational evidence streams, this consensus pipeline provides a reproducible framework for prioritising compounds in extrapolative GPCR chemical space. Experimental validation through radioligand binding and functional assays is required to confirm DRD2 activity.