Using machine learning in the in-silico design of selective dopamine receptor ligands: Advancements in targeted therapies for neurological and psychiatric disorders

Accurately predicting ligand-receptor binding affinities is crucial for advancing structure-based drug design. Here, we present a novel graph neural network (GNN) model employing edge-conditioned message passing (NNConv), enabling the integration of bond-specific features into molecular graph representations. Using a curated dataset of approximately 24,000 ligands targeting dopamine receptor subtypes D1-D5, our GNN models achieved strong predictive performance with low error metrics (RMSE ~0.6–0.8). Model interpretability was assessed using the GNNExplainer algorithm, which identified chemically meaningful substructures driving predictions. Complementary molecular docking and pharmacophore analyses revealed distinct binding profiles across bioactivity classes. Active ligands consistently engaged key conserved residues, notably Asp 3.32 , Phe 6.52 , and Ser 5.42/5.43 , through directional hydrogen bonds and π-π interactions, resembling the binding patterns of Rotigotine. Inactive ligands exhibited superficial hydrophobic contacts with limited polar anchoring, while borderline ligands showed partial engagement with receptor cavities. ADMET and Lipinski's rule assessments confirmed favorable pharmacokinetic properties for the most potent ligands. Finally, molecular dynamics simulations of the top-ranked ligands for each dopamine receptor subtype confirmed the dynamic stability and persistent interactions of these complexes at their respective binding sites, which validate the results of the structural study. These findings demonstrate that edge-aware GNN models, combined with interpretability and structural validation, offer a powerful strategy for decoding ligand-receptor interactions and accelerating neuropharmacological drug discovery. • Approach enhances identification of dopamine receptor-targeting compounds. • Novel GNN model integrates bond-specific features for predictions. • The Model achieved low RMSE (0.6–0.8) in binding affinity predictions. • Active ligands form strong interactions with conserved receptor residues. • ADMET analysis confirmed favorable pharmacokinetic properties of ligands.