Machine learning-driven drug repurposing for GPR17: activity prediction via graph neural networks and multistage computational validation.

To identify novel GPR17-targeting ligands with potential relevance to multiple sclerosis (MS) therapy, we developed an integrated computational workflow combining graph neural network (GNN)-based prediction with multistage structure-based validation. GPR17 is a class A G protein-coupled receptor implicated in oligodendrocyte differentiation, neuroinflammation, and remyelination, making it a remarkable target for myelin repair. A curated dataset of 323 chemically diverse ligands with experimentally reported pIC50 values was assembled from the literature and used to train an NNConv-based GNN model. The model demonstrated robust predictive performance across training, validation, and independent test sets (R2 = 0.868, 0.727, and 0.790, respectively). To further assess model generalizability, scaffold-based validation was additionally performed to evaluate prediction across structurally distinct chemotypes. Analysis of the learned chemical space revealed enrichment of antihistamine- and leukotriene-related scaffolds among highly ranked compounds, motivating a drug-repurposing strategy. Candidate molecules, including barmastine, astemizole, sulukast, and iralukast, were evaluated alongside reference ligands pranlukast and cangrelor. Predicted hits were subsequently refined through pharmacophore-guided virtual screening, molecular docking, ADMET profiling, molecular dynamics simulations, and MM-PBSA/MM-GBSA free-energy calculations. Several ligands exhibited stable binding modes supported by persistent hydrophobic interactions and hydrogen-bond occupancies reaching 99.8%. Importantly, this study proposes computationally prioritized candidates rather than experimentally validated inhibitors. Overall, the proposed hybrid workflow provides a practical strategy for early-stage GPR17 ligand discovery and future CNS-oriented drug development.