AI-assisted pharmacophore modeling of lead molecules targeting soluble epoxide hydrolase for anti-inflammatory drug discovery

Inflammatory diseases pose ongoing clinical challenges due to the limitations of current treatments such as NSAIDs. Targeting soluble epoxide hydrolase (sEH), which deactivates anti-inflammatory epoxyeicosatrienoic acids (EETs), represents a promising therapeutic strategy. In this study, we developed an optimized dual-feature ensemble directed message passing neural network (ODFE-DMPNN) framework to model sEH inhibitory activity using a small dataset of 138 compounds. By integrating d-MPNN-derived molecular graph representations with RDKit physicochemical descriptors and Morgan fingerprints, ODFE-DMPNN achieved a mean R² of 0.492 and an MSE of 0.026 in stratified 10-fold cross-validation. Guided by this model, we designed and synthesized two novel inhibitors, DJ-22 and LSM-36, with IC₅₀ values of 11.1 nM and 1.17 nM, respectively, demonstrating significantly higher activity than TPPU (IC₅₀ = 43.8 ± 1.65 nM). Structural analysis revealed that urea moieties mediate key hydrogen bonds, while flanking hydrophobic groups contribute to π-interactions with critical residues, consistent with pharmacophore features. Cellular assays confirmed significant anti-inflammatory effects via IL-1β suppression. This work demonstrates a powerful AI-guided strategy for pharmacophore-based drug discovery, accelerating the development of next-generation sEH inhibitors.