AI-enhanced virtual screening approach to hit identification for GluN1/GluN3A NMDA receptor

N-methyl-D-aspartate receptors (NMDARs) are calcium-permeable ionotropic glutamate receptors broadly expressed throughout the central nervous system, where they play crucial roles in neuronal development and synaptic plasticity. Among the various subtypes, the GluN1/GluN3A receptor represents a unique glycine-gated NMDAR with notably low calcium permeability. Despite its distinctive properties, GluN1/GluN3A remains understudied, particularly with respect to pharmacological tools development. This scarcity poses challenges for deeper investigation into its physiological functions and therapeutic relevance. In this study, we employed a hybrid virtual screening (VS) pipeline that integrates ligand-based and structure-based approaches for the efficient and precise identification of small-molecule candidates targeting GluN1/GluN3A. A large compound library comprising 18 million molecules was screened using an AI-enhanced multi-stage method. The initial phase utilized shape similarity ranking via ROCS-BART, followed by refinement with a graph neural network (GNN)-based drug-target interaction model to enhance docking accuracy. Functional validation using calcium flux (FDSS/μCell) identified two compounds with IC₅₀ values below 10 μM. Of these, one candidate exhibited potent inhibitory activity with an IC₅₀ of 5.31 ± 1.65 μM, which was further confirmed through manual patch-clamp recordings. These findings highlight an AI-enhanced VS workflow that achieves both efficiency and precision, providing a promising framework for exploring elusive targets such as GluN1/GluN3A.