Abstract
N-methyl-D-aspartate receptors (NMDARs) are critical mediators of excitatory neurotransmission and are composed of seven subunits (GluN1, GluN2A-D, and GluN3A-B) that form diverse receptor subtypes. While GluN1/GluN2 subtypes have been extensively characterized and have led to approved therapeutics, the GluN1/GluN3A subtype remains underexplored despite emerging evidence of its involvement in neuropsychiatric disorders. Efficient identification of modulators requires accurate prediction of drug-target affinity (DTA), particularly for challenging targets such as GluN1/GluN3A. In this study, we applied the ImageDTA model, which is a multiscale 2D convolutional neural network (CNN), to virtually screen 18 million small molecules for GluN1/GluN3A inhibitors. This artificial intelligence (AI)-driven approach prioritized 12 compounds, three of which demonstrated potent inhibitory activity (IC₅₀ < 30 µM) in experimental validation. The most potent hit, with an IC(50) of 4.16 ± 0.65 µM, revealed a novel structural scaffold, thus highlighting the potential of AI in accelerating drug discovery for underexplored receptor subtypes. These findings establish a robust framework for advancing GluN1/GluN3A-targeted therapeutics.