Ligand-based drug discovery methods typically utilize pharmacophore similarities among molecules to screen for potential active compounds. Among these, scaffold hopping is a widely used ligand-based lead identification strategy that facilitates clinical candidate discovery by seeking inhibitors with similar biological activity yet distinct scaffolds. In this study, we employed GeminiMol, a deep learning-based molecular representation framework that incorporates bioactive conformational space information. This approach enables ligand-based virtual screening by referencing known active compounds to identify potential hits with similar structural and bioactive conformational features. Using GeminiMol-based ligand screening method, we discovered a potent GluN1/GluN3A inhibitor, GM-10, from an 18-million-compound library. Notably, GM-10 features a completely different scaffold compared to known inhibitors. Subsequent validation using whole-cell patch-clamp recording confirmed its activity, with an IC of 0.98 ± 0.13 μM for GluN1/GluN3A. Further optimization is required to enhance its selectivity, as it exhibited IC values of 3.89 ± 0.79 μM for GluN1/GluN2A and 1.03 ± 0.21 μM for GluN1/GluN3B. This work highlights the potential of AI-driven molecular representation technologies to facilitate scaffold hopping and enhance similarity-based virtual screening for drug discovery.