Abstract
GluN3A-containing NMDA receptors have recently emerged as promising therapeutic targets for neurological disorders. However, discovering potent modulators remains a significant challenge, primarily due to the limitations of traditional high-throughput screening methods. In this study, we introduce a novel drug-target affinity prediction method, CLG-DTA, designed to enhance drug discovery for the GluN1/GluN3A receptor. This graph contrastive learning-based method incorporates natural language supervision by transforming regression labels into textual representation, and integrating them with traditional affinity data to enhance molecular representation. Additionally, a numerical knowledge graph is employed to refine continuous text embeddings, enabling precise modeling of complex drug-target interactions across diverse data modalities. Using CLG-DTA, we screened a library of 18 million compounds and identified 12 candidates for experimental validation. Among them, five compounds exhibited significant activity, with Boeravinone E demonstrating the highest potency ( IC50 = 3.40 ± 0.91 μM). These findings highlight the potential of CLG-DTA in accelerating the identification of promising GluN1/GluN3A modulators and lay a robust foundation for future therapeutic development.