<p>Drug repositioning (DR) is a highly promising research strategy aimed at discovering new therapeutic indications for existing drugs. Current computational DR methods have become effective tools for uncovering drug-disease associations, yet they suffer from three critical limitations: most models can only extract either local or global embeddings of node features, traditional methods often construct shallow networks due to the vanishing gradient problem, making it difficult to capture the complex multi-level relationships between drugs and diseases, and they struggle to mine meaningful information from small-scale negative samples. To overcome these limitations, we propose an innovative method named GADRC, which employs a synergistic architecture of graph convolutional networks and graph attention networks to simultaneously capture local structural features of drug molecules and global pathway features of diseases for the first time. Additionally, we introduce a biologically interpretable deep residual network, whose cross-layer identity connection mechanism effectively addresses the depth degradation problem in traditional graph neural networks, enabling the model to stably learn multi-level interactions between drug targets and disease markers. Finally, we develop a feature-guided undersampling strategy combined with a weighted cross-entropy loss function, which constructs biologically similar subgroups through positive sample feature clustering and dynamically selects hard negative samples with weighted importance, significantly improving the utilization efficiency of negative samples. Experimental results on three benchmark datasets demonstrate that GADRC consistently outperforms most methods in DR tasks. Moreover, case and molecular docking studies on Alzheimer’s disease and breast cancer further validate its effectiveness and provide new insights into GADRC’s ability to identify novel drug-disease associations.</p>