<p>Drug repurposing has emerged as a promising strategy to identify new therapeutic uses for already approved drugs, accelerating drug discovery and significantly reducing development time and costs. Current computational methods for drug-disease association predictions are vulnerable to shallow feature representation, inefficient attention mechanisms, and poor generalization capability across datasets. We propose a hybrid graph attention model using graph embedding techniques for drug-disease association prediction to overcome the existing limitation. Our model integrates sparse and dense convolutional layers with adaptive attention mechanisms to enhance feature selection and improve predictive accuracy. Compared to existing models such as TL-HGBI, DRRS, DeepDR, NIMGCN, and LAGCN, our approach demonstrates substantial improvements across key performance metrics. Specifically, it achieves an AUC of 0.8945, surpassing the best-performing baseline and an accuracy of 98.56%, which is significantly higher than all compared models. Additionally, our model attains a specificity of 99.13% and a recall of 42.51%, indicating better identification of true positives and a reduced false positive rate. These results confirm the robustness and generalizability of the proposed architecture. Further, the use of Optuna-based hyperparameter tuning, optimized layer embeddings, and dynamic attention weight initialization contributes to making this model interpretable and scalable traits that are critically needed in current computational drug discovery frameworks. Overall, the method offers a data-driven solution for identifying therapeutic candidates for complex diseases and guiding focused treatment strategies in modern healthcare contexts.</p>