Drug repurposing offers a cost-effective and time-efficient strategy for discovering new therapeutic applications for approved drugs, reducing development timelines from ۱۰-۱۵ years to ۳-۵ years. Knowledge graphs (KGs) have emerged as powerful tools for representing complex biomedical relationships, integrating molecular interactions, pathway information, and clinical outcomes. Their ability to capture multifaceted drug-disease-target interactions makes them particularly valuable for drug repurposing, as they can reveal hidden patterns and potential therapeutic applications through network analysis. However, conventional approaches, particularly random walk-based methods, face significant limitations: they are inherently stochastic, lack comprehensive contextual understanding, and often fail to fully utilize the topological and semantic richness of KGs, especially in sparse graph regions. We propose a novel framework that harnesses Graph Neural Networks (GNNs) for
drug repurposing applications. GNNs can effectively learn hierarchical representations by systematically aggregating both local and global graph information through multiple message-passing layers, enabling the capture of complex interaction patterns across biological scales. To enhance node embeddings, we integrate semantic features extracted from
large language models (LLMs), including BioBERT (Lee et al., ۲۰۱۹) and GPT (Yenduri et al., ۲۰۲۳), addressing a critical gap in traditional approaches by incorporating unstructured textual information from biomedical literature. Our validation uses
Alzheimer's disease as a case study, chosen for its complex pathophysiology and urgent need for effective treatments. The model was evaluated on two benchmark datasets, MSI (Ruiz, Zitnik and Leskovec, ۲۰۲۱) and PrimeKG (Chandak, Huang and Zitnik, ۲۰۲۳), achieving a ۶% improvement in F۱ score compared to baseline methods in predicting drug-disease associations. Pathway analysis using t-tests on the top ۱۰ ranked drugs revealed statistically significant differences (p < ۰.۰۰۳) between high-ranked and lower-ranked drugs, specifically in pathways implicated in Alzheimer's disease, including amyloid-beta processing and neuroinflammation. Ablation studies demonstrated that LLM derived features contributed to a ۴% improvement in prediction accuracy compared to using graph structural features alone. Our integrated GNN-LLM framework presents a robust solution for computational drug repurposing, with potential applications across diseases with complex pathological mechanisms. Future work will focus on incorporating temporal dynamics and patient-specific factors for personalized
drug repurposing strategies.
