A graph-based knowledge distillation framework for drug repurposing via multi-task learning

Biomedical Knowledge Graphs (BKGs) capture intricate interactions between biological entities, playing a crucial role in the repurposing of drugs. However, current BKG completion methods often face challenges in scalability, predictive performance, and computational efficiency. We propose a novel Graph-based Knowledge Distillation approach for Drug Repurposing via a Multi-Task Learning framework (GKDRMTL), to address these limitations. By leveraging a teacher–student knowledge distillation strategy, our model not only enhances predictive accuracy but also substantially reduces computational demands. Compared to the state-of-the-art baselines, the student model consistently demonstrates substantial efficiency gains, achieving ∼30–93 % faster training time per epoch, ∼75–99 % lower memory usage, ∼46–88 % faster inference time, while maintaining competitive accuracy. Evaluated on an extended HetioNet, a heterogeneous biomedical knowledge graph, GKDRMTL reached state-of-the-art results across multiple link prediction tasks, including drug–disease associations, drug-drug similarity, disease-disease similarity, and disease–gene associations. The teacher achieves near-perfect performance in Area under the Receiver Operating Characteristic Curve (ROC-AUC) of 0.9889, Area Under the Precision-Recall Curve (AUPR) of 0.9875, and Accuracy of 0.9876. While the student approximates teacher performance with ROC-AUC of 0.9739, AUPR of 0.9704, and Accuracy of 0.9673, despite its simplified architecture. These findings underscore the importance of integrating knowledge distillation with multi-task learning for efficient and high-performance biomedical link prediction. The code of the proposed method and data are available here: https://github.com/Zahra-Alaeddini/GKDRMTL.