"Drug Repurposing Based On Heterogenous Networks: A Multi-Modal Graph Based Approach" (Preprint)
Jiale Nan, Yuanyuan Sun, Meiling Che, Jianhai Lin, Dongping Gao · 2025
BACKGROUND To date, most successful cases of drug repurposing and the methods employed have been built on an understanding of pharmacology or retrospective analyses of clinical outcomes of the original indications. More systematic approaches to identify potential repurposing drugs are required. OBJECTIVE For drug repurposing, computational methods offer a relatively rapid and cost-effective approach, encompassing structure-based methods, ligand-based methods, and data-driven approaches, and it is necessary for binding heterogeneous information to explore drug repurposing by this multi-modal computational method. METHODS We propose a drug repurposing model based on multi-modal graph learning, which offers a computational way along with a prediction list of 1500+ FDA-approved ‘old’ drugs and corresponding indications. It can hopefully provide an integrative toolkit to infer insightful drug-target interactions, identify potential treatment candidates, understand the heterogeneity of clinical drug use, and inspire ‘wet’ experiment-driven pre-clinical study. RESULTS Our study conducted ablation experiments by replacing the modality representation learning part with a MLP and direct concatenation method, and substituting the adaptive graph learning module with a graph structure based on a RBF kernel. The results are shown in Table IV. In terms of feature ablation, the model without the modality representation learning module experienced significant drops across all metrics, with the most severe declines being 14.4, 11.03, 12.11, 11.58, and 11.9 points on A, P, R, F1, and AUC, respectively. Removing inter-modality shared information and intra-modality specific information both led to decreased model performance. Regarding self-adaptive graph learning, the absence of a learnable matrix function caused reductions of 11.83, 18.2, 14.63, 16.42, and 19.82 points on A, P, R, F1, and AUC, respectively. Since the matrix function primarily learns from graph structure data, which is closely related to the subsequent graph neural network prediction module, lacking an adaptive learning module significantly discounts the effect of graph representation, having a substantial impact on downstream tasks. CONCLUSIONS This study constructed a large-scale heterogeneous network and proposed a multimodal graph learning framework that considers the interactions among different modal (network) datasets. The effectiveness of the model was validated using the constructed dataset, showing superior performance over baseline models on four metrics: A, P, R, and F1. Ablation studies were conducted to verify the role of modal representation learning and adaptive graph learning modules in the overall algorithm, experimentally demonstrating that this part of the model architecture has good performance. In case analysis, the potential uses of Macitentan for multiple indications and the drug repurposing results for Alzheimer’s disease were validated through databases and literature. The prediction results indicate strong evidence support, with the potential to complement and improve drug-disease association information in public databases in the future.