Heterogeneous and Similarity Network Data

Precise prediction of potential drug-disease associations (DDAs) is essential for enhancing treatment strategies and expediting drug development. However, current methods often rely on single-modal data and fail to effectively integrate multimodal information when representing node attributes. Furthermore, many feature extraction processes neglect the integration of node attribute features with topological features.To address these limitations, we propose MedPathEx, a method that integrates multimodal data fusion with metapath feature extraction techniques. First, we constructed a biomedical heterogeneous network comprising three entities—drugs, genes, and diseases—along with their interrelationships. By incorporating multimodal data, we generated similarity networks for the nodes within this heterogeneous network. We then used a graph convolutional network to extract the node attribute features from these similarity networks. Simultaneously, meta-paths enhanced with a multi-head attention mechanism capture local topological features from the heterogeneous network, whereas a global attention mechanism further refines global topological features, enabling a seamless fusion of local and global features. Finally, MedPathEx effectively combined these node attribute features with network structural features to create a comprehensive feature representation that was used to calculate the potential association probabilities between drugs and diseases.The experimental results indicate that MedPathEx surpasses current methods in terms of critical metrics, including AUC, AP, and F1 scores. MedPathEx effectively identified novel candidate drugs in case studies of coronary artery disease and hypertension, underscoring its substantial potential for practical applications.

精确预测潜在药物-疾病关联（DDAs）对于优化治疗策略和加速药物研发至关重要。然而，现有的方法往往依赖于单一模态数据，并在表征节点属性时未能有效整合多模态信息。此外，许多特征提取过程忽视了将节点属性特征与拓扑特征进行整合。为解决这些局限性，我们提出了MedPathEx方法，该方法将多模态数据融合与元路径特征提取技术相结合。首先，我们构建了一个包含药物、基因和疾病三个实体及其相互关系的生物医学异构网络。通过整合多模态数据，我们为该异构网络中的节点生成了相似性网络。随后，我们运用图卷积网络从这些相似性网络中提取节点属性特征。同时，借助多头注意力机制的元路径捕捉异构网络中的局部拓扑特征，而全局注意力机制则进一步细化全局拓扑特征，从而实现了局部与全局特征的完美融合。最终，MedPathEx有效地将节点属性特征与网络结构特征相结合，构建了一个全面的特征表示，用于计算药物与疾病之间潜在关联的概率。实验结果表明，MedPathEx在AUC、AP和F1等关键指标上优于现有方法。在冠状动脉疾病和高血压的案例研究中，MedPathEx有效地识别了新的候选药物，凸显了其在实际应用中的巨大潜力。