Enhanced drug disease association prediction through multimodal data integration and meta path guided global local feature fusion.

Accurately predicting Drug-Disease Associations (DDAs) is of great significance for drug repurposing and new drug development. Although existing methods have promoted the development of this field to a certain extent, most of them are still limited to single-modal data and cannot fully characterize the complex features of drugs, diseases, and genes. At the same time, many methods only focus on either local neighborhoods or global structures during feature extraction, lacking the organic combination of the two, which limits the accuracy and generalization of predictions. To address this, this paper proposes MedPathEx, a drug-disease association prediction method that combines multi-modal data integration and local-global feature learning. Specifically, we first construct a drug-gene-disease heterogeneous network and fuse multi-modal attributes such as drug chemical structures, ATC classifications, side effects, disease phenotypes and semantic information, as well as gene function annotations to generate more comprehensive node representations. Subsequently, we use graph convolutional networks to extract the attribute features of nodes themselves, capture local semantic relationships through meta-path modeling with a multi-head attention mechanism, and introduce a global attention mechanism to extract overall topological patterns, thereby achieving “micro-macro complementary” feature learning. Finally, by fusing node attributes and structural features, MedPathEx obtains a more discriminative comprehensive representation for the prediction of potential DDAs. Experimental results show that MedPathEx outperforms existing methods in key indicators such as AUC, AP, and F1. Moreover, it successfully identifies new candidate drugs in cases of coronary artery disease and hypertension, demonstrating its great potential in practical applications.