Line Graph Neural Network for Drug-Disease Association Prediction

Drug-Disease Association (DDA) plays a crucial role in drug discovery by identifying the efficacy, mechanisms, or potential therapeutic effects of drugs for specific diseases. However, traditional wet-lab experiments for identifying such associations are often costly and inefficient. Graph Neural Networks (GNNs), as an emerging technology, have demonstrated remarkable performance in DDA prediction tasks. However, in the process of predicting DDAs, GNNs often employ graph pooling layers to extract fixed-size features for classification, which inevitably leads to information loss. To address this challenge, this paper explores a novel approach by leveraging the concept of line graphs in graph theory. Specifically, in a line graph, each node corresponds to a unique edge in the original graph. By transforming the DDA prediction task in the original graph into a node classification task in the corresponding line graph, GNNs can better capture relationships between edges. Experimental results on two publicly available datasets demonstrate that the proposed model demonstrates superior performance, achieving significant improvements in overall effectiveness compared to baseline methods.