SLGCN: Structure-enhanced line graph convolutional network for predicting drug–disease associations

Drug repositioning is a rapidly growing strategy in drug discovery, as the time and cost needed are considerably less compared to developing new drugs. In addition to traditional wet experiments, designing effective computational methods to discover potential drug–disease associations is an attractive shortcut in drug repositioning. Most current methods based on graph neural networks ignore the heterophily of the constructed drug–disease network, resulting in inefficient predictions. In this paper, a novel structure-enhanced line graph convolutional network (SLGCN) is proposed to learn comprehensive representations of drug–disease pairs, incorporating structural information to conduct heterophily. First, line graphs centered around drug–disease pairs are extracted. This process turns the association prediction task into a node classification problem, which better displays the learning ability of SLGCN. Then, in message aggregation, a relation matrix is proposed to mark the structural importance of neighboring nodes. In this way, messages from nodes with lower structural importance can be assigned small weights. Unlike vanilla GCN, which adds self-loops to average ego representations and aggregated messages, an update gate is proposed to integrate biology information contained in ego representations with topology information contained in aggregated messages. Extensive experiments show that SLGCN achieves better performance than other advanced methods among the two datasets.