Label-aware Dual-view Graph Neural Network for Protein-Protein Interaction Classification

Protein -protein interaction (PPI) plays an important role in various biological processes of organisms, and is beneficial for the development of relevant drugs, the diagnosis of diseases, etc. On this account, an increasing number of studies on PPI have been proposed. However, existing works neither construct the model with respect to the characteristics of downstream tasks (e.g. the potential relation between labels) nor fully explore the raw data (e.g. the confidence of interactions between proteins). To address this issue, we propose a novel Label -aware Dual -view Graph Neural Network for Protein -Protein Interaction Classification (LADV-PPI) method. Specifically, we construct a multi -head feature similarity graph by exploiting the rich information within the node features, which will be combined with the topology graph to generate the consensus graph. Based on this, we design a dual -view graph neural network with parameter sharing to learn PPI representations jointly from both topology graph and consensus graph. In addition, we introduce a self -adaptive label relation graph aggregation module to generate effective label representations, which will be incorporated into the dualchannel label -aware multi -layer perceptrons to guide the learning process of PPI representations. Extensive experimental results on three real -world datasets demonstrate that the proposed LADV-PPI achieves superior performance as compared to eight state-of-the-art baseline methods.