A protein-protein interactions prediction method based on label guided multi-scale graph neural network is proposed, which not only enhances the representation ability of data, but also introduces label information to guide learning. Firstly, the multi-scale graph representation is obtained by graph data augmentation, and the multi-scale graph representation is input into graph neural network to obtain multi-scale protein representation, and comparative learning is introduced to further improve the protein characterization ability. Secondly, the self-learning label relation graph is constructed to learn the relationship between labels and obtain the feature representation of labels. Finally, the prediction of protein-protein interactions is guided by the feature representation of labels. Experiments are carried out on three public datasets. Compared with the optimal benchmark method, the proposed method has better performance. Specifically, compared with the best baseline method, the micro-F1 scores on the three datasets SHS27k, SHS148k and STRING increase by 2.01%, 0.94% and 0.93% respectively.
