Predicting microbe-disease associations via attention guided graph convolutional networks

It is important to predict microbe-disease associations, as it helps to understand the cause of diseases episodes, the prevention of diseases, among other roles. Traditionally, the study of microbe-disease association is mainly based on laboratory purified culture for research and observation, which is highly scientific but has some limitations, such as some microbes needing to be cultured in a laboratory environment and costs heavily. This study proposes an attention guided graph convolutional network d approach (AGGCN) to predict possible potential associations between microbes and diseases. In the first place, AGGCN constructs four different similarity networks of microbes and diseases, which are combined with association networks to form a microbe-disease heterogeneous network. Subsequently, it is used as an input to the graph convolutional neural network to obtain each node embedding of this heterogeneous network. In this study, a three-layer network is used, so three different types of node embeddings can be obtained, and these three different types of node embedding representations are integrated via the multichannel graph attention mechanism to obtain the final node embeddings. Finally, the node embeddings obtained above are used for association prediction of microbes-diseases in heterogeneous networks. Experimentally verified, the AUC of AGGCN is 0.8972, and the AUPR is 0.5387. Comparative experimental results indicated that AGGCN shows better than the other five mainstream methods. It provides a new insight to explore network potential feature representation and predict associations between microbe and disease.