Predicting Microbe-Disease Associations via Multiple Layer Graph Convolutional Network and Attention Mechanism

Abstract

Recently clinical evidences have confirmed that human diseases are affected by the microbes inhabiting human bodies. Identifying latent microbe-disease associations can provide a deep insight into the pathogenesis of diseases. However, traditional biological experiments are inefficient and expensive to achieve pathogenic microbes for diseases, computational approaches become a new alternative choice. In this work, we introduce a graph neural network method (MLAGCNMDA) with multiple layers of graph convolutional network and attention mechanism to predict potential microbe-disease pairs. In MLAGCNMDA, a heterogeneous network is constructed based on the known microbe-disease associations and multiple similarities between microbes and diseases. Moreover, nodes embedding of the heterogeneous network are learned by a multi-layer graph convolutional network model, in which the attention mechanism is introduced in each graph convolutional layer to distinguish the importance of neighbor nodes. Finally, a bilinear decoder is used to decode the node embedding to reconstruct microbe-disease associations. The experiments show that our method outperforms the baseline methods with reliable average AUCs of 0.945 and 0.946 in the Leave-one-out and 5-fold cross validation assessment framework. Case studies on two diseases, i.e., colorectal carcinoma and liver cirrhosis, further confirm the reliability and effectiveness of our method.