Predicting microbe-disease associations via graph neural network and contrastive learning

Abstract

In the contemporary field of life sciences, researchers have gradually recognized the critical role of microbes in maintaining human health. However, traditional biological experimental methods for validating the association between microbes and diseases are both time-consuming and costly. Therefore, developing effective computational methods to predict potential associations between microbes and diseases is an important and urgent task. In this study, we propose a novel computational framework, called GCATCMDA, for forecasting potential associations between microbes and diseases. Firstly, we construct Gaussian kernel similarity networks for microbes and diseases using known microbe-disease association data. Then, we design a feature encoder that combines graph convolutional network and graph attention mechanism to learn the node features of networks, and propose a feature dual-fusion module to effectively integrate node features from each layer's output. Next, we apply the feature encoder separately to the microbe similarity network, disease similarity network, and microbe-disease association network, and enhance the consistency of features for the same nodes across different association networks through contrastive learning. Finally, we pass the microbe and disease features into an inner product decoder to obtain the association scores between them. Experimental results demonstrate that the GCATCMDA model achieves superior predictive performance compared to previous methods. Furthermore, case studies confirm that GCATCMDA is an effective tool for predicting microbe-disease associations in real situations.

Keywords: contrastive learning; graph attention mechanism; graph convolutional network; gut microbial metagenomics; microbe-disease associations.