Microbe-Disease Association Prediction Based On Multi-Data Fusion And Graph Neural Network

With the development of the Human Microbiome Project,researchers began to gradually realize that microbes are not only an integral part of the human body,but also closely related to human health and diseases.Identifying potential associations between microbes and diseases is crucial for in-depth understanding of the pathogenic mechanisms of diseases and providing a theoretical basis for disease diagnosis,treatment,and prevention.However,traditional biological experimental methods are time-consuming,labor-intensive and expensive,so efficient computational models to mine potential microbe-disease associations are an effective auxiliary means.This paper mainly uses graph neural network method to study the relationship between microbes and diseases on the biological network.The main research work is summarized as follows:(1)For the problem of low reliability of the existing microbe-disease associations,we analyzed 6 public datasets and collated,screened and reconstructed a new benchmark dataset(multi MDA).On the other hand,most current computational methods use a single prior information and are highly dependent on known associations,leading to inaccurate predictions of new diseases/microbes.We proposed a method for constructing the heterogeneous network,which constructed similarity networks that not only considered the fusion between network structure information and biological function information,but also considered the fusion of multiple information within them.(2)Considering the difficulty of capturing complex microbe-disease associations,we proposed a deep learning framework based on graph convolutional attention network(MDAGCAN)to identify potential microbe-disease associations.In MDA-GCAN,we leveraged multiple graph convolutional layers and graph attention layers to learn embedding representations for nodes in the heterogeneous network.Then,unknown microbe-disease associations were reconstructed by decoding the embedding representations using a bilinear decoder.Experimental results show that MDA-GCAN has better predictive performance than most algorithms.(3)To solve the problem of the model reliability and low accuracy of the association relationship applied to new diseases/microbes,we designed a deep learning method(MDAIMCGAE)based on inductive matrix completion and graph autoencoder to predict potential associations between microbes and diseases.The model learned node representations of the heterogeneous network through graph autoencoder,and then reconstructed unknown microbe-disease associations in a deep learning end-to-end framework.Finally,the optimal model is collaborative trained by combining the reconstruction loss and the inductive matrix completion loss to infer the relationship between microbe-disease pairs.Experimental results in different experimental scenarios show that MDA-IMCGAE outperformed other state of the art methods and can be effectively applied to the prediction of unrelated new diseases and novel microbes.This paper will start from the perspective of bioinformatics,based on rich biomedical data,utilize graph neural network technology for modeling and analysis,and propose two computational models that can be applied to the prediction of microbe-disease associations.These models are of great significance for advancing the understanding of diseases pathogenic mechanisms and promoting the application of microbes in the fields of disease diagnosis,treatment,and prevention.