Utilizing Dual-Channel Graph and Hypergraph Convolution Network to Discover Microbes Underlying Disease Traits.

Discovering microbes underlying disease traits opens up opportunities for the diagnosis and effective treatment of diseases. However, traditional methods are often based on biological experiments, which are not only time-consuming but also costly, driving the need for computational frameworks that can accelerate the discovery of these associations. Motivated by these challenges, we propose an innovative prediction algorithm named dual-channel graph and Hypergraph Convolutional Network (DCGHCN) to discover microbes underlying disease traits. First, based on the K-Nearest Neighbors (KNN) principle, we constructed attribute graphs for microbes and diseases, respectively. Next, Graph Convolutional Networks (GCNs) are used to capture homogeneous level implicit representations from attribute graphs of microbes and diseases. We used the output of the GCN layer as input to construct a hypergraph convolutional layer of microbes and diseases, to evaluate the impact of the confirmed microbes and diseases associations (MDAs) on the prediction results. Perform scalar product calculation on the microbe and disease features to determine the predicted score. The innovation of DCGHCN lies in employing the KNN algorithm to handle missing values in the correlation matrix during preprocessing and the use of a dual-channel structure to combine the advantages of GCNs and Hypergraph Convolutional Networks (HGCNs). We used 5-fold cross-validation (CV) to evaluate the performance of DCGHCN. The results showed that the DCGHCN model achieved AUC (Area Under the ROC Curve), AUPR (Area Under the PR Curve), F1-score and accuracy of 0.9415, 0.7637, 0.7515, and 0.9818. We selected two diseases for case studies, and a large number of published literature conclusions confirmed the prediction results of DCGHCN, thus proving that DCGHCN is an effective tool for discovering microbes underlying disease traits.