Prediction of circRNA-disease association based on attention fusion graph–hypergraph convolutional networks and contrastive learning

Circular RNA (circRNA) plays a key regulatory role in the pathogenesis of various complex diseases. Due to the limitations of traditional biological experiments, which are costly and time-consuming, efficient computational methods have become an effective supplement for predicting the association between circRNA and diseases. However, most current computational methods fail to utilize the complementarity among multi-level features in the data and lack effective modeling of information interaction across views. Therefore, we propose an AGHCL model based on attention fusion graph–hypergraph convolutional networks and contrastive learning for predicting circRNA-disease associations. Firstly, AGHCL constructs corresponding graphs and hypergraphs for each type of similarity data for circRNA and diseases. The graph convolution is used on the graph to obtain low-order feature representations of circRNA and diseases, and hypergraph convolution is used on the hypergraph to obtain high-order feature representations. To complement the feature representations obtained from different convolutional networks, AGHCL uses an attention aggregation mechanism to dynamically fuse these feature representations for more accurate integration. Additionally, a contrastive learning module is used to obtain consistent circRNA and disease feature representations across different views. Then, the variational auto-encoder is employed to extract features of circRNA and diseases from known circRNA-disease associations. Finally, multiple circRNA and disease features are integrated to obtain a circRNA-disease score matrix for association prediction. The experimental results demonstrate that our model significantly outperforms existing methods across all evaluation metrics on benchmark datasets.