Predicting disease associations based on the higher order structure of ceRNA networks.

Competitive endogenous RNA (ceRNA) network regulation is an important posttranscriptional regulatory mechanism that plays an important role in physiological and pathological processes, and has been widely used in biomarker screening and regulatory factor studies of disease-related genes. However, existing studies have mainly focused on the association of a single type of RNA with disease, while studies targeting the application of ceRNA networks in disease prediction are still limited, so it is crucial to explore the potential of ceRNA networks in disease prediction. In this study, we propose CERDA-HOSR, a computational method for mining ceRNA network-disease associations based on higher order graph attention networks. The method uses higher order graph convolutional networks to aggregate neighborhood information to generate representations of different RNAs and diseases. Given the higher order complexity of biological networks and sample imbalance problem, traditional random negative sampling is difficult to effectively capture global information; for this reason, a higher order negative sampling strategy is designed to optimize the quality of negative samples by combining the network structure and higher order neighborhood relations to improve the generalization ability and prediction accuracy of the model. Finally, LightGBM calculates the ceRNA network-disease association probability based on the learned embedding. A large number of simulation experiments validate the superiority of CERDA-HOSR, and its practical application is further demonstrated by case studies of cardiovascular disease, acute myeloid leukemia, and papillary thyroid cancer. In addition, ablation experiments and exploratory analyses further enhance its robustness and provide an effective tool for disease prediction and biomarker screening.