Abstract
BACKGROUND: Biological experiments have demonstrated that circRNA plays an essential role in various biological processes and human diseases. However, it is time-consuming and costly to merely conduct biological experiments to detect the association between circRNA and diseases. Accordingly, developing an efficient computational model to predict circRNA-disease associations is urgent. METHODS: In this research, we propose a multiple heterogeneous networks-based method, named XGBCDA, to predict circRNA-disease associations. The method first extracts original features, namely statistical features and graph theory features, from integrated circRNA similarity network, disease similarity network and circRNA-disease association network, and then sends these original features to the XGBoost classifier for training latent features. The method utilizes the tree learned by the XGBoost model, the index of leaf that instance finally falls into, and the 1 of K coding to represent the latent features. Finally, the method combines the latent features from the XGBoost with the original features to train the final model for predicting the association between the circRNA and diseases. RESULTS: The tenfold cross-validation results of the XGBCDA method illustrate that the area under the ROC curve reaches 0.9860. In addition, the method presents a striking performance in the case studies of colorectal cancer, gastric cancer and cervical cancer. CONCLUSION: With fabulous performance in predicting potential circRNA-disease associations, the XGBCDA method has the promising ability to assist biomedical researchers in terms of circRNA-disease association prediction.