Weighted Nonnegative Matrix Factorization Based on Multi-source Fusion Information for Predicting CircRNA-Disease Associations

Abstract

Evidences increasingly have shown that circular RNAs (circRNAs) involve in various key biological processes. Because of the dysregulation and mutation of circRNAs are close associated with many complex human diseases, inferring the associations of circRNA with disease becomes an important step for understanding the pathogenesis, treatment and diagnosis of complex diseases. However, it is costly and time-consuming to verify the circRN-disease association through biological experiments, more and more computational methods have been proposed for inferring potential associations of circRNAs with diseases. In this work, we developed a novel weighted nonnegative matrix factorization algorithm based on multi-source fusion information for circRNA-disease association prediction (WNMFCDA). We firstly constructed the overall similarity of diseases based on semantic information and Gaussian Interaction Profile (GIP) kernel, and calculated the similarity of circRNAs based on GIP kernel. Next, the circRNA-disease adjacency matrix is rebuilt using K nearest neighbor profiles. Finally, nonnegative matrix factorization algorithm is utilized to calculate the scores of each pairs of circRNA and disease. To evaluate the performance of WNMFCDA, five-fold cross-validation is performed. WNMFCDA achieved the AUC value of 0.945, which is higher than other compared methods. In addition, we compared the prediction matrix with original adjacency matrix. These experimental results show that WNMFCDA is an effective algorithm for circRNA-disease association prediction.