Abstract:
The miRNA is a single-stranded and small non-coding RNA, which is closely related to human diseases. The prediction on miRNA-disease associations can help understand the pathogenesis of diseases at the molecular level and provide a basis of studying the prognosis, diagnosis, evaluation and treatment of diseases. In the miRNA-disease association prediction, most methods take miRNA functional similarity and disease semantic similarity as input and don’t consider the similarity information including miRNA sequence, disease functional and hamming. Moreover, in the feature extraction process, they ignore the information complementarity between the linear features and nonlinear features such that the quality of feature extraction of miRNA and disease may be affected. Aiming at the above shortcomings, this paper proposes a novel miRNA-disease association prediction model, GCNMSF. First, we introduce the miRNA sequence similarity, disease semantic similarity and hamming similarity, and use similarity kernel fusion method to integrate multi-source similarities of miRNA and disease, respectively. Then, we use the graph convolutional network to learn nonlinear features and embed the convolutional attention block into GCN to optimize feature distribution. At the same time, the non-negative matrix factorization method is introduced to learn linear features of miRNA and disease for enriching the feature space and improving the ability of predicting miRNA-disease associations. Besides, we integrate the linear and nonlinear features of miRNA and disease to predict miRNA-disease associations. Finally, it is shown via experimental results on five-fold-cross validation that the proposed model in this work has better performance than some existing methods. It is also verified from the results of ablation experiment and case studies that the fusion of multi-source similarity information and the combination of linear and nonlinear features are helpful for miRNA-disease association prediction. In addition, the case studies of lung and breast cancers further confirm that GCNMSF can not only predict the potential miRNA-disease associations, but also discover the miRNA-disease associations of unknown diseases.