Abstract
The miRNA plays a key role in the biological process and it has close relationship with disease. The wet experiment to test the association between miRNA and disease is time-consuming and costly, so semi-supervised learning based computational methods have been proposed to predict potential miRNAs associated with disease. However, these methods cannot fully utilize the local structure and global structure information of miRNA-disease association data, and the prediction performance can be improved further. In this work, we propose a novel approach for miRNA-disease association prediction by using multi-layer linear projection (MLPMDA). Firstly, we use the top n neighbors of miRNA and disease to update miRNA-disease association matrix, respectively, to employ the local structure and reduce the sparsity effect. Secondly, we define a computing layer to which a heterogeneous matrix composing of the updated association matrix and integrated miRNA similarity and integrated disease similarity is fed, and output predicted scores for miRNA-disease associations by linear projection method. Thirdly, we design multiple computing layers where the heterogeneous matrix input for the current layer is constructed based on the predicted miRNA-disease association scores from the last layer. Finally, we capture possible missing miRNA-disease associations by integrating prediction scores from each single view. We obtain AUC values 0.9847, 0.9883, 0.9899 for one dataset under 2-fold, 5-fold and 10-fold cross-validations, respectively, and their corresponding AUPR values are 0.7777, 0.7806 and 0.7518, which outperforms seven state-of-the-art methods. At last, we predict the potential miRNAs associated with three diseases, most of which are verified with some evidence. (C) 2020 Elsevier B.V. All rights reserved.