Predicting lncRNA-protein interactions based on graph autoencoders and collaborative training

Abstract

Long non-coding RNAs(lncRNAs) play an important role in various biological processes. lncRNAs usually perform their molecular functions by interacting with proteins. Therefore, it is essential to predict potential lncRNA-protein associations for disease prevention and disease treatment. Label-propagation-based methods are widely used for predicting associations among biological entities. However, in these approaches, similarity computation and label propagation are separate procedures, which decrease the effectiveness of label propagation. Moreover, the prediction accuracy of existing models is also not ideal. In this paper, we proposed an end-to-end deep learning lncRNA-protein Interaction predictor through Graph Autoencoders and Collaborative training (LPIGAC). Different from previous studies, our model implemented two graph autoencoders on lncRNA graph and protein graph respectively, and trained these two graph autoencoders collaboratively. Graph autoencoders on lncRNA graph and protein graph are competent to reconstruct score matrix through initial association matrix, which is equivalent to propagate labels on graphs. This end-to-end framework can strengthen the robustness and precision of the label propagation procedure. Cross validations indicate that LPIGAC outperforms current lncRNA-protein associations prediction methods. Case studies demonstrate that LPIGAC is competent to detect potential lncRNA-protein associations. Source code of our paper is available at https://github.com/zhanglabNKULPIGAC.