CLTDA: Identifying Associations Between tsRNAs and Diseases Based on Contrastive Learning

Increasing evidences have highlighted the significant association between tsRNAs and diseases. Predicting potential tsRNA-disease associations based on computational methods can effectively reduce human and resource consumption. However, there is a scarcity of computational methods for predicting tsRNA-disease associations. Therefore, we propose Contrastive Learning-based prediction of tsRNA-Disease Associations (CLTDA). It reconstructs known associations between tsRNAs and diseases based on adaptive Singular Value Decomposition (SVD). Then, we employ Graph Convolutional Networks (GCNs) for feature extraction from both the original and reconstructed tsRNA-disease associations, and optimize the GCNs by using contrastive learning loss and Bayesian Personalized Ranking (BPR) loss. In addition, the Bayesian negative sampling method is used to select high-quality negative samples for learning the features of tsRNA and disease. Finally, a Multi-Layer Perceptron (MLP) is utilized to calculates the score of potential association. We conduct five-fold cross-validation and denovo experiments on a manually collected tsRNA-disease association dataset, and the experimental results show that CLTDA outperforms the other six state-of-the-art methods. We also perform a case study on lung cancer and experimental results show that CLTDA is an effective tool for predicting potential associations between tsRNAs and diseases.