Integrating BERT pre-training with graph common neighbours for predicting ceRNA interactions.

Introduction: Predicting interactions between microRNAs (miRNAs) and competing endogenous RNAs (ceRNAs), including long non-coding RNAs (lncRNAs) and circular RNAs (circRNAs), is essential for understanding gene regulation. With the development of Graph Neural Networks (GNNs), existing works have demonstrated the ability to capture information from miRNA-ceRNA interactions to predict unseen associations. However, current deep GNNs only leverage node-node pairwise features, neglecting the information inherent in the RNA chains themselves, as different RNAs possess chains of varying lengths. Methods: To address this issue, we propose a novel model termed the BERT-based ceRNA Graph Predictor (BCGP), which leverages both RNA sequence information and the heterogeneous relationships among lncRNAs, circRNAs, and miRNAs. Our BCGP method employs a transformer-based model to generate contextualized representations that consider the global context of the entire RNA sequence. Subsequently, we enrich the RNA interaction graph using these contextualized representations. Furthermore, to improve the performance of association prediction, BCGP utilizes the Neural Common Neighbour (NCN) technique to capture more refined node features, leading to more informative and flexible representations. Results: Through comprehensive experiments on two real-world datasets of lncRNA-miRNA and circRNA-miRNA associations, we demonstrate that BCGP outperforms competitive baselines across various evaluation metrics and achieves higher accuracy in association predictions. In our case studies on two types of miRNAs, we show BCGP's remarkable performance in predicting both miRNA-lncRNA and miRNA-circRNA associations. Discussion: Our findings demonstrate that by integrating RNA sequence information with interaction relationships within the graph, the BCGP model significantly enhances the accuracy of association prediction. This provides a new computational tool for understanding complex gene regulatory networks.