MSGTrans: circRNA-Disease Association Prediction With Cluster-Based Negative Sampling and Multi-Scale Graph Transformer

Circular RNAs (circRNAs) contribute significantly to various biological processes, with their stable structures and regulatory functions, circRNAs are gaining attention for their biomarker potential and therapeutic applications in human diseases. However, the wet-lab experiments for identifying the associations between circRNAs and diseases are time-consuming and expensive, necessitating the development of efficient computational methods. Here, a novel computational method, MSGTrans, is developed for predicting circRNA-disease associations, specifically, the model integrates multi-source nodes and cluster-based negative sampling is applied to address the imbalance of associations, then multi-scale graph transformers and cross-scale attention mechanism are designed to capture both local and global features of nodes, finally, LightGBM is introduced to predict the associations between circRNAs and diseases. Experimental results demonstrate that MSGTrans outperforms several state-of-the-art methods, offering reliable circRNA candidates for specific diseases.