Different Graph-Level Attention Based on Multi-Scale for Predicting lncRNA-Disease Associations

The relationship between long non-coding RNAs (lncRNAs) and diseases is crucial for understanding biological processes, as well as the onset, progression, prevention, and treatment of diseases. Accurate prediction of associations between lncRNAs and diseases holds significant potential, offering new insights for biological research and identifying novel therapeutic targets for clinical applications. These predictions enhance research efficiency, reduce unnecessary experimental costs, and improve diagnostic and treatment precision. However, existing methods often fail to simultaneously consider global structural information, local subgraph details, and multi-scale graph information. In this study, we design a graph neural network prediction framework based on multi-scale graph-level attention, designed to predict disease-related candidate lncRNAs, named GLALDA. It employs a dual-level attention mechanism to integrate both global structural and local subgraph information. To enhance the ability of the model to capture the graph structure, edge feature information is incorporated into the attention calculations. Furthermore, we utilize a cross-attention mechanism to deeply fuse feature representations from graphs of different scales, effectively combining local node details with global context. The resulting integrated features are then fed into a scoring network for evaluation. Experimental results on public datasets demonstrate that GLALDA achieves an AUC of 0.949 and an AUPR of 0.947, outperforming six other state-of-the-art methods. Furthermore, through in-depth case studies of three cancers, we further validate GLALDA's capability to identify potential disease-related lncRNA candidates. These findings underscore the framework's potential to advance both biological research and clinical applications.