EG-TransAttention: Edge-aware multiscale attention network for predicting circRNA-disease associations
Abstract
Studying the interactions between circRNA and diseases is crucial for exploring the pathogenesis of diseases. However, existing circRNA-disease association methods are constrained by their inability to encode multiscale hierarchical relationships and nonlinear features embedded in graph structures. Additionally, the complex interactions between nodes are not effectively utilized in the feature extraction. Specifically, we introduce a novel computational framework (EG-TransAttention) for predicting unknown circRNA-disease associations. The framework introduces three key innovations: (1) An edge-aware message passing mechanism leverages edge attributes as adaptive weights to guide the aggregation of neighbor information, thereby preserving the hierarchical topological structure of the graph. (2) A Light Multi-scale Linear Attention (LightMLA) module employs linear operators and multi-scale convolutions to efficiently model local-global dependencies without quadratic complexity. (3) A ParNet cross-domain attention architecture fuses heterogeneous circRNA-disease features through parallelized hierarchical learning. Finally, the learned representations are fed into a random forest classifier, which robustly infers potential associations. Extensive experiments on the CircR2Disease and CircRNADisease public databases demonstrate that EG-TransAttention outperforms five state-of-the-art baselines, while ablation studies, robustness tests, and in-depth case analyses on breast cancer, colorectal cancer, esophageal squamous cell carcinoma and gastric cancer, further validate the performance of the proposed framework.