Beyond structural bias: Improving circRNA-disease association prediction with Multi-Hop Neighborhood Hierarchical...

Aberrant expression of circular RNAs (circRNAs) is closely associated with various diseases, underscoring the importance of accurate association prediction for advancing diagnostics and therapeutics. Given that biological experiments are often costly and time-consuming, computational approaches have become indispensable for uncovering circRNA-disease associations (CDAs). Graph Neural Networks (GNNs)-based approaches show particular promise. However, existing methods often rely on coupled shallow GNNs, which struggle to capture higher-order information effectively. Additionally, the sparsity of known associations introduces structural unfairness, limiting nodes' access to optimal personalized information. To address these issues, we propose Multi-Hop Neighborhood Hierarchical Fusion (MNHF), a novel computational framework for CDAs prediction. Specifically, MNHF constructs adjacency matrices for circRNA and disease based on known associations, followed by multi-hop propagation within each view using a decoupled architecture to capture multi-scale features. Shared informativeness estimators evaluate the quality of propagated features, while gated mechanisms control information flow. Neighborhood-level information is fused and reduced via a shared aggregator. Finally, a unified classifier combines the outputs from both views to produce the final prediction. We conducted extensive experiments on five benchmark CDAs datasets and compared MNHF with state-of-the-art methods. The results demonstrate that MNHF achieves superior performance and exhibits strong generalization capabilities, highlighting its potential for advancing CDAs studies.