PEGNet-CDA: A Propagation-Enhanced Graph Network for CircRNA-Disease Association Prediction

Circular RNAs are endogenous non-coding molecules with high stability and important regulatory roles in human diseases, yet experimentally validated circRNA-disease associations remain sparse. This work presents PEGNet-CDA, a propagation-enhanced graph network for CDA prediction. PEGNet-CDA encodes circRNA sequences using a pre-trained language model and derives dis-ease representations from Gaussian interaction profile kernel similarities, then projects heterogeneous features into a shared latent space. On the aligned bipartite graph, the model sequentially integrates diversified propagation behaviors, including diffusion-style propagation, neighborhood aggregation, and topology-aware filtering, to learn discriminative node embeddings. Association likelihoods are estimated through a dot-product matching function. Experiments on three public benchmarks, circAtlas 3.0, CircR2Disease 2.0, and circRNADisease 2.0, demonstrate stable and competitive performance across both classification and ranking metrics. On CircR2Disease, PEGNet-CDA achieves 84.37% accuracy with an AUC of 0.9169 and an AUPR of 0.9124. Additional ablation studies, hyperparameter analyses, matching-function comparisons, and case studies further support the effectiveness and biological relevance of the proposed framework.