NAF-CDA: A node-adaptive fusion model for circRNA-disease association prediction.

Abstract

Accurate prediction of circRNA-disease associations is essential for uncovering disease mechanisms and prioritizing potential biomarkers. However, existing computational approaches are still challenged by sparse known associations, heterogeneous biological similarity information, and the uncertainty of unobserved associations used as negative samples. Here, we propose NAF-CDA, a node-adaptive robust fusion framework for circRNA-disease association prediction. Unlike conventional similarity integration strategies that assign fixed weights to different information sources, NAF-CDA learns node-specific similarity contributions by considering the heterogeneous characteristics of individual circRNAs and diseases. The framework constructs multi-source similarity networks from association profiles, circRNA functional information, and disease semantic information, followed by an adaptive fusion strategy to integrate complementary biological evidence. To alleviate the influence of potential false negatives on threshold determination, a similarity-constrained reliable negative sampling strategy is employed to exclude high-risk unknown pairs from the training-negative candidate pool. Furthermore, a robust prediction refinement module combining graph inference and dynamic-rank matrix completion is introduced to recover latent association patterns while determining the retained rank adaptively according to the singular-value distribution under predefined rank-related constraints. Extensive experiments on four benchmark datasets, including CircR2Disease, CircRNADisease, Circ2Disease, and CircR2Disease2, demonstrate that NAF-CDA achieves competitive performance, obtaining AUC values of 99.14%, 97.83%, 97.96%, and 98.66%, respectively. In controlled imbalance evaluations, NAF-CDA showed a smaller AUPR degradation than the compared methods when the negative-to-positive ratio increased. Ablation studies further confirm that adaptive similarity fusion and dynamic-rank refinement substantially improve prediction performance, while reliable negative sampling improves threshold-dependent classification performance. Overall, NAF-CDA provides an effective computational framework for prioritizing potential circRNA-disease associations under sparse and heterogeneous biological network conditions.