CurvEdgeNet: edge feature integration with Ricci curvature for circRNA-disease association prediction

Circular RNAs (circRNAs) are crucial regulators in disease pathogenesis, making accurate circRNA-disease association prediction essential for understanding disease mechanisms and therapeutic development. Current computational methods rely on simple network topologies and linear feature extraction, failing to capture the non-Euclidean geometric properties and complex interaction patterns in biological networks, which significantly impairs prediction accuracy and biological interpretability. To address this challenge, we propose CurvEdgeNet, a novel heterogeneous graph neural network framework that integrates discrete curvature and edge-enhancement mechanisms. CurvEdgeNet employs a multi-stage architecture: first constructing a heterogeneous network using multiple similarity measures and known associations, then designing a curvature-enhanced graph attention network that introduces Ollivier-Ricci curvature from discrete geometry into circRNA-disease prediction, embedding geometric edge properties into graph neural network message-passing to effectively capture local topological features, while global feature dependencies are captured through a BiLSTM module, and finally fusing these complementary features using multi-head attention before feeding them into a multilayer perceptron for prediction. Five-fold cross-validation on three independent datasets—CircR2Disease V2, CircRNA2Disease V2, and CircFunBase—demonstrates excellent predictive performance, achieving AUC values of 0.9854±0.0031, 0.9900±0.0020, and 0.9932±0.0016 respectively. This geometry-enhanced approach provides new insights into biological networks' non-Euclidean properties and offers a powerful tool for discovering potential circRNA-disease associations, facilitating biomarker identification and therapeutic target development.