Researchers from Xinjiang University and collaborating institutions have introduced a computational framework named CurvEdgeNet to predict associations between circular RNAs and diseases. Published in Frontiers in Genetics in November 2026, the study presents a geometry-enhanced approach that integrates graph-theoretic curvature with deep learning techniques. The primary objective of the research is to systematically identify potential links between circular RNAs and various pathological conditions, thereby aiding in the elucidation of disease mechanisms and the prioritization of candidates for biological validation. Circular RNAs are endogenous non-coding RNAs with closed circular structures that have emerged as key molecules bridging gene regulation and disease mechanisms. By acting as microRNA sponges, regulating gene splicing, or interacting with proteins, these molecules are deeply involved in complex processes such as tumorigenesis and neurodegenerative diseases.
Existing computational methods for predicting circRNA-disease associations often rely on conventional similarity networks or node-centered feature learning. The authors argue that these approaches may inadequately capture the non-Euclidean geometry and edge-level interactions inherent in biological networks. To address these limitations, CurvEdgeNet constructs a heterogeneous circRNA-disease network by integrating multiple circRNA and disease similarity measures with known associations. A central innovation of the framework is the incorporation of a discrete Ollivier-Ricci curvature-based edge descriptor. This descriptor is integrated into a curvature-enhanced graph attention mechanism to characterize local network geometry, allowing the model to distinguish between high-curvature edges that form community structures and low-curvature edges that connect different functional modules.
The architecture of CurvEdgeNet combines local geometric features with global feature dependencies. A bidirectional long short-term memory module is employed to capture global feature dependencies within the network. Multi-head attention mechanisms are then used to integrate complementary local geometric and global representations for association prediction. The model projects heterogeneous structural features into a unified hidden dimension and utilizes a three-layer Graph Attention Network to capture topological relationships. The curvature-enhanced attention mechanism adjusts information flow based on edge curvature, with a hyperparameter gamma set to 0.3 to control the influence of curvature. This dynamic adjustment aims to improve the accuracy and efficiency of information propagation across the network.
The study utilized three publicly available benchmark datasets for training and evaluation: CircR2Disease V2, CircRNADisease V2, and CircFunBase. These datasets contain varying numbers of circRNAs and diseases, providing diverse training samples to enhance the model’s generalization capability. CircRNADisease V2 includes 2,284 circRNAs and 253 diseases, while CircR2Disease V2 contains 2,053 circRNAs and 224 diseases. CircFunBase comprises 2,597 circRNAs and 67 diseases. The evaluation process employed five-fold cross-validation, where the ratio of the training set to the test set was 4:1. Negative samples were generated by randomly selecting unobserved circRNA-disease pairs to ensure a balanced dataset, with strict exclusion of test associations from feature construction to prevent information leakage.
Under five-fold cross-validation, CurvEdgeNet achieved high area under the receiver operating characteristic curve values across all three datasets. On the CircR2Disease V2 dataset, the model achieved an AUC of 0.9854 with a standard deviation of 0.0031. On the CircRNADisease V2 dataset, the AUC was 0.9900 with a standard deviation of 0.0020. On the CircFunBase dataset, the model achieved an AUC of 0.9932 with a standard deviation of 0.0016. These results demonstrate the effectiveness of the proposed framework in predicting circRNA-disease associations. The area under the precision-recall curve values were also high, reaching 0.9853, 0.9897, and 0.9931 on the respective datasets, indicating strong performance under imbalanced label distributions.
To validate the performance of CurvEdgeNet, the researchers conducted comprehensive comparisons with eleven state-of-the-art methods. These baseline methods represent diverse algorithmic approaches, including graph attention networks, matrix factorization, deep learning frameworks, and traditional machine learning methods. The comparison included models such as MOPSOGAT, GATCL2CD, and ICDMOE. CurvEdgeNet attained top-tier performance across all datasets and core metrics, surpassing the strongest competing methods in AUC, AUPR, and Matthews correlation coefficient. For instance, on the CircR2Disease V2 dataset, CurvEdgeNet achieved clear gains of 0.0070 in AUC and 0.0557 in MCC relative to MOPSOGAT. Two-sided Wilcoxon signed-rank tests confirmed that these performance improvements were statistically significant, with p-values below 0.001 for all pairwise comparisons against the top three baselines.
The robustness of the model was further assessed by repeating experiments under five distinct random seeds. The results showed that CurvEdgeNet consistently occupied the uppermost performance interval across all three datasets and core metrics. The data points corresponding to CurvEdgeNet clustered tightly across different seeds, demonstrating insensitivity to random initialization and data partitioning. Pairwise seed-wise comparisons confirmed that CurvEdgeNet outperformed the competing baselines under every matched random seed setting. This stability suggests that the performance lift is not an artifact of a single favorable random seed but reflects reproducible predictive superiority.
An ablation study was conducted to evaluate the contribution of each core component in CurvEdgeNet. The study used a progressive cumulative removal strategy, starting from the full model and removing modules step by step. Removing the Ollivier-Ricci curvature-enhanced feature extraction module resulted in an average AUC decrease of 0.0448. Further removing the multi-head attention fusion mechanism led to an additional average AUC decrease of 0.0156. The removal of the BiLSTM feature extraction module caused a significant average AUC decrease of 0.0860, highlighting its importance in modeling long-range dependencies. Finally, removing the curvature-enhanced attention mechanism resulted in the largest average marginal AUC decrease of 0.1097, indicating that this component makes the most substantial contribution under the progressive ablation setting.
Case studies were performed on breast cancer and colorectal cancer to evaluate the practical utility of CurvEdgeNet in prioritizing disease-associated circRNAs. For breast cancer, 10 out of the top 20 predicted circRNAs had been experimentally validated in previous literature, corresponding to a validation rate of 50%. Among these, hsa_circ_0000284, also known as circHIPK3, received the highest prediction score and has been shown to promote breast cancer progression through the regulation of miR-326. For colorectal cancer, 14 out of the top 20 predicted associations were literature-validated, resulting in a validation rate of 70%. The highest-scoring prediction, hsa_circ_0000467, has been reported to promote colorectal cancer progression by enhancing eIF4A3-mediated c-Myc translation.
The geometric perspective provided by Ollivier-Ricci curvature allows the model to interpret complex biological relationships more deeply. High-curvature edges may represent stable, cohesive, and functionally coordinated relationships within local biological modules, potentially connecting circRNAs and diseases that share similar functional annotations or participate in related pathways. In contrast, low-curvature edges may indicate sparse, bridge-like, or heterogeneous connections between distinct functional modules. These edges may represent cross-module communication or latent links between seemingly unrelated biological processes, capturing non-obvious relationships such as circRNAs involved in multi-disease mechanisms or diseases sharing molecular features with other disease classes.
Despite its strong performance, the study acknowledges several limitations that require further investigation. The computational efficiency of CurvEdgeNet for large-scale data needs improvement, as increasing data volumes may lead to greater computational resource consumption and time costs. Additionally, the model’s ability to capture weak signals associated with rare circRNAs or diseases during feature extraction remains to be enhanced. Future work will focus on optimizing the model through compression and parallel computing to improve efficiency. The researchers also plan to explore more advanced feature-mining algorithms to better capture weak signals and extend the framework to traditional Chinese medicine research by integrating associations among medicinal herbs, bioactive compounds, molecular targets, circRNAs, and diseases.
The source code and datasets used in this study are publicly available, facilitating reproducibility and further research. By integrating network geometric properties with deep learning techniques, CurvEdgeNet provides a novel computational approach for prioritizing candidate disease-associated circRNAs. This framework may facilitate further investigations into disease mechanisms and biomarker discovery, offering a valuable tool for the systematic identification of circRNA-disease associations in the era of precision medicine.
Subject of Research: Computational prediction of circRNA-disease associations using graph curvature
Article Title: CurvEdgeNet: edge feature integration with Ricci curvature for circRNA-disease association prediction
Article References: Fan, W., Liu, H., Lv, X., Chen, C., Chen, C., & Xu, Y. (2026). CurvEdgeNet: edge feature integration with Ricci curvature for circRNA-disease association prediction. Frontiers in Genetics, 17, Article 1909079. https://doi.org/10.3389/fgene.2026.1909079
Image Credits: AI Generated
DOI: 10.3389/fgene.2026.1909079
Keywords: circRNA, disease association prediction, Ollivier-Ricci curvature, graph neural networks, computational biology, biomarker discovery, heterogeneous networks, deep learning, Frontiers in Genetics, CurvEdgeNet, edge, feature
News Source: Drew Townsend. (November 9, 2026). CurvEdgeNet Predicts CircRNA-Disease Links Using Network Geometry. Scienmag.
