What question did this study set out to answer?

To develop a novel model for predicting miRNA-disease associations using advanced computational techniques.

March 14, 2026

A Novel Mirna-Disease Association Prediction Model Based on an Adaptive Content-Guided Fusion Mechanism and Dilateformer

Key Points

To develop a novel model for predicting miRNA-disease associations using advanced computational techniques.
Proposed CGDAMDA integrates adaptive content-guided fusion and DilateFormer architecture.
Constructed a miRNA-disease heterogeneous network and applied positional encoding techniques.
Utilized sliding-window attention and multiscale dilated attention for deep feature representation extraction.
Employed XGBoost classifier for final MDA predictions.
Achieved AUC scores of 0.9585±0.0041 and 0.9632±0.0021 on HMDD v2.0 and v3.2 datasets, respectively.
Outperformed several state-of-the-art models in predictive accuracy.

Abstract

Introduction: MicroRNAs (miRNAs) are crucial in regulating gene expression. Identifying miRNA-disease association (MDA) is essential for disease diagnosis and treatment. Due to the high cost of traditional biological experiments, computational models have emerged as an effective alternative approach for MDA prediction. However, existing methods often lack flexibility in feature fusion and processing and cannot capture long-range dependencies, thereby limiting the accuracy of predictions. Methods: In this paper, we propose a novel MDA prediction model, CGDAMDA, which integrates an adaptive content-guided fusion mechanism with a DilateFormer architecture. First, homologous information of miRNAs and diseases is independently fused via the adaptive content-guided fusion mechanism to generate comprehensive similarity representations. A miRNA–disease heterogeneous network is then constructed, followed by Laplacian positional encoding and Weisfeiler–Lehman absolute role encoding. Deep feature representations are extracted using a combination of slidingwindow attention and multiscale dilated attention, enabling the model to capture complex feature dependencies. Finally, the learned miRNA and disease embeddings are fed into an XGBoost classifier to predict potential MDAs. Results: In 5-fold cross-validation, CGDAMDA achieves AUC scores of 0.9585±0.0041 and 0.9632±0.0021 on the HMDD v2.0 and HMDD v3.2 datasets, respectively, outperforming several state-of-the-art models. Discussion: CGDAMDA improves MDA prediction by introducing adaptive fusion and Dilate- Former encoder. In the future, consideration is given to integrating multi-omics data to further enhance the generalization ability of the model. Conclusion: In conclusion, our proposed CGDAMDA obtains high-quality node feature representations through adaptive feature fusion and shallow-to-depth feature extraction, which effectively improves the prediction performance of MDA.

Demander à l'IA

Bookmark

View Full Paper