Scholarlink

<p>Recent studies have shown that miRNAs undergo dynamic expression changes under pathological conditions and play diverse regulatory roles in disease progression. Accurately identifying their specific regulatory association types is essential for understanding disease mechanisms. However, most existing computational models mainly focus on association existence prediction or rely on node-centric representation learning, while insufficiently modeling fine-grained regulatory types and the semantic information carried by association edges. To address these limitations, we propose BGMMDA, a computational model for predicting multicategory miRNA-disease associations based on a bidirectional hypergraph attention network and a gated convolutional strategy. Specifically, multi-source similarity information and known miRNA-disease associations are first integrated to construct a weighted heterogeneous association graph. Then, candidate miRNA-disease associations are modeled as semantic hyperedges, and a bidirectional hypergraph attention network is designed to establish a closed-loop information propagation mechanism, enabling collaborative optimization between node representations and edge-level semantic representations. In addition, a gated convolutional strategy is introduced to selectively enhance informative pairwise features while suppressing noisy or redundant signals from the original association space. Finally, a unified multi-task loss function is used to improve type aware discrimination and representation stability. Experiments were conducted on the HMDD v3.2 dataset under two five-fold cross-validation settings. In the primary CVtype experiment, BGMMDA achieved Top-1 precision, Top-1 recall, and Top-1 F1 of 0.8711, 0.8700, and 0.8694, respectively. In the CVtriplet experiment, BGMMDA obtained an AUC of 0.9508 and an AUPR of 0.9503. Comparative experiments with five state-of-the-art methods demonstrate that BGMMDA achieves superior and more balanced performance in both regulatory type classification and potential association identification, confirming its effectiveness and practical applicability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of multicategory miRNA-disease associations based on bidirectional hypergraph attention network and gated convolutional strategy

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Yan Sun,
  • Xiaoqi Tang,
  • Junliang Shang,
  • Hanxiang Wang,
  • Defu Qiu,
  • Yuanke Zhang,
  • Jin-Xing Liu

摘要

Recent studies have shown that miRNAs undergo dynamic expression changes under pathological conditions and play diverse regulatory roles in disease progression. Accurately identifying their specific regulatory association types is essential for understanding disease mechanisms. However, most existing computational models mainly focus on association existence prediction or rely on node-centric representation learning, while insufficiently modeling fine-grained regulatory types and the semantic information carried by association edges. To address these limitations, we propose BGMMDA, a computational model for predicting multicategory miRNA-disease associations based on a bidirectional hypergraph attention network and a gated convolutional strategy. Specifically, multi-source similarity information and known miRNA-disease associations are first integrated to construct a weighted heterogeneous association graph. Then, candidate miRNA-disease associations are modeled as semantic hyperedges, and a bidirectional hypergraph attention network is designed to establish a closed-loop information propagation mechanism, enabling collaborative optimization between node representations and edge-level semantic representations. In addition, a gated convolutional strategy is introduced to selectively enhance informative pairwise features while suppressing noisy or redundant signals from the original association space. Finally, a unified multi-task loss function is used to improve type aware discrimination and representation stability. Experiments were conducted on the HMDD v3.2 dataset under two five-fold cross-validation settings. In the primary CVtype experiment, BGMMDA achieved Top-1 precision, Top-1 recall, and Top-1 F1 of 0.8711, 0.8700, and 0.8694, respectively. In the CVtriplet experiment, BGMMDA obtained an AUC of 0.9508 and an AUPR of 0.9503. Comparative experiments with five state-of-the-art methods demonstrate that BGMMDA achieves superior and more balanced performance in both regulatory type classification and potential association identification, confirming its effectiveness and practical applicability.