Scholarlink

<p>Drug-disease association prediction is crucial for accelerating drug repurposing, yet existing computational models largely focus on integrating biological attributes while neglecting topological regulatory patterns within heterogeneous biological networks. To address this, we propose MsDGCN, a novel multi-scale diffusion graph convolutional network. The model is built upon two key innovations: a multi-scale architecture that captures complex topological signals through meta-path guided diffusion and cross-path attention mechanisms, and a dedicated hard negative mining strategy that selects biologically plausible confounders to sharpen the decision boundary. To handle severe data imbalance, an XGBoost classifier with dynamic class weighting is integrated to model nonlinear feature interactions. Extensive ten-fold cross-validation demonstrates that MsDGCN significantly outperforms state-of-the-art baselines across key metrics, confirming its capability to effectively leverage network topology for improved drug-disease association prediction. MsDGCN provides a robust and generalizable framework for predicting drug-disease associations. Code and data are available at: <a href="https://github.com/CDMBlab/MsDGCN.">https://github.com/CDMBlab/MsDGCN.</a></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

MsDGCN: multi-scale diffusion graph convolutional network for the prediction of drug-disease association

  • 期刊论文

  • 发表日期：

    2026年4月7日

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

摘要

Drug-disease association prediction is crucial for accelerating drug repurposing, yet existing computational models largely focus on integrating biological attributes while neglecting topological regulatory patterns within heterogeneous biological networks. To address this, we propose MsDGCN, a novel multi-scale diffusion graph convolutional network. The model is built upon two key innovations: a multi-scale architecture that captures complex topological signals through meta-path guided diffusion and cross-path attention mechanisms, and a dedicated hard negative mining strategy that selects biologically plausible confounders to sharpen the decision boundary. To handle severe data imbalance, an XGBoost classifier with dynamic class weighting is integrated to model nonlinear feature interactions. Extensive ten-fold cross-validation demonstrates that MsDGCN significantly outperforms state-of-the-art baselines across key metrics, confirming its capability to effectively leverage network topology for improved drug-disease association prediction. MsDGCN provides a robust and generalizable framework for predicting drug-disease associations. Code and data are available at: https://github.com/CDMBlab/MsDGCN.