Scholarlink

<p>Accurate identification of mutation-driven cancer driver genes is vital for cancer biology and targeted therapy. To address noise in protein–protein interaction (PPI) networks, we propose DGCL_DWA, a novel method utilizing graph diffusion and contrastive learning. DGCL_DWA first employs personalized PageRank to generate a diffusion graph, revealing hidden biological connections. Chebyshev graph convolution extracts features from both the PPI and diffusion networks, and neighborhood contrastive learning harmonizes gene representations, reducing noise. The network-specific features are refined via Chebyshev graph convolutions, which are constrained via node classification and link prediction. A dynamic weight adjustment strategy balances task-specific losses during training. Finally, logistic regression is used to predict driver genes. The experimental results demonstrate the superior performance in pan-cancer and specific cancer driver gene identification compared with state-of-the-art methods. Ablation studies confirm the positive impact of the diffusion graph, contrastive learning, and dynamic weight adjustment on predictive accuracy. The source codes are available at <a href="https://doi.org/10.57760/sciencedb.31933">https://doi.org/10.57760/sciencedb.31933</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

Predicting Cancer Driver Genes via Contrastive Graph Diffusion and Dynamic Weighting

  • 期刊论文

  • 发表日期：

    2025年11月1日

查看全文
  • Wei Peng,
  • Xin-Ping Xu,
  • Wei Dai,
  • Zheng-Nan Zhou,
  • Xiao-Dong Fu,
  • Li Liu,
  • Li-Jun Liu

摘要

Accurate identification of mutation-driven cancer driver genes is vital for cancer biology and targeted therapy. To address noise in protein–protein interaction (PPI) networks, we propose DGCL_DWA, a novel method utilizing graph diffusion and contrastive learning. DGCL_DWA first employs personalized PageRank to generate a diffusion graph, revealing hidden biological connections. Chebyshev graph convolution extracts features from both the PPI and diffusion networks, and neighborhood contrastive learning harmonizes gene representations, reducing noise. The network-specific features are refined via Chebyshev graph convolutions, which are constrained via node classification and link prediction. A dynamic weight adjustment strategy balances task-specific losses during training. Finally, logistic regression is used to predict driver genes. The experimental results demonstrate the superior performance in pan-cancer and specific cancer driver gene identification compared with state-of-the-art methods. Ablation studies confirm the positive impact of the diffusion graph, contrastive learning, and dynamic weight adjustment on predictive accuracy. The source codes are available at https://doi.org/10.57760/sciencedb.31933.