Scholarlink

<p>Accurate histological grading of uterine corpus endometrial carcinoma (UCEC) is essential for personalized cancer treatment, yet remains challenging due to limited precision and poor interpretability of existing methods. We propose an interpretable multilevel attention graph adversarial autoencoder (IMA-GAAE), integrating multidimensional feature selection based on XGBoost and gene expansion via GO functional similarity, to enhance molecular-level grading diagnosis. Using UCEC transcriptome data, the model achieved 88.10% accuracy in binary classification and 73.35% in ternary classification, outperforming baseline models. Moreover, interpretability analyses (SHAP, saliency graph, and t-SNE) aligned with clinical knowledge. Importantly, seven potential genes were identified as biomarkers for UCEC grading, with three genes linked to potential antitumor drugs, providing insights into drug resistance and therapeutic strategies. The proposed method not only improves UCEC grading accuracy but also uncovers clinically relevant biomarkers and drug associations, offering a valuable tool for precision oncology.</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

IMA-GAAE: an accurate grading framework for uterine corpus endometrial carcinoma based on multidimensional interpretable feature selection and graph adversarial network

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Jie Shi,
  • Ying Su,
  • Hao Liu,
  • Chen Chen,
  • Xuecong Tian,
  • Wenjia Guo,
  • Cheng Chen,
  • Xiaoyi Lv

摘要

Accurate histological grading of uterine corpus endometrial carcinoma (UCEC) is essential for personalized cancer treatment, yet remains challenging due to limited precision and poor interpretability of existing methods. We propose an interpretable multilevel attention graph adversarial autoencoder (IMA-GAAE), integrating multidimensional feature selection based on XGBoost and gene expansion via GO functional similarity, to enhance molecular-level grading diagnosis. Using UCEC transcriptome data, the model achieved 88.10% accuracy in binary classification and 73.35% in ternary classification, outperforming baseline models. Moreover, interpretability analyses (SHAP, saliency graph, and t-SNE) aligned with clinical knowledge. Importantly, seven potential genes were identified as biomarkers for UCEC grading, with three genes linked to potential antitumor drugs, providing insights into drug resistance and therapeutic strategies. The proposed method not only improves UCEC grading accuracy but also uncovers clinically relevant biomarkers and drug associations, offering a valuable tool for precision oncology.