天津医药 ›› 2026, Vol. 54 ›› Issue (11): 1215-1221.doi: 10.11958/20260165

• 应用研究 • 上一篇    下一篇

基于坏死性凋亡基因构建前列腺癌的机器学习诊断模型及潜在药物筛选

吴军1(), 冷雪峰1, 宗丹1, 刘政江2,△()   

  1. 1 中国人民解放军海军第九七一医院泌尿外科 (邮编 266071)
    2 中国人民解放军联勤保障部队第九七〇医院(威海院区)泌尿外科
  • 收稿日期:2026-01-05 修回日期:2026-07-03 出版日期:2026-11-15 发布日期:2026-11-10
  • 通讯作者: 刘政江 E-mail:15953242411@163.com;lunwenzy666@126.com
  • 作者简介:吴军(1983),男,主治医师,主要从事泌尿生殖肿瘤方面研究。E-mail:15953242411@163.com
  • 基金资助:
    青岛市医疗卫生科研项目(2025-WJKY238)

Construction of a machine learning diagnostic model for prostate adenocarcinoma based on necroptosis-related genes and screening of potential therapeutic agents

WU Jun1(), LENG Xuefeng1, ZONG Dan1, LIU Zhengjiang2,△()   

  1. 1 Department of Urology, the 971st Hospital of the Chinese People's Liberation Army Navy, Qingdao 266071, China
    2 Department of Urology, the 970th Hospital of the Chinese People's Liberation Army Joint Logistics Support Force (Weihai Campus)
  • Received:2026-01-05 Revised:2026-07-03 Published:2026-11-15 Online:2026-11-10
  • Contact: LIU Zhengjiang E-mail:15953242411@163.com;lunwenzy666@126.com

摘要:

目的 探讨坏死性凋亡相关基因在前列腺癌(PRAD)中的表达特征,筛选关键诊断基因,构建并验证机器学习诊断模型,并预测潜在靶向治疗药物。方法 从癌症基因组图谱(TCGA)数据库下载497例PRAD组织和52例正常前列腺组织的RNA测序数据作为训练队列,从基因表达数据库(GEO)下载GSE46602(50例)和GSE69223(30例)数据集作为验证队列。收集GeneCards数据库中相关性得分≥1的坏死性凋亡相关基因(303个),与TCGA表达数据匹配后,筛选PRAD组织与正常组织间的差异表达基因(DEGs)。通过最小绝对值收缩和选择算子(LASSO)回归分析筛选最终纳入模型的诊断基因,并分别构建随机森林(RF)、支持向量机(SVM)和决策树(DT)3种机器学习诊断模型。采用受试者工作特征(ROC)曲线评估模型的诊断效能,并通过200次重复10倍交叉验证模型稳定性。基于oncoPredict包结合CTRP2数据库预测潜在靶向药物。结果 共筛选出33个坏死性凋亡相关DEGs,其中PRAD组织中上调基因9个、下调基因24个,经LASSO回归进一步筛选得到14个诊断基因。3种机器学习模型中,RF模型的诊断效能最优,训练队列曲线下面积(AUC)为0.993(95%CI:0.987~1.000),敏感度为99.6%,特异度为98.4%,准确率为98.1%,Kappa值为0.954;10倍交叉验证平均AUC为0.974(95%CI:0.967~0.981)。外部验证结果显示,RF模型在第1验证队列和第2验证队列的AUC分别为0.954(95%CI:0.896~1.000)和0.932(95%CI:0.879~0.985)。药物筛选共获得287对显著相关的基因-药物对,其中靶向增强子同源物2(EZH2)的药物218种、靶向基质金属蛋白酶9(MMP9)的药物22种。靶向MMP9的药物中与之相关性最强的为BRD.K90370028(r=-0.962),靶向EZH2的药物中与之相关性最强的为BRD.K19103580(r=-0.956)。结论 基于坏死性凋亡相关基因构建的RF模型诊断效能优异且稳定性良好,筛选出的靶向药物为PRAD的精准治疗提供了潜在方向。

关键词: 前列腺肿瘤, 坏死性凋亡, 机器学习, 诊断模型

Abstract:

Objective To investigate the expression profiles of necroptosis-related genes in prostate adenocarcinoma (PRAD), identify key diagnostic genes, construct and validate machine learning diagnostic models, and predict potential targeted therapeutic agents. Methods RNA sequencing data of 497 PRAD tissue samples and 52 normal prostate tissue samples were downloaded from the Cancer Genome Atlas (TCGA) database as the training cohorts, while GSE46602 (n=50) and GSE69223 (n=30) datasets were retrieved from the Gene Expression Omnibus (GEO) database as the validation cohort. Necroptosis-related genes with a relevance score ≥1 were collected from the GeneCards database (n=303) and matched with TCGA expression data. Differentially expressed genes (DEGs) between PRAD tissue and normal tissue were identified using the limma package. Least absolute shrinkage and selection operator (LASSO) regression analysis was applied to select the final diagnostic genes incorporated into the models. Three machine learning diagnostic models were subsequently constructed, including random forest (RF), support vector machine (SVM) and decision tree (DT). The diagnostic performance of each model was evaluated using receiver operating characteristic (ROC) curves, and model stability was assessed via 200 iterations of 10-fold cross-validation. Potential targeted therapeutic agents were predicted using the oncoPredict package in combination with the CTRP2 database. Results A total of 33 necroptosis-related DEGs were identified, including 9 upregulated genes and 24 downregulated genes in PRAD tissue. And a total of 14 diagnostic genes were ultimately selected through LASSO regression. Among the three machine learning models, the RF model demonstrated the best diagnostic performance, achieving an area under the ROC curve (AUC) of 0.993 (95%CI: 0.987-1.000) in the training set, with a sensitivity of 99.6%, specificity of 98.4%, accuracy of 98.1% and Kappa coefficient of 0.954. The average AUC of 10-fold cross-validation was 0.974 (95%CI: 0.967-0.981). External validation revealed that the RF model achieved AUCs of 0.954 (95%CI: 0.896-1.000) and 0.932 (95%CI: 0.879-0.985) in the first and second validation cohorts, respectively. Drug screening identified 287 significantly correlated gene-drug pairs, comprising 218 drugs targeting enhancer of zeste homolog 2 (EZH2) and 22 drugs targeting matrix metalloproteinase-9 (MMP9). Among drugs targeting MMP9, BRD. K90370028 showed the strongest correlation (r=-0.962), while BRD. K19103580 exhibited the strongest correlation between EZH2-targeting drugs (r=-0.956). Conclusion The RF model constructed based on necroptosis-related genes demonstrates excellent diagnostic performance and satisfactory stability, while the screened targeted drugs provide potential directions for the precision treatment of PRAD.

Key words: prostate adenocarcinoma, necroptosis, machine learning, diagnostic model

中图分类号: