[摘 要] 目的：基于干性相关基因鉴定食管鳞状细胞癌（ESCC）干性相关分子亚型，并构建亚型分类器，以期为预测ESCC患 者治疗反应提供参考。方法：从StemChecker数据库获取26个干性相关基因集，对源自TCGA-ESCC及GSE53625数据库的259 例ESCC样本转录组数据采用单样本基因集富集分析算法进行量化，随后通过一致性聚类法鉴定ESCC干性亚型，并系统评估了 不同亚型在预后特征、肿瘤微环境、干性生物学过程及治疗反应等方面的差异。通过加权基因共表达网络分析（WGCNA）鉴定 ESCC 干性亚型核心基因集，并基于随机森林、XGBoost 和神经网络三种机器学习算法构建并验证ESCC 干性亚型分类器。结 果：基于26个干性基因集的单样本基因集富集分析量化结果，将ESCC划分为两种干性亚型（命名为C1与C2），259例ESCC样 本中，109 例归为C1 亚型，150 例归为C2 亚型，二者在细胞分子特征、免疫微环境及治疗反应方面呈现显著差异。与C1 亚型相 比，干性C2亚型表现为不良预后[Kaplan-Meier分析显示C2亚型总生存期显著短于C1亚型（log-rank P = 0.028）]、更强干性特征、 更高水平的M2型巨噬细胞浸润、癌症相关成纤维细胞含量及间质评分；治疗策略方面，TIDE算法预测C1亚型免疫治疗应答率 为54.13%（59/109），显著高于C2亚型的24.00%（36/150） （P < 0.01），表明C1亚型患者更可能从免疫治疗获益；而C2干性亚型患 者对顺铂、吉西他滨及多西他赛的敏感性更高。WGCNA在TCGA-ESCC队列中鉴定出蓝色模块，在GSE53625队列中筛选出黑 色模块，这两个核心基因模块与C2干性亚型最具相关性。整合双队列模块基因集后，最终确定7个基因作为C2干性亚型的核心 基因。基于7 个基因转录组数据在训练集构建了机器学习分类模型，并在独立验证集中通过混淆矩阵结果证实了相较于 XGBoost和神经网络模型，建立的随机森林干性分型模型准确率达86.92%，AUC为0.91（95% CI：0.86～0.96），显著优于XGBoost （P < 0.001）和神经网络（P < 0.01）。结论：揭示了ESCC干性异质性，所构建的随机森林干性亚型分类器有助于实现ESCC患者 的分子分型及个体化治疗方案的精准选择。

[Abstract] Objective: Esophageal squamous cell carcinoma (ESCC) exhibits pronounced inter- and intra-tumoral heterogeneity, posing substantial challenges for prognostic prediction and therapeutic response evaluation. Cancer stem cells (CSCs) function as pivotal drivers underlying oncogenesis, metastasis, and therapy resistance. In this study, we aim to identify stemness-related molecular subtypes, and construct a subtype classifier to further assist in predicting ESCC patients' therapeutic response. Methods: Twenty-six stem-associated gene sets were retrieved from the StemChecker database. The transcriptome data of 259 ESCC samples from the TCGA-ESCC and GSE53625 databases were quantified using single-sample gene set enrichment analysis. Consensus clustering was then employed to identify stemness subtypes. Subtype-specific differences in prognosis, tumor microenvironment, stemness-related biological process, and therapeutic responses were predicted and analyzed. By applying weighted gene co-expression network analysis (WGCNA), the stemness subtype-related hub genes were identified. Subsequently, the ESCC stemness subtype classifier was constructed and evaluated using three machine learning algorithms: random forest, XGBoost, and the neural network. Results: Based on the single-sample gene set enrichment analysis (ssGSEA) of 26 stemness gene sets, two stemness subtypes (named C1 and C2) were determined, each exhibiting distinct stem-related molecular, immunological and therapeutic characteristics. C2 subtype exhibited adverse prognosis, stronger stemness, higher levels of M2 macrophages and cancer-associated fibroblasts, as well as higher stromal scores. In terms of therapeutic options, immunotherapy might be more appropriate for C1 ESCC patients, while cisplatin, gemcitabine · · 510 郑航, 等. 生物信息学整合机器学习算法鉴定食管鳞状细胞癌干性分子亚型 and docetaxel might be more suitable for C2 ESCC patients. WGCNA identified the blue module in TCGA-ESCC cohort and the black module in GSE53625 cohort as hub gene modules most relevant to the stemness C2 subtype. By integrating the gene sets screened from these two modules, a total of 7 genes were obtained as the final hub genes of the stemness C2 subtype. From the results of the confusion matrix, the validity of the model for predicting stemness subtype using the 7-gene transcriptome data and the machine learning model proposed in this study was proven. The random forest-based stemness classification model achieved higher accuracy, precision and AUC values compared with XGBoost and neural network model. Conclusion: This study delineates valuable perspectives on ESCC stemness heterogeneity, and the developed random forest-based stemness subtype classifier could aid in molecular stratification and personalized therapeutic regimen selection for ESCC patients.

[基金项目] 北京市临床重点专科建设项目；北京朝阳医院金种子科研基金项目（CYJZ202307）