引用文本：钱焱, 林青雨, 龚智霖, 等. 基于机器学习的胃癌氨死亡相关预后风险预测模型的构建及其肿瘤微环境研究[J/CD]. 消化肿瘤杂志（电子版）, 2025, 17(4): 499-508.

作者：钱焱1，林青雨1，龚智霖2，蔡世荣1，陈思乐1

单位：1. 中山大学附属第一医院胃肠外科中心，广东 广州 510080;2. 中山大学附属第一医院泌尿外科，广东 广州 510080

Authors：Qian Yan1, Lin Qingyu1, Gong Zhilin2, Cai Shirong1, Chen Sile1

Unit：1. Gastrointestinal Surgery Center, the First Affiliated Hospital of Sun Yat-sen University, Guangzhou 510080, Guangdong, China；2. Department of Urology, the First Affiliated Hospital of Sun Yat-sen University, Guangzhou 510080, Guangdong, China

摘要：

目的 基于机器学习建立胃癌预后的氨死亡相关风险预测模型，并探讨氨死亡与肿瘤微环境的关系，为胃癌患者的预后分析及治疗提供新的视角。方法 通过对癌症基因组图谱（the cancer genome atlas, TCGA）数据库中胃癌肿瘤组织（n=375）及正常癌旁组织（n=21）的转录组比较，鉴定氨死亡相关差异表达基因（differentially expressed genes, DEGs）。通过单因素Cox回归分析筛选与总生存相关的DEGs。通过随机抽样的方式将TCGA数据库中具有生存信息的胃癌样本按7∶3划分为训练集（n=190）和测试集（n=82），2个数据集合并作为总体样本（n=272）。基于训练集，将筛选得到的DEGs纳入机器学习并比较不同算法的预后预测效能，选取预测效能最优的算法建立风险评分模型。根据风险评分模型计算出的中位风险评分，将训练集、测试集、总体样本分别划分为高风险组和低风险组。采用单因素和多因素Cox回归分析确定胃癌患者预后的独立影响因素。利用风险分层与年龄、是否R0切除、肿瘤分期构建可视化列线图预测胃癌患者预后。通过受试者操作特征曲线（receiver operating characteristic curve, ROC曲线）和校准曲线进行模型和列线图的验证。通过ESTIMATE和CIBERSORT等算法对高、低风险组的肿瘤微环境进行分析和评价。结果 筛选出8个与预后相关的氨死亡相关DEGs。随机森林模型是预后预测效能最优的模型（C-index：训练集为0.891，测试集为0.713）。基于该模型的风险分层是胃癌患者预后的独立影响因素（P<0.001）。构建的列线图能有效预测胃癌患者1、2、3年总生存率。高风险组的肿瘤微环境呈现出更为显著的“免疫抑制”和“促肿瘤”特征。结论 本研究建立了1个基于8个氨死亡相关DEGs的胃癌预后风险预测模型，并初步探讨了氨死亡与肿瘤微环境的关系，为胃癌的预后评估和免疫治疗提供了新的探究方向。

关键词：胃癌；氨死亡；风险预测模型；肿瘤微环境

Abstract：

Objective To establish a machine learning-based prognostic risk prediction model for gastric cancer associated with ammonia-induced cell death, and to explore the relationship between ammonia-induced cell death and the tumor microenvironment, thereby providing new insights for prognosis analysis and treatment of gastric cancer. Method Ammonia-induced cell death related differentially expressed genes (DEGs) were identified by comparing transcriptome of gastric cancer tissues (n=375) and adjacent normal tissues (n=21) from the cancer genome atlas (TCGA) database. Univariate Cox regression analysis was used to select DEGs significantly associated with overall survival (OS). By random sampling, gastric cancer samples with survival information from the TCGA database were divided into a training set (n=190) and a test set (n=82) at a ratio of 7∶3, and the two datasets were combined as overall cohort (n=272). Based on the training set, the screened DEGs were incorporated into machine learning, and the prognostic prediction performance of different algorithms was compared. The algorithm with the best predictive performance was selected to establish a risk scoring model. Based on the median risk score calculated from the risk scoring model, the training set, test set, and overall cohort were divided into high-risk and low-risk groups. Univariate and multivariate Cox regression analyses were conducted to identify independent influencing factors for prognosis. A visual nomogram was constructed using the risk category, age, R0 resection, and tumor stage to predict the prognosis of patients. The model and nomogram were validated using receiver operating characteristic (ROC) curves and calibration curves. The tumor microenvironment of different risk groups were subsequently analyzed and evaluated using algorithms such as ESTIMATE and CIBERSORT. Result Eight ammonia-induced cell death related DEGs associated with prognosis were screened. After comparing multiple machine learning algorithms, the random forest algorithm, which demonstrated the optimal predictive performance with the C-index of 0.891 in the training set and 0.713 in the test set, was selected to establish the ammonia-induced cell death-related prognostic risk prediction model. The risk category based on the model was an independent influencing factor for prognosis of gastric cancer patients (P<0.001). The constructed nomogram effectively predicted the 1-, 2-, and 3-year OS rates of patients. Analysis of the tumor immune microenvironment revealed that the high-risk group exhibited more significant characteristics of “immunosuppression” and “tumor promotion”. Conclusion This study established a prognostic risk prediction model for gastric cancer based on eight ammonia-induced cell death related DEGs and preliminary explored the relationship between ammonia-induced cell death and the tumor microenvironment, providing a new direction for prognostic assessment and immunotherapy research in gastric cancer.

Key words：Gastric cancer; Ammonia-induced cell death; Risk prediction model; Tumor microenvironment

注：网络优先发布