基于机器学习的免疫细胞浸润分类模型预测胰腺癌患者生存期的价值分析  

Value analysis of machine learning-based dichotomous model of immune cell infiltration in predicting the survival of patients with pancreatic adenocarcinoma

作　　者：谢辉[1,3] 邓祖亮 Xie Hui;Deng Zuliang(Department of Radiation Oncology,Affiliated Hospital of Xiangnan University,Chenzhou,Hunan 423000,China)

机构地区：[1]湘南学院附属医院放疗科 [2]湘南学院附属医院科教部 [3]医学影像人工智能湖南省重点实验室,湖南郴州423000

出　　处：《湘南学院学报（医学版）》2021年第4期19-27,共9页Journal of Xiangnan University（Medical Sciences）

基　　金：湖南省卫生健康委科研基金(20201718);湖南省教育厅科研基金(19a458)。

摘　　要：目的对胰腺癌22种免疫细胞浸润占比进行分析,探讨利用机器学习算法,构建预后二分类模型对胰腺癌患者生存期评价的可行性。方法从TCGA数据库,下载胰腺癌患者的转录组数据和相关的临床数据。通过“CIBERSORT”算法得到入组胰腺癌患者22种免疫细胞浸润占比特征。利用生物信息学方法对免疫细胞浸润占比特征进行筛选,构建预后模型。以受试者工作特征曲线下面积(AUC)、准确度、灵敏度、特异度、阳值预测值和阴值预测值作为预后二分类模型优劣的考核指标。行免疫细胞相关性分析及免疫功能分析,进一步阐述预后二分类模型的可信度与胰腺癌中的免疫景观。结果从22种免疫细胞中筛选出Macrophages M1、T cells CD8、B cells nave 3种免疫细胞浸润占比特征作为胰腺癌患者生存是否超过1年二分类预后模型的构建因子。采用9种机器学习分类模型算法构建二分类初始模型。按AUC排序,在训练集中表现最佳者为Extreme gradient boost(XGB)算法,在训练集中表现最佳者为Random Forest(RF)算法。RF稳定性相对较好,最终选取RF算法构建模型,具有较高的精准度。基于危险评分的免疫细胞相关性分析中,B cells nave和T cells CD8在低危组中免疫浸润高于高危组(P<0.001),Macrophages M1在高危组中免疫浸润高于低危组(P<0.001);与模型构建因子有关的免疫功能分析中,APC_co_inhibition、MHC_class_Ⅰ、Type_Ⅰ_IFN_Reponse在高危组中免疫浸润评分高于低危组(P<0.05);在免疫浸润分型分析中,C1,C2,C3,C6在高、低危组中显示出差异性(P<0.05)。结论基于机器学习构建的胰腺癌患者预后二分类模型,能较准确地评估胰腺癌患者生存是否超过1年,为临床对胰腺癌患者进行个体化治疗提供参考。Objective To investigate the feasibility of using machine learning algorithm to construct a prognostic dichotomous model to evaluate the survival of patients with pancreatic cancer,the proportion of 22 types of immune cell infiltration was analyzed.Methods The transcriptome data and related clinical data of pancreatic adenocarcinoma patients were downloaded from the TCGA database.The“CIBERSORT”algorithm was used to obtain the characteristics of infiltration proportion of 22 types of immune cells in the patients with pancreatic adenocarcinoma.Bioinformatics methods was used to screen infiltration proportion of the immune cell and construct a prognostic dichotomous model.AUC,accuracy,sensitivity specificity,positive predictive value and negative predictive value were used to evaluate the quality of the prognosis dichotomous model.Immune cell correlation analysis and immune function analysis further illustrate the reliability of prognostic dichotomous model and the immune landscape in pancreatic adenocarcinoma.Results Macrophages M1,B cells nave and T cells CD8 were screened out from 22 types of immune cells as the construction factors of prognosis dichotomous model of survival of patients with pancreatic adenocarcinoma.Nine machine learning classification model algorithms were used to construct the dichotomous model.Sorted by AUC,Extreme Gradient Boost(XGB)algorithm performed best in the training set,and Random Forest(RF)algorithm performed best in the test set.Since RF has relatively good stability,RF algorithm was finally selected to build the model.In the correlation analysis of immune cells based on risk scores,B cells nave and T cells CD8 was found to be more infiltrate in the low-risk group than that in the high-risk group(P<0.001),while Macrophages M1 was found to be more infiltrate in the high-risk group than that in the low-risk group(P<0.001).In the immune function analysis related to model building factors,the immune infiltration scores of APC_co_inhibition,MHC_class_Ⅰ,CD8+_T_cells in the hig

关 键 词：机器学习 免疫细胞 浸润 胰腺癌 分类模型 

分 类 号：R735.9[医药卫生—肿瘤]