基于生物信息学和机器学习算法构建肾移植术后抗体介导排斥反应的诊断预测模型  

Constructing a diagnostic prediction model for antibody-mediated rejection after kidney transplantation based upon bioinformatics and machine learning algorithms

作　　者：吴极岳 孙泽家 毕青 邱雪萌 王玮[1] Wu Jiyue;Sun Zejia;Bi Qing;Qiu Xuemeng;Wang Wei(Department of Urology,Affiliated Beijing Chaoyang Hospital,Capital Medical University,Beijing 100020,China)

机构地区：[1]首都医科大学附属北京朝阳医院泌尿外科,北京100020

出　　处：《中华器官移植杂志》2024年第10期718-727,共10页Chinese Journal of Organ Transplantation

摘　　要：目的构建肾移植术后抗体介导排斥反应(antibody-mediated rejection,AMR)的诊断预测模型,并初步筛选AMR的潜在治疗药物。方法收集并整理GEO数据库中7个有关AMR的大型肾移植队列数据集,差异表达分析用于鉴定AMR与正常受者的差异表达基因。分别使用随机森林(random forest,RF)、极端梯度提升、支持向量机和广义线性模型4种机器学习算法构建肾移植受者术后AMR的诊断模型,并绘制受试者操作特征(receiver operating characteristic,ROC)曲线比较各模型的准确性。选择最优模型中核心基因进行整合以构建AMR受者的诊断预测列线图,进行校准曲线绘制及决策曲线分析以评估列线图的准确性。将AMR受者活检组织中差异表达基因上传至关联性图谱(connectivity map,CMap)数据库进行检索,筛选前5种与AMR具有相反表达模式的化合物作为AMR的潜在治疗药物。结果CXCL10、FCGR1B、GBP5、CD69、LY96、BCL2A1和EVI2A 7个基因在AMR受者的外周血和活检组织中高表达(FDR<0.05)。基于RF算法的AMR诊断模型在多种机器学习算法中AUC值最高(0.904),且在外部数据集GSE50084和GSE175718中其AUC值分别为0.876和0.824。对于整合RF模型中BCL2A1、CXCL10、FCGR1BP、CD69和EVI2A五个核心基因所构建的AMR诊断预测列线图,校准曲线显示该列线图的预测结局与实际结局接近;决策曲线表明在较大横坐标范围内,列线图的净获益率较极端曲线更高。CMap数据库预测结果显示排名前5位的化合物为雷特格韦、利美尼定、白毛茛碱、美替拉酮和丙戊酸。结论基于外周血基因表达谱所构建的列线图对AMR的诊断具有较高的准确性和普适性,CMap数据库预测的5种化合物雷特格韦、利美尼定、白毛茛碱、美替拉酮和丙戊酸可能是AMR的潜在治疗药物。Objective To construct a diagnostic prediction nomogram for antibody-mediated rejection(AMR)after kidney transplantation(KT)based upon peripheral blood gene expression profiling and preliminarily screening potential drugs for AMR.MethodsSeven large kidney transplant cohort datasets related to AMR were retrieved from the database of GEO.Differential expression analysis was utilized for identifying differentially expressed genes between AMR and normal recipients.Multiple machine learning algorithms of random forest(RF),extreme gradient boosting(XGB),support vector machine(SVM)and generalized linear model(GLM)were employed for constructing diagnostic models for AMR after kidney transplantation.Receiver operating characteristic(ROC)curve was plotted for comparing the accuracy of each model.The key genes of optimal model were integrated for creating a diagnostic prediction nomogram for AMR.Calibration curve and decision curve analyses were employed for evaluating the accuracy of nomogram.The differentially expressed genes from biopsy tissues of AMR recipients were uploaded to the database of CMap for identifying potential therapeutic drugs through screening Top 5 compounds with opposite expression patterns to AMR.ResultsSeven genes of CXCL10,FCGR1B,GBP5,CD69,LY96,BCL2A1 and EVI2A were over-expressed in both peripheral blood and biopsy tissues of AMR recipients.There were statistically significant differences with recipients without AMR(FDR<0.05).The AMR diagnostic model based upon RF algorithm demonstrated the highest AUC value(0.904)among various machine learning algorithms.Its AUC values were 0.876 and 0.824 in external datasets of GSE50084 and GSE175718.As for the diagnostic prediction nomogram constructed through integrating five key genes of BCL2A1,CXCL10,FCGR1BP,CD69&EVI2A from RF model,calibration curve indicated that the predicted outcomes of nomogram approximated actual outcomes.Decision curve indicated that net benefit rate of nomogram was higher than that of extreme curves over a wide range of horizontal ax

关 键 词：肾移植 抗体介导排斥反应 生物信息学 机器学习 诊断模型 

分 类 号：R699.2[医药卫生—泌尿科学]