摘要
采用气相色谱-质谱联用技术结合化学计量学,针对高维小样本的疾病代谢组学图谱建立高性能的戊二酸血症Ⅰ型(GA-Ⅰ)早期检测模型。基于偏最小二乘判别分析(PLS-DA)的共线性处理和数据解释优势,自助抽样法(Bootstrap)通过数据扰动方式集成多个模型的变量选择能力,挑选出能够持续被筛选的变量实现稳健特征筛选(BS-PLSDA)。对于GA-Ⅰ的尿液代谢组学图谱,在两种逐步增大训练集之间样本差异的比例划分(7:3和6:4)下,载荷(LW)、变量投影重要性(VIP)、显著性多元相关(sMC)3种信息向量对应的BS-PLSDA均优于其单独PLS-DA建模的特征变量筛选稳健性。在样本划分比例为7:3时,BS-VIP-PLSDA的Kuncheva指数高达0.807 5。筛选出的稳健特征变量与文献报道的诊断指标一致,不仅真正解释组别间的差异与GA-Ⅰ的代谢机理密切相关,且BS-LW-PLSDA、BS-VIP-PLSDA和BS-sMC-PLSDA展示了良好的预测性能,受试者工作特征曲线下面积均值分别为0.773 9、0.854 8和0.847 1,马修斯相关系数均值分别为0.671 9、0.783 8和0.801 3。与支持向量机递归特征消除法(SVM-RFE)相比,在采用相同的集成特征选择策略下,尽管非线性径向基核函数对应的BS-RBF-SVMRFE可获得高建模性能,但数据解释能力较低。该研究提出的BS-PLSDA可兼顾建模性能和模型解释能力,符合实际临床需求,对GA-Ⅰ早期检测、辅助诊断和疾病机理研究具有很好的指导意义。
An efficient early detection framework for glutaric acidemia type Ⅰ(GA-Ⅰ)was developed by utilizing urinary metabolomics analysis based on gas chromatography-mass spectrometry(GC-MS)coupled with chemometrics,aiming to overcome small samples and high dimension modeling problems.In the proposed framework,assisted by the capability of partial least squares discriminant analysis(PLS-DA)in collinearity processing and data interpretation,bootstrap was introduced to perform data perturbation and induce multiple base classifiers,integrating their feature selection strengths and forming a novel algorithm of BS-PLSDA.Based on three informative vectors of loading weights(LW),variable importance in the projection(VIP)and significance multivariate correlation(sMC),the formed novel algorithm BS-PLSDA enabled the screening of discriminative features that were so strong to survive across multiple base classifiers.Investigated by GC-MS urinary metabolomic profiling of GA-Ⅰ,the results showed that BS-PLSDAs of three informative vectors all outperformed their corresponding PLS-DAs modeled by single classifier in selection stability,even if the ratio of sample partitioning was altered from 7:3 to 6:4,gradually increasing the sample difference among training sets.When the ratio of sample partitioning was 7:3,the Kuncheva index of BS-VIP-PLSDA could reach to 0. 807 5.Furthermore,the screened stable discriminative features exhibited close biological correlations to the metabolic mechanism of GA-Ⅰ,in which several reported diagnostic organic acids were searched.Meanwhile,they yielded desired predictive powers that the averages of area under receiver operating characteristic curve(AUC)were 0. 773 9,0. 854 8and 0. 847 1,while Matthews correlation coefficient(MCC)were 0. 671 9,0. 783 8 and 0. 801 3 for BS-LW-PLSDA,BS-VIP-PLSDA and BS-sMC-PLSDA,respectively.Finally,a comparison was performed between PLS-DA and support vector machine recursive feature elimination(SVM-RFE).Equipped with the same ensemble feature selection strategy,the model BS-RBFSVMRFE using nonlinear radial basis function(RBF)was superior to BS-PLSDAs in classification performance.Nevertheless,it obtained poor model interpretability.All the results revealed that the proposed BS-PLSDA exhibited its modeling feasibilities both in classification performance and data interpretation, resulting in good meet in clinical demand. It suitably guided the early detection,and aided clinical diagnosis and disease mechanism understanding for GA-Ⅰ.
作者
肖雯
牛芊芊
孙智勇
杨琴
吴本清
XIAO Wen;NIU Qian-qian;SUN Zhi-yong;YANG Qin;WU Ben-qing(School of Physics and Optoelectronic Engineering,Yangtze University,Jingzhou 434023,China;Rare Disease Engineering Research Center of Metabolomics in Precision Medicine,Shenzhen Aone Medical Laboratory Co.,Ltd.,Shenzhen 518000,China;Shenzhen Hospital,University of Chinese Academy of Sciences,Shenzhen 518000,China)
出处
《分析测试学报》
CAS
CSCD
北大核心
2022年第11期1577-1583,共7页
Journal of Instrumental Analysis
基金
国家自然科学基金项目(21803009)
湖北省高等学校优秀中青年科技创新团队计划项目(T2020008)
深圳市工程研究中心(工程实验室)组建项目(F-2020-Z99-502615)
深圳市科技创新委员会基础研究学科布局项目(JCYJ20180507183428877:20180253)
大学生创新创业训练计划项目(Yz2021267)。
关键词
戊二酸血症Ⅰ型
早期检测
气相色谱-质谱
偏最小二乘判别分析
自助抽样法
稳健特征筛选
glutaric acidemia typeⅠ
early detection
gas chromatography-mass spectrometry
partial least squares discriminant analysis
bootstrap
stable feature selection
