【摘要】 为拓展挥发组学在血液种属鉴别（BSI）领域的应用，该研究测定了多种物种血液的挥发性有机物（VOCs），构建了机器学习分类模型并评估其性能。使用顶空固相微萃取/气相色谱-质谱（HS-SPME/GC-MS）检测8种常见物种血液中的VOCs，利用偏最小二乘判别分析（PLS-DA）和正交偏最小二乘判别分析（OPLSDA）筛选潜在标志物，使用9种常见机器学习分类算法构建一对多分类模型，进行基准测试、超参数调优、算法性能评估，并评估重采样方法和标志物变量选择的影响。筛选出人类与7种不同动物血液种属特征相关的潜在标志物共17种，基准测试获得最优算法为K-最近邻（KNN）算法，分类模型最优超参数组合为：K值为5，距离加权核函数为三权函数，闵氏距离参数p为0.324 0；验证集下最优模型准确度、曲线下面积、布里尔分数分别为0.928 4、0.997 0、0.057 6。使用所有成分变量的模型与仅使用潜在标志物变量模型的结果无显著差异（t检验p> 0.05），不同重采样方法模型结果无显著差异（t检验p> 0.05）。该研究表明，挥发组学分析在血液种属鉴别方面极具潜力，潜在标志物可靠性强、模型准确度高、抗干扰能力强。更多还原

【Abstract】 Blood species identification(BSI) plays a significant role in criminal investigation,import and export inspection,animal protection and other fields. Volatilomics analysis of blood volatile organic compounds(VOCs) is a novel approach for blood species identification. To screen potential biomarkers of blood from different species, the study established multiple machine learning(ML) classification algorithms, and compared the predictive value of different classification models for blood species identification. Headspace solid-phase microextraction(HS-SPME) coupled with gas chromatography-mass spectrometry(GC-MS) was used to analyze VOCs in the blood of eight common species. Partial least squares discriminant analysis(PLS-DA) and orthogonal partial least squares discriminant analysis(OPLS-DA) were employed to screen potential biomarkers. Samples were randomly divided into training and testing sets at a ratio of 7∶ 3. Nine common classification models were established, and the best algorithm was selected and optimized by comparing all models. A model using all VOCs as variables was constructed to verify the reliability of the potential biomarkers,and different resampling methods were used to assess the impact of training and testing set division on the model. A total of 17 VOCs related to species characteristics of human and seven different animal bloods were screened. The accuracy of the multi-layer perceptron,naive Bayes algorithm,multinomial logistic regression algorithm,K-nearest neighbor(KNN) algorithm,Gaussian kernel function support vector machine,polynomial kernel function support vector machine,decision tree,random forest model, and extreme gradient boosting tree model were 0. 859 7, 0. 575 1, 0. 859 7,0. 942 1, 0. 815 0, 0. 734 2, 0. 842 9, 0. 923 1 and 0. 872 9, respectively. Among them, the accuracy,area under the receiver operating characteristic curve(AUC),and Brier score of the KNN model under the testing set were 0. 918 4,0. 999 0 and 0. 037 6. KNN was selected as the optimal algorithm,and the best model’s hyperparameter combination was:K value of 5,distance-weighted kernel function of the triweight function,and the Minkowski distance parameter p of 0. 324 0. The best-performing model on the validation set achieved an accuracy of 0. 928 4, with corresponding metrics of 0. 997 0 for the AUC and 0. 057 6 for the Brier score. There was no significant difference in the results between the model using all component variables and the potential biomarker variable model(t-test p>0. 05), and there was no significant difference in the results of models using different resampling methods(t-test p>0. 05). Volatilomics analysis shows great potential in blood species identification,with strong reliability of potential biomarkers,high model accuracy,and strong antiinterference ability.更多还原