Department of CBRN Defence & Security, The Swedish Defence Research Agency (FOI), Cementvägen 20, Umeå SE-901 82, Sweden.
Anal Chem. 2021 Mar 23;93(11):4850-4858. doi: 10.1021/acs.analchem.0c04555. Epub 2021 Mar 12.
Route determination of sulfur mustard was accomplished through comprehensive nontargeted screening of chemical attribution signatures. Sulfur mustard samples prepared via 11 different synthetic routes were analyzed using gas chromatography/high-resolution mass spectrometry. A large number of compounds were detected, and multivariate data analysis of the mass spectrometric results enabled the discovery of route-specific signature profiles. The performance of two supervised machine learning algorithms for retrospective synthetic route attribution, orthogonal partial least squares discriminant analysis (OPLS-DA) and random forest (RF), were compared using external test sets. Complete classification accuracy was achieved for test set samples (2/2 and 9/9) by using classification models to resolve the one-step routes starting from ethylene and the thiodiglycol chlorination methods used in the two-step routes. Retrospective determination of initial thiodiglycol synthesis methods in sulfur mustard samples, following chlorination, was more difficult. Nevertheless, the large number of markers detected using the nontargeted methodology enabled correct assignment of 5/9 test set samples using OPLS-DA and 8/9 using RF. RF was also used to construct an 11-class model with a total classification accuracy of 10/11. The developed methods were further evaluated by classifying sulfur mustard spiked into soil and textile matrix samples. Due to matrix effects and the low spiking level (0.05% w/w), route determination was more challenging in these cases. Nevertheless, acceptable classification performance was achieved during external test set validation: chlorination methods were correctly classified for 12/18 and 11/15 in spiked soil and textile samples, respectively.
通过全面的非靶向化学归因特征筛选，确定了芥子气的合成路线。采用气相色谱/高分辨质谱法分析了通过 11 条不同合成路线制备的芥子气样品。检测到大量化合物，并对质谱结果进行多元数据分析，发现了特定路线的特征谱图。采用正交偏最小二乘判别分析（OPLS-DA）和随机森林（RF）两种有监督机器学习算法对回溯合成路线归属进行性能比较，使用外部测试集。使用分类模型解析以乙烯为起始原料的一步法路线和两步法中的硫代二甘醇氯化方法的测试集样品（2/2 和 9/9），实现了完全分类准确性。在经过氯化后，追溯初始硫代二甘醇合成方法在芥子气样品中的应用更为困难。然而，使用非靶向方法检测到的大量标记物使得使用 OPLS-DA 正确分配了 5/9 的测试集样品，使用 RF 则正确分配了 8/9。RF 还用于构建总分类准确率为 10/11 的 11 类模型。通过对土壤和纺织品基质样品中添加的芥子气进行分类，进一步评估了所开发的方法。由于基质效应和低添加水平（0.05%w/w），在这些情况下，路线确定更具挑战性。然而，在外部测试集验证期间，仍实现了可接受的分类性能：在添加到土壤和纺织品样本中的芥子气中，氯化方法的分类正确率分别为 12/18 和 11/15。
