Synthesis of HD

Crude samples of HD representing 11 production methods (Figure ​Figure11) were prepared in house. HD was prepared via the intermediate thiodiglycol (TDG) in routes 1 to 9 (R1–R9). The TDG was produced by three methods and was subsequently transformed into HD using three chlorination protocols, resulting in nine two-step HD production methods that are collectively referred to as TDG routes. Two additional routes (R10 and R11) were included in which gaseous ethylene is directly transformed into HD. Four replicate batches of HD were synthesized by each route and were used to construct attribution models. Crude HD for a test set (one replicate of each of the TDG routes, R1–R9) was synthesized independently, approximately two years after the crude HD training set. However, test set samples for the ethylene routes were obtained from pooled training set samples representing R10 and R11. All synthesis HD batches were stored at room temperature for 1 week and then diluted to 1 mg/mL in dichloromethane, followed by storage at −20 °C.

Open in a separate window

Figure 1

Schematic overview of HD synthesis routes. Two-step routes (R1–9) proceed via the intermediate thiodiglycol (TDG routes). HD is produced directly from ethylene in the one-step ethylene routes (R10 and R11).

Sample Preparation

An overview of the samples included in this study is presented in Table 1, and an overview of the study’s workflow is shown in Figure ​Figure33. Crude HD samples, used for training and test sets, were diluted in dichloromethane to a concentration of 500 ng/μL. An internal standard, dibenzothiophene (1 ng/μL), was added to all samples and used to evaluate performance in terms of peak integration parameters and isotope ratio filters, and to enable semiquantification of CAS by TIC area comparison.

Open in a separate window

Figure 3

Flowchart illustrating sample use and workflow.

All matrix training samples were spiked with pooled crude HD samples. Pooled crude HD samples representing each route (R1–R11) were prepared by mixing 250 μL of each replicate sample for the route in question. The use of pooled crude HD samples made it possible to add exactly the same solution to all six soil and textile matrix samples. Spiked matrix samples were prepared by adding 50 μL of the appropriate pooled crude HD sample (1 mg/mL) to 0.1 ± 0.01 g of soil or textile, theoretically resulting in the addition of HD to a level of 0.05% w/w. Spiked matrix training samples were prepared in triplicate on two occasions, giving six replicates for each material. The matrix test set was prepared by adding 50 μL of the crude HD test set batches (R1–R9, 1 mg/mL) to 0.1 g ± 0.01 g of soil and textile samples in duplicates. All spiked matrix samples were stored in closed glass vials at room temperature overnight. Spiked soil samples were extracted by adding 1 mL of ethyl acetate and shaking (VWR Mini Shaker, PA, U.S.) at 200 rpm for 30 min, followed by centrifugation (Allegra 25 F Beckman Coulter, Bromma, Sweden) at 2500 rpm for 3 min. Ethyl acetate was then decanted, and the extraction process was repeated once, after which the two ethyl acetate extracts were combined. Spiked textile samples were extracted by ultrasonication (Sonorex Digitec Bandlin, Berlin, Germany) in 3 mL of ethyl acetate for 10 min. The ethyl acetate phase was then removed, and the material was extracted once more, after which the two extracts were combined. The combined extracts were filtered (Chromacol PTFE 1 μm, Sigma–Aldrich, MO, U.S.) and concentrated to a volume of 50 μL under a stream of nitrogen at 40 °C. At least one blank sample of each matrix was prepared on each sample preparation occasion. The blank matrix samples were spiked with 50 μL DCM and thereafter treated identically to the spiked matrix samples.

Chemical Analysis

Samples were analyzed with a Trace 1310 gas chromatograph coupled to an Exactive GC Orbitrap mass spectrometer (Thermo Scientific, MA, U.S.). A DB-5MS column (25 m, ID 0.25, 0.25, Agilent, CA, U.S.) was used to separate the compounds. The sample (1 μL) was injected splitless at 200 °C with helium as the carrier gas at a constant flow of 1.2 mL/min. The GC program started at 40 °C for 1 min, followed by a 10 °C/min increase to 300 °C and a hold at 300 °C for 5 min, giving a runtime of 32 min. Mass spectrometric scans were performed in the range of 30–750 m/z with a resolution of 30,000. The temperatures of the transfer line and ion source were set at 250 and 230 °C, respectively. Daily tuning and calibration ensured the quality of the mass spectrometry, and the instrument’s performance was monitored by daily analyses of QC samples.²⁶ Crude HD and spiked matrix samples were analyzed in random order, and every sixth analysis run was done using a solvent blank sample. When analyzing spiked matrix samples, a solvent blank sample and a sample preparation blank were analyzed after every sixth nonblank sample.

Nontargeted Data Processing

The chromatographic data was processed by peak detection, retention time alignment, and peak integration followed by isotope ratio filtering. This resulted in a processed dataset with peak areas from extracted ions at specific retention times corresponding to different compounds. Data processing was done in Tracefinder (version 4.1, Thermo Scientific, MA, U.S.) using the analysis mode for unknown screening, which enables nontargeted screening of data. Peak picking was done with the deconvolution plugin (version 1.3, Thermo Scientific, MA, U.S.). The retention time alignment window was set to 10 s, the accurate mass tolerance to 10 ppm, the signal-to-noise (s/n) threshold to 5, the total ion-chromatogram intensity threshold to 500,000, the ion overlap window to 90–99%, and the response threshold to 10,000.

The extracted peaks were automatically time-aligned and integrated in the unknown screening view using the Avalon detection algorithm and the nearest RT detection method with seven smoothing points. Data representing compounds present in blank samples were manually removed from each crude HD dataset. The datasets for spiked soil and textile samples were processed and manually merged after removing peaks found in blank samples. The variation in the spiked matrix data was higher than in the crude HD data, so the m/z deviation threshold was increased to 0.01 to permit merging.

Isotope ratio filters

Data analysis was done both with and without isotope ratio filtration using sulfur and/or chlorine isotope filters, 1.9958 ± 10 ppm and 1.99705 ± 10 ppm, respectively. The internal standard was used as a control compound to ensure that peak detection was performed correctly and the sulfur isotope ratio filter settings were appropriate. Filtration was done to extract all peaks corresponding to analytes containing chlorine, sulfur, or both. It thus removed all other peaks, including matrix-associated peaks irrelevant to HD route classification. Mass defect filters have previously been used to select for metabolites based on their specific isotope ratios.³⁵

Machine Learning

Data from the crude HD samples were initially analyzed by principal component analysis (PCA) to get an overview of the variation in the data and detect potential outliers. Subsequent analyses were performed in parallel using two machine learning algorithms: orthogonal partial least squares discriminant analysis (OPLS-DA)³⁶ and random forest (RF),³⁷ which were performed in Simca (version 15, Sartorius Stedim Biotech, Germany) and the R software (R³⁸ version 4.0.3 with a random forest 4.6–14³⁹ package), respectively.

OPLS-DA

Prior to analysis, the data was normalized by total area normalization, log-transformed, and scaled to unit variance. The data was too complex to allow classification of 11 routes with a single OPLS-DA model; full classification required a hierarchic decision tree guided by multiple OPLS-DA models. A similar approach was previously used to classify production routes of the nerve agent Russian VX, based on data for spiked food samples.²⁸ The hierarchic decision tree method (Figure ​Figure22) first uses a classification model, M1, which distinguishes between the TDG routes (R1–R9) and the ethylene routes (R10 and R11). A second classification model, M2, was constructed to classify HD samples attributed to the TDG two-step routes (by M1) to the method of chlorination used in their synthesis, i.e., to attribute samples to one of the groups R(1, 4, 7), R(2, 5, 8), or R(3, 6, 9). Finally, data from HD samples sharing the same chlorination method was modeled to classify them based on the method used to synthesize TDG (M3a–c). Five models were constructed for the crude HD samples and another five for the spiked matrix samples. The OPLS-DA models were evaluated by cross-validation with all sample replicates included in the same cross-validation group to avoid overfitting. The classification models were also assessed in terms of their numbers of latent variables (OPLS components) and using three measures of performance: the variation in the data matrix explained by the model (R²X), the variation in the response matrix explained by the model (R²Y), and the variation in the response matrix predicted by the model (Q²). CAS important for class separation were identified by their predictive variable importance in projection (VIP_predictive) values.⁴⁰

Open in a separate window

Figure 2

Hierarchical decision tree where the first model (M1) differentiates between ethylene route and TDG route samples. The second model (M2) differentiates between the three chlorination methods, while M3a–M3c differentiate between methods of TDG synthesis. This model tree was applied to the datasets for the crude HD samples and the spiked matrix samples

Crude HD Samples: Route Attribution through Nontargeted HRMS

The isotope ratio filtered data contained 714 potential CAS whose normalized peak areas were modeled in parallel using two machine learning algorithms, OPLS-DA and RF. The nontargeted method gave classification models of high quality (M1_crude and M2_crude, Table 2). Both OPLS-DA and RF successfully distinguished ethylene routes (R10 and R11) from TDG routes (R1–R9) and predicted the chlorination methods of TDG samples with 100% classification accuracy (Table 2). These results are consistent with previous findings.²¹ A significant advantage of nontargeted data processing in this context is that there is no need to create a target library of the detected compounds for each route.

Crude HD Samples: TDG Attribution through Nontargeted HRMS

Synthesis routes R1 to R9 involve the intermediate TDG, which can be produced by three different methods (Figure ​Figure11). It is challenging to classify the method of TDG synthesis in crude HD samples because the TDG-related CAS are present at levels of <100 pg./μL, approximately 10–100 times lower than those indicative of the chlorination method. This is probably due to the high overall purity of the TDG produced (89–99.5%), and the majority of the CAS have been transformed during the chlorination step. The improved detection of CAS by nontargeted HRMS significantly increased the scope for TDG attribution. Both OPLS-DA and RF models could differentiate the TDG synthesis methods based on 714 potential CAS. The external validation results showed that classification models M3a–c_crude successfully classified TDG synthesis methods in HD samples (OPLS-DA, 56%; RF; 89%, Table 2), showing that the nontargeted HRMS method detected sufficient TDG-related CAS to enable comprehensive route resolution.

CAS Determination and Compound Identity

Chemical identification of CAS is important for understanding their formation during HD production and for verifying that relevant CAS were extracted during the analysis. CAS can be identified in several ways, for example, by searching for specific compounds expected to be present based on prior knowledge, by visual comparison of detected compounds in samples, or by looking at variables found to be important in the classification models. The first two targeted approaches have limitations that restrict the amount of CAS information that can be extracted from the sample data. In this nontargeted work, we identified key CAS that were important variables in the route classification models. While some CAS were clearly associated with the chemistry of HD production, others had no obvious origin. However, evaluations of the effects of different variables on classification model performance clearly showed that successful attribution depended on complete CAS profiles rather than individual compounds (Figure S1, Table S1, S2 and S3).

The CAS found to be related to the chlorination method in two-step routes were largely identical to those previously found by GC–MS.²¹ PCl₃ is used for chlorination in R2, R5, and R8, and gave rise to phosphorous-containing cyclic adducts, notably those with ID numbers M2_crude_11 and M2_crude_14 (Table S2). Chlorination with SOCl₂ (R3, 6, and 9) produced HD containing chlorinated derivatives of amylene, a stabilizer in the dichloromethane used as the reaction solvent. Chlorination with HCl (R1, 4, and 7) yielded few CAS of low intensity. CAS associated with routes 10 and 11 included polysulfides and vinyl chlorides. Dithiane (ID number M3_crude_6, Table S3) is a CAS related to TDG synthesis. Most of the compounds relevant to TDG synthesis method attribution using models M3a_crude, M3b_crude, and M3c_crude could not be identified due to their low abundance (Table S3).

Spiked Matrix Samples: HD Route Attribution through Nontargeted HRMS

Because chemical forensic investigations of incidents involving chemical warfare agents may require analysis of samples taken from the environment, we also investigated the extraction of HD-associated CAS from spiked matrix samples. Route determination based on analysis of CWAs in environmental matrix samples may be complicated by interactions with the sample matrix and/or degradation of the chemicals that constitute the CAS. A low spiking level was chosen to further challenge the analytical method and classification models. Thus, 0.05% w/w of the HD produced by the various routes was spiked to the matrix samples, resulting in CAS concentrations in the range of 0.05–50 ppm. As shown in Figure S2, the most abundant peaks in the chromatograms, aside from the HD peak, originated from the matrix. The use of isotope ratio filters was essential when processing data for these samples because it extracted information on compounds containing chlorine and/or sulfur from the complex HRMS data. Despite the low spiking level, CAS profiles related to specific routes were detected in spiked matrix samples. However, there were large differences between the CAS profiles of crude HD and spiked matrix samples (Figure ​Figure44,) and the crude classifications models could not correctly classify spiked matrix samples. HD and many of the synthesis by-products are clearly highly reactive towards several matrix components, which dramatically altered the CAS attribution profiles. Only 48 of the 714 potential CAS in crude HD were chemically stable in both spiked matrices, and new potential CAS were formed in spiked soil and textile, respectively (Figure ​Figure44). The 48 CAS were not equally distributed between routes and most of them were not important for route separation. Although the CAS profiles of spiked soil and textile samples differed, sufficient common CAS were found to enable the construction of matrix models using the combined soil and textile data. Highly significant classification models could be created for the separation of ethylene routes and TDG routes, and for discriminating between chlorination methods, using either RF or OPLS-DA (Table 3). The quality (R²X, R²Y, Q², and OOB) of spiked matrix models 1 and 2 (M1_matrix and M2_matrix) was comparable to the corresponding crude HD models (M1_crude and M2_crude, Tables 2 and 3). The chosen analytical technique and peak picking method thus made it possible to extract CAS from spiked matrix data despite the low spiking level. The most important variables in the RF and OPLS-DA M2_matrix models were related to PCl₃ chlorination (R2, 5, and 8), with a few relating to SOCl₂ chlorination (R3, 6, and 9), as shown in Table S4. Of the 12 most important variables, only one was related to HCl chlorination. The chlorination method was correctly predicted in 70% of the spiked matrix test set samples independent of classification algorithm (Table 3). Adequate classification models for predicting TDG synthesis methods could be built, but their prediction performance was slightly lower than the corresponding M3a–c_crude models. Spiked matrix test sets representing R7, R8, and R9 were easiest to predict, and the number of correct predictions in spiked soil samples was equal to that in textile samples, indicating that the classification models handled both matrices equally well.

Open in a separate window

Figure 4

Distribution of the 1097 potential CAS found in crude HD and spiked matrices.

Comparison of Classification Tools

The parallel use of OPLS-DA and RF in this study allowed us to benchmark the performance of the two classification tools. When the multimodel hierarchical decision tree (Figure ​Figure22) was used, the two methods achieved very similar predictive accuracy (Table 2). Many of the CAS important for separation by OPLS-DA and RF were identical, especially in M2_crude (Table S2). However, since RF is a tree-based algorithm, it does not require the use of such a decision tree; the tree was only used with RF models to permit detailed performance comparison with OPLS-DA. RF could also be used to resolve all of the classes in a single model; accordingly, 11-class models were constructed for both crude samples (M4_crude) and matrix samples (M4_matrix). The M4_crude model outperformed the OPLS-DA decision tree models for crude samples, with 10/11 correct classifications (Tables 2 and 4).

Conversely, the 11-class RF model for spiked matrix samples (M4_matrix) performed less well than the hierarchical decision tree models generated using both RF and OPLS-DA. In the test set validation of model M4_matrix (Table 4), 18 samples out of 33 were classified correctly.

While the performance differences between the classification methods were small, they differed in their handling of misclassifications: OPLS-DA models assign samples that cannot be classified as not belonging to any class (no class-category in Table 5), whereas RF models force samples into one class. This property was shown when relevant CAS could not be detected in matrix samples spiked with low levels of HD and the samples were incorrectly assigned to the HCl chlorination routes by RF (Table 5). The CAS profile associated with HCl chlorination (R1, 4, 7) is sparse, making it sensible to incorrect classifications of samples with low CAS levels. This RF misclassification problem also affected the 11-class models.

The authors thank Johan Dahlén, Linköpings University, Sweden, for support and valuable discussions. This work was funded by the Swedish Ministry of Defence and the Swedish Civil Contingencies Agency.