Headspace Gas Chromatography Coupled to Mass Spectrometry and Ion Mobility Spectrometry: Classification of Virgin Olive Oils as a Study Case

Due to its multiple advantages, ion mobility spectrometry (IMS) is being considered as a complementary technique to mass spectrometry (MS). The goal of this work is to investigate and compare the capacity of IMS and MS in the classification of olive oil according to its quality. For this purpose, two analytical methods based on headspace gas chromatography (HS-GC) coupled with MS or with IMS have been optimized and characterized for the determination of volatile organic compounds from olive oil samples. Both detectors were compared in terms of sensitivity and selectivity, demonstrating that complementary data were obtained and both detectors have proven to be complementary. MS and IMS showed similar selectivity (10 out of 38 compounds were detected by HS-GC-IMS, whereas twelve compounds were detected by HS-GC-MS). However, IMS presented slightly better sensitivity (Limits of quantification (LOQ) ranged between 0.08 and 0.8 µg g−1 for HS-GC-IMS, and between 0.2 and 2.1 µg g−1 for HS-GC-MS). Finally, the potential of both detectors coupled with HS-GC for classification of olive oil samples depending on its quality was investigated. In this case, similar results were obtained when using both HS-GC-MS and HS-GC-IMS equipment (85.71 % of samples of the external validation set were classified correctly (validation rate)) and, although both techniques were shown to be complementary, data fusion did not improve validation results (80.95% validation rate).


Introduction
Ion mobility spectrometry (IMS) is a cutting-edge technique that, coupled with gas chromatography (GC), is proving to be a powerful analytical tool in a wide range of research fields, such as foodomics [1,2]. IMS is based on gas phase ion separation inside a drift tube under the influence of a constant electric field at atmospheric pressure. Neutral and gaseous molecules are ionized through the ionization source, and generated ions travel to the drift tube through the shutter grid. The drift time of the ions is characteristic of the analyte because of their different size and shape and is measured in milliseconds (ms). IMS is able to provide analytical information of a great number of samples due to it being a fast and very sensitive technique and, moreover, it requires minimal or no sample preparation; due to its multiple advantages, it is being considered as a complementary technique for mass spectrometry

Samples
A total of 160 olive oil samples (52 EVOO, 56 VOO, and 52 LOO) were analyzed and used for training and evaluation models. In addition, a set of 21 samples (7 EVOO, 7 VOO, and 7 LOO) was used as an external validation set. All samples were from different geographical areas of Spain and were supplied by Sovena S.A. (Sevilla, Spain).
Samples were stored in the freezer in individual bottles without headspace. One gram of these samples was placed in a 20 mL vial closed with a magnetic cap and silicone septum and stored at 4 • C before their analysis (no more than 24 h). Finally, samples were defrosted at room temperature for 30 min, shaken by vortex for 1 min and submitted to HS-GC-MS and HS-GC-IMS analysis.

Instrumentation and Software
Analyses were performed on a multipurpose autosampler (MPS) headspace unit provided by Gerstel and coupled to an Agilent Technologies 6890N (N.05.05 version) gas chromatograph (Agilent, Waldbronn, Germany), which was coupled with a 5973N simple quadrupole mass selective spectrometer equipped with an inert ion source, or with an IMS module from G.A.S (Gesellschaft für Analytische Sensorsysteme mbH, Dortmund, Germany) equipped with a Tritium source and drift tube of 9.8 cm.
In both instruments, analytes were separated in a non-polar GC column HP-5MS UI (Agilent), 30 m length, 0.25 mm internal diameter, and 0.25 µm film thickness.
A vortex from Heathrow Scientific (Heathrow Scientific LLC, Vernon Hills, IL, USA) was also used during the sample treatment.

HS-GC-IMS Analysis
A sample of 1 g was incubated at 90 • C for 3 min (750 rpm). Injection of 750 µL from headspace was carried out using a 2.5 mL syringe at 90 • C in splitless mode. Nitrogen of 99.99% purity (supplied by Air Liquide, Madrid, Spain) was used as the carrier gas at a constant flow rate of 1 mL min −1 . The oven was set as follows: initial temperature of 50 • C held for 3 min, which was increased from 50 • C to 120 • C at 5 • C min −1 and held at 120 • C for 3 min (total run: 20 min). Analytes were driven to the IMS module and ionized by a Tritium source at atmospheric pressure in a positive ion mode. Nitrogen was also used as a drift gas at a constant flow of 150 mL min −1 . The IMS was operated at a constant voltage of 500 V cm −1 and a temperature of 45 • C. Each spectrum was acquired with an Foods 2020, 9, 1288 4 of 18 average of 32 scans, obtained using a repetition rate of 30 ms, a grid pulse width of 150 µs, and drift and blocking voltages of 241 V and 70 V, respectively.

HS-GC-MS Analysis
A sample of 1 g was incubated at 90 • C for 15 min (750 rpm), no magnetic bar was used for vortexing. Then, 750 µL were injected at 90 • C in splitless mode (using a 2.5 mL syringe). Helium (99.9999% purity) supplied by Air Liquide (Madrid, Spain) was used as the carrier gas at a constant flow rate of 1 mL min −1 . The oven was set at an initial temperature of 40 • C held for 3 min and heated from 40 • C to 250 • C at a rate of 10 • C min −1 and held at 250 • C for 6 min (total run: 30 min). The MS was operated in electron impact mode at 70 eV; ionization energy and data were collected in the range of 20-400 m/z. The ion source, transfer line, and quadrupole temperatures were 230, 300, and 150 • C, respectively.
An HS-GC-IMS analysis results in a tri-dimensional map in which the Y axis represents the retention time in the chromatographic column (in seconds), the X axis represents the drift time in the drift tube (in milliseconds), and the Z axis represents the intensity value (in V) of each compound. For data processing, an initial step of peak alignment was carried out, it was performed only in a retention time scale using a reactant ion peak (RIP) as reference and LAV software. Then, markers were selected by visual exploration of the topographic plots of each sample and their intensities were selected as the analytical signals.
The dataset was formed with all samples and all selected markers. This strategy has been already proposed for IMS data processing, specifically for olive oil analysis [20,22,23], although, it has been also used for other food applications [1].
Data from HS-GC-MS were processed following two different strategies: the use of the whole chromatographic profile, i.e., the total ion chromatogram (TIC), or the use of the areas of the main chromatograph peaks. No data pre-processing was needed for the dataset of peak integration. However, in the case of the TIC dataset, baseline correction was necessary. It was corrected by subtracting the mean value of the background.
After data pre-processing, the tree different datasets were divided into two groups: the training set for the construction of the chemometric models (128 samples) and evaluation set validation (32 samples) for the optimization of method parameters. In addition, an external test set (21 samples) was used for method validation. The constructions of chemometric models were carried out based on previous reported methods for the classification of olive oil [20,22]. A non-supervised PCA analysis using auto-scales was carried out in order to reduce the dimensionality. Then, PCA scores were used to carry out an LDA. Finally, k-NN, using k = 3, was applied to classify the samples.

Optimization of HS-GC-MS and HS-GC-IMS Methods
The optimization of both analytical methods was carried out with the aim of achieving the best results in terms of intensity and separation between peaks, from the previously described conditions [6][7][8][9][10][20][21][22][23]. With this purpose, the variable injection volume, time, and temperature of incubation, and injector temperature were investigated using both techniques.
The effect of the injection volume was studied between 500 µL and 750 µL, obtaining higher intensity signals with 750 µL. The effect of the sample incubation temperature was studied between 60 • C and 90 • C. High temperatures facilitate the release of volatile organic compounds with high boiling points. This caused an increase in the number of signals and their intensities in the 60-90 • C Foods 2020, 9, 1288 5 of 18 range, therefore, 90 • C was selected as the optimum. Then, the sample incubation time was studied between 5 and 20 min. For HS-GC-MS, the signal intensities increased as the incubation time increases, however no significant differences were found between 15 and 20 min, and therefore, sample incubation time was set at 15 min (Supplementary Materials Figure S1). In the case of HS-GC-IMS, no significant differences were appreciated between 5 and 20 min, thus the incubation time was investigated in the range of 1 to 5 min, selecting 3 min as the optimum, since higher temperatures did not improve the spectrum (Supplementary Materials Figure S2). Finally, the injector temperature was studied between 70 • C and 90 • C and no significant differences were found. For this reason, the injector temperature was set to 90 • C (temperature of sample incubation). Based on previous work, the salt addition was not considered since it does not cause any increase in VOC signals and in contrast, a reduction of some signals is observed when the saturated salt solution is added to the olive oil [23].
The oven program was also studied for both methods in order to achieve optimal conditions. The best peak separation in HS-GC-MS was obtained with the following conditions: initial temperature of 40 • C held for 3 min, increased to 250 • C at 10 • C min −1 and held at 250 • C for 6 min. For HS-GC-IMS, temperatures higher than 120 • C were not recommended by the manufacturer of the IMS, since the drift tube has a temperature limitation of 100 • C, therefore the oven was set as follows: initial temperature of 50 • C were held for 3 min, which was increased from 50 • C to 120 • C at 5 • C min −1 and held at 120 • C for 3 min.
Finally, the drift tube temperature of HS-GC-IMS was investigated between 45 and 75 • C and no significant differences were obtained, therefore 45 • C was selected for further experiments.
In order to evaluate both analytical methods in term of validation, calibration curves were established for ethyl acetate, 1-penten-3-one, 2-pentanone, 4-methyl-pentan-2-one, hexanal, trans-2-pentenal, trans-2-hexen-1-al, heptanal, 6-methyl-5-hepten-2-one, 3-hexenyl acetate, nonanal, decanal, trans-2-decenal, and hexyl acetate using refined oil spiked at six concentration levels between 0.05 and 50 µg g −1 . Each concentration level was injected twice. The statistical parameters were calculated by least-square regression for HS-GC-MS. In the case of HS-GC-IMS data, because each compound can generate two signals corresponding to monomers and dimers, different calibration graphs were considered. Specifically, the least-square and logarithmic regressions were tested. In both cases, graphics were constructed using the signal of the monomer and the dimer or the sum of both monomer and dimer signals. The best results were obtained using the logarithmic regressions and the sum of monomer and dimer signals. As can be seen in Table 1, satisfactory determination coefficients were obtained in all the cases, although they were slightly better in MS (R 2 > 0.98 for MS and R 2 > 0.92 for IMS).
Limits of detection (LODs) and quantification (LOQs) were estimated as 3 x signal-to-noise ratio (S/N) and 10 × S/N, respectively. The LOQs ranged between 0.2 and 2.1 µg g −1 for HS-GC-MS and between 0.08 and 0.82 µg g −1 for HS-GC-IMS. Trans-2-hexen-1-al, hexanal, and nonanal could be detected and quantified for both methods, although slightly better LOQ were obtained with HS-GC-IMS.
In order to compare the two different analytical methods, a simple regression using Statgraphic software was carried out. This study was performed with the compound that could be detected and quantified by both techniques, named trans-2-hexen-1-al, hexanal, and nonanal. In all the cases, P-value was greater than 0.05, so there was not a statistically significant relationship between IMS and MS data at the 95.0% confidence level. This could be justified by the difference in the analytical response obtained since, at the range of concentrations studied, MS data were adjusted to a linear regression while a logarithmic adjustment was necessary with the IMS data.
Calibration curves were used to quantify the identified compounds by both techniques in the olive oil samples and this information was used to compare the three categories applying ANOVA and Tukey's test as post hoc test. Table 2 shows the concentration average of each analyte for each category and the Tukey's test results using both techniques.
As has been shown, certain compounds can be associated with a particular category. However, the high variability observed within the same group did not allow the establishment of a compound concentration limit in each category and therefore chemometric models are necessary. Table 1. Calibration curves and performance characteristics of the headspace-gas chromatography (HS-GC) coupled with mass spectrometry (MS) and ion mobility spectrometry (IMS) methods.

Chemometrics for Olive Oil Classification According to Its Quality
Chemometric models were constructed using the data obtained by HS-GC-MS and HS-GC-IMS. Different chemometric approaches were investigated to process HS-GC-MS data, specifically, the selection of significant marker peaks (known and unknown compounds) and the use of the total ion chromatogram (TIC).
As described in Section 2.6, HS-GC-IMS data processing has been well-studied and it was performed following the indication of Contreras et al. [23] In this case, chemometric models based on PCA-LDA were preferred to the use of modelling techniques such as SIMCA or QDA. Several papers have demonstrated the poor performance of SIMCA as compared to other methods, e.g., to LDA. The fact that LDA was developed by statisticians, whereas SIMCA was developed by chemists (chemometricians) might contribute to the characteristic differences between their theoretical backgrounds. For example, SIMCA does not require any distributional assumptions, whereas LDA assumes normal distribution and equal variances for each class [37]. In addition, Nikita et al. observed that QDA does not give better results than LDA and does not offer an alternative to LDA [38].
A total of 86 markers were selected by visual exploration of the topographic plots obtained by HS-GC-IMS and their intensities were used as a dataset. Therefore, the data matrix had a dimension of 160 (samples) × 86 (markers). This matrix was split into two datasets: calibration samples (80%) and evaluation samples (20%), i.e., 128 samples were used for model construction and 32 for the optimization of K in the K-NN model. After PCA, 29 principal components were obtained including the 99.07% of variance cumulative. The results of PCA are shown in Supplementary Materials Figure S3. The PCA-LDA model showed a clear separation between each group of samples ( Figure 3). As has been shown, certain compounds can be associated with a particular category. However, the high variability observed within the same group did not allow the establishment of a compound concentration limit in each category and therefore chemometric models are necessary.

Chemometrics for Olive Oil Classification According to Its Quality
Chemometric models were constructed using the data obtained by HS-GC-MS and HS-GC-IMS. Different chemometric approaches were investigated to process HS-GC-MS data, specifically, the selection of significant marker peaks (known and unknown compounds) and the use of the total ion chromatogram (TIC).
As described in Section 2.6., HS-GC-IMS data processing has been well-studied and it was performed following the indication of Contreras et al. [23] In this case, chemometric models based on PCA-LDA were preferred to the use of modelling techniques such as SIMCA or QDA. Several papers have demonstrated the poor performance of SIMCA as compared to other methods, e.g., to LDA. The fact that LDA was developed by statisticians, whereas SIMCA was developed by chemists (chemometricians) might contribute to the characteristic differences between their theoretical backgrounds. For example, SIMCA does not require any distributional assumptions, whereas LDA assumes normal distribution and equal variances for each class [37]. In addition, Nikita et al. observed that QDA does not give better results than LDA and does not offer an alternative to LDA [38].
A total of 86 markers were selected by visual exploration of the topographic plots obtained by HS-GC-IMS and their intensities were used as a dataset. Therefore, the data matrix had a dimension of 160 (samples) × 86 (markers). This matrix was split into two datasets: calibration samples (80%) and evaluation samples (20%), i.e., 128 samples were used for model construction and 32 for the optimization of K in the K-NN model. After PCA, 29 principal components were obtained including the 99.07% of variance cumulative. The results of PCA are shown in Supplementary Materials Figure  S3. The PCA-LDA model showed a clear separation between each group of samples ( Figure 3). The k-NN was then optimized using the evaluation set. A classification rate of 87.50% and 84.50% were obtained for k = 3 and k = 5, respectively, so k = 3 was selected as optimum. Finally, the chemometric models were applied to classify other 20 samples (external validation set), obtaining a validation rate (percentage of samples of the external validation set classified correctly) of 85.71% owing to three samples being incorrectly classified: one EVOO as VOO, one VOO as EVOO, and one VOO as LOO (Table 3). Table 3. Validation matrix by k-NN using HS-GC-IMS data. The k-NN was then optimized using the evaluation set. A classification rate of 87.50% and 84.50% were obtained for k = 3 and k = 5, respectively, so k = 3 was selected as optimum. Finally, the chemometric models were applied to classify other 20 samples (external validation set), obtaining a validation rate (percentage of samples of the external validation set classified correctly) of 85.71% owing to three samples being incorrectly classified: one EVOO as VOO, one VOO as EVOO, and one VOO as LOO (Table 3). As mentioned above, two different chemometric strategies were investigated to process HS-GC-MS data. The first strategy consisted of using the chromatographic peak areas, including known and unknown compounds, in order to obtain the maximum possible information of the chromatogram. To do so, a total of 95 peaks were integrated and processed. The second strategy consisted of the use of the TIC and, firstly, the need for an alignment step was evaluated; no misalignment was detected between samples. However, a baseline shift was observed, and therefore a pre-processing step consisting of a baseline correction was carried out. Baseline was corrected by subtracting the mean value of background (an empty section of peaks, between 14.9 and 15.34 min). In this case, the dimensions of the data matrix were 160 × 95 for the first strategy and 160 × 5272 for the second strategy, since TIC was composed of 5272 values.
These matrices were also split into two datasets: a training set (128 samples) and an evaluation set (32 samples). The PCA allowed a reduction of the dimensionality to 58 and 122 principal components (99% variance cumulative) for the first and second strategy, respectively. The results of the PCA are shown in Supplementary Materials Figures S4 and S5. PCA-LDA models are shown in Figures 4 and 5. Foods 2020, 9,  As mentioned above, two different chemometric strategies were investigated to process HS-GC-MS data. The first strategy consisted of using the chromatographic peak areas, including known and unknown compounds, in order to obtain the maximum possible information of the chromatogram. To do so, a total of 95 peaks were integrated and processed. The second strategy consisted of the use of the TIC and, firstly, the need for an alignment step was evaluated; no misalignment was detected between samples. However, a baseline shift was observed, and therefore a pre-processing step consisting of a baseline correction was carried out. Baseline was corrected by subtracting the mean value of background (an empty section of peaks, between 14.9 and 15.34 min). In this case, the dimensions of the data matrix were 160 × 95 for the first strategy and 160 × 5272 for the second strategy, since TIC was composed of 5272 values.
These matrices were also split into two datasets: a training set (128 samples) and an evaluation set (32 samples    Different K-NN models, using k = 3 and k = 5, were applied to the evaluation set. The same classification rate was obtained for peak integration data (78.13%), however better results were obtained with k = 3 for TIC data (87.50% for k = 3 and 84.40 % for k = 5). The application of the k-NN method (k = 3) to the external validation set is shown in Tables 4 and 5, and as can be seen, better validation rates (85.71%) were obtained using the TIC, similar to that obtained by IMS. In this case, all LOO samples were also correctly classified. However, one EVOO and two VOO were also misclassified. The results obtained with only one chemometric model are comparable to those previously reported by means of two sequential models [10]. The use of the chromatographic peak areas showed validation rates of 76.19%, where two EVOO samples were classified as VOO, two VOO samples as EVOO and LOO, and one LOO sample as VOO.

Data Fusion of MS and IMS
HS-GC-MS and HS-GC-IMS showed the same validation rate, demonstrating that both techniques are appropriate for the classification of olive oil samples. In addition, they have been shown to be complementary, since they allowed the detection and quantification of different Different K-NN models, using k = 3 and k = 5, were applied to the evaluation set. The same classification rate was obtained for peak integration data (78.13%), however better results were obtained with k = 3 for TIC data (87.50% for k = 3 and 84.40 % for k = 5). The application of the k-NN method (k = 3) to the external validation set is shown in Tables 4 and 5, and as can be seen, better validation rates (85.71%) were obtained using the TIC, similar to that obtained by IMS. In this case, all LOO samples were also correctly classified. However, one EVOO and two VOO were also misclassified. The results obtained with only one chemometric model are comparable to those previously reported by means of two sequential models [10]. The use of the chromatographic peak areas showed validation rates of 76.19%, where two EVOO samples were classified as VOO, two VOO samples as EVOO and LOO, and one LOO sample as VOO.

Data Fusion of MS and IMS
HS-GC-MS and HS-GC-IMS showed the same validation rate, demonstrating that both techniques are appropriate for the classification of olive oil samples. In addition, they have been shown to be complementary, since they allowed the detection and quantification of different compounds. Therefore, in an attempt to improve the validation rate, the construction of chemometric models using data fusion was investigated.
With that purpose, the 95 peak areas selected from HS-GC-MS data and the 86 markers from topographic maps of HS-GC-IMS were united to characterize each olive oil sample, i.e., a final data matrix with dimensions of 160 (samples) × 181 (markers) was used to carry out the chemometric treatment. After PCA, the dataset was decreased to 45 principal components (99.10% variance cumulative). Better results were obtained using k = 3 when the K-NN method was applied to the evaluation set (k = 3, 81.25%; k = 5, 71.85%), so k = 3 was also selected as optimum. A validation success rate of 80.95% (Table 6) was obtained despite the good separation obtained with PCA-LDA ( Figure 6). One EVOO sample was classified as VOO and one VOO was upgraded to EVOO. The most important mistake was made when classifying the LOO sample, since one of them was classified as VOO and the other one as EVOO. For all these reasons, data fusion did not improve the results obtained and, in this case, HS-GC-MS and HS-GC-IMS could not be used as complementary techniques, but rather as alternatives. compounds. Therefore, in an attempt to improve the validation rate, the construction of chemometric models using data fusion was investigated.
With that purpose, the 95 peak areas selected from HS-GC-MS data and the 86 markers from topographic maps of HS-GC-IMS were united to characterize each olive oil sample, i.e., a final data matrix with dimensions of 160 (samples) × 181 (markers) was used to carry out the chemometric treatment. After PCA, the dataset was decreased to 45 principal components (99.10% variance cumulative). Better results were obtained using k = 3 when the K-NN method was applied to the evaluation set (k = 3, 81.25%; k = 5, 71.85%), so k = 3 was also selected as optimum. A validation success rate of 80.95% (Table 6) was obtained despite the good separation obtained with PCA-LDA ( Figure 6). One EVOO sample was classified as VOO and one VOO was upgraded to EVOO. The most important mistake was made when classifying the LOO sample, since one of them was classified as VOO and the other one as EVOO. For all these reasons, data fusion did not improve the results obtained and, in this case, HS-GC-MS and HS-GC-IMS could not be used as complementary techniques, but rather as alternatives.

Discussion
In this work, two coupling techniques, HS-GC-MS and HS-GC-IMS, have been evaluated for the classification of olive oil samples.
Both techniques have proven to be complementary for identification and quantification of

Discussion
In this work, two coupling techniques, HS-GC-MS and HS-GC-IMS, have been evaluated for the classification of olive oil samples.
Both techniques have proven to be complementary for identification and quantification of characteristic olive oil VOCs, since they were able to monitor different compounds. Specifically, ten compounds were identified in olive oil samples using HS-GC-IMS, whereas twelve compounds were identified using HS-GC-MS. Only trans-2-hexen-1-al, hexanal, and nonanal were identified by both techniques. Regarding sensitivity, HS-GC-IMS showed LOQ slightly better.
To obtain calibration curves, a least-square regression was applied in HS-GC-MS; however, logarithmic regressions considering the sum of monomer and dimer signals as analytical responses had to be used in HS-GC-IMS. Satisfactory determination coefficients were obtained in all cases, although they were slightly better in MS.
Certain compounds detected for IMS and/or MS techniques can be associated with a particular category. Specifically, heptanal, 6-methyl-5-hepten-2-one, nonanal, and trans-2-decenal allowed differentiation between LOO and the edible olive oil samples; whereas ethyl acetate, 3-hexenyl acetate, and trans-2-hexen-1-al enabled the differentiation between non-defective (EVOO) and defective (non-EVOO) olive oil samples in concordance with Romero et al. [5] The ketone 1-penten-3-one was the only compound that allowed the differentiation between the three categories, also in concordance with Garrido-Delgado et al. [36]. However, the high variability observed within the same group did not allow the establishment of a compound concentration limit in each category, making necessary the use of chemometric models.
Finally, HS-GC-MS and HS-GC-IMS models showed the same validation rate (85.71%) for classification of olive oil samples, although it was necessary to use the entire chromatographic profile obtained by MS to achieve the results obtained by IMS. The validation results of HS-GC-MS are similar to other previous reported work, although in those cases, an SPME step was used instead of HS. Quintanilla-Casas et al. obtained results of 89.2% using two sequential PLS discriminant analysis models and an untargeted fingerprinting strategy [10]. Sales et al. obtained 70 and 85% of classification rate monitoring 15 VOCs [7] or using untargeted fingerprinting strategies, respectively [11]. The other work proposed for olive oil classification do not distinguish between the three classes of olive oil so they could not be compared. In the case of HS-GC-IMS, classification rates ranged between 94% and 100% using a 60 m capillary column, ramped temperature, and a supervised method such as OPLS-DA [23]. Gerhardt et al. obtained 83.3% using untargeted fingerprinting and LDA chemometric models [27] and Valli et al. obtained 77% using a features signal of 15 VOCs and PCA and PLS-DA models [25].
Therefore, both techniques are appropriate for the classification of olive oil samples and could be used as complementary or screening support to the test panel. However, both techniques should not be applied together since data fusion did not improve the results.
Supplementary Materials: The following are available online at http://www.mdpi.com/2304-8158/9/9/1288/s1, Figure S1: HS-GC-MS chromatogram obtained for 5, 10, 15 y 20 min of incubation time, Figure S2. HS-GC-IMS spectra obtained for 1, 3 y 5 min of incubation time, Figure S3. PCA scores and loading plot using the HS-GC-IMS markers, Figure S4. PCA scores and loading plot using the peak integration of the HS-GC-MS method, Figure S5. PCA scores and loading plot using the TIC of the HS-GC-MS method Table S1. Chemical information of the VOCs.