Bottles of mezcal came from producers located in Oaxaca, San Luis Potosí, and Puebla (Table 1). Fifteen of them (ID: 301, 310, 450, 460, 470, 580, 601, 630, 640, 690, 701, 710, 760, 790, and 920) were collected in 2021, and six (ID: 239, 245, 373, 688, 863 and 949) in 2022. The mezcales were made with the species Agave angustifolia, Agave potatorum, Agave salmiana, and Agave karwinskii, which in turn comprised two elaboration processes: artisanal and ancestral. The experimental units for this study were mezcal bottles of 250 mL. The experiment was carried out in triplicate.

Table 1

ID	Agave species	Geographical origin	Process	State
301	A. angustifolia	Sierra Sur	Ancestral	Oaxaca
310	A. angustifolia	Valles Centrales	Artisanal	Oaxaca
450	A. angustifolia	Valles Centrales	Artisanal	Oaxaca
460	A. potatorum	Valles Centrales	Artisanal	Oaxaca
470	A. angustifolia	Valles centrales	Artisanal	Oaxaca
580	A. angustifolia	Valles Centrales	Artisanal	Oaxaca
601	A. potatorum	Sierra Sur	Ancestral	Oaxaca
630	A. angustifolia	Sierra Sur	Ancestral	Oaxaca
640	A. angustifolia	Valles Centrales	Artisanal	Oaxaca
690	A. potatorum	Valles Centrales	Artisanal	Oaxaca
701	A. potatorum	Sierra Sur	Ancestral	Oaxaca
710	A. salmiana	Centro-SLPz	Artisanal	SLP
760	A. angustifolia	Valles Centrales	Artisanal	Oaxaca
790	A. angustifolia	Sierra Sur	Ancestral	Oaxaca
920	A. angustifolia	Valle de Atlixco y Matamoros	Artisanal	Puebla
239	A. potatorum	Sierra Sur	Ancestral	Oaxaca
245	A. potatorum	Valles Centrales	Artisanal	Oaxaca
373	A. angustifolia	Valles centrales	Artisanal	Oaxaca
688	A. angustifolia	Sierra Sur	Ancestral	Oaxaca
863	A. karwinskii	Three regionsy	Artisanal	Oaxaca
949	A. karwinskii	Sierra Sur	Artisanal	Oaxaca

^z

SLP: San Luis Potosí. ^yValles Centrales, Centro and Sierra Sur.

The ethanol percentage (% v/v), based on NMX-V-013-NORMEX-2019 (Dirección General de Normas, 2019), was determined with a set of certified alcoholometers OIML-ISO4801-NF-B-35-515 (Alla France, Chemillé, France), one with a measuring scale of 30 - 40 (% v/v) and another with a scale of 40 - 50 (% v/v), calibrated at 20 °C, with graduation of 0.10 and accuracy of 0.10.

FT-MIR spectroscopy was performed with a spectrophotometer (Thermo Scientific^TM Nicolet iS5, MA, USA) using a diamond crystal with ATR, the angle of incidence was 45°. Each spectrum was collected in absorbance mode, over the 650.00 - 4000.00 cm^-1 range, with a resolution of 4 cm^-1. The temperature control was at 25 ± 1 °C. The sample (100 µL) was carefully placed on the diamond crystal to avoid air bubbles and immediately covered with the lid to reduce mezcal evaporation (Quintero-Arenas et al., 2020). Sixty-four scans were taken and the average for each mezcal was calculated using the OMNIC™ spectroscopy program (Thermo ScientificTM, USA). Prior to multivariate analysis, the data were pre-processed and treated with spectral transformations. An overview of the treatments applied is shown in Figure 1.

Figure 1

The mezcal samples were subjected to FT-MIR spectroscopy with ATR in the mid-IR region, from 650-4000 cm^-1 (Figure 2), to discriminate mezcal according to agave species. Although Silva et al. (2014) revealed that from 3627-2971 cm^-1 was not useful for the multivariate statistical analysis and Lachenmeier et al. (2005) excluded the regions 1887-1447 and 3696 - 2971 cm^-1. In this study, just removing the 3110 to 3600 cm^-1 region was enough to obtain robust results compared to the analysis of the whole region (analysis not shown).

Figure 2

Datasets obtained from the FT-MIR spectroscopy were exported and transformed using Unscrambler software. The transformations applied (Figure 1) included: baseline offset, Multiplicative Scatter Correction (MSC), unit vector normalization, Savitzky-Golay smoothing with nine symmetry points, and finally the first (order one) and second (order two) derivatives of Savitzky-Golay with polynomial order two and six symmetry points.

In general, the elimination of baseline shifts helps to reveal hidden information and emphasize small spectral variations (Formosa et al., 2020). The MSC normalization was performed to correct each spectrum based on the mean value (average of all spectra) (Windig et al., 2008). The unit vector normalization and Savitzky-Golay smoothing were implemented to improve the signal by removing the noise of each spectrum (Haq et al., 2018; Wu et al., 2015).

Treatments 1 and 2 described in Figures 1 and 3 were evaluated and compared with the raw data using a multivariable approach, consisting of PCA, PLS-DA, and OPLS-DA routines. Figure 3 shows the plots of spectral data transformed by the first and second derivatives. Figure 3A, like Figure 3B, shows three regions, designed Region 1 from 1360 - 1766 cm^-1, Region 2 from 1908 - 2172 cm^-1, and Region 3 from 3361 - 3746 cm^-1, indicating the difference between the data of both treatments 1 and 2. These differences could be reflected in the multivariate analysis.

Figure 3

Figure 4 shows the loadings plots based on PCA of the FT-MIR data matrix of mezcales produced with A. angustifolia, A. potatorum, A. salmiana, and A. karwinskii. Two principal components described 57.6 % of the spectral variation for FT-MIR raw data (Figure not shown); 51.3 % (treatment 1) and 27.9 % (treatment 2) for the pre-treated datasets (figures not shown). Although the spectral variation is greater for the FTMIR raw data, it is suspicious for two main reasons. The first reason is that a sample of mezcal elaborated with A. salmiana is located inside the ellipse of A. angustifolia, and the second reason is related to the advantage of analyzing a pre-treated matrix since the spectral transformation is necessary to minimize or eliminate physical effects. Specifically, a derivation procedure is applied to eliminate additive and multiplicative effects, and in addition, both transformations remove the baseline of the spectra (Anjos et al., 2016).

Figure 4

Additionally, the trends in Figure 4 correspond to the distribution of the loadings, which shows a homogeneity of the values in Figure 4(B and C), in contrast to Figure 4A. The distribution of the loadings around the four quadrants

(I, II, III, and IV) allows us to obtain reliable and robust models (López-Aguilar et al., 2021). A similar pattern is observed in the loadings of the PLS-DA model for species (Figure 5), Figure 5F shows better behavior and homogeneous data of transformed spectra with second derivative than raw spectra (Figure 5D) of the FT-MIR spectroscopy.

Figure 5

However, the spectral variation represented by the scores plots in Figure 5C is only 23.3 % (component 1: 11.7 % and component 2: 11.6 %). Although the variation values are higher for the first derivative [Figure 5A; 42.8 % (component 1: 34.4 % and component 2: 8.4 %)] and raw data [Figure 5B; 37.5 % (component 1: 27.9 % and component 2: 9.6 %)], the scores plot shows a better separation for the second derivative data. The validation allowed us to evaluate the predictive power and accuracy (Ghosh et al., 2020) of the three models (raw spectra and transformed data with first and second derivatives) using leave one out-cross validation (LOO-CV) with 100 permutation routines and the value Q² denotes the ability of the analysis (Westerhuis et al., 2008).

The results (Figure 6A) showed that the PLS-DA analysis derived from the raw spectra indicated low and negative Q² values for the components, which means that the model is not prognostic at all or is overfitted (Szymańska et al., 2012). Furthermore, the observed p-value is not significant (Figure 6D; p = 0.8) by presenting 80 % of incorrect permutations, in contrast to other treatments, where the first and second derivative data had positive Q² values for the five components [Figure 6(B and C) ] and the performance during the permutation tests was significant for the two models [Figure 6 (E and F)]. The permutation trials assume that there is no discrimination between the pair of randomly formed groups (Westerhuis et al., 2008).

Figure 6

Table 2 shows the capacity of the OPLS-DA on the FTMIR spectra according to applied transformation. The three pairwise comparisons from A. potatorum, A. angustifolia and A. salmiana [Figure 7(A, B, and C) ] made with raw data, presented Q² values below 0.2 [Figure 8(A, B, and C)], in addition, the p-values showed no significant differences in the analysis (p ≥ 0.05). Apart from that, of the three comparisons made with data of the first derivative [Figure 7(D, E, and F) ], only the two comparisons between A. karwinskii and A. potatorum [Figure 7E], and A. angustifolia and A. karwinskii [Figure 7F], showed significance (p = 0.02) of the test for Q² [Figure 8(E and F) ].

Table 2

Pretreatment	Statistical	A. angustifolia and A. potatorum	A. karwinskii and A. potatorum	A. angustifolia and A. karwinskii
Raw data	Q2	-0.0858	0.134	0.149
	p-value	0.22	0.11	0.06
First derivative	Q2	0.162	0.55	0.376
	p-value	0.1	0.02	0.02
Second derivative	Q2	0.232	0.654	0.563
	p-value	0.04	< 0.01	0.01

Figure 7

Figure 8

The three pairwise comparisons between A. angustifolia, A. potatorum and A. karwinskii performed with the second derivative data matrix are also shown in Table 2 and Figure 7(G, H, and I). Interestingly, the three comparisons presented Q² values statistically significant [Figure 8(G, H, and I)], for example, the comparison between A. angustifolia and A. potatorum [Figure 7G] had a Q² value [Figure 8G] of 0.232 and a p-value of 0.04 (p < 0.05). Similarly, the discrimination between A. karwinskii and A. potatorum [Figure 7H] showed a Q² value [Figure 8H] of 0.654 and a p-value less than 0.01 (p < 0.05). Finally, the discrimination between A. angustifolia and A. karwinskii [Figure 7I] showed a Q² value [Figure 8I] of 0.563 and a p-value of 0.01.

Importantly, the lowest Q² value was for the discrimination between A. angustifolia and A. potatorum [Figure 8G]. This suggests that of the three pairs of comparisons, it is more difficult to distinguish mezcales between these species, possibly because they are highly commercialized species and the agaves are mixed during the cooking step. Mezcal production includes more than 50 species of agave, of these, 22 are the most used, and A. angustifolia becomes important for presenting high yields (Castañeda-Nava et al., 2019; Sánchez-Gómez et al., 2022). Particularly, in localities from Sierra Sur, Oaxaca, the producers use 40 % of A. angustifolia and 15 % of A. potatorum of the total cultivated agave (Rios-Colín et al., 2022).

The PLS-R model applied to the FT-MIR data matrix of the mezcales analyzed in 2021, was used to predict the ethanol percentage (% v/v) of a set of samples obtained in 2022 (Table 6) and subjected to FT-MIR analysis. The region evaluated during testing (Figure 11A) was the same as that used during training (1500-1700 cm^-1), which was also subjected to a spectral transformation with the Savitzky-Golay first derivative of polynomial order two, and six symmetry points (Figure 11B).

Table 6

Mezcal	Ethanolz	Absorbancey
		1500.35x	1500.83	1502.28	….	1700.91/0.05
239-1	47.0	0.04	0.04	0.04		0.05
239-2	47.1	0.04	0.04	0.04		0.05
239-3	47.4	0.04	0.04	0.04		0.05
245-1	46.2	0.04	0.04	0.04		0.05
245-2	46.6	0.04	0.04	0.04		0.05
245-3	46.6	0.04	0.04	0.04		0.04
373-1	49.3	0.04	0.04	0.04		0.05
373-2	49.2	0.04	0.04	0.04		0.05
373-3	49.4	0.01	0.01	0.01		0.01
688-1	49.4	0.04	0.04	0.04		0.05
688-2	49.4	0.04	0.04	0.04		0.04
688-3	49.4	0.04	0.04	0.03		0.04
863-1	39.9	0.04	0.04	0.04		0.05
863-2	39.6	0.02	0.02	0.02		0.03
863-3	40.0	0.04	0.04	0.04		0.05
949-1	48.9	0.03	0.03	0.03		0.04
949-2	49.1	0.04	0.04	0.04		0.04
949-3	49.3	0.03	0.03	0.03		0.04

^z

Experimental ethanol percentage (% v/v).

Figure 11

The prediction of the ethanol percentage (% v/v) for the new set of samples obtained in 2022 (Table 7) and subjected to FT-MIR analysis in triplicate, based on the PLS-R model performed with the samples evaluated in 2021, allowed us to obtain the ethanol percentage in an approximate way. For example, samples 373-1 and 688-1 showed the highest difference between the reference value and predicted value (8.3 y 5.8 %), but it is also important to highlight that samples 245-1 and 863-3 showed the lowest difference value (0.6 and 0.7 %, respectively). As a result, the RMSEP and R² values were 2.32 and 0.55, respectively, which showed an opposite behavior to that obtained in training, since the RMSP was greater and the R² was lower. Moreover, the RPD was 1.27 which is lower than the value obtained in the validation (2.29); since the RPD indicates the predictability of a model (Kamruzzaman, 2021), then high RPD values are required.

Table 7

Mezcal	Reference Y	Predicted Y	Difference
239-1	47	44.6	2.4
239-2	47.1	44.4	2.7
239-3	47.4	44.4	3.0
245-1	46.2	46.8	0.6
245-2	46.6	44.5	2.1
245-3	46.6	48.1	1.5
373-1	49.3	41.0	8.3
373-2	49.2	46.6	2.6
373-3	49.4	47.3	2.1
688-1	49.4	43.6	5.8
688-2	49.4	46.7	2.7
688-3	49.4	48.4	1.0
863-1	39.9	40.9	1.0
863-2	39.6	38.4	1.2
863-3	40.0	39.3	0.7
949-1	48.9	46.6	2.3
949-2	49.1	45.3	3.8
949-3	49.3	46.1	3.2

Reference Y: indicates the experimental ethanol percentage (reported as % v/v).

Despite the behavior observed in the prediction (three samples presented the highest dissimilarity between the estimated and the reference value), this proposal can be the basis for determining the ethanol percentage (% v/v) by FT-MIR. This technique is a simple, fast, accurate, reliable, and viable alternative compared to traditional methods such as GC or densitometry which are more laborious (Debebe et al., 2017; Quintero-Arenas et al., 2020).

The PLS-DA with transformed data by both first (p = 0.01) and second (p < 0.01) derivatives, allowed the differentiation of mezcales produced from Agave angustifolia, Agave potatorum, Agave salmiana, and Agave karwinskii, while OPLS-DA was more robust when analyzed with second-derivative data. Pairwise comparisons by OPLS-DA, allowed us to discriminate the mezcales between A. angustifolia and A. potatorum (Q² = 0.232, p-value = 0.04; R²Y = 0.969, p-value = 0.01), A. karwinskii and A. potatorum (Q² = 0.654, p-value < 0.01; R²Y = 0.985, p-value < 0.01), and A. angustifolia and A. karwinskii (Q² = 0.563, p-value = 0.01; R²Y = 0.989, p-value = 0.01). The FT-MIR and multivariate analysis allowed the prediction of the ethanol percentage (% v/v) of the mezcales obtained in 2022, based on the PLS-R model previously run on the samples evaluated in 2021. During training, the R² values in the calibration and validation of the PLS-R analysis were 0.98 and 0.81, respectively. The RMSEC and RMSEP coefficients were 0.41 and 1.68, respectively. However, in testing, the behavior was the opposite, since the R² value (0.55) was lower and the RMSEP was higher (2.08). Finally, although we declare that the capacity of the PLS-R model was acceptable, it can be the basis for generating new ones, as more samples will be analyzed.