An assessment of voltammetry on disposable screen printed electrodes to predict wine

17 The present work aimed at determining the applicability of linear sweep voltammetry coupled to 18 disposable carbon paste electrodes to predict chemical composition and wine oxygen 19 consumption rates (OCR) by PLS-modeling of the voltammetric signal. Voltammetric signals 20 were acquired in a set of 16 red commercial wines. Samples were extensively characterized 21 including SO 2 , antioxidant indexes, metals and polyphenols measured by HPLC. Wine OCRs 22 were calculated by measuring oxygen consumption under controlled oxidation conditions. 23 Chemical variables and wine OCRs were predicted from first order difference voltammogram 24 curves by PLS-regression. 25 A significant number of fully validated models predicting chemical variables from voltammetric 26 signals were obtained. This fast, cheap and easy-to-use approach presents an important potential 27 to be used in wineries for rapid wine chemical characterization. 28

Even if the combination of voltammetric signals with multivariate statistical tools has been little explored, principal component analysis (Gonzalez, Vidal, & Ugliano, 2018;Ugliano, 2016) and partial least square regression modeling (Martins, et al., 2008) have been suggested to be interesting approaches to provide valuable information when monitoring wine oxidation effects or providing wine fingerprinting.
In this context, it was hypothesized that relationships between voltammogram regions and specific phenolic compounds as well as overall wine oxygen consumption rates (OCR) could be established by multivariate analysis following an untargeted voltammetric approach.Thus, the present work aimed at evaluating the applicability of linear sweep voltammetry coupled to disposable carbon paste sensors to predict chemical composition and wine oxygen consumption rates (OCR) by PLS-modeling in a set of commercial red wines.

Wine samples
A set of 16 red Spanish wines were studied.They were all purchased at a local store and were from different regions, grape varieties and vintages (detailed information is provided in Table S1 of Supporting Information).

Oxidation experiment
Oxygen consumption rates of wines were determined from data collected in an oxidation experiment consisting of five consecutive air-saturation cycles as described in Ferreira, Carrascon, Bueno, Ugliano, and Fernandez-Zurbano (2015).Air saturations were carried out by gentle shaking 500 mL of wine contained in a closed 1-liter glass bottle, then the cap was opened to allow fresh air to enter the bottle.This procedure was repeated for each saturation until a final concentration of 5.6±0.1 mg L -1 of dissolved oxygen was reached.Then, wine samples were incubated in the dark (25±0.5 ºC) and dissolved oxygen was monitored at least once a day with a non-destructive Nomasense oxygen analyzer (Nomacorc S.A., Thimister-Clermont, Belgium) until 90% of oxygen was consumed or during 7 days.This cycle was repeated five times.

Voltammetric measurements
Electrochemical measurements were performed with a commercial Nomasense Polyscan electrochemical analyzer (Nomacorc, Belgium) using disposable screen printed sensors.The system consisted in three sensors: working and counter electrodes both screen printed carbon paste electrodes and reference electrode consisting of an Ag/AgCl electrode.A drop of sample was loaded onto the sensor, and linear sweep voltammograms were acquired between 0 and 1200 mV at a scan rate of 100 mV s -1 .A total of 122 voltammetric signals for each wine in duplicate were recorded, and further worked with averaged data.A new sensor was used for each measurement.Repeatability of the measurement was tested by three consecutive measurements of the same wine.

Chemical characterization
Metals.Fe, Cu, Mn, Zn and Al were quantified by inductively coupled plasma optical emission spectroscopy (ICP-OES) with previous microwave-assisted digestion of samples as described by Gonzálvez, Armenta, and De La Guardia (2008).
Absorbance measurements.Absorbance at 420, 520 and 620 nm of undiluted wine was measured using glass cuvettes with optical paths of 1, 2, 5 or 10 mm.Measurement which provided absorbance readings between 0.3 and 0.7 were considered as recommended by the OIV (2009a).

Conventional oenological parameters. pH was determined by Infrared Spectrometry with Fourier
Transformation (IRFT) with a WineScanTM FT 120 (FOSS), which was calibrated with wine samples analyzed in accordance with official OIV (International Organization of Vine and Wine) practices; free and total sulfur dioxide were determined by the aspiration/titration method (Rankine method) recommended by the OIV.
Measured Redox potential.This parameter, which is not a truly redox potential as recently discussed (Danilewicz, Tunbridge, & Kilmartin, 2019), was measured using a Pt electrode fitted to a Ag/AgCl reference electrode model 50 58 from Crison (Alella, Barcelona) and a microprocessor 6230N from Jenco Instruments (San Diego, CA).Measurements were recorded in a glove chamber (Jacomex, France) with a level below 0.002% (v/v) of oxygen in gas phase.
Therefore, wine was firstly poured in a 4 mL vial where the electrode was introduced (with no agitation) and measurement was recorded after 35 min.Then, the electrode was cleaned with milliQ water and introduced in a solution containing equimolar amounts (0.01 M) of ferro-and ferricyanide supplied by Panreac (Barcelona, Spain).This solution has a known redox potential of 220 ±10 mV a 25°C (vs.Ag/AgCl(s)).If the measured redox potential was in this range, the electrode was rinsed again with water and was then ready for subsequent measurements.In case the measured redox potential differed more than 10 mV from the expected 220 mV value, the diaphragm of the electrode was cleaned with a solution of thiourea (<6%) and HCl (<2%) (Crison, Alella, Barcelona).All analyses were performed in duplicate.
Chemical data (average, maximum and minimum) are presented in Table S2 of Supporting Information.

Determination of wine oxygen consumption rates
The oxygen consumed in the five saturation cycles was calculated for each wine (as the average among three independent saturation cycles per sample) as the difference between the dissolved oxygen at the beginning and at the end of each cycle.Then, the oxygen consumed for each saturation was plotted against the days employed to consume the oxygen.The five points (accumulated O2 consumed at the end of each saturation, time in which saturation ended) followed a straight line which was adjusted by least square regression.The ordinate at time 1 day was taken as the initial oxygen consumption rate.The slope was taken as the average oxygen consumption rate (Ferreira, Carrascon, Bueno, Ugliano, & Fernandez-Zurbano, 2015).Data are available in Table S3 of Supporting Information.

Exploration of raw voltammetric signals
First derivative voltammograms allow to improve the separation between anodic waves in comparison with raw voltammograms (Gonzalez, Vidal, & Ugliano, 2018).Thus, first order difference voltammograms curves were calculated for all wines.Further Principal Component Analysis (PCA) was calculated in order to analyze the dominating types of variability for these curves and, if possible, to reduce the initial number of variables.

Modeling OCRs and chemical variables from voltammetric signals
The main purpose was the prediction by regressing calibration of the chemical variables from the voltammograms.The general model is given by where, for a sample size  ( = 16),  (16,121) represents the input matrix with the differences between two consecutive voltammetric measurements,  (16,97) the output matrix with the chemical variables,  (121,97) is the matrix of regression coefficients and  (16,97) the matrix of residuals.
Single response models are analyzed.Then, single  -variable Partial Least Square regression method is used for every chemical variable and the whole spectrum of voltammograms ().
Therefore, the prediction by regressing for one single  data on  was as follows: where,   (16,1) are the vectors that represent every one of the chemical variables 1 ≤  ≤ 97 for the red wine sample set and,   (121,1) and   (16,1) are respectively, the vectors of regression coefficients and residuals.
Firstly, the input variables  are enhanced in two ways, they have been filtered applying a 7 points window Stavizki-Golay smoothing; and, on the other hand they have been standardized to comparable noise levels.Likewise, chemical variables  ;1≤≤92 have been standardized.
With this considerations, a first PLS model was computed.Taking the ratio between sample size and number of variables into account, variable selection has not been considered, in order to avoid the problem of overfitting.Therefore, for every single chemical variable, the whole spectrum on the X has been considerate in one PLS model.The model was validated using full cross validation.
Then, those models with validated explained variance greater than 25% and presenting root mean squared error (RMSE) between the 9% and the 12% of the range were considered.Considering the size of the sample, and the number of factors that explain the main information of the  −variables, only models with less than or equal to four PLSs, have been considered.
All the analyses have been carried out with Unscrambler X 10.5.1, Matlab R2018a, R 4.0 and XLStat v2018.

Voltammogram profiles
Figure 1 shows the first derivative voltammograms for the sample set.Two characteristic anodic waves with two maximal points and a minimal can be observed.The first maximal point and the minimal are around 420 mV and 600 mV, respectively.Differently, the second maximal point is around 730 mV.The derivative curve displays maximum values in the first maximal point (around 420 mV) with a derivative current reaching values of 220 nA/mV.This can be explained because red wines contain high levels of components that are rapidly involved in oxidative reactions such as anthocyanins, ortho-diphenols and triphenols of gallic acids (Table S2), which usually occur at low potential (Kilmartin, Zou, & Waterhouse, 2002) and thus can be associated with this first anodic wave.The derivative current of the second anodic wave, which corresponds to less readily oxidizable compounds (Ugliano, 2016), has been associated with vanillic or coumaric acids, the meta-diphenols on the A ring of flavonoids such as catechin, SO2, certain amino acids and brown pigments related to oxidation reactions (Kilmartin, Zou, & Waterhouse, 2002;Makhotkina & Kilmartin, 2013).
In order to shed light on specific linkages between compounds and voltammetric signals, PLSmodels have been built and discussed.

Predicting OCR from voltammetric signals
PCA was calculated with the derivative voltammetric signals.The first three PCs retain 91% (82% in validation) of original variance.This result shows that voltammetric information can be retained by three independent and non-correlated variables.Remarkably, even big efforts were invested in building PLS-models predicting chemical variables and OCRs from these three PCs, validated models could not be obtained, which could have simplified the prediction task.A possible explanation is that because we have no guarantee that the selected principal components are associated with the outcome.In fact, it is a possible drawback of PCR method (PCA + regression), where the selection of the principal components to incorporate in the model is not supervised by the outcome variable.
As detailed in the material and methods section and in a previous reference (Ferreira, Carrascon, Bueno, Ugliano, & Fernandez-Zurbano, 2015), two different OCRs were defined for red wines: the initial OCR, that corresponds to the rate of oxygen consumption during the first 24 h, and the average OCR, that refers to the average rate of consumption for the rest of the experiment.Initial OCRs are significantly faster and far more variable (0.54 -8.22 mg O2/L/day) than the average rates (0.365 -0.792 mg O2/L/day).Interestingly, potentials in the first anodic wave, specifically in the 355-475 mV range (marked in green in Figure 2), present a significant negative correlation with the initial OCR (r < -0.54; P < 0.05 in all cases), while for the average OCR no significant correlation with potentials (i.e., X variables) could be established.This is a surprising result, because we had expected that higher potential signals would be related to higher contents of readily oxidizable substrates and thus to higher oxygen consumption rates.However, this result is completely equivalent to that obtained in a previous paper, in which chemical compositional parameters were just poorly positively correlated or not correlated at all with initial and average OCRs, respectively; while significant negative correlations with some chemicals were observed (Ferreira, Carrascon, Bueno, Ugliano, & Fernandez-Zurbano, 2015).In a further attempt to investigate the relationship between OCRs (initial and average) and voltammetric signals (first derivative), PLS models were calculated.Unfortunately, modeling failed to capture validated models for initial and average OCRs, thus we could not validate one of our initial hypothesis.
Conversely, if a previous step consisting in the prediction of initial OCR from voltammetric potentials, but not considering the second voltammetric wave (600-1000 mV), which corresponds to less readily oxidizable compounds (Ugliano, 2016), a validated model explaining 62% of original variance for initial OCR was obtained.The model included 8 voltammetric signals with half of them displaying positive (at 20, 100, 1050 and 1130 mV: marked in orange in Figure 2) and the other half negative (300, 440, 520 and 1140 mV: marked in blue in Figure 2) relationships with initial OCR (Figure S4 of Supporting Information).Not surprisingly, the highest positive contributions to initial OCRs correspond to voltammetric signals measured at very low potentials (10 and 100 mV).It is not clear to which species can correspond signals at 10 mV, although results derived from white wines (unpublished data) suggest that it may be copper, but this result should be further validated in future research.On its side, the signal at 100 mV could be related to the beginning of the anodic curve for ascorbic acid (Kilmartin, Zou, & Waterhouse, 2002;Makhotkina & Kilmartin, 2013).It has to be highlighted that the modeling of initial OCR from voltammetric signals omitting voltammetric signals belonging to the second anodic wave (not based on variables selection in PLS) has to be considered with caution.Given the low number of samples and high number of predicting variables, overfitting can be occurring, thus this model only establishes preliminary relationships between voltammetric signals and initial OCR.This hypothesis should be confirmed in further investigations.

Predicting chemical compositional variables from voltammetric signals
Table 1 shows the chemical variables that could be satisfactorily modeled from voltammetric signals (RMSE between the 9% and the 12% of the range) (29 out of 95).Validated models explain between 23% and 74% (average = 47%) of original variance by full-cross validation, which correspond to moderate-high correlation coefficients ranging from 0.5 to 0.9 (average = 0.7).Explained variances by calibration reach values in the range of 48-99% and corresponding to correlation coefficients between 0.7 and 0.9 (average = 0.9). Figure 3 shows the voltammetric signals (in nA of anodic current per increment of mV in the working electrode) included in models and the sign and magnitude of their coefficients following a color code.Figure 4 shows some examples of line plots representing the X-loadings corresponding to the first two PLSs (for the plots of the rest of models see Figure S5 of Supporting information).These representations are useful in the interpretation and for confirming the validity of the predictive models.These plots represent the variables (potentials of the voltammograms) that are important for predicting the variables studied such as the concentration of the compounds.
In the case of flavonols, leaving aside quercetin, myricetin-3-galactoside and myricetin, relevant derivatives from the quantitative point of view were modelled.In the case of flavanols and anthocyanins, all the most relevant quantitatively were satisfactorily modelled.By contrast, the ability to model cinnamic, hydroxicinnamic acids and their derivatives was very poor, and only two out of 24 components could be satisfactory modelled.Most remarkably, models for predicting compositional data for metals and for absorbance at 620 nm could not be derived from the voltammetric signals.
It is interesting to note that models (Figure 3, Figure 4 and Figure S5 of Supporting Information) for flavonols, gallic acid ethyl ester, flavanols, and monomeric anthocyanins, including the overall measure of bleachable anthocyanins (MP), present positive coefficients for potentials belonging to the first anodic wave of voltammograms (mainly 140-600 mV), which is supported by the fact that these compounds are most readily oxidizable molecules of wines and thus involved in most rapid oxidative reactions (Ugliano, 2016).Differently, non-bleachable anthocyanins, named polymeric pigments (both small and large PP), can be predicted mainly from higher potentials, belonging mainly to the second wave of the first derivative of voltammograms (840-1160 mV).
Among flavanols, epigallocatechin and gallocatechin show positive coefficients for lower potentials (180-250 mV) than the rest of flavanols measured (catechin, epicatechin, procyanidins B1 and B2) (270-520 mV).This is well in accordance with previous reported results, that show that gallocatechins oxidize at the surface of carbon electrodes earlier than other readily oxidizable compounds, such as monomers and dimmers of (epi)catechin (Kilmartin, 2016).Remarkably is that non-acylated antocyanins present similar models positively contributed by positive voltametric signals at low (160-240 mV) and high (680-800 mV) potentials, while the models for coumaroly anthocyanins, mainly those with higher prediction ability (delphinidin and peonidin-3-O-(6-O-p-coumaroylglucosides)), show positive coefficients mainly in the first anodic wave (180-480 mV), and thus they are more readily oxidizable.
In summary, our results suggest that the voltammetric signal in disposable carbon paste electrodes is mainly the result of wine major flavonols, flavanols, anthocyanins, polymeric pigments, pH and free SO2, being poorly contributed by phenolic acids, metal cations or sulphite adducts.
Conversely, it can be also suggested that voltammetric information is highly multidimensional and therefore can be satisfactorily used to predict many relevant chemical compositional data.

Conclusions
The voltammetric signals recorded from wines with disposable carbon paste electrodes are extraordinarily rich in compositional information from a relatively wide range of chemical species and parameters, which are suggested to be satisfactorily extracted using PLS.The best performance in modelling terms was in all cases obtained from the 1 st derivative of the voltammograms.The voltammetric signals seem to be mainly influenced by major flavonols, flavanols, anthocyanins, polymeric pigments and free SO2, all of which could be satisfactorily modelled.Although oxygen consumption rates (OCR) could not be satisfactorily modelled, positive correlations with voltammetric signals and satisfactory models obtained after selection of variables for initial OCR (based on prior knowledge and not on PLS variable selection), allow to draw the hypothesis that OCRs have a potential of being satisfactorily predicted and thus voltammetry could be also a suitable rapid tool for predicting OCR.
The results presented in this work suggest that disposable carbon paste sensors measuring voltammetric signals and coupled to PLS-modeling have an important potential to be used in wineries for rapid, cheap and easy-to-use approach for wine chemical characterization and oxidation-related control.It is important to emphasize that the number of samples is quite low and also that only the best models are selected for presentation in Table 1.Therefore, the present work is a feasibility study and models must be validated on new data to confirm the results.

Figure captions Figure 1 .
Figure captions

Figure 2 .Figure 3 .
Figure 2. First derivative voltammograms of wines with highest and lowest oxygen consumption

Figure 4 .
Figure 4.The X-loadings for the two first PLS components based on the PLS model for a) quercetin-3-O-galactoside, b) quercetin-3-glucoronide, c) catechin, d) epigallocatechin, e) malvidin-3-O-glucoside, and petunidin-3-O-glucoside.The red line represents the first PLS and the blue line the second PLS Figure 3 a)

Figure S4 .
Figure S4.Map with coefficients of variables included in validated PLS-model predicting initial oxygen consumption rate (initial OCR) from voltammetric signals.

Table 1 .
Variables successfully modeled in the set of red wines (n=16) from voltammetric signals by PLS regression, % of explained variance by full cross validation (and the % of explained variance), the number of PLSs included in each model and the root mean squared error of prediction.

Table S1 .
Information of wines employed in the study.

Table S2 .
Chemical characterization of the 16 red wines studied (data expressed as micrograms per liter, otherwise it is specified).Compounds marked in red were satisfactorily modelled from voltammograms.

Table S3 .
Initial and average oxygen consumption rates for red wines (OCR) expressed as mg O2/L/day (average of three independent replicates)