Glyphosate Detection by Means of a Voltammetric Electronic Tongue and Discrimination of Potential Interferents

2.1. Sample Preparation

Two types of analysis have been carried out: discrimination of samples and prediction of concentration levels. Discrimination studies were carried out using two sets of samples. The first set contained four different samples consisting of: (i) phosphate buffer (0.1 mol·dm⁻³ at pH 6.7); (ii) phosphate buffer and 20 mg·dm⁻³ of a domestic plant fertilizer containing humic acids (manufacturer Massó Garden, Valencia, Spain; composition: 4.5% w/w total N, 1.75% w/w organic N, 2.75% w/w ammoniacal N, 3% w/w K₂O, 22% w/w organic C); (iii) phosphate buffer and 40 mg·dm⁻³ of CaCl₂ (with 96% of purity from Aldrich, Valencia, Spain); and (iv) phosphate buffer, 20 mg·dm⁻³ of humic acids and 40 mg·dm⁻³ of CaCl₂. The second set of samples was prepared similarly, but also contained glyphosate at a concentration of 5 × 10⁻⁴ mol·dm⁻³.

In the first part of the prediction of glyphosate concentration levels, 11 glyphosate samples were used, in a concentration range of between 0 and 5 × 10⁻³ mol·dm⁻³ (2.5 × 10⁻⁵, 5 × 10⁻⁵, 6.25 × 10⁻⁵, 1.25 × 10⁻⁴, 2.5 × 10⁻⁴, 5 × 10⁻⁴, 1.25 × 10⁻³, 2.5 × 10⁻³, 5 × 10⁻³). These solutions were buffered with phosphate at a concentration of 0.1 mol·dm⁻³ and at pH 6.7. Each sample was measured three times. PLS analyses were performed on the 33 measured samples (11 samples × three repetitions). The analysis was also made for each measuring strategy (rotating or static electrodes). A set of 21 samples (three trials of seven samples) was used as the training set, whereas the remaining 12 samples were used to validate the model. Calibration and validation samples were selected randomly. Each sample was measured five times and the mean was applied in a pre-processing step.

In a second part, studies of the potential interferents were done and the prediction of the glyphosate concentration levels was carried out in the presence of humic acids and calcium. Samples preparation was performed using the MODDE 8.0 experimental design program (Umetrics, Umea, Sweden). With this program, a system of three compounds (glyphosate, humic acids and Ca²⁺) was established at three concentration levels. The final set included 19 samples prepared by addition of three selected compounds (glyphosate, humic acids and calcium chloride) into water (phosphate buffer 0.1 mol·dm⁻³, pH 6.7). The lower level was 1 × 10⁻⁴ mol·dm⁻³ for glyphosate, and 0 and 10 ppm for humic acids and calcium, respectively. The medium level contained 5 × 10⁻⁴ mol·dm⁻³ for glyphosate, 10 ppm of humic acids and 25 ppm of calcium. The higher level consisted in 1 × 10⁻³ mol·dm⁻³ for glyphosate, and in 20 and 40 ppm for humic acids and calcium, respectively (see Table 1).

For this study, the 19 solutions were measured three times, thus obtaining a total of 57 samples, 39 (three trials of 13 samples) of which were used for training 18 samples for the validation experiments.

2.2. Electrochemical Studies

Electrochemical experiments were performed using an Autolab PGSTAT100 instrument (Echo Chemie, Utrecht, The Netherlands). Electrochemical characterization was carried out using aqueous solutions of 2.5 mmol·dm⁻³ of glyphosate. The solutions at pH 6.7 were buffered with phosphate buffer at a concentration of 0.1 mol dm⁻³. All the measurements were taken at room temperature (25 ± 2 °C) using Pt, Cu and Co as the working electrodes. A platinum electrode and a saturated calomel electrode were used as the counter and reference electrode, respectively. The rotating disk electrodes (RDE) studies were performed at a scan rate of 10 mVs⁻¹ and at a rotation velocity of 1,500 rpm.

2.3. The Electronic Tongue

3.1. RDE Voltammetry Studies

Prior to the electronic tongue experiments, an electrochemical study was carried out to determine the electrochemical behaviour of glyphosate using the metallic electrodes Co, Cu and Pt. The study was conducted in water at room temperature and at pH 6.7 using 0.1 mol·dm⁻³ of phosphate as the buffer. As is well-known, the electrochemical response of a given compound depends on the intrinsic chemical nature of both the electrode and the redox behaviour of the product itself. In addition, each electrode’s transient response depends on the diffusion coefficients of the oxidized or reduced species, and also on the possible presence of specific chemical or electrochemical reactions between the electrode and the redox-active species. In fact, it has been reported that glyphosate forms complexes with Cu(II) and Co(II) metal cations. Indeed, as Figure 1 shows, the chemical structure of glyphosate contains three potentially coordinating groups: amine, carboxylate and phosphonate moieties; therefore, the herbicide is able to form chelating structures with different metal ions and surfaces [46,47].

In the electrochemical experiments, the rotating disk voltammogram of the copper electrode showed an irreversible oxidation process at −80 mV corresponding to Cu²⁺ formation (see Figure 2). Two peaks appeared when potential was reversed; the first one corresponded to the reduction of Cu²⁺ to Cu⁺, and the second to reduction to Cu⁰. In the presence of glyphosate, the corrosion current of copper increased significantly, probably due to the Cu²⁺ complexation by glyphosate forming soluble species to result in the electrochemical dissolution of copper. The rotating disk voltammogram of the cobalt electrode in the absence of glyphosate showed how oxidation of the metal started at −500 mV. In this case however, a complex RDE curve was obtained at higher potentials, suggesting intricate electrochemical behaviour with the formation of several species at the electrode, most likely involving different hydroxide and oxide cobalt species. In the presence of glyphosate, the corrosion current increased, which was in agreement with the Co²⁺-glyphosate complexes formation [46]. When using a rotating disk of platinum electrode, glyphosate did not display a well-defined redox process in water (electrochemical window in the −0.8 to +1.0 V range). Nevertheless, this electrode was used in the electronic tongue to obtain information about non-Faradic processes.

The wave form for the electronic tongue (see Figure 3(a)) was designed by taking into account some of the redox processes observed in the electrochemical studies. Pulses 2 and 6 corresponded to the irreversible reduction peak of Cu⁺ to Cu⁰. Peaks at 600 mV (pulse 4) and at −500 mV (pulse 5) corresponded to the processes Cu⁰ to Cu⁺² and the corresponding reduction Cu²⁺ to Cu⁰, respectively. At 200 mV, a cross appears in the baseline in the cobalt RDE (pulse 8). Moreover, pulses 1, 3, 7 and 9 were set to 0 mV, which corresponded to the maximum difference between the baseline and the glyphosate response in the Cu voltammogram. Figure 3(b) shows a typical response of both electrodes (Co and Cu) when applying the pulse sequence after subtracting the baseline. The figure shows two different responses corresponding to studies using static electrodes and rotating electrodes (1,500 rpm). The difference in the current signal between the rotating and static conditions is about 300 μA, indicating a greater response in the presence of glyphosate when rotating conditions are used.

3.2. PCA Analysis

The response of the voltammetric electronic tongue was studied in water (phosphate buffer, 0.1 mol·dm⁻³, pH 6.7) using different samples. Studies were also conducted in the presence of potential interferents, such as humic acids and Ca²⁺. Two sets of samples were prepared; the first consisted in four samples, i.e., (i) phosphate buffer (0.1 mol·dm⁻³ at pH 6.7); (ii) phosphate buffer and 20 mg·dm⁻³ of humic acids; (iii) phosphate buffer and 40 mg·dm⁻³ of CaCl₂; and (iv) phosphate buffer, 20 mg·dm⁻³ of humic acids and 40 mg·dm⁻³ of CaCl₂; the second set of samples comprising the same solutions, but they also contained glyphosate at a concentration of 5 × 10⁻⁴ mol·dm⁻³. Each sample was measured three times. Moreover, two different electrochemical measurements were taken with the static and rotating (1,500 rpm) electrodes. In order to show the differential response of the tongue in the presence or absence of glyphosate, the electrochemical data was combined and studied using PCA analysis.

The PCA score plot obtained for the rotating electrodes for the complete set of samples is provided in Figure 4. The PC1 contained only 52.87% and 59.24% of variance, and the first two components represented 79.12 and 79.23% of the total variance of the data when using the static (not shown) and rotating electrodes, respectively, whereas four PCs were needed to account for 95% of variance. Nevertheless, two dimensions (PC1 and PC2) in the PCA plots were enough to differentiate samples. In fact, the electronic tongue was able to discriminate the presence of glyphosate and those samples containing this herbicide appeared well separated in both assays using either the rotating (see Figure 4) or the static (not shown) electrodes. This is a promising result as it suggests that the differential response in the voltammetric technique using a set of electrodes might be a suitable method to detect glyphosate in aqueous samples.

3.3. PLS Analysis

The PCA study has shown that the data from the electronic tongue clustered according to glyphosate presence help obtain a good classification model, even in the presence of potential interferents such as humic acids and Ca²⁺. In this section, we were also interested in analyzing whether the data taken from the electronic tongue could be used to predict glyphosate concentration levels. In order to achieve this goal, the Partial Least Square (PLS) regression technique was used. In a pre-processing step, the baseline signal was subtracted from each sample signal. For this study, there were five latent variables, which were determined by cross-validation studies using one third of the whole data set used to validate the model. Eleven solutions of glyphosate within a concentration range of 0 and 5 × 10⁻³ mol·dm⁻³ were used. These solutions were buffered with phosphate buffer (pH 6.7). Each solution was measured three times. PLS analyses were performed with 33 samples (11 solutions × 3 repetitions) for each set of measurements using the rotating or static electrodes. In both cases, 21 samples (three trials of seven samples) were used as the training set, whereas the remaining 12 samples were used to validate calibration. The PLS model consists of a set of regression coefficients (βi) and residual value (E). This allows to obtain the value of glyphosate concentration from the measurement data. If we consider that x₁, x₂, …, x_n are the values corresponding to the temporal evolution of the current signal at the working electrodes for a certain sample then the value of glyphosate concentration for this sample can be calculated using Equation (1):

Figure 3 shows the results of applying the PLS application to the calibration and validation sets when using the rotating or static electrodes. Additionally, Table 2 displays the adjusting parameters (r², p1, p2 and RMSEP) from the PLS prediction model applied to the validation set for each created model.

Figure 5 allows a simple visual inspection of the accuracy and precision created in the prediction model applied to the validation set of glyphosate. Additionally, a more rigorous analysis was achieved by linearly fitting experimental points. We used a simple linear model (i.e., y = p1x + p2) to adjust the predicted versus real concentration, and we obtained the parameters p1 (slope of the fitting line), p2 (intercept with the y axis) and RMSEP (root mean square of prediction). Parameters p1 and p2 related with accuracy in prediction, whereas RMSEP deals with the precision of the model. Thus, the PLS model is much better as the p1 value approached 1, while the RMSEP values came close to 0. Figure 5(A,B) depicts that the correlation obtained for the PLS model was better when using the rotating working electrodes (see also Table 2).

This observation could be confirmed if we compared the r² values. For the models created with the rotating electrodes, r² was 0.876, whereas r² was 0.809 for the models created with the static electrodes. These results are in agreement with our first observation (see above) where we considered that an increase in the current signal when using the rotating electrodes could improve the accuracy and precision of the created system.

Encouraged by these results obtained while predicting the concentration levels of glyphosate in phosphate buffer, further experiments were carried out using a new set of samples containing humic acids and Ca⁺² for the purpose of evaluating a possible matrix effect and for detecting glyphosate in the presence of these potential interferents. By bearing this aim in mind, samples selection was performed using the MODDE 8.0 experimental design programme (Umetrics). With this programme, a system of three compounds was established at three levels. The final set included 19 samples, prepared by the addition of different glyphosate, humic acids and calcium chloride concentrations in water with phosphate buffer at 0.1 mol·dm⁻³, pH 6.7 (see the Experimental Section for details). It has been shown above that the models created under rotating conditions provide better results. For this reason, only the studies using the rotating Pt, Co and Cu electrodes were carried out in this part of our work. Nineteen solutions were prepared and each solution was measured three times. From these solutions, 38 samples were used as the training set, whereas the remaining 19 samples were used to validate calibration. In the cross-validation studies, four latent variables were used. In this case, the glyphosate concentration was calculated as: $C=E+{\sum }_{i=1}^{621}{\beta }_i\cdot x_i$.

Table 3 shows the adjusting parameters from the PLS models created, whereas Figure 6 illustrates the PLS model obtained for glyphosate when using the electronic tongue. As seen in Table 3, adjusting parameters r² and p1 came closer to 1 for glyphosate, demonstrating a good correlation between the electrochemical response of the electrodes and the different amounts of glyphosate in the water samples. Moreover, the attempts made to predict the concentrations of humic acids or Ca²⁺ were unsuccessful and poor r² and values p1 were obtained in both cases (see Table 2). This absence of interference in the presence of these species may relate with a larger complexation constant between glyphosate and copper than with the complexation constants between humic acids and copper, and between glyphosate and calcium cations. These results indicate that the electronic tongue is quite capable of detecting this analyte, even in the presence of different concentrations of the potential interferents, such as humic acids and Ca²⁺.