Development of a Direct Spectrophotometric and Chemometric Method for Determining Food Dye Concentrations

Introduction

Food coloring dyes are widely used in food products, pharmaceuticals, cosmetics, and medical devices. ¹ However, despite their wide array of uses, studies have shown that food coloring dyes have adverse effects. Tartrazine and carmoisine azo dyes, for example, are known even at low doses to adversely affect and alter biochemical markers in vital organs (e.g., liver and kidney) in young male rats. ² Another study by Basak et al. suggested that maternal exposure to artificial food colors and additive plays a role in the mucosal defense system and possibly in carcinogenesis of rats. ³ Gao et al. studied the effect of food azo tartrazine dye on learning and memory functions in mice and rats, concluding that dose levels of tartrazine may adversely affect brain tissue by inhibiting antioxidant defense systems, allowing the accumulation of reactive oxygen species and lipid peroxidation products. ⁴

Dyes are complex organic substances and have been derived from coal tar, petroleum, and other natural substances. Their affordability, stability, and vibrancy compared to natural colors have engendered significant commercial interest. The U.S. Food and Drug Administration, however, requires that many dyes, unlike other food additives, be batch certified prior to commercial use. While studies have been conducted to determine the carcinogenic or other harmful effects of long-term casual exposure to dyes, significant biases and limitations may have affected experimental results. For example, most of these studies were performed or funded by dye manufacturers, suggesting the possibility of bias. Second, many of these studies were conducted on a short-term basis. Lastly, studies have sought to evaluate the safety of individual dyes without considering that food samples often contain a mixture of dyes. ⁵

Commonly used methods for the direct quantitative determination of artificial dyes in food samples include liquid chromatography–mass spectrometry, green liquid chromatography, high performance liquid chromatography, and tandem mass spectrometry.^6,7 While quantitative synthetic dye determination methods are used by other scientists, newly developed methods for qualitative determination are also available, such as comprehensive two-dimensional (2D) liquid chromatography, which combines ion-exchange chromatography and fast ion-pair reversed–phased chromatography. ⁸ Despite the advantages of these methods for directly evaluating colored dyes in food samples, these techniques are considered laborious, expensive, and time-consuming. Further, while recent studies have reported multicomponent analysis of food dyes using spectrophotometry, the majority of results were focused on binary mixture analysis and did not implement the partial least squares (PLS) approach nor structured experimental designs.^9–11

The current study is aimed at developing a novel alternative method for directly quantitating dye mixtures in food samples using a design of experiments, ultraviolet visible (UV-Vis) spectroscopy, and chemometrics. We focused our analysis on a three-component mixture consisting of Erythrosine B (red), Brilliant Blue, and tartrazine (yellow) dyes.

Results of our study show the potential of using chemometrics and UV-Vis spectroscopy to simultaneously quantitate the aforementioned food dyes in synthetic mixtures. Further, the spectra derived from actual food samples were comparable to those of their synthetic counterparts, illustrating the capability of this method to be applied to food samples containing the aforementioned dye colors.

Experimental

Procedure

Mixture designs were prepared using three components, erythrosine B (red), Brilliant Blue (blue), and tartrazine (yellow), in aqueous solutions. Distilled water was used as a blank for the analyses of prepared mixtures. A full factorial design consisting of 64 mixtures was used as a training set (Supplemental Material Table S1). A Box–Behnken design consisting of 13 mixtures was used as a test set (Table S2). The test set was an independent set used to determine the accuracy of the calibration model. All mixture designs were prepared in order to limit the absorbance spectra below 1.0 absorbance units and were generated using the “DoE.base” and “rsm” packages under the R Program^14,15 (Figures 1 and 2). A general overview of the analysis is provided in Figure 3. Briefly, our datasets consisted of training and independent test sets. The training set was used to develop a calibration model via the PLS chemometric technique. We then used the developed model to predict the unknown concentrations of these dyes in both the test set and food (candy) samples. The candy samples were prepared by dissolving these in 100 mL of methanol. Approximately 2 g of each sample was dissolved in methanol to obtain an absorbance spectrum of less than 1.0 absorbance units. The candy samples were left in methanol until the outer coating of the candy’s dye was completely dissolved, which took approximately one week to be completed. The molar concentrations of the red, blue, and yellow dyes in each candy sample were predicted using our previously generated calibration model. Once the unknown concentrations of the dyes in our candy samples were predicted using our aforementioned developed model, the predicted concentrations were replicated by preparing ternary mixtures of these dye standards. The overlap and similarity between the replicated and the actual sample spectra were examined by visual inspection.

OPEN IN VIEWER

Figure 1.

Absorbance spectra for the full factorial design of experiments (n = 64) for the training set.

OPEN IN VIEWER

Figure 2.

Absorbance spectra for the Box–Behnken design of experiments (n = 13) for test set.

OPEN IN VIEWER

Figure 3.

Overview of the chemometric analysis.

Methanol was used as a blank for the analysis of food samples. All spectra generated in this experiment were acquired using a UV-Vis HP8452A spectrophotometer in the wavelength range of 350–674 nm at every 2 nm wavelength scan.

The initial data were preprocessed by autoscaling using the PLS package under the Program R. ¹⁶ The PLS algorithm was utilized for the development of our calibration models. Partial least squares is a powerful multivariate statistical technique that has been successfully applied in many areas. Details of the technique can be referred to in Otto. ¹⁷ Briefly, it involves the decomposition of A (absorbance matrix) and C (concentration matrix) as follows:

A = TP T + E

(1)

C = UQ T + F

(2)

where T and U are the n × d scores matrices containing orthogonal rows; P comprises the p × d loadings of the A matrix; E is the n × p error (residual) of A matrix; Q is the m × d loadings of the C matrix; and F is the n × m error (residual) of the C matrix. Computation of the B-coefficients is then given by the general model:

B = W ( P T W ) - 1 Q T

(3)

with W as d × p matrix of PLS weights.

To assess calibration model performance, the model from the training set was applied to the test set, and the root mean square error (RMSE) was calculated. The RMSE quantifies the extent to which predicted concentrations vary from the actual concentrations in the test set. It is therefore useful as a means of assessing the prediction accuracy of the regression model: ¹⁸

RMSE = ∑ i = 1 n ( y i - y i ' ) 2 N

(4)

where y and y′ are the predicted and actual concentrations, respectively, and N is the number of samples. The performance of our calibration model was further assessed using the relative error of prediction (REP). This can be estimated according to:

REP = ( 100 × RMSE ) / C ¯

(5)

where C ¯ is the mean calibration concentration. All chemometric analyses were performed using the PLS Package under the R Program.^16,19

Results and Discussion

In order to optimize the number of components to be used for the PLS model, the root mean squared error of cross-validation (RMSECV) as a function of the number of components using the training set (64 leave-one-out [LOO] segments) was determined (Figure 4). There were two CV estimates: CV is the ordinary CV estimate and adjCV is a bias-corrected CV estimate. For LOO CV, there is virtually no difference. As evident in Figure 4, approximately three components (for each colored dye) can sufficiently be used in building the PLS model. We also took into account any possible noise in the analysis. Thus, four factors were utilized for calculating the concentrations of each dye in synthetic mixtures and in the unknown candy samples. These took into account the three components and a noise.

OPEN IN VIEWER

Figure 4.

Root mean squared error of cross-validation (RMSECV) as a function of the number of components using the training set. There are two CV estimates: CV is the ordinary CV estimate and adjCV is a bias-corrected CV estimate.

The results in the training set were cross-validated to see the linearity and accuracy of our calibration model. The predicted versus measured concentration plot shows how well the model for each dye was able to predict the measured concentrations of the training set using four components. The model performed well overall, with the exception of only a few outliers (Figure 5). The training set RMSE and R² results using four components for the blue, red, and yellow dyes showed a lower R² value for the red component due to few outliers (Table 1). Removing the outliers did not improve the model performance as measured by the RMSE and REP for the test set.

OPEN IN VIEWER

Figure 5.

Comparison of predicted concentrations from model and measured concentrations (in mol/L) from training set (n = 64) for blue, red, and yellow dyes.

OPEN IN VIEWER

Table 1.

Training set RMSE and R² results using four components for the blue, red, and yellow dyes.

Dyes	RMSE	R2
Blue	2.90 × 10–7	0.82
Red	5.32 × 10–7	0.68
Yellow	4.23 × 10–7	0.97

Table 2 shows the results for the predicted versus measured concentrations of each dye in the test set. It can be seen that the predicted versus measured concentrations were both in the same order of magnitude.

OPEN IN VIEWER

Table 2.

Comparison of predicted and measured molar concentrations (in mol/L) of the test set (n = 13) from wavelengths 350–674 nm.

	Blue concentration		Red concentration		Yellow concentration
Mixtures	Predicted	Measured	Predicted	Measured	Predicted	Measured
1	7.03 × 10–7	5.55 × 10–7	1.11 × 10–6	1.00 × 10–6	5.84 × 10–6	5.89 × 10–6
2	2.29 × 10–6	2.32 × 10–6	1.34 × 10–6	1.00 × 10–6	5.90 × 10–6	5.89 × 10–6
3	6.12 × 10–7	5.55 × 10–7	3.09 × 10–6	3.51 × 10–6	6.25 × 10–6	5.89 × 10–6
4	1.92 × 10–6	2.32 × 10–6	3.11 × 10–6	3.51 × 10–6	5.63 × 10–6	5.89 × 10–6
5	6.37 × 10–7	5.55 × 10–7	2.32 × 10–6	2.51 × 10–6	3.05 × 10–6	2.62 × 10–6
6	2.04 × 10–6	2.32 × 10–6	2.52 × 10–6	2.51 × 10–6	2.83 × 10–6	2.62 × 10–6
7	6.16 × 10–7	5.55 × 10–7	2.21 × 10–6	2.51 × 10–6	8.39 × 10–6	8.90 × 10–6
8	2.02 × 10–6	2.32 × 10–6	2.36 × 10–6	2.51 × 10–6	8.22 × 10–6	8.90 × 10–6
9	1.56 × 10–6	1.67 × 10–6	2.36 × 10–6	1.00 × 10–6	6.48 × 10–6	2.62 × 10–6
10	1.53 × 10–6	1.67 × 10–6	3.20 × 10–6	3.51 × 10–6	2.86 × 10–6	2.62 × 10–6
11	1.60 × 10–6	1.67 × 10–6	1.21 × 10–6	1.00 × 10–6	8.41 × 10–6	8.90 × 10–6
12	1.49 × 10–6	1.67 × 10–6	3.11 × 10–6	3.51 × 10–6	8.72 × 10–6	8.90 × 10–6
13	1.61 × 10–6	1.67 × 10–6	2.47 × 10–6	2.51 × 10–6	6.16 × 10–6	5.89 × 10–6

The predicted and measured concentrations from the test set (Table 2) were used to calculate the RMSE and REP for the model correlating to each color (Table 3). The values calculated show relatively small RMSE and REP for the generated model.

OPEN IN VIEWER

Table 3.

Test set error including RMSE and REP for the blue, red, and yellow dyes.

Test set error
Dyes	RMSE	REP
Blue	1.83 × 10–7	12.01%
Red	4.59 × 10–7	19.52%
Yellow	1.12 × 10–6	19.37%

To evaluate the usefulness of our ternary mixture calibration model to predict unknown concentrations, we tested several candy samples and predicted the concentrations of blue, red, and yellow dye color components in them (Table 4). It must be noted that this is a preliminary study to determine how well the method would perform in predicting the unknown dye concentrations in selected food samples.

OPEN IN VIEWER

Table 4.

Predicted molar concentrations of the dye colors in the unknown candy samples (in mol/L).

Colors	Jelly Bean Pink concentration	Jelly Bean Purple concentration	Jelly Bean Green concentration	Sprinkles Yellow concentration
Blue	1.31 × 10–7	9.48 × 10–7	1.19 × 10–6	2.20 × 10–7
Red	3.36 × 10–6	3.61 × 10–6	4.91 × 10–7	1.07 × 10–7
Yellow	8.82 × 10–7	1.77 × 10–7	2.45 × 10–6	1.43 × 10–5

In Table 4, the “predicted molar concentrations of unknown samples” refers to the concentrations of the dyes in the candy/methanol solutions, as predicted by the PLS calibration model. Using standard dye solutions (red, blue, and yellow dyes in methanol), we then produced “replication” solutions matching the predicted concentrations from the candy samples. The intent of this experiment was to compare the spectra between the unknown candy/methanol solutions and the corresponding replicate dye mixtures given by the PLS calibration model. While we acknowledge the possibility of spectral interference due to other candy ingredients or dyes, the results show a comparable overlay between the actual candy samples and replicated sample spectra (Figure 6, Supplemental Material Figures S1–S3).

OPEN IN VIEWER

Figure 6.

Overlay plots between the recreated (standard) and actual spectra for the Jelly Bean Pink candy.

The replication of the Jelly Bean Pink sample (Figure S1) had a higher absorbance than the actual sample, with the replication wavelength of maximum absorbance (λ_max) at 532 nm and the actual sample λ_max at 534 nm. The replication of the Jelly Bean Purple sample showed local λ_max at 532 and 624 nm, compared to the λ_max of the actual Jelly Bean Purple sample at 536 and 622 nm (Figure S2). The Jelly Bean Green sample and its replication both exhibited λ_max at 624 nm (Figure S3). Lastly, the replication of the Decorating Yellow Sprinkles sample showed λ_max at 426 nm, near that of the actual Yellow Sprinkles sample at 428 nm (Figure S4). Thus, although small differences are observed between spectra, the spectra are generally comparable. The observed discrepancies in absorption and λ_max may be due to interfering ingredients in the actual Jelly Bean and Sprinkles samples.

Overall, the good spectral overlaps between the replicated and the actual samples suggest that our methods allow for accurate, simultaneous quantification of Erythrosine B (red), Brilliant Blue, and tartrazine (yellow) food dyes in prepared mixtures and food samples, requiring no purification or separation. This may be promising for future industrial use and warrants further investigation. There are, however, some limitations. It should be noted that some of the obtained yellow dye concentrations in the unknown candy samples are higher than the values of the same component in the training or prediction sets. Expanding the number of training samples to include a wide range of concentration matrices might lead to the development of a more robust PLS model, detecting varied concentrations for each sample component. As is typical with such analyses, this method should be validated with other commonly used techniques, such as HPLC or GC and consequently determine recovery results. Additionally, matrix complexity might confound results in food samples with more complex dye mixtures and overlapping absorbance peaks. To handle such complexity, though, the PLS calibration model can be extended to include other dyes of interest in food industries. Thus, this study and future studies may offer a direct, simple, and inexpensive method for the simultaneous quantitation of multiple dyes in various food samples.

Conclusions

A potential alternative method for the quantitative determination of food dye mixtures consisting of Erythrosine B (red), Brilliant Blue, and tartrazine (yellow) was developed. The REP was relatively low in the test set samples and there was good spectral overlap between the replicated and the actual samples. Thus, with independent validation, this study may offer a promising industrial tool that could potentially benefit the food enterprise. To better handle foods with more diverse, complex dye mixtures, further studies will examine the performance of a PLS model calibrated with additional dyes of interest in food industries.