Sage Journals: Discover world-class research

Abstract

The contamination of maize, a major staple food in South Africa, with fumonisin B1 (FB1), has become a major food safety concern. The regulation of this mycotoxin is extremely important and requires efficient detection methods. Near infrared (NIR) spectroscopy has gained widespread interest as a rapid and non-destructive mycotoxin analysis method. The purpose of this study was, therefore, to determine the NIR absorbance bands of FB1. The spectra of 30 FB1 solutions, constituted in methanol, as well as 30 methanol-only samples were recorded in the spectral range of 1000–2500 nm (10,000 – 4000 cm⁻¹). The data was pre-processed with multiplicative scatter correction (MSC) and a partial least squares discriminant analysis (PLS-DA) model was computed. The variable importance in projection (VIP) scores and selectivity ratio (SR) values were used for wavelength selection. A new PLS-DA model was computed with 454 chosen wavelengths and the regression vector of this model was investigated to further identify and remove irrelevant wavelengths. The final model was computed with 150 wavelengths and nine latent variables (LVs) and obtained a classification accuracy of 100% for both the calibration and external validation sets. By investigating the regression vector of the final PLS-DA model, potential FB1 absorbance bands were identified at 1446 nm, 1453 nm, 1891 nm, 2036 nm, 2046 nm, 2148 nm, 2224 nm, 2262 nm and 2273 nm. This study was therefore able to identify the previously unknown NIR absorbance bands of FB1 at 100 ppm.

Keywords

Chemometrics FB1 absorbance bands fumonisin B1 maize mycotoxin detection near infrared spectroscopy variable selection

Introduction

Fumonisins are toxic secondary metabolites produced by the fungi Fusarium verticillioides and Fusarium proliferatum and are the most common mycotoxins which accumulate in maize and maize by-products.¹ Since its discovery in 1988, more than a thousand fumonisins have been identified, with the most important being fumonisin B1 (FB1) and fumonisin B2 (FB2) of which FB1 is the most toxic and the most abundant in maize.^2,3 FB1 has a main chain of 20 carbon atoms to which one amine group and several carboxylic acid, ester, hydroxyl, methyl and methylene groups are attached.² The chemical structure of FB1 is presented in Figure 1.

Figure 1.

The chemical structure of Fumonisin B1.

Fumonisins are toxic to both humans and animals. In animals it has been responsible for various chronic illnesses such as pulmonary edema in pigs, leukoencephalomalacia in horses, and liver cancer in rats.^2,4,5 In humans, fumonisins have been associated with an increase in oesophageal cancer cases in certain regions of South Africa, China and Iran, and has been classified as a group 2B human carcinogen by the International Agency for Research on Cancer.^6–11 Due to its toxicity, its presence in maize, which is a major staple food for many South Africans, must be closely regulated.¹² The Codex Alimentarius Commission has set maximum levels (MLs) for FB1 and FB2 to 2 ppm in maize flour and maize meal and 4 ppm in raw maize.¹³ These levels have been legislated in South Africa.¹⁴ To ensure that regulations are met, rapid and accurate mycotoxin analysis methods with the possibility of automated monitoring are required.

Currently, the most popular conventional analysis methods include high performance liquid chromatography (HPLC) coupled with fluorescence detection and/or mass spectrometry (MS) and enzyme-linked immunosorbent assay (ELISA).^15–17 Although these methods are accurate and precise, they are unable to provide real time feedback and require extensive sample preparation which is destructive, time consuming, expensive and results in a large amount of chemical waste.^18,19

Near infrared (NIR) spectroscopy has gained attention due to its ability as a rapid, non-destructive detection method which provides real time feedback and reduces chemical waste. NIR spectroscopy exposes food constituents to electromagnetic waves in the region of 780–2500 nm.¹⁸ The food constituents interact with the electromagnetic waves and these interactions are recorded as a spectrum. To obtain quantitative and/or qualitative information from the spectra, different mathematical and statistical techniques are used which are defined as chemometrics.¹⁸

In 2002, NIR spectroscopy was used to successfully differentiate between fumonisin positive (>100 ppm) and negative (<10 ppm) maize kernels, but was unable to discriminate between kernels which were contaminated at the regulatory limits (4 ppm).²⁰ However, in 2022, a smartphone controlled portable NIR spectrometer, combined with chemometric models, was used to classify naturally contaminated samples of ground maize at the FB1 regulatory limit and achieved a classification accuracy of 86%.²¹ The improved classification results at lower detection levels within only 20 years, emphasises the rapid evolution of NIR spectroscopy. Thus, the further development of instrumentation and improvement of chemometric algorithms, will potentially result in the greater implementation of NIR spectroscopic methods for mycotoxin detection when compared to conventional analysis methods.

To develop NIR calibration models, the substance to be measured must be able to absorb in the NIR region.²² These substances are organic in nature and can be detected with direct measurement.²² However, if the substance is inorganic, or present at levels too low for direct measurement, indirect measurement can be used.²² During this measurement, the substance of interest is predicted through covariation with other absorbing substances.²²

Although fumonisins are organic compounds, the first study that used NIR spectroscopy to detect it, stated that fumonisins absorb negligible amounts of NIR energy at the ppm level.²⁰ However, there was no basis for this statement, since the absorbance bands of fumonisins had not yet been determined and, therefore, it was not possible to conclude at which level fumonisins could be detected. Nevertheless, due to this statement, subsequent studies focussed their attention on the use of indirect measurement for the detection and quantification of fumonisins.^15,16

One study used fungal count and ergosterol levels to classify maize kernels according to their fumonisin content.¹⁵ The study reported that ergosterol levels were not proportional to the fumonisin content and could, therefore, not be used for classification.¹⁵ The study also chose a threshold value of 200 000 CFU/g for the fungal biomass, beyond which the fumonisin content would most likely exceed 4 ppm.¹⁵ By using this threshold, they were able to correctly classify 96% and 82% of maize kernels in the calibration and external validation sets, respectively.¹⁵ Other studies depended on chemometric models for analysis and not spectral interpretation. These studies concluded that the wavelengths responsible for classification and/or quantification, were related to changes in the colour, vitreousness and nutrient content (protein, starch and fatty acids) of maize kernels.^16,20,21

Although these studies obtained relatively good classification accuracies, the factors and conditions required by which fungi produces FB1, is not fully understood. Often the fungi can be present in maize grain without producing FB1 or the toxin is present without visible symptoms of infection, and this will decrease the accuracy of these methods.²³ However, direct measurement methods have the potential to increase the accuracy and allow the detection of FB1 in maize kernels even after the fungi is removed.

In 2009, Peiris et al. examined the spectra of another mycotoxin, deoxynivalenol (DON), in acetonitrile and on sound and infected maize kernels.²⁴ A dilution series of 0.5–2000 ppm deoxynivalenol in acetonitrile were prepared and after the spectra were pre-processed with second derivatives and spectral subtraction was applied, differences in the absorbance bands intensities were identified at 1408 nm, 1904 nm and 1919 nm.²⁴ Spectral shifts were observed at similar wavelengths (1425 – 1440 nm and 1915 – 1930 nm) when the spectra of sound and infected kernels were compared. The study was able to identify the absorbance bands of DON at the ppm level and confirmed that, contrary to previous published statements, certain mycotoxins can absorb detectable amounts of NIR energy at the ppm level.²⁴

Based on these findings, it was decided that the statement made against the direct detection of FB1, needs to be reinvestigated. The aim of this project was, therefore, to determine the feasibility of using NIR spectroscopy for the direct detection of FB1. This was done by determining the FB1 absorbance bands at the ppm level, which, to the best of our knowledge, has not yet been identified.

Materials and methods

Sample preparation

Forty milligrams (40 mg) of FB1 was obtained from Merck (Darmstadt, Germany) in pure crystalline form. Since the amount of FB1 purchased was not sufficient to obtain the absorbance spectra of the FB1 powder in the reflectance mode, it was dissolved in methanol and its spectra were obtained in transmittance mode.

FB1 was dissolved in 99.9% HPLC grade methanol (Merck, Darmstadt, Germany) to prepare a 5000 ppm FB1 stock solution. From the stock solution, two samples were further diluted with methanol to obtain two 100 ppm FB1 solutions. For each of the 100 ppm FB1 solutions, a vial containing methanol only was prepared and was used as a blank. The samples were prepared in a vacuum hood at room temperature in the Department of Food Science, Stellenbosch University. Both the methanol-only and the FB1 solution samples were stored at −20°C. These two repeats of the FB1 solution and methanol-only samples are referred to as the experimental repeats and were prepared to account for the variation introduced during sample preparation.

Before the spectra were recorded, the samples were removed from the freezer, the room temperature was set to 21°C and the spectrometer was switched on and allowed to heat up for an hour. From the first experimental repeat, the methanol was pipetted into a 0.5 mm quartz cuvette (Hellma, Germany), placed into the spectrometer and three scans were taken without removing the cuvette from the instrument. These scans were referred to as the instrumental repeats and were used to investigate the instrumental drift and the effect of NIR energy on the samples. After the spectra of the methanol instrumental repeats were obtained, the methanol was discarded, the cuvette was rinsed with distilled water and ethanol. The cuvette was then dried with a glass specific cloth and allowed to dry to ensure that no ethanol or water is left behind. The FB1 solution of the experimental repeat was then pipetted into the cuvette and the spectra were recorded in triplicate. This process was repeated five times (sample repeats) for each of the experimental repeats. The sample repeats ensured that the variation introduced by changes in the temperature and the volume of the samples between measurements, were accounted for. Thus, a total of 60 repeats were used in this study to ensure that the within sample variation was accounted for and to prevent the incorrect identification of FB1 absorbance bands. The samples were scanned at the Vibrational Spectroscopy Unit in the Department of Food Science, Stellenbosch University. The preparation of the repeats is illustrated in Figure 2.

Figure 2.

An illustration of how the experimental repeats, sample repeats and instrumental repeats were prepared and recorded. This illustration was created with BioRender.com.

NIR instrumentation

A Büchi NIRFlex N-500 NIR spectrometer (BÜCHI Labortechnik AG, Switzerland) with an InGaAs (Indium Gallium Arsenide) detector and a tungsten halogen lamp was used. The spectrometer had a wavelength range of 1000–2500 nm (10,000 – 4000 cm⁻¹) and a spectral resolution of 0.4 nm and 16 scans were taken and averaged to obtain the spectra. The spectrometer’s temperature was set to 35°C and a maximum temperature drift of ± 0.3°C was allowed. The spectrometer was connected to a computer and the NIRWare software package (BÜCHI Labortechnik AG, Switzerland) was used to import the data as transmittance and convert it to absorbance. The data was then exported as an excel file.

Data analysis

MATLAB R2021 (MathWorks Inc., Natick, MA, USA) was used to analyse and visualize the raw spectral data and the PLS_Toolbox 9.0 (Eigenvector Research Inc., Wenatchee, WA, USA) extension was used for multivariate data analysis.

Spectral analysis

The raw absorbance spectra of the FB1 solutions and methanol-only samples were evaluated and the differences between the spectra were investigated to determine if the presence of FB1 had a visible effect on the spectra. However, methanol is an organic molecule which absorbs NIR energy and, therefore, the changes in methanol may result in spectral changes which could affect the identification of the FB1 absorbance bands.²⁵ Thus, the bands affected by the changes in methanol were identified by subtracting the mean methanol spectrum from each of the methanol spectra using equation 1.

m_{d i f f} = m_{i j k} - \bar{m}

(1)

Where:

$m_{d i f f}$ = methanol difference spectra

$m_{i j k}$ = methanol spectra for experimental repeat i, biological repeat $j$ and instrumental repeat k

$\bar{m}$ = mean methanol spectrum

Muncan and Tsenkova²⁶ (2019) discussed the effect of NIR energy on the spectra of water and referred to this effect as illumination perturbation. It was observed that illumination perturbation affects the hydrogen bonds in water and since methanol also forms hydrogen bonds, it is likely that this phenomenon will also affect the methanol spectra.²⁷ Therefore, equation (2) was used to determine the effect of illumination perturbation on the methanol spectra by subtracting the mean methanol spectrum from the mean methanol spectra of each of the instrumental repeats.

m_{k_d i f f} = {\bar{m}}_{k} - \bar{m}

(2)

Where:

$m_{k_d i f f}$ = methanol difference spectra for instrumental repeat k

${\bar{m}}_{k}$ = mean methanol spectrum for instrumental repeat $k$

$\bar{m}$ = mean methanol spectrum

After the absorbance bands affected by illumination perturbation were identified, the possible FB1 absorbance bands were determined by subtracting each methanol spectra from its corresponding FB1 solution spectra using equation (3).

x_{d i f f} = f_{i j k} - m_{i j k}

(3)

Where:

$x_{d i f f}$ = difference spectra

$f_{i j k}$ = FB1 solution spectra for experimental repeat i, sample repeat $j$ and instrumental repeat k

$m_{i j k}$ = methanol spectra for experimental repeat i, sample repeat $j$ and instrumental repeat $k$

Multivariate analysis

Spectral pre-processing

Several pre-processing methods were evaluated and the method which resulted in the best model performance was selected. In this study, standard normal variate (SNV) and multiplicative scatter correction (MSC) were used to reduce scattering effects.^28,29 Savitzky-Golay (SG) second derivative pre-processing was used for the removal of noise, and the reduction of additive and multiplicative effects.³⁰ SG was used with a second order polynomial and a window width of 21 points.

Principal component analysis

The goal of principal component analysis (PCA) is to reduce data dimensionality by using a small number of principal components (PCs) to capture the variation present in a large dataset.³¹ The PCA models were constructed for the raw as well as the pre-processed data. The models were then compared and the model which was able to achieve the best separation between the FB1 solution samples and the methanol-only samples in the score plots, using the lowest number of PCs, was selected. The loadings of the chosen model were investigated to identify which bands were responsible for the separation in the scores plot.

Partial least squares discriminant analysis

Partial least squares discriminant analysis (PLS-DA) combines the regression properties of PLS with the classification properties of DA and can provide information on the variables which are important for classification, making it perfect for this study.³² PLS-DA was performed on the pre-processed as well as the raw data. The Kennard-Stone algorithm was used to split the data into 70% calibration and 30% external validation samples.³³ For both the calibration and the external validation sets, the samples were selected in a 1:1 ratio of FB1 solution to methanol samples. The venetian blinds cross-validation method was used to select the optimal number of latent variables (LVs) that corresponded with the lowest root mean squared error of cross-validation (RMSECV). The performance of the models was evaluated by calculating the classification accuracy, classification error, specificity, and sensitivity.²¹ The model which obtained the highest classification accuracy with the lowest number of LVs, was selected.

The importance of each band on classification was investigated by inspecting the vector of the regression coefficients with respect to the FB1 solution samples. Bands which had a large positive regression coefficient were recognised as bands that can potentially be used for the identification of FB1. However, there is a possibility that these bands may not be related to FB1. This is because the use of LVs, instead of the original variables for classification, increases the model complexity.³⁴ The LVs are constructed by not only maximizing the covariance between the independent variables (X) and the dependant variables (Y), but also the explained variance in the X variables.³⁵ This often results in the inclusion of irrelevant information, especially when the dominant source of variation is not related to Y, which causes the interpretation of regression coefficients to be very complex.³⁴

Variable selection

Variable selection methods can be used to remove irrelevant variables which will typically improve the model performance and interpretation, and reduce model complexity.³⁶ One of the most used variable selection methods is the variable importance in projection (VIP).³⁷ The VIP scores are useful for understanding the X prediction variables which best predicts the variation in the Y variables.³⁸ The scores are calculated as the sum of squares of the PLS weights (w*) which accounts for the amount of Y variation explained for each LV.³⁸ The VIP scores summarizes the influence of each X variable on the PLS model. The X variables which have the largest influence on the model are then chosen according to which variables have a VIP score greater than or equal to a certain threshold value.³⁹ Since the sum of squares of the VIP has an average value of 1, it is usually used as the threshold value, however this value is not statistically justified and is allowed to be altered.³⁴

Target projection (TP) with Selectivity Ratio (SR) is another variable selection method suggested for multivariate data.⁴⁰ SR ranks the X variables in relation to the variance explained in the Y variables. The SRs are determined by calculating the ratio of the explained to the residual variance of the X variables on the Y target-projected components.³⁸ The F-test was suggested to select the most important variables which are related to the investigated property.⁴¹ For a variable to be selected, the SR value must exceed the critical value of the F distribution (F_crit). The F_crit value is calculated from the F distribution table.⁴² A significance level ( $α$ ) is chosen and for the degrees of freedom, the numerator and denominator is calculated as $N - 2$ and $N - 3$ respectively, where $N$ is the sample size.

S R_{i} > F_{c r i t} = F (α, N - 2, N - 3)

(4)

For this study, the threshold value of the VIP scores was set to one and a significance level of 0.05 was selected to determine the F_crit value which was used as the threshold value for the SR. The automatic VIP and SR function of the PLS toolbox was used to select the wavelengths of interest. During the execution of this function, wavelength selection was performed using both the VIP scores and the SRs simultaneously and the wavelengths that produced the model with the best RMSECV, was included in the final model. In the first run of both the VIP and SR variable selection methods, the wavelengths with the lowest VIP or SR values were eliminated, this was repeated until the model showed no improvement. For each run, a certain fraction of the wavelengths had to be removed. To determine which fraction was the most appropriate to remove, the function evaluated the outcome of several fractions. These fractions included: 2%, 5%, 8%, 10%, 15%, 20%, 30%, 35%, 40% and 45%.

Results and discussion

Spectral analysis

The raw absorbance spectra of the 100 ppm FB1 solution samples and the methanol-only samples are presented in Figure 3 and differences between these spectra were observed at approximately 1880–2020 nm and 2270– 2280 nm. These differences may have been caused by the presence of FB1 in methanol or by the changes of methanol itself. The spectra of the methanol-only samples were investigated to determine the cause of these spectral differences. The mean spectrum of the recorded methanol spectra is given in Figure 4 and the absorbance bands and its assignments, are indicated and summarised in Table 1.

Figure 3.

The raw absorbance spectra of the 30 FB1 solution samples and the 30 methanol-only samples.

Figure 4.

The mean absorbance spectrum of methanol and the assignments of its different absorbance bands. The symbols $v$ and $δ$ are used to indicate stretching and bending bands respectively and the subscript $a$ is used to indicate antisymmetric stretching/bending and $s$ indicates symmetric stretching/bending. Combination bands are indicated as the sum of two vibrations and the level of the overtone bands are indicated by a coefficient with the order n + 1.

Table 1.

A summary of the methanol absorbance bands and their assignments based on the work done by Fernandes et al. (2008) and Workman and Weyer (2012).^25,43

Wavelengths	Molecular vibration	Functional group
1195 nm (8368 cm ⁻¹ )	Second overtone stretch of the CH bond	-CH₃
1350 – 1670 nm (7407 – 5988 cm⁻¹) - 1353 nm and 1380 nm (7407 - 7246 cm⁻¹) - 1415 nm (7067 cm⁻¹) - 1450 nm (6896 cm⁻¹) - 1460 – 1600 nm (6849 – 6250 cm⁻¹)	Overlapped bands of - Second overtone stretching and bending combination band of the CH bond - First overtone of free OH bond - First overtone stretch of OH in primary alcohol - First overtone broad band of bonded OH	-CH₃ Free -OH -OH alcohol Bound -OH
1700 nm (5882 cm ⁻¹ ) 1730 nm (5780 cm ⁻¹ )	First antisymmetric overtone stretch of the CH bond First symmetric overtone stretch of the CH bond	-CH₃
1800 – 2200 nm (5555 – 4545 cm⁻¹) - 1930 nm (5181 cm⁻¹) - 1965 nm (5089 cm⁻¹) - 2016 nm (4960 cm⁻¹) - 2080 nm (4808 cm⁻¹) - 2123 nm (4710 cm⁻¹)	Broad combination region of OH group - OH stretching and HOH bending - OH stretching and CH bending combination band - OH stretching and bending combination band - OH stretching and CO bending combination band - OH stretching and CO stretching combination band	-OH -OH and -CH -OH -OH and -CO -OH and -CO
2200 – 2500 nm (4545 – 4000 cm⁻¹) - 2280 nm (4386 cm⁻¹) - 2338 nm (4277 cm⁻¹)	First combination region of CH bond - Antisymmetric stretching and bending combination band of the CH bond - Symmetric stretching and bending combination band of the CH bond	-CH₃
2410 – 2460 nm (4149 – 4065 cm ⁻¹ )	Second overtone symmetric bending umbrella vibration of the CH bond	-CH₃

The absorbance spectra of the methanol-only samples are presented in Figure 5 and spectral differences were observed at approximately the same wavelengths as identified in Figure 3. To determine the specific wavelengths at which the spectra differed, a very similar method to that used by Peiris et al. was used.²⁴ The difference spectra in Figure 6 were calculated with equation (1). From the peaks of the difference spectra, it was determined that the largest differences between the raw methanol spectra occurred at 1450 nm, 1938 nm, 2277 nm, 2339 nm and 2432 nm. It was therefore concluded that these bands are affected by changes in methanol.

Figure 5.

The raw absorbance spectra of the methanol-only samples. For each instrumental repeat, 10 methanol-only samples were scanned. The different instrumental repeats are indicated.

Figure 6.

The difference spectra after the mean methanol spectrum was subtracted from each methanol spectra. For each instrumental repeat, 10 methanol-only samples were scanned. The different instrumental repeats are indicated.

The spectral differences at 1450 nm, 1938 nm and 2277 nm were associated with the first overtone stretch of the OH bond, the OH stretching and CH bending combination band and the antisymmetric stretching and bending combination band of the CH bond.^25,43 The symmetric stretching and bending combination band of the CH bond was responsible for the spectral differences at 2339 nm and the differences at 2432 nm were caused by the second overtone symmetric bending umbrella vibration of the CH bond.^25,43 These differences between the methanol spectra may have been caused by changes in the volume of the samples, temperature, evaporation and/or the effect of illumination perturbation.

The spectra in Figure 7 were calculated using equation (2) and the peaks of the difference spectra illustrate the effect of illumination perturbation on the absorbance spectra of methanol. The methanol spectra changed with each subsequent exposure to NIR energy, and the largest change occurred between the spectra of instrumental repeat 1 (red) and instrumental repeat 2 (blue). The spectra of instrumental repeat one was taken before the sample was exposed to NIR energy and the spectra of instrumental repeat two was taken after the sample was exposed to NIR energy for the first time, which explained the larger difference between these spectra. With each exposure to NIR energy, the absorbance at wavelengths associated with bound OH groups (1595 nm) decreased and increased at the wavelengths associated with free OH groups (1420 nm).⁴³ These changes occurred in the first overtone band region of the OH group and subsequently the combination band region of the OH group was also affected. The changes in the combination region of the OH groups were visible at 2030 nm and 2094 nm.⁴³ Thus, the exposure of methanol to NIR energy could have resulted in the breakage of its hydrogen bonds.

Figure 7.

The difference spectra after the mean spectrum of methanol were subtracted from the mean methanol spectra of instrumental repeat 1, instrumental repeat 2 and instrumental repeat 3.

A decrease in absorbance at 2277 nm, 2341 nm and 2437 nm with each subsequent exposure to NIR energy was also observed and was associated with the change in the CH groups.⁴³ The change in these absorbance bands were caused either by the breakage of the CH bonds or a decrease in the amount of methanol due to evaporation. It was concluded that illumination perturbation could have influenced the methanol spectra and therefore these changes must be considered when the FB1 absorbance bands are identified.

The difference between the FB1 solution spectra and the spectra of the methanol-only samples were calculated using equation (3) and are presented in Figure 8. The difference spectra had peaks at 1442 nm, 1938 nm, 2277 nm and 2427 – 2460 nm. These peaks occurred at similar wavelengths as those affected by changes in methanol (Figure 6). It was therefore concluded that the spectral differences observed in Figure 8 were the result of the changes in methanol between the samples and not the presence of FB1. Thus, methanol was responsible for the largest variation between the spectra and the absorbance bands of FB1 could not be determined by simple spectral subtraction. More advanced chemometric methods were therefore required to identify the unique FB1 absorbance bands.

Figure 8.

The difference spectra after the spectra of each methanol-only sample were subtracted from its corresponding FB1 solution samples.

Principal component analysis

A PCA model was constructed on the raw spectral data and the scores plots of the model are presented in Figure 9. From the scores plots it was clear that PC 7 (0.02%) obtained the best separation between the methanol-only and the FB1 solution samples and the scores plot of PC 6 (0.08%) against PC 7 obtained an even better separation (Figure 10).

Figure 9.

The scores plot of PC1 against PC2 – PC7 of the PCA model computed on the raw data.

Figure 10.

The scores plot of PC 6 and PC 7 of the PCA model computed on the raw data of the 30 FB1 solution samples and the 30 methanol-only samples.

MSC, SNV and SG were applied to the data. The SNV and MSC pre-processed PCA models delivered very similar results and were able to obtain the best separation between the samples when plotting the scores of PC 5 (0.06%) against those of PC 6 (0.04%) (Figure 11). Three FB1 solution samples did, however, occur in the same area as the methanol-only samples and were identified as potential outliers, requiring further investigation.

Figure 11.

The scores plot of PC 5 and PC 6 of the PCA model computed on the MSC pre-processed spectra of the 30 FB1 solution samples and the 30 methanol-only samples.

Despite the potential outliers, the MSC and SNV pre-processed models were still able to separate the samples well with a smaller number of PCs than the model constructed on the raw data. The SG pre-processed model was not able to separate the samples as well as the SNV and MSC pre-processed models and therefore the MSC pre-processed PCA model was chosen for further analysis. Although the PCs used for separation in this model explained a very small amount of the variation (0.10%), this was expected since FB1 was present at very low levels and the changes in methanol were responsible for the largest differences between the spectra.

The loadings of PC 5 and PC 6 are given in Figure 12. Since most of the FB1 solution samples had negatives score values for PC 5 and PC 6, the wavelengths which had negative loading values for each of these PCs, were identified as potentially important wavelengths at which FB1 absorbs NIR energy. These wavelengths included 1400 – 1600 nm, 2024 – 2082 nm and 2240 – 2274 nm. Both loadings had a large peak at 2262 nm, which indicated that this wavelength has a particularly strong correlation with FB1. To confirm that the FB1 bands does occur at these wavelengths, PLS-DA models were computed.

Figure 12.

The loadings of PC 5 and PC 6 of the PCA model computed on the MSC pre-processed spectra of the FB1 solution samples and the methanol-only samples.

Partial least squares discriminant analysis

PLS-DA models were computed on the raw and pre-processed spectral data and the model performances and parameters are summarised in Table 2. The model of the raw data gave perfect classification results for both the calibration and external validation sets, however the SNV and MSC pre-processed models gave the same results with less LVs. The SG pre-processed model delivered the same results with a larger number of LVs. The MSC pre-processed PLS-DA model was therefore selected for further investigation. Each of the models had a lower specificity of cross validation, which was an indication that there were FB1 solution samples present in the calibration sets that had a large influence on the models.

Table 2.

The model performances of the different PLS-DA models used to distinguish between the FB1 solution samples and methanol-only samples in which 42 samples were used for calibration and 18 samples were used for validation.

Pre-processing		None	SNV	MSC	SG
Number of LVs		8	7	7	9
Calibration	Sensitivity	100	100	100	100
	Specificity	100	100	100	100
	Classification accuracy	100	100	100	100
Cross validation	Sensitivity	100	100	100	100
	Specificity	95	95	95	95
	Classification accuracy	98	98	98	98
External validation	Sensitivity	100	100	100	100
	Specificity	100	100	100	100
	Classification accuracy	100	100	100	100

Three samples were identified as possible outliers in Figure 11 and the model was reconstructed without these samples. The new PLS-DA model was computed with 6 LVs and a perfect classification accuracy of cross validation was obtained. Due to the lower number of LVs and higher classification accuracy of cross validation, the samples were confirmed as outliers and the regression coefficients of the model computed without these samples (Figure 13), were used to determine the wavelengths at which FB1 absorbs.

Figure 13.

The regression vector of the PLS-DA model computed on the MSC pre-processed spectra of the FB1 solution samples and the methanol samples, after the three outlier samples of the FB1 solution were removed.

The wavelengths in the ranges of 1350–1580 nm, 1810 – 1930 nm, 2015 – 2170 nm and 2200 – 2280 nm was selected as possible FB1 band positions. Of these, 2262 nm was especially important since it had a very large regression coefficient and was also identified as an important wavelength in the loadings of both PC 5 and PC 6 (Figure 12). Although it was possible to identify that the band at 2262 nm played an important role in the detection of FB1, it was difficult to identify other important bands, since the regression vector included several irrelevant wavelengths. Variable selection methods were therefore required to discard irrelevant wavelengths and improve the identification of important bands.

Variable selection

The automatic VIP and/or SR ratio function successfully reduced the number of wavelengths from 1501 to 454 and selected the wavelengths: 1000 – 1048, 1050 – 1053, 1060, 1062 – 1064, 1408 – 1460 nm, 1788 – 1905, 2016 – 2056, 2134 – 2159, 2166 – 2168, 2191 – 2354, and 2374 – 2500. The function removed the wavelengths which had the lowest VIP score and/or SR ratio and whose inclusion resulted in a lower model performance. A new model was computed with the selected wavelengths. Six LVs were computed and a classification accuracy of 100% was obtained for both the calibration and external validation sets. The regression vector of the model after wavelength selection is presented in Figure 14.

Figure 14.

The regression vector of the PLS-DA model computed on the 454 wavelengths which were selected with the VIP and/or SR function.

From the regression vector, it was determined that the wavelengths at 1424 – 1460 nm, 1871 – 1906 nm, 2016 – 2056 nm, 2135 – 2159 nm, 2167 – 2168 nm and 2203 – 2280 nm were related to FB1. To further investigate the relationship of these wavelengths with FB1, the final PLS-DA model was computed with only these wavelengths. The model was computed with nine LVs and obtained a 100% classification accuracy for the calibration and external validation sets. The excellent classification accuracy obtained by this model, with a relatively small number of LVs, confirmed that the essential wavelengths for FB1 detection were included in the model. The regression vector of this model is presented in Figure 15.

Figure 15.

The regression vector of the final PLS-DA model computed with the 150 wavelengths which had been selected from regression vector in Figure 15.

The regression vector had peaks at 1446 nm, 1453 nm, 1891 nm, 2036 nm, 2046 nm, 2148 nm, 2224 nm, 2262 nm and 2273 nm. These wavelengths differ from those which were influenced by the changes in methanol (Figure 6). Most of these wavelengths fell within the important wavelength ranges identified in the loadings (Figure 12), and all of them had a positive regression coefficient for the model before wavelength selection (Figure 13), indicating that they are related to FB1. Therefore, it was concluded that FB1 absorbs NIR energy at approximately these wavelengths.

The amine, carboxylic acid, ester, hydroxyl, methyl and methylene groups of FB1 absorbs NIR energy. The NIR band positions of these functional groups are very dependent on the environment in which it occurs. Since FB1 is a very complex molecule and its interaction with methanol is not well known, the assignment of band positions was very complicated. To aid with band assignment, the NIR characteristics of the different functional groups were investigated. In addition, the band positions were compared to those of the same functional groups found in other materials.⁴³

Primary amines have doublet peaks in the first overtone region at approximately 1400–1700 nm and can either have a doublet or a single peak in its first combination region at 2000 nm, depending on the polarity of the solvent in which it was dissolved.⁴³ It is therefore most likely that the doublet peaks identified at 1446 nm and 1453 nm belongs to the first overtone band and the peaks at 2036 nm and 2046 nm are part of the first combination band of the primary amine group. The second overtone of the C = O bond in carboxylic acids have been identified at approximately 1900 nm and could therefore be responsible for the peak at 1891 nm.⁴⁴

Similar peaks at 2148 nm, 2242 nm, 2262 nm and 2273 nm have, to the best of our knowledge, not yet been assigned to the specific functional groups present in FB1 in previous research. However, these bands fall within the combination band region of most of the functional groups and are therefore most likely the result of the combination bands of the methylene, carbonyl and/or primary amine groups. Table 3 summarises the potential FB1 bands and their assignments.

Table 3.

The wavelengths associated with fumonisin B1 and their band assignments.

FB1 bands	Molecular vibration	Functional group
1446 nm and 1453 nm (6915 cm ⁻¹ – 6882 cm ⁻¹ )	Antisymmetric and symmetric first overtone N-H stretch of the -NH₂ group	-NH₂
1891 nm (5288 cm ⁻¹ )	Second overtone band of C = O in carboxylic acid	-COOH
2036 nm and 2046 nm (4911 cm ⁻¹ and 4888 cm ⁻¹ )	First combination band of -NH₂ due to NH stretching and the deformation or scissoring of the -NH₂ group	-NH₂
2148 nm, 2242 nm, 2262 nm and 2273 nm (4655 cm ⁻¹ , 4460 cm ⁻¹ , 4421 cm ⁻¹ and 4399 cm ⁻¹ )	Combination bands of the carbonyl, methylene and/or amine groups	-COOR, -COOH, -NH₂, -CH₂

Conclusion

This study identified the potential absorbance bands of FB1, which were previously unknown. Furthermore, the bands of 100 ppm FB1, dissolved in methanol, were identified, demonstrating that FB1 does absorb detectable levels of NIR energy at the ppm level. Thus, NIR spectroscopy has the potential to be used as a rapid and non-destructive method for the direct measurement of FB1. The regulatory limits of FB1 in maize (4 ppm) is, however, much lower than the levels at which FB1 was detected in this study, and we also concluded that the band positions of the FB1 absorbance bands are extremely dependant on the environment in which it is measured. Therefore, further research will focus on the determination of the absorbance bands of FB1 in naturally and artificially contaminated maize kernels. It will also focus on determining the limits of detection and quantification of FB1 with NIR spectroscopy by creating dilution series.

Footnotes

Acknowledgements

The authors would like to acknowledge Dr T. Tait at the Department of Food Science, Stellenbosch University for his contributions to the project. The authors also acknowledge Dr J Colling from the Vibrational Spectroscopy Unit of the Central Analytical Facility at Stellenbosch University for use of the Buchi NIRFlex N-500.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Research Foundation of South Africa [grant number 137998]; Postgraduate Scholarship Program of Stellenbosch University.

ORCID iDs

Anja Laubscher

Paul J. Williams

References

Miller

. Mycotoxins in small grains and maize : Old problems, new challenges. Food Addit Contam 2008; 25: 219–230.

Gelderblom

WCA

Jaskiewicz

Marasas

WFO

, et al. Fumonisins - novel mycotoxins with cancer-promoting activity produced by Fusarium moniliforme. Appl Environ Microbiol 1988; 54(7): 1806–1811.

Kamle

Mahato

Devi

, et al. Fumonisins: Impact on Agriculture, Food, and Human Health and their Management Strategies. Toxins (Basel) 2019; 11(6): 328.

Harrison

Colvin

Greene

, et al. Pulmonary edema and hydrothorax in swine produced by fumonisin B1, a toxic metabolite of Fusarium moniliforme. J Vet Diagn Invest 1990; 2: 217–221.

Wilson

Ross

Rice

, et al. Fumonisin B 1 levels associated with an epizootic of equine leukoencephalomalacia. J Vet Diagnostic Investig 1990; 9: 213–216.

Marasas

Jaskiewicz

Venter

, et al. Fusarium moniliforme contamination of maIze in oesophageal cancer areas in Transkei. South African Med J 1987; 74(3): 110–114.

Rheeder

Marasas

WFO

Thiel

, et al. Fusarium moniliforme and fumonisins in corn in relation to human esophageal cancer in Transkei. Postharvest Phytopathol Mycotoxins 1992; 82(3): 357.

Yoshizawa

Yamashita

Luo

. Fumonisin occurrence in corn from high- and low-risk areas for human esophageal cancer in China. Appl Environ Microbiol 1994; 60(5): 1626–1629.

Alizadeh

Roshandel

Roudbarmohammadi

, et al. Fumonisin B1 contamination of cereals and risk of esophageal cancer in a high risk area in Northeastern Iran. Asian Pacific J Cancer Prev 2012; 13(6): 2625–2628.

10.

Marasas

WFO

Gelderblom

WCA

Shephard

, et al. Mycotoxins: A global problem. In: Leslie

Bandyopadhyay

Viconti

(eds). Mycotoxins: Detection Methods, Management, Public Health and Agricultural Trade. London: CAB international, 2008, pp. 29–40.

11.

Selvaraj

Zhou

Wang

, et al. Mycotoxin detection — Recent trends at global level. J Integr Agric 2015 Nov 1; 14(11): 2265–2281.

12.

Shephard

. Impact of mycotoxins on human health in developing countries. Food Addit Contam 2008; 25: 146–151.

13.

Shephard

Burger

Rheeder

, et al. The effectiveness of regulatory maximum levels for fumonisin mycotoxins in commercial and subsistence maize crops in South Africa. Food Control 2019 Mar 1; 97: 77–80.

14.

DoH SAND of H . Regulations governing tolerances for fungus-produced toxins in foodstuffs: amendment: Government Notices, Government Gazett, South - Africa, 2016, pp. 4–5.

15.

Levasseur-Garcia

Bailly

Kleiber

, et al. Assessing risk of fumonisin contamination in maize using near-infrared spectroscopy. J Chem 2015; 1–10.

16.

Tyska

Mallmann

Vidal

, et al. Multivariate method for prediction of fumonisins B 1 and B 2 and zearalenone in Brazilian maize using near infrared spectroscopy (NIR). PLoS One 2021; 16(1): 1–14.

17.

Chavez

Cheng

Herrman

, et al. Single kernel aflatoxin and fumonisin contamination distribution and spectral classification in commercial corn. Food Control 2022; 131: 108393.

18.

Pasquini

. Near infrared spectroscopy: fundamentals, practical aspects and analytical applications. J Braz Chem Soc 2003; 14(2): 198–219.

19.

Krska

Schubert-ullrich

Molinelli

, et al. Mycotoxin analysis : An update. Food Addit Contam 2008; 25(2): 152–163.

20.

Dowell

Pearson

Maghirang

, et al. Reflectance and transmittance spectroscopy applied to detecting fumonisin in single corn kernels infected with Fusarium verticillioides. Cereal Chem 2002; 79(2): 222–226.

21.

Shen

Kang

, et al. Rapid detection of fumonisin B1 and B2 in ground corn samples using smartphone-controlled portable near-infrared spectrometry and chemometrics. Food Chem 2022 Aug 1; 50: 384.

22.

Manley

. Near-infrared spectroscopy and hyperspectral imaging: non-destructive analysis of biological materials. Chem Soc Rev 2014; 43: 8200–8214.

23.

Bacon

Bennet

Hinton

, et al. Scanning electron microscopy of Fusarium moniliforme within asymptomatic corn kernels and kernels associated with Equine Leukoencephalomacia. Plant Dis 1992; 76(2): 144–148.

24.

Peiris

KHS

Pumphrey

Dowell

. NIR absorbance characteristics of deoxynivalenol and of sound and Fusarium - damaged wheat kernels. J Near Infrared Spectrosc 2009; 17(4): 213–221.

25.

Fernandes

Raimundo

IMJ

Pasquini

, et al. Simultaneous determination of methanol and ethanol in gasoline using NIR spectroscopy: Effect of gasoline composition. Talanta 2008; 75: 804–810.

26.

Muncan

Tsenkova

. Aquaphotomics — from innovative knowledge to integrative platform in science and technology. Molecules 2019; 24(15): 2742.

27.

Bands

Kabisch

Pollmer

. Hydrogen bonding in methanol—organic solvent and methanol—water mixtures as studied by the vco and voh. J Mol Struct 1982; 81(1–2): 35–50.

28.

Martens

Jensen

Geladi

. Multivariate linearity transformation for near - infrared reflectance spectrometry. In: Proceedings of the Nordic symposium on applied statistics, Stavnager, Norway” Stokkand Forlag Publishers, 1983, pp. 205–234.

29.

Barnes

Dhanoa

Lister

. Standard normal variate transformation and de-trending of near-infrared diffuse reflectance spectra. Appl Spectrosc 1989; 43(5): 772–777.

30.

Savitzky

Golay

MJE

. Smoothing and Differentiation. Anal Chem 1964; 36(8): 1627–1639.

31.

Hotelling

. Analysis of a complex of statistical variables into principal componentS. J Educ Psychol 1933; 24(417–41): 498–520.

32.

Brereton

Lloyd

. Partial least squares discriminant analysis: taking the magic away. J Chemom 2014; 28(4): 213–225.

33.

Kennard

Stone

. Computer aided design of experiments. Technometrics 1969; 11(1): 137–148.

34.

Tran

Lee

Buydens

LMC

, et al. Interpretation of variable importance in partial least squares with significance multivariate correlation (sMC). Chemom Intell Lab Syst [Internet] 2014; 138: 153–160. Available from: DOI: 10.1016/j.chemolab.2014.08.005.

35.

Martens

Naes

. Multivariate calibraion. Chichester, UK: John Wiley & Sons, 1992.

36.

Andersen

Bro

. Variable selection in regression — a tutorial. J Chemom 2010; 24(11–12): 728–737.