2020 Volume 26 Issue 3 Pages 389-397
This study aimed to develop a method for the detection of food residues on the surface of stainless steel plates using fluorescence fingerprint (FF). Extracts of 20 food products were dropped on stainless steel plates as a model of food residues. Fluorescence fingerprint measurement of the food residues was carried out; the food residue samples were subsequently collected by swabbing the surface to determine adenosine triphosphate (ATP) luminescence. Partial least squares regression (PLSR) models were constructed to predict ATP luminescence (R2 = 0.60; RMSE = 1.57×105 RLU) and solid content (R2 = 0.46; RMSE = 0.89×10−4 g) on stainless steel surfaces from the FF data. From the coefficients of the prediction model, NADH and NADPH showed the greatest contribution to the prediction of solid content.
The cleanliness of food contact surfaces is one of the major concerns in the food industry. The accumulation of food residues on food contact surfaces is a potential source of microbial contamination (Silva et al., 2008) and can lead to food safety issues. Thus, the detection of food residues on surfaces before the growth of microorganisms is important to evaluate the cleanliness of surfaces.
Adenosine triphosphate (ATP) luminescence monitoring has been widely used to quantitatively evaluate the cleanliness of food contact surfaces, and utilizes bioluminescence produced by firefly (Photinus pyralis) luciferase through oxidative decarboxylation of luciferin in the presence of ATP (Hawronskyj and Holah, 1997). This method is simple and rapid, allowing real-time measurement of the cleanliness of surfaces in the food industry. However, ATP content does not indicate the amount of food residue, since ATP content varies among food products. Moreover, sampling methods involving swabbing require training of workers or experimenters. Therefore, alternative methods are needed to determine food residues, taking into account the extensive properties of foods, and minimize errors due to human handling.
Various studies have been carried out to measure food residues on food contact surfaces. For rapid in situ measurement, UV irradiation has been used to detect residual cells and soiling on stainless steel plates (Whitehead et al., 2008), aluminum and fiber re-enforced plastic (Abban et al., 2014). This method could detect several food constituents at various concentrations (0.0001–1%) (Whitehead et al., 2008); however, it could not quantify food residues.
On the other hand, harmful substances or bacteria related to food safety have been measured by various methods employing light spectrum technologies such as fluorescence. Imaging of nicotinamide adenine dinucleotide (NADH) and riboflavin on biofilms of E. coli O157:H7 and Salmonella on several handling surface materials was performed based on autofluorescence (Jun et al., 2009; Jun et al., 2010). Detection systems using fluorescence have been developed to measure the efficacy of sanitation procedures (Lefcourt et al., 2013; Wiederoder et al., 2012; Wiederoder et al., 2013).
Fluorescence fingerprint (FF), which is also known as an excitation-emission matrix, has been used to discriminate and quantify the constituents of different food products (ElMasry et al., 2015; ElMasry et al., 2016; Fujita et al., 2013; Kokawa et al., 2011, 2012; Kokawa et al., 2013; Shibata et al., 2018; Shibata et al., 2011). Fluorescence fingerprint comprises fluorescence intensities with comprehensive excitation and emission wavelengths, forming three-dimensional data of excitation wavelength×emission wavelength×fluorescence intensities. As FF contains vast amounts of fluorescence information compared with conventional fluorescence measurement performed at a single excitation wavelength condition (Jun et al., 2009), it could be more effective for detecting various food residues. In addition, considering that ATP is autofluorescent and FF records vast amounts of additional fluorescence information, measurement of different food residues could be achieved by FF, whereas previous methods are limited in their application to specific food products. The objective of this study is to detect food-residues on stainless steel plates by FF analysis.
Sample preparation Twenty kinds of food products, purchased at a supermarket in Tokyo, were selected from four food categories: 1) agricultural products (tomato, bean sprout, potato, lettuce, apple, mandarin orange, grapefruit and pineapple), 2) beverages (three kinds of commercial beer, red tea and green tea), 3) marine products (squid, shrimp, mackerel, tuna and scallop) and 4) livestock products (beef and chicken). One hundred grams of each food was homogenized with 100 mL of deionized water using a blender (TK440, TESCOM Denki Co., Ltd., Tokyo, Japan) for 2 minutes. Then, the homogenized sample was centrifuged at 1 000 G for 15 minutes at 20 °C (MX-305, TOMY SEIKO CO., LTD., Tokyo, Japan) to obtain the supernatant. This purification step was repeated three times to produce food extract samples diluted 1 (undiluted), 2, 4 and 8 times with deionized water. Then, 10 µL of the food extract samples were dropped on 50 mm × 50 mm stainless steel SUS304 plates. The diameter of extract droplets was approximately 1.6 mm. Prior to use, the stainless steel plates were sterilized by autoclaving at 121 °C. After the extract samples were added to the plates, they were dried at 25 °C for 30 minutes in an incubator (LTI-600SD, TOKYO RIKAKIKAI CO., LTD., Tokyo, Japan). The dried food extracts on the stainless steel surfaces were used as a model of food residues on food contact surfaces.
Solid content The solid content of food residue samples was measured using an oven drying method. The food extract samples (each sample weighed approximately 1 g) were placed in weighing bottles and then dried at 110 °C for 24 h. Next, the weighing bottles containing the dried samples were cooled in a desiccator. Finally, the solid content of the extract samples was calculated from the change in weight before and after drying.
Measurement of fluorescence fingerprint Fluorescence fingerprint measurement of the food residues on stainless steel plates was carried out using a fluorescence spectrophotometer (F7000, Hitachi High-technologies Corporation, Tokyo, Japan) equipped with a 150 W Xe arc lamp. As illustrated in Fig. 1, the excitation light emitted from the light source was at a 45° angle of incidence to the food residue on the stainless steel plate, and the emission light or scattering light was measured by the detector at a 45° angle from the plate. We ensured that all food residues on the plates were covered by the incident excitation light and all measurements were performed inside the spectrophotometer under a completely dark environment. Furthermore, the slit widths of both excitation and emission wavelengths were fixed at 10 nm. The measurement ranges for the excitation and emission wavelengths were 200–500 nm and 200–700 nm, respectively, with a wavelength increment of 10 nm. The wavelength scanning speed and photomultiplier voltage were set to 60 000 nm/min and 400 V. The fluorescence spectra, including scattered light, were collected by fixing the wavelength of the excitation light and scanning the wavelength of the light emitted from the food residue samples.
Schematic views of fluorescence measurement of food residue on a stainless steel plate
ATP luminescence of food residues on stainless steel plates After FF measurements, food residues on the surface of stainless plates were collected by swabbing the surface using a swabbing kit (Lucipack pen, Kikkoman Biochemifa Company, Tokyo, Japan). Then, ATP luminescence, which indicates the amount of ATP in the food residue, was quantified using a lumitester PD-30 (Kikkoman Biochemifa Company). This system utilizes the principle that chemical luminescence occurs when ATP reacts with luciferin, luciferase, and magnesium ions and is also capable of measuring adenosine monophosphate (AMP) content. Thus, the measured intensity of luminescence was attributed to ATP and AMP generated by ATP degradation. In this study, we term the intensity as ATP luminescence, which was expressed as relative light units (RLU).
Preprocessing of FF data Preprocessing of FF data was carried out to exclude non-fluorescent data based on previous studies (Fujita et al., 2013; Shibata et al., 2011; Yoshimura et al., 2014). Since fluorescence is the emission of light with a longer wavelength than excitation light, data where the emission wavelength was shorter than the excitation wavelength were removed. Furthermore, the ridge of high intensity in the FF contour map, where the excitation wavelength was equal to the emission wavelength, represents the intensity of scattered light, and ridges extending from 400 and 600 nm on the emission axis are second- and third-order light, respectively. Those data generated by light scattering from the surface of the diffraction grating were not fluorescence and were, therefore, omitted. The omitted data width was as follows: first-order light, ± 30 nm, second-order light, ± 30 nm from the excitation wavelength. Finally, a total of 717 fluorescence intensity values were acquired with the combination of excitation and emission wavelengths from one sample, this total being set as independent variables for subsequent statistical analyses.
Statistical analysis Initially, the FF dataset was divided into a calibration dataset of 176 FFs to construct statistical models and a validation dataset of 89 FFs to evaluate the PLSR models. The division of calibration and validation sets was determined so that the samples were distributed equally in both datasets. PLSR models were constructed by setting the FF data of the calibration dataset as independent variables and the ATP luminescence and the solid contents after normalization as the dependent variables. The number of latent variables in the models was determined by leave-one-out cross-validation of the calibration dataset. The models were then applied to the validation dataset to confirm their accuracy. Evaluation of fitting of the calibration models to the calibration and validation datasets was performed using coefficient of determination (R2) and root mean squared error (RMSE).
Figure 2A–C shows the FF contour maps of food residue samples from the four categories. Scattering and non-fluorescent light were removed by preprocessing as described above. In the FFs of most agricultural products in Fig. 2A, an intense peak was observed at Ex. 330 nm and Em. 400 nm, which could be assigned to NADH (Cantor and Schimmel, 1980; Leblanc and Dufour, 2002; Rahman et al., 2019). The peak was also observed in other samples in Fig. 2B and C, respectively. Further, the FFs of marine and livestock samples showed strong fluorescence ranging from Ex. 250 to 330 nm and Ex. 300 to 400 nm. The fluorescence might be caused by aromatic amino acids (Trivittayasil et al., 2014) and nucleic acids such as ATP, ADP (adenosine diphosphate) and AMP (Rahman et al., 2019). A previous study indicated that the intensity from nucleic acids was weaker than that from tryptophan (Cantor and Schimmel, 1980); thus, the signal appears to be mainly attributable to that of tryptophan.
FF contour maps of agricultural product samples
Ex and Em denote excitation and emission wavelengths, respectively. All samples are undiluted ones.
FF contour maps of beverage samples
Ex and Em denote excitation and emission wavelengths, respectively. All samples are undiluted ones.
FF contour maps of marine and livestock product samples
Ex and Em denote excitation and emission wavelengths, respectively. All samples are undiluted ones.
Figure 3 shows the prediction of ATP luminescence by FF. The model was constructed with 5 latent variables, and the R2 for calibration and validation were 0.61 and 0.60, with RMSE of 1.48×105 and 1.57×105 RLU, respectively. The model tended to underestimate ATP luminescence in the higher ATP luminescence region.
Prediction of ATP luminescence by FF
To show the contribution of each fluorescence to the model, regression coefficients of the model were calculated from the normalized variables. Figure 4 shows the regression coefficient of each variable in the PLS model. The regression coefficient values were plotted as a contour against excitation and emission wavelengths. Wavelength conditions with higher absolute values contributed more to the PLS model. The coefficient in Ex. 230 nm and Em. 300 nm showed the highest value (0.343) and those in the conditions near the wavelength presented relatively high coefficients, which indicates a greater contribution to the estimation of ATP luminescence value in the model. The wavelength condition of Ex. 230 nm and Em. 300 nm is attributed to tyrosine (Zhang et al., 2013) and tryptophan (Trivittayasil et al., 2014). However, coefficients of the conditions attributed to ATP, whose peak intensity was Ex. 290 and Em. 380 (Rahman et al., 2019), showed a smaller value. This indicates that ATP showed a slight contribution to the model.
Regression coefficients of ATP luminescence prediction model
ATP luminescence is not sufficient for evaluating cleanliness or food residues because it varies among food products. In this study, some samples showed high solid content with lower ATP luminescence and vice versa. As illustrated in Fig. 5, ATP luminescence showed a low correlation with solid content, namely the amount of food residue. For example, undiluted samples of pineapple and grapefruit showed low luminescence with high solid content (1.2×105 RLU and 5.0×10−4 g, respectively), whereas an undiluted tomato sample showed high luminescence with low solid content (9.0×105 RLU and 8.0×10−5 g, respectively). Since not all substances emit fluorescence, it would be difficult to estimate the amount of food residue by ATP luminescence alone.
Relationships between ATP luminescence and solid content
Figure 6 shows the prediction of solid content on stainless steel surfaces by FF. The model was constructed with 6 latent variables, and showed R2 of 0.50 and 0.46, and RMSE of 0.81×10−4 and 0.89×10−4 g for the calibration and validation datasets, respectively.
Prediction of solid content by FF
Figure 7 shows the regression coefficient of each variable in the PLS model for the prediction of solid content. The contour map of the regression coefficients was drawn as for the previous ATP luminescence model. A prominent peak was located in the area of Ex. 310 to 370 nm and Em. 400 to 440 nm, which is attributed to NADH and NADPH. On the other hand, the coefficients corresponding to ATP fluorescence were smaller than those of NADH and NADPH. Therefore, it was concluded that fluorescence from NADH and NADPH had the greatest positive contribution to the model for the prediction of solid content.
Regression coefficients of solid content prediction model
In this study, NADH and NADPH showed the greatest contribution to the prediction model of solid content, whereas ATP fluorescence did not show a correlation with solid content or even ATP luminescence.
Possible reasons for this are as follows: 1) The effects of sample properties, and 2) masking of ATP fluorescence by aromatic amino acids. First, of the 20 samples, 13 were derived from agricultural products and beverages, which emitted intense NADH fluorescence. Table 1 shows the Cook's distance (D) (Neter et al., 1996) of the samples that exceeded by three-fold the mean of the calibration dataset for the prediction of solid content. Cook's D indicates the degree of influence of each sample on the model. The results indicated that several tomato and beer samples were found to be dominant in the model construction, while one scallop and one lettuce sample had high Ds. The tomato and beer samples tended to show intense fluorescence of NADH and NADPH; thus, their properties had a strong effect on the prediction model for solid content.
| Sample | Dilusion ratio | Cook's Distance (× 10−3) |
|---|---|---|
| Beer A | 2 | 19.1 |
| Beer B | 1 | 17.6 |
| Beer C | 1 | 3.68 |
| Beer C | 2 | 7.60 |
| Scallop | 1 | 7.97 |
| Tomato | 1 | 7.77 |
| Tomato | 1 | 17.2 |
| Tomato | 2 | 14.3 |
| Tomato | 2 | 15.8 |
| Lettuce | 1 | 4.36 |
| Sample mean | N/A | 1.17 |
The samples shown had more than three times of mean of the calibration samples' values.
Second, the peak wavelength conditions of ATP overlap those of aromatic amino acids (tyrosine and tryptophan), which have greater fluorescence compared to ATP (Lakowicz, 2006), leading to masking of ATP fluorescence by aromatic amino acids. Thus, the fluorescence of ATP was not proportional to the amount of ATP in this study.
Notably, the fitting of models was considered insufficient. Generally, the main ingredients of foods are carbohydrates, proteins and fats. Although proteins could be determined from the fluorescence of tryptophan, and fat was removed during sample preparation, carbohydrates were not measured by FF, as they do not show autofluorescence. Recently, several carbohydrates such as total soluble sugars and starch have been measured using near-infrared (NIR) spectroscopy (Shetty et al., 2012; Lohr et al., 2017). Thus, prediction of the amount of food residue might be improved by combining NIR spectral analysis with FF measurement. Moreover, application of a nonlinear model may improve the prediction between FFs and the responses (ATP luminescence and solid content).
In this study, we described the quantification of food residues as possible hazardous substances using FF. The advantages of our FF method are that it is capable of generally predicting the amount of food and could have greater sensitivity with a more sensitive detector or stronger light source. On the other hand, water-soluble substances were mainly extracted and used for the experiment. Thus, this method is applicable to food constituents from water-soluble substances, which also show autofluorescence. Furthermore, non-fluorescent substances such as oil are difficult to detect using this method. However, measurements could be conducted in cases where autofluorescent pigments are present (Sikorska et al., 2004).
Fluorescence fingerprint measurement allowed the detection of vast characteristic fluorescence information from food residues. Food residues from 20 kinds of food products on stainless steel plates were measured by FF. The prediction models developed from the FFs and PLSR showed tendency of an underestimation of both ATP luminescence and solid content. Although ATP showed a slight contribution to the ATP prediction model, NADH and NADPH showed the greatest contribution to the prediction model of solid content.
Acknowledgements This work was supported by a Grant-in-Aid for Scientific Research (C), Grant Number 18K05900.
fluorescence fingerprint
ATPadenosine triphosphate
AMPadenosine monophosphate
NADHnicotinamide adenine dinucleotide
