Article pubs.acs.org/JAFC
Differentiating Milk and Non-milk Proteins by UPLC Amino Acid Fingerprints Combined with Chemometric Data Analysis Techniques Weiying Lu,†,∥ Xiaxia Lv,†,∥ Boyan Gao,§ Haiming Shi,*,† and Liangli (Lucy) Yu*,§ †
Institute of Food and Nutraceutical Science, School of Agriculture and Biology, Shanghai Jiao Tong University, Shanghai 200240, China § Department of Nutrition and Food Science, University of Maryland, College Park, Maryland 20742, United States S Supporting Information *
ABSTRACT: Amino acid fingerprinting combined with chemometric data analysis was used to differentiate milk and non-milk proteins in this study. Microwave-assisted hydrolysis and ultraperformance liquid chromatography (UPLC) were used to obtain the amino acid fingerprints. Both univariate and multivariate chemometrics methods were applied for differentiation. The confidence boundary of amino acid concentration, principal component analysis (PCA), and partial least-squares-discriminant analysis (PLS-DA) of the amino acid fingerprints demonstrated that there were significant differences between milk proteins and inexpensive non-milk protein powders from other biological sources including whey, peanut, corn, soy, fish, egg yolk, beef extract, collagen, and cattle bone. The results indicate that the amino acid compositions with the chemometric techniques could be applied for the detection of potential protein adulterants in milk. KEYWORDS: protein differentiation, amino acid fingerprinting, univariate and multivariate data analyses
■
INTRODUCTION Milk is one of the most consumed foods worldwide, with an estimated annual global production of 179 million tons in 2010.1 Milk also has been a top subject for food fraud in the past three decades.2 One potential milk-related fraud is counterfeiting milk with inexpensive proteins such as soy protein. These non-milk proteins may alter the nutritional value of milk and induce possible allergic reactions or other serious health problems.3,4 Consequently, a differentiation technique for milk and non-milk proteins is needed for food safety and quality to ensure consumer well-being. Differentiation of proteins is challenging because they may exhibit similar physical and chemical properties. Several chromatographic and spectroscopic techniques combined with chemometric approaches have been examined for their potential in differentiating milk and non-milk proteins.5−9 For example, infrared (IR) spectroscopy was applied to detect soy milk adulteration in a mixed cow and buffalo milk combined with a principal component analysis (PCA) and partial leastsquares regression (PLSR) approach.5 IR was also reported in detecting whey-adulterated milk using a soft independent modeling of class analogy (SIMCA) and PLSR.6,7 Although IR is a convenient technique, recent studies have indicated that unknown chemical or instrumental variances could reduce its effectiveness in protein differentiation.10 Liquid chromatography−mass spectrometry (LC-MS) was used to differentiate the cow and water buffalo milk by comparing the mass of βlactoglobulin variants.8 Because the protein sequences of βlactoglobulin from bovine and bubaline are different, bovine and bubaline β-lactoglobulin have different molecular masses that can be used as unique mass spectral markers. However, the protein marker approach for LC-MS may not be suitable for detecting unknown adulterants because a specific standard is required for each target adulterant. Fingerprinting with © 2015 American Chemical Society
chemometric techniques may be applied to differentiate proteins of food without standards. For instance, ultraperformance liquid chromatography (UPLC) fingerprinting with chemometrics was reported to detect soy, pea, brown rice, and wheat proteins in skim milk powder.11 The protein fingerprints were measured without hydrolysis. The point-bypoint data analysis, which compared the total number of chromatographic points falling outside the 99% confidence bounds to the given threshold, and chemometrics methods (PCA and SIMCA) were applied to the protein UPLC fingerprints to build the differentiation model. This protein fingerprinting technique was able to differentiate skim milk powder from skim milk powder adulterated with soy protein isolate at 1% w/w. However, because the UPLC fingerprints for some other proteins such as brown rice protein had only one characteristic peak, the differentiation power of these protein fingerprints was very limited. Additionally, it was difficult to calibrate the peak intensity changes over time and between batches for the whole protein UPLC fingerprints. Consequently, novel approaches are required to achieve an accurate and effective protein differentiation. This study aimed to differentiate milk and non-milk proteins using a novel UPLC amino acid fingerprinting combined with chemometric analysis approach. Unlike previous methods that detect proteins directly,8,11 the proteins were first hydrolyzed to amino acids by microwave-assisted digestion with hydrochloric acid and analyzed for their amino acid fingerprints in this approach. Afterward, the confidence boundary using mole ratios of amino acids were constructed to differentiate milk Received: Revised: Accepted: Published: 3996
February 5, 2015 April 1, 2015 April 3, 2015 April 3, 2015 DOI: 10.1021/acs.jafc.5b00702 J. Agric. Food Chem. 2015, 63, 3996−4002
Article
Journal of Agricultural and Food Chemistry
quaternary solvent system including ATU solvent A (A), 10% (v/v) ATU solvent B (B), water (C), and ATU solvent B (D) was applied. The gradient elution program is listed in Table 1. The flow rate was
from non-milk protein samples. In addition, detailed differences between milk and non-milk proteins were demonstrated by PCA and partial least-squares-discriminant analysis (PLS-DA) using all 15 mole ratios of amino acids. This approach may provide a scientific foundation toward developing an effective nontargeted adulteration detection of foreign protein in milk and maybe other dairy products.
■
Table 1. UPLC Gradient Elution Program for Amino Acid Analysisa time (min)
A (%)
B (%)
C (%)
D (%)
gradient curve
0.00 7.10 7.30 7.69 7.99 8.59 9.00 12.00
10.0 8.0 8.0 7.8 4.0 4.0 10.0 10.0
0.0 15.6 15.6 0.0 0.0 0.0 0.0 0.0
90.0 57.9 57.9 70.9 36.3 36.3 90.0 90.0
0.0 18.5 18.5 21.3 59.7 59.7 0.0 0.0
11 7 6 6 6 6 6 6
MATERIALS AND METHODS
Materials and Chemicals. Raw milk samples were respectively collected in local dairy farms from Shandong province (192 samples) and Shanghai suburbs (171 samples) during May 2012−May 2013. Between 0.3 and 1 L of milk was collected for each sample. All milking cows were of the Holstein breed. The samples were freeze-dried once transferred to the laboratory, and the dry milk powders were stored in zip-lock bags at −20 °C. Each dried milk powder weighed between 30 and 130 g. The dried milk powders were mixed by shaking the zip-lock bags. Other non-milk protein powders were byproducts from foodprocessing factories. Whey protein was purchased from Aocheng Trading (Zhuhai, Guangdong, China). Peanut protein was provided by Changshou Food (Qingdao, Shandong, China). Corn 1, corn 2, soy, fish, egg yolk, beef extract, collagen, and cattle bone proteins were provided by Yuantai Biotech (Shanghai, China). Among the non-milk protein powders, six (whey, fish, egg yolk, beef extract, collagen, and cattle bone) were made from animals and the other four (corn 1, corn 2, soy, and peanut) were from plants. The infant formula powder purchased from a local grocery store was used as a reference sample of processed milk product. Water was purified using a Milli-Q Advantage A10 system (EMD Millipore, Darmstadt, Hesse, Germany). HPLC grade acetonitrile was purchased from Merck KGaA (Darmstadt, Hesse, Germany). The mixed amino acid standard AccQTag Ultra (ATU) eluent A and B and ATU derivatization kit were purchased from Waters Corp. (Milford, MA, USA). Other amino acid standards used in the recovery test were purchased from Sigma-Aldrich (St. Louis, MO, USA). Analytical grade hydrochloric acid was purchased from Linfeng Chemical Reagent (Shanghai, China). Other chemicals or solvents were of the highest commercial grade and used without further purification. Sample Preparation. Each dried milk powder (30 mg) was hydrolyzed directly with 10 mL of 6 M hydrochloric acid by a Discover SP-D pressurized microwave digestion system (CEM, Matthews, NC, USA). The power was 200 W. The ramp time and the cooling time were 15 and 20 min, respectively. The high prestirring was 15 s, with a temperature of 160 °C and a pressure at 250 psi. After hydrolysis, the samples were mixed for 15 s using a vortexer. Then, 0.4 mL hydrolysates were evaporated using a Buchi R-215 rotary evaporator (Flawil, St. Gallen, Switzerland) and redissolved in 0.5 mL of water. The solution was filtered through a 0.22 μm membrane filter (ANPEL Scientific Instrument, Shanghai, China). The Waters ATU derivatization kit was used for precolumn sample derivatization. The borate buffer (70 μL), the hydrolysates (10 μL), and the reconstituted reagent (20 μL) were sequentially added to a vial. The solution was mixed using a vortexer for 5 s and allowed to stand for 1 min at room temperature. Then, the solution was immediately transferred into an oven and heated at 55 °C for 10 min. The solution was stored at 4 °C and subjected to UPLC analysis. UPLC Analysis. The concentrations of 17 amino acids, including histidine, serine, arginine, glycine, aspartic acid, glutamic acid, threonine, alanine, proline, cysteine, lysine, tyrosine, methionine, valine, isoleucine, leucine, and phenylalanine, were quantified. Three amino acids, tryptophan, asparagine, and glutamine, were not quantified. Tryptophan decomposed, and asparagine and glutamine were respectively converted to aspartic and glutamic acids under acidic conditions during hydrolysis. Cysteine, due to its low concentration, was excluded from further data analyses. The hydrolysates were analyzed by using a Waters Acquity H-class UPLC system equipped with a photodiode array (PDA) detector. A Waters ATU C18 column (2.1 mm i.d. × 100 mm, 1.7 μm) was used. The column temperature was 43 °C, and the sample manager temperature was 20 °C. A
a
A, AccQTag Ultra (ATU) solvent A; B, 10% (v/v) ATU solvent B; C, water; D, ATU solvent B.
0.5 mL/min. The injection volume was 1.0 μL. The PDA detection wavelength was 260 nm. All samples were analyzed in a random order. Triplicate measurements were performed for each hydrolysate, and the mean concentrations were used for further statistical and chemometric analyses. The concentrations of calibration standards were 0.03, 0.05, 0.10, 0.25, 0.50, and 1.00 mmol/L for all amino acids. Univariate Data Analysis. Univariate analysis classifies milk and non-milk protein powders using a single variable. The chromatographic peak areas were integrated by Empower 3 software (Waters) using a traditional integration algorithm with peak width at 10 and threshold at 300. Manual identification of amino acid peaks was performed after integration. The results were reported as millimoles of each amino acid per gram of dry milk powder. To reduce undesired variances introduced by protein concentration variances between samples, sample moisture, preprocessing, and injection, etc., the relative mole ratios were applied. Mole ratios of individual amino acids to glutamic acid, the most abundant amino acid in milk protein, were calculated for each sample. The sample was classified as a non-milk protein if the relative mole ratio of an amino acid to glutamic acid was outside the 99.7% confidence boundaries of the milk mean value. The boundaries represent the mean ratio ± 3 times the standard deviation. Chemometric Data Analysis. The relative mole ratio data were processed with multivariate chemometric data analysis methods using a MATLAB R2013b (The MathWorks, Inc., Natick, MA, USA) inhouse program. PCA and PLS-DA were performed by the “svd” and “plsregress” (available in statistics toolbox) MATLAB functions, respectively. The data were preprocessed by autoscaling (meancentering and scaling to unit variance). Typically in PLS-DA, data are divided as training set for building the model and test set for quantitative validation of model. In this study, two-thirds of the samples were divided into a training set (249 samples), and the rest were put into an independent test set (124 samples). The test set included three non-milk samples, which were peanut, corn 2, and beef extract proteins. The milk and non-milk protein samples were coded “1” and “0” in the response variable, respectively.
■
RESULTS AND DISCUSSION Method Development. A milk sample was selected randomly for method development. The hydrolysis is an important step that may influence the accuracy and precision when the amino acid concentration is quantified. The ramp time and the container used for hydrolysis were optimized. Three ramp times including 15, 30, and 40 min were compared. There is no significant difference between the hydrolysis effectiveness between 15 and 30 min of reaction, whereas 40 min of hydrolysis resulted in significantly lower amino acid contents, indicating a possible decomposition of amino acids with longer heating. A 15 min ramp time for hydrolysis was 3997
DOI: 10.1021/acs.jafc.5b00702 J. Agric. Food Chem. 2015, 63, 3996−4002
Article
Journal of Agricultural and Food Chemistry
of the surrounding baseline noise. LOD and LOQ were 0.63 and 2.50 μmol/L for all amino acids, respectively. The linearity ranges were 0.01−1.00 mmol/L for all amino acids. For intraand interday precision evaluations, all samples were independently sampled and hydrolyzed. The intraday precision was calculated by four samples, and the interday precision was measured over two consecutive days with three samples each day. All intra- and interday variations were 80% except that of histidine and lysine (Table 2). UPLC Chromatograms. The representative UPLC amino acid chromatograms are plotted in Figure 1. The chromatograms were plotted sequentially toward the top-right corner for better demonstration. 6-Aminoquinoline (AMQ) was a byproduct of derivatization. All amino acid peaks were baseline separated and eluted within 9 min. Glutamic acid, proline, and leucine had the highest concentrations among the 16 quantified amino acids in the milk powders. The reported amino acid concentrations were in good agreement with that obtained from previous work using the conventional heating hydrolysis method.12 The tested non-milk proteins had unique profiles of amino acids. For instance, there was a significantly higher amount of glycine in the collagen. Table 3 lists the relative mole amino acid ratios of the milk and non-milk proteins. Each protein material had a characteristic set of amino acid compositions. For example, beef extract, cattle bone, and collagen had a high concentration of glycine; egg yolk had high ratios of aspartic acid and serine. The amino acid ratios were different for proteins from different biological sources, consistent with previous findings in other food products such as in rice wines and liquors.13,14 Due to the limited amount of non-milk proteins, these data alone may not be suitable for predicting other proteins outside the scope of this study. However, because these fingerprints exhibited significant dissimilarities from one another, it can be assumed that other
selected for further study. In addition, the 10 and 25 mL hydrolysis vessels in the microwave digestion instrument were compared. There was no significant difference between the analysis results using both vessels, and the 25 mL vessel was employed for subsequent analyses. The UPLC method was developed on the basis of the procedure recommended by Waters. The method was validated for linearity, repeatability, and stability. The correlation coefficients for all amino acids were >0.98 (Table 2). The Table 2. Calibration Parameters and UPLC Method Validation Resultsa
His Ser Arg Gly Asp Glu Thr Ala Pro Lys Tyr Met Val Ile Leu Phe
RT (min)
R2
intraday RSD (%)
interday RSD (%)
recovery (%)
2.67 3.87 4.10 4.29 4.70 5.23 5.66 6.12 6.73 7.69 7.86 7.98 8.05 8.65 8.73 8.80
0.997 0.997 0.998 0.997 0.988 0.989 0.997 0.990 0.997 0.984 0.997 0.998 0.996 0.997 0.997 0.998
7.6 6.4 6.2 6.4 5.5 5.5 6.2 6.7 6.4 6.6 6.5 23.0 5.8 6.2 6.2 6.6
7.6 4.5 6.9 5.5 25.3 5.0 5.9 5.5 4.7 22.6 5.6 19.4 6.6 8.8 6.1 10.7
74.9 84.6 112.8 135.8 89.9 86.6 97.3 104.4 93.3 72.8 96.9 93.7 104.3 92.0 96.4 98.2
a RT, retention time; R2, correlation coefficient; RSD, relative standard deviation. The calibrations were based on millimoles of each amino acid per gram of dry milk powder.
limit of detection (LOD) and limit of quantification (LOQ) were determined using a chromatogram of amino acid standard at 12.5 μmol/L by comparing the peak height with the height
Figure 1. Representative UPLC amino acid chromatograms: (A) amino acid standard (0.50 mmol/L); (B) milk protein; (C) collagen. AMQ, 6aminoquinoline; DP, derivatization peak. The chromatograms were plotted sequentially toward the top-right corner for better demonstration. 3998
DOI: 10.1021/acs.jafc.5b00702 J. Agric. Food Chem. 2015, 63, 3996−4002
Article
Journal of Agricultural and Food Chemistry Table 3. Mole Ratios of Amino Acids in Protein Powdersa milk n His Ser Arg Gly Asp Thr Ala Pro Lys Tyr Met Val Ile Leu Phe
363 0.09 0.34 0.12 0.17 0.41 0.23 0.24 0.58 0.35 0.18 0.07 0.32 0.20 0.47 0.20
± ± ± ± ± ± ± ± ± ± ± ± ± ± ±
0.06 0.17 0.08 0.10 0.06 0.14 0.04 0.16 0.09 0.12 0.07 0.10 0.07 0.15 0.15
IM
whey
EY
BE
CB
CO
fish
soy
P
C1
C2
1 0.09 0.34 0.11 0.18 0.43 0.22 0.24 0.59 0.37 0.15 0.01 0.32 0.19 0.47 0.16
1 0.08 0.45 0.11 0.24 0.71 0.48 0.55 0.47 0.60 0.12 0.10 0.35 0.32 0.66 0.14
1 0.16 0.93 0.46 0.51 0.92 0.48 0.80 0.44 0.66 0.26 0.13 0.58 0.36 0.78 0.30
1 0.21 0.59 0.41 2.51 0.64 0.45 1.10 1.02 0.49 0.29 0.19 0.63 0.36 0.80 0.43
1 0.29 0.67 0.87 7.30 0.59 0.45 1.67 2.54 0.26 0.20 0.15 0.58 0.31 0.68 0.46
1 0.16 0.95 1.56 10.23 0.60 0.57 2.50 3.38 0.31 0.10 0.21 0.37 0.14 0.50 0.44
1 0.06 0.29 0.48 4.05 0.53 0.10 1.59 1.56 0.32 0.05 0.00 0.28 0.13 0.37 0.16
1 0.12 0.43 0.35 0.47 0.64 0.26 0.39 0.38 0.37 0.15 0.02 0.27 0.20 0.43 0.21
1 0.11 0.42 0.57 0.63 0.65 0.16 0.37 0.32 0.20 0.17 0.01 0.26 0.15 0.39 0.24
1 0.17 0.47 0.35 0.67 0.59 0.30 0.63 0.60 0.31 0.18 0.05 0.42 0.25 0.58 0.26
1 0.16 0.48 0.40 0.88 0.37 0.40 1.32 1.01 0.25 0.10 0.17 0.61 0.26 0.73 0.22
a
IM, infant milk; EY, egg yolk; BE, beef extract; CB, cattle bone; CO, collagen; P, peanut; C1, corn 1; C2, corn 2; n, number of samples. Data were expressed as relative mole ratios to the glutamic acid. The data were reported as mean ± 3 times standard deviation for milk.
acid yielded poor abilities in distinguishing milk and non-milk proteins (Table 4). In particular, only one non-milk protein was differentiated according to that of tyrosine. To conclude, the confidence bounds calculated from the mole ratios of alanine, arginine, glycine, and aspartic acid to glutamic acid were able to differentiate milk and non-milk proteins with at least 90% correct identification. However, other amino acid ratios, including histidine, serine, threonine, proline, lysine, tyrosine, methionine, valine, isoleucine, leucine, and phenylalanine to glutamic acid, were unable to be used alone as differentiation criteria. The selection of amino acid ratios greatly influenced the classification results in the univariate analysis approach. Therefore, the multivariate techniques PCA and PLS-DA were used in an attempt to better classify samples as milk or nonmilk. The multivariate methods use latent variables constructed from combinations of the original variables (amino acid mole ratios) to extract additional information from the data set. PCA of Milk and Non-milk Proteins. PCA is a mathematical approach that is able to transform the highdimensional multivariate data into a lower number of dimensions. Figure 2 is the principal component (PC) scores and loading plots for milk and non-milk powders using all amino acid composition ratios. The first and second PCs that explain the top two largest percentages of total variances were plotted. The data points representing milk samples form a defined cluster, and the infant milk powder was inside this cluster (Figure 2A). Contrarily, data points representing nonmilk proteins were positioned relatively outside the cluster of milk samples, suggesting that the amino acid compositions of milk samples measured by this method were unique compared to the non-milk proteins. In comparison with the univariate method, PCA could avoid the selection of amino acid ratios and yielded a direct plot for differentiation. The amino acid fingerprints contain much quantitative information on amino acid composition, and PCA conveniently combines and presents all of this information altogether. Besides the differentiation of milk and non-milk proteins, the relationships between different non-milk proteins can also be studied by PCA to examine whether the sampling of the nonmilk proteins was unbiased. All non-milk proteins were located in the lower region of Figure 2A. Although the model was
unknown proteins outside this study may also have unique amino acid fingerprints. To establish objective, quantitative, and detailed criteria for differentiation, both univariate and multivariate data modelings were performed. Univariate Analysis of Milk and Non-milk Proteins. The confidence boundaries are reported in Table 3. The differentiation results by univariate analysis are reported in Table 4. Generally, confidence boundaries calculated from Table 4. Univariate Analysis Results of Milk and Non-milk Proteinsa milk (n = 363) His Ser Arg Gly Asp Thr Ala Pro Lys Tyr Met Val Ile Leu Phe
non-milk (n = 10)
pos
neg
pos
neg
353 352 352 352 361 352 358 353 357 352 358 351 355 351 351
10 11 11 11 2 11 5 10 6 11 5 12 8 12 12
6 4 9 9 9 6 10 7 5 1 4 4 4 5 3
4 6 1 1 1 4 0 3 5 9 6 6 6 5 7
a
n, total number of samples; pos, number of correctly differentiated samples; neg, number of incorrectly differentiated samples.
different amino acid ratios yielded significantly different correct identification rates for non-milk proteins. From the 15 confidence boundaries of amino acid ratios calculated, only the alanine to glutamic acid ratio correctly classifies all 10 nonmilk proteins (Table 4). The confidence boundaries calculated by the mole ratio of arginine, glycine, or aspartic acid to glutamic acid yielded 90% correct identifications for non-milk protein samples. On the contrary, mole ratios of tyrosine, methionine, valine, isoleucine, and phenylalanine to glutamic 3999
DOI: 10.1021/acs.jafc.5b00702 J. Agric. Food Chem. 2015, 63, 3996−4002
Article
Journal of Agricultural and Food Chemistry
Figure 2. Principal component scores (A, top) and loading (B, bottom) plots for all milk and non-milk proteins. IM, infant milk; C1, corn 1; C2, corn 2; P, peanut; EY, egg yolk; BE, beef extract; CB, cattle bone; CO, collagen.
calculated from only 10 foreign protein samples, these samples spanned a much larger region than 363 milk samples, suggesting a large diversity in the amino acid compositions from different sources of proteins. The diversities in non-milk protein powders indicated that the sampling of the non-milk proteins was unbiased. Among these samples, whey protein was the nearest to milk, probably because whey is a major protein in milk. The plant proteins including corn, soy, and peanut proteins were clustered closely, in comparison with the animal proteins including beef, cattle bone, collagen, and egg yolk, which spanned a wider area. The fish protein was not located in
either cluster. The tested animal proteins were generally farther away from the milk samples compared with the plant proteins. The relative importance of each amino acid ratio for differentiation could be obtained from the loading plot (Figure 2B). According to the loading of the second PC, the mole ratios of tyrosine to glutamic acid were typically high in the milk samples, whereas the ratios of arginine, proline, alanine, glycine, and aspartic acid to glutamic acid were high in the non-milk samples, especially the animal proteins. The conclusion was consistent with the data in Table 4, which reported at least 70% correct prediction by any one of these amino acid ratios. It was also consistent with a previous study that reported aspartic acid, 4000
DOI: 10.1021/acs.jafc.5b00702 J. Agric. Food Chem. 2015, 63, 3996−4002
Article
Journal of Agricultural and Food Chemistry
application. The multivariate approaches such as PCA and PLS-DA build robust and informative models, although they depend on relatively complicated and specialized dataprocessing techniques. Both PCA and PLS-DA approaches were able to differentiate all non-milk protein samples from milk protein samples using the amino acid fingerprints. PCA was able to achieve a graphical overview of the relationships between different groups of samples. In PLS-DA, with the additional separation of data into training and test sets, the differentiation could be performed automatically and quantitatively. In summary, amino acid fingerprinting combined with univariate or multivariate data analyses successfully differentiated milk protein from inexpensive non-milk protein powders. Amino acid fingerprints and chemometric techniques could be potentially developed as a nontargeted foreign protein detection technique for detecting milk adulteration.
proline, and alanine were marker amino acids to differentiate non-fat dry milk protein and whey protein.15 Compared with previous IR spectroscopic studies with multivariate analysis,6,7,10 spectroscopic methods are superior in analysis time and ease of application. However, the amino acid fingerprints furnish fewer variables and a compact data set, thus obtaining a simple differentiation model. Also, each coefficient in the amino acid fingerprint model can be clearly explained by the amounts of specific amino acid in the sample, in contrast with the IR spectroscopy where significant chemical variances appeared in different milk powders.10,16 Consequently, the amino acid model might be less influenced by instrumental, environmental, or other sources of variances. In summary, PCA of UPLC amino acid fingerprints is able to demonstrate the differences of all milk and non-milk proteins, as well as to investigate the detailed relationships among non-milk proteins according to the PC scores plot without the selection of any particular amino acid. PLS-DA of Milk and Non-milk Proteins. PLS-DA is a widely applied modeling technique.17,18 It was applied to quantitatively evaluate the multivariate differentiation approach. Figure 3 demonstrated the percent of variance explained in the
■
ASSOCIATED CONTENT
S Supporting Information *
PLS-DA model prediction table and plot. This material is available free of charge via the Internet at http://pubs.acs.org.
■
AUTHOR INFORMATION
Corresponding Authors
*(H.S.) Phone: (86) 21-3420-4041. Fax: (86) 21-3420-4107. E-mail:
[email protected]. *(L.Y.) Phone: (301) 405-0761. Fax: (301) 314-3313. E-mail:
[email protected]. Author Contributions ∥
W.L. and X.L. contributed equally to this work.
Funding
This work was supported by the National High Technology Research and Development Program of China (Grants 2013AA102202 and 2013AA102207); a special fund for Agro-scientific Research in the Public Interest (Grant 201203069); SJTU startup fund for young talent (Grant 13X100040047); and SJTU 985-III disciplines platform and talent fund (Grants TS0414115001 and TS0320215001).
Figure 3. PLS-DA-explained variance plot.
response variable of the training set with respect to PLS-DA models calculated by different numbers of latent variables. In this study, the minimum number of latent variables was determined to be the number required to reach >90% of variances explained in the response variable; meanwhile, the percent variances did not increase greatly with further addition of latent variables. From Figure 3, six latent variables that explained 91.1% of the variance in the response variable were chosen for PLS-DA modeling. All 124 samples in the test set were correctly differentiated as containing milk or non-milk proteins, consistent with the PCA result. Table S1 in the Supporting Information is the number of PLS-DA misclassifications in the test set with respect to different numbers of latent variables. It can be observed that the PLS-DA models built by 5−15 latent variables performed equally in prediction. The predicted and measured responses of the test set by PLS-DA with six latent variables are presented in Figure S1 in the Supporting Information. Comparison of the uni- and multivariate approaches shows that the univariate approach offered intuitive confidence ranges. However, different amino acid ratios yielded different differentiation capabilities. Consequently, careful and thorough research about the selection of amino acids in a larger sample set in univariate approaches is necessary before actual
Notes
The authors declare no competing financial interest.
■
REFERENCES
(1) World Food and Agriculture. FAO Statistical Yearbook 2013; Food and Agriculture Organization of the United Nations: Rome, Italy, 2013. (2) Moore, J. C.; Spink, J.; Lipp, M. Development and application of a database of food ingredient fraud and economically motivated adulteration from 1980 to 2010. J. Food Sci. 2012, 77, R118−R126. (3) Sanchez-Monge, R.; Lopez-Torrejón, G.; Pascual, C. Y.; Varela, J.; Martin-Esteban, M.; Salcedo, G. Vicilin and convicilin are potential major allergens from pea. Clin. Exp. Allergy 2004, 34, 1747−1753. (4) Caubet, J. C.; Ford, L. S.; Sickles, L.; Järvinen, K. M.; Sicherer, S. H.; Sampson, H. A.; Nowak-Węgrzyn, A. Clinical features and resolution of food protein-induced enterocolitis syndrome: 10-year experience. J. Allergy Clin. Immunol. 2014, 134, 382−389. (5) Jaiswal, P.; Jha, S. N.; Borah, A.; Gautam, A.; Grewal, M. K.; Jindal, G. Detection and quantification of soymilk in cow−buffalo milk using attenuated total reflectance Fourier transform infrared spectroscopy (ATR−FTIR). Food Chem. 2015, 168, 41−47. (6) Santos, P. M.; Pereira-Filho, E. R.; Rodriguez-Saona, L. E. Rapid detection and quantification of milk adulteration using infrared microspectroscopy and chemometrics analysis. Food Chem. 2013, 138, 19−24. 4001
DOI: 10.1021/acs.jafc.5b00702 J. Agric. Food Chem. 2015, 63, 3996−4002
Article
Journal of Agricultural and Food Chemistry (7) Santos, P. M.; Pereira-Filho, E. R.; Rodriguez-Saona, L. E. Application of hand-held and portable infrared spectrometers in bovine milk analysis. J. Agric. Food Chem. 2013, 61, 1205−1211. (8) Czerwenka, C.; Mű ller, L.; Lindner, W. Detection of the adulteration of water buffalo milk and mozzarella with cow’s milk by liquid chromatography−mass spectrometry analysis of β-lactoglobulin variants. Food Chem. 2010, 122, 901−908. (9) Scholl, P. F.; Farris, S. M.; Mossoba, M. M. Rapid turbidimetric detection of milk powder adulteration with plant proteins. J. Agric. Food Chem. 2014, 62, 1498−1505. (10) Botros, L. L.; Jablonski, J.; Chang, C.; Bergana, M. M.; Wehling, P.; Harnly, J. M.; Downey, G.; Harrington, P.; Potts, A. R.; Moore, J. C. Exploring authentic skim and nonfat dry milk powder variance for the development of nontargeted adulterant detection methods using near-infrared spectroscopy and chemometrics. J. Agric. Food Chem. 2013, 61, 9810−9818. (11) Jablonski, J. E.; Moore, J. C.; Harnly, J. M. Nontargeted detection of adulteration of skim milk powder with foreign proteins using UHPLC−UV. J. Agric. Food Chem. 2014, 62, 5198−5206. (12) Marino, R.; Iammarino, M.; Santillo, A.; Muscarella, M.; Caroprese, M.; Albenzio, M. Technical note: Rapid method for determination of amino acids in milk. J. Dairy Sci. 2010, 93, 2367− 2370. (13) Shen, F.; Ying, Y.; Li, B.; Zheng, Y.; Zhuge, Q. Multivariate classification of rice wines according to ageing time and brand based on amino acid profiles. Food Chem. 2011, 129, 565−569. (14) Xiao, X.; Hou, Y.; Liu, Y.; Liu, Y.; Zhao, H.; Dong, L.; Du, J.; Wang, Y.; Bai, G.; Luo, G. Classification and analysis of corn steep liquor by UPLC/Q-TOF MS and HPLC. Talanta 2013, 107, 344− 348. (15) Greenberg, R.; Dower, H. J. Detection of added whey protein concentrate in nonfat dry milk by amino acid analysis. J. Agric. Food Chem. 1986, 34, 30−32. (16) Harnly, J. M.; Harrington, P. B.; Botros, L. L.; Jablonski, J.; Chang, C.; Bergana, M. M.; Wehling, P.; Downey, G.; Potts, A. R.; Moore, J. C. Characterization of near-infrared spectral variance in the authentication of skim and nonfat dry milk powder collection using ANOVA-PCA, pooled-ANOVA, and partial least-squares regression. J. Agric. Food Chem. 2014, 62, 8060−8067. (17) Ellis, D. I.; Brewster, V. L.; Dunn, W. B.; Allwood, J. W.; Golovanov, A. P.; Goodacre, R. Fingerprinting food: current technologies for the detection of food adulteration and contamination. Chem. Soc. Rev. 2012, 41, 5706−5727. (18) Lu, W.; Jiang, Q.; Shi, H.; Niu, Y.; Gao, B.; Yu, L. Partial leastsquares-discriminant analysis differentiating Chinese wolfberries by UPLC−MS and flow injection mass spectrometric (FIMS) fingerprints. J. Agric. Food Chem. 2014, 62, 9073−9080.
4002
DOI: 10.1021/acs.jafc.5b00702 J. Agric. Food Chem. 2015, 63, 3996−4002
