Article Cite This: J. Agric. Food Chem. 2018, 66, 8478−8488
pubs.acs.org/JAFC
Variance of Commercial Powdered Milks Analyzed by Proton Nuclear Magnetic Resonance and Impact on Detection of Adulterants James Harnly,*,† Marti Mamula Bergana,‡ Kristie M. Adams,‡ Zhuohong Xie,‡ and Jeffrey C. Moore‡ †
Downloaded via TUFTS UNIV on August 19, 2018 at 09:50:11 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
Food Composition and Methods Development Laboratory, Beltsville Human Nutrition Research Center, Agricultural Research Service, United States Department of Agriculture, Building 161, BARC-East, Beltsville, Maryland 20705, United States ‡ United States Pharmacopeia, 12601 Twinbrook Parkway, Rockville, Maryland 20852, United States ABSTRACT: Proton nuclear magnetic resonance spectra for 66 commercial powdered milk samples were analyzed by principal component analysis, soft independent modeling of class analogy, and pooled, crossed analysis of variance. It was found that the sample type (skim milk powder or non-fat dry milk), the supplier, the production site, the processing temperature (high, medium, or low temperature), and the day of analysis provided statistically significant sources of variation. Interestingly, inexact alignment (deviations of ±0.002 ppm) of the spectral reference peak was a significant source of variation, and fine alignment was necessary before the variation arising from the other experimental factors could be accurately evaluated. Using non-targeted analysis, the lowest detectable adulteration for dicyandiamide, melamine, and sucrose was 0.05%, the lowest detectable adulteration for maltodextrin and urea was 0.5%, the lowest detectable adulteration for ammonium sulfate and whey was 5%, and the lowest adulteration for soy protein isolate was undetectable using methods described herein. The measurement of variance and detection of adulteration were relatively unaffected by the resolution. Similar results were obtained with unbinned data (0.0003 ppm resolution) and binning of 333 data points (0.1 ppm resolution). KEYWORDS: NMR, milk powder, chemometrics, ANOVA, adulteration, binning
■
INTRODUCTION United States Pharmacopeia (USP) initiated a project in 2011 to develop strategies for the detection of adulterants in foods using powdered milk samples as a model.1 Initial efforts focused on characterizing commercial powdered milks in hopes of establishing a global reference material and developing a library of suitable methods. These studies resulted in publication of results using near-infrared (NIR) spectrometry,2,3 highperformance liquid chromatography with ultraviolet detection (HPLC−UV),4 Raman spectroscopy,5 and non-protein nitrogen.6 On the basis of these studies, USP developed guidance for the development and validation of non-targeted methods for detection of adulteration.7 Research on the adulteration of powdered milk samples expanded with the development of a series of skim milk powders adulterated with eight different compounds. The initial collection of commercial samples and the new adulterated samples was analyzed by proton nuclear magnetic resonance (1H NMR) spectroscopy.8 1 H NMR has several advantages with respect to authentication of foods and botanicals.9 First, it detects all of the sample components simultaneously, thus providing an inherently comprehensive survey of the sample metabolome. Second, the acquired intensities of the proton spectra are extremely reproducible, offering the possibility of using archived spectra for authentication rather than constantly reanalyzing (and using up) valuable reference materials. A potential disadvantage of 1H NMR is its lack of sensitivity compared to mass spectrometry (MS).9 However, this is not a major obstacle with respect to economically motivated adulteration because significant levels of added adulterants are generally needed to achieve an economic advantage. © 2018 American Chemical Society
NMR has been previously applied to the detection of a variety of compounds in milk.10−15 Lachenmeier was the first to specifically measure melamine. Most approaches were interested in specific compounds. Sundekilde et al.15 were the first to take a metabolomics approach. While these NMR studies were termed non-targeted as a result of their acquisition of data for all components in a sample, none used non-targeted chemometric approaches to process the data. Harnly et al. showed that spectral fingerprints acquired by infrared (IR), MS, NIR, NMR, and ultraviolet (UV) can be used with chemometric methods, specifically principal component analysis (PCA), for identification and authentication of foods and botanical supplements.3,16−18 In combination with sample metadata, pooled, crossed analysis of variance (pcANOVA) can be used to determine the fraction of variance associated with each experimental factor.18,19 Authentication is best approached using a one-class model [i.e., only authentic or reference samples are modeled, the first step in soft independent modeling of class analogy (SIMCA)] and establishing selected statistical limits for determining if an unknown sample is similar to or different from the reference samples.20 Because any spectral feature that is different from the average features of the reference samples will result in the unknown sample being deemed adulterated, this chemometrics approach is truly non-targeted. The collection of commercial powdered milk samples (skim milk powder and non-fat dry milk) analyzed by 1H NMR in the Received: Revised: Accepted: Published: 8478
January 24, 2018 April 19, 2018 April 26, 2018 April 26, 2018 DOI: 10.1021/acs.jafc.8b00432 J. Agric. Food Chem. 2018, 66, 8478−8488
Article
Journal of Agricultural and Food Chemistry Table 1. Samples Analyzed sample code
supplier, production locationa
milk powder class
S021 S022 S023 S024 S030 S031 S032 S033 S047 S051 S053 S054 S055 S061 S068 S070 S076 S077 S080 S081 S082 S084 S085 S086
B, 1 B, 1 B, 1 B, 1 B, 2 B, 2 B, 2 B, 2 B, 1 B, 1 B, 1 B, 1 B, 1 C, 3 C, 3 C, 3 B, 2 B, 1 A, 4 A, 5 A, 5 A, 4 A, 4 A, 5
N S N N N N N N N N N N N N N N N S S N N S S N
a
heat process temperature
production date (month/day/year)
sample code
LH
07/12/10 02/27/10 05/05/10 05/05/10 07/18/10 11/16/09 06/19/10 02/26/10 06/07/10
S087 S089 S091 S093 S094 S095 S096 S097 S098 S106 S107 S108 S110 S116 S117 S136 S145 S147 S149 S154 S157 S170 S171 S172
MH MH HH HH LH LH LH LH LH LH LH LH LH LH HH LH LH HH LH MH MH HH
08/26/08 03/08/11 02/21/11 02/07/11 01/14/11 02/21/11 03/27/11 03/09/11 02/27/11 01/30/11 03/08/11 01/15/11
supplier, production location
milk powder class
A, 5 A, 6 A, 6 B, 1 B, 1 D, 7 A, 8 A, 8 A, 8 E, 9 E, 9 F F G, 10 G, 10 H, 12 I, 11 I, 11 I, 11 J, 13 J, 13
N N N N N S S S S S S N N S S S N N N N N N N N
K, 14
heat process temperature
production date (month/day/year)
LH MH MH LH MH MH MH LH LH MH MH
03/07/11 03/12/11 12/26/10 02/01/11 02/13/11 10/20/10 02/08/11 03/12/11 08/29/11 08/17/10 01/15/10
MH MH
04/02/11 03/15/11 11/01/11 05/12/12 07/05/12 05/15/12 07/06/12 07/06/12
LH LH HH LH LH LH LH LH
Samples were collected from 11 suppliers (A−K) with 14 processing locations (1−14).
Table 2. Spiked Adulterants of Sample S097 whey
urea
sucrose
soy protein isolate
melamine
maltodextrin
dicyandiamide
ammonium sulfate
S097 whey
S097 urea
S097 sucrose
S097 SPI
S097 mel
S097 malto
S097 DCD
S097 AS
urea_ref
sucrose_ref
SPI_ref
mel_ref
malto_ref
DCD_ref
AS_ref
urea urea urea urea urea urea
sucrose sucrose sucrose sucrose sucrose sucrose
SPI SPI SPI SPI SPI SPI
mel mel mel mel mel mel
malto malto malto malto malto malto
DCD DCD DCD DCD DCD DCD
AS AS AS AS AS AS
sample adulterant whey_ref sample + adulterant 5.000% whey 5 1.000% whey 1 0.500% whey 05 0.050% whey 005 0.010% whey 001 0.005% whey 0005
5 1 05 005 001 0005
5 1 05 005 001 0005
5 1 05 005 001 0005
current study was previously analyzed by NIR. 2,3 In combination with chemometric methods, the NIR studies demonstrated that compound identification was not needed to detect differences in the chemical composition of samples.2,3 NIR spectra provide harmonic bands for vibrational transitions and, thus, do not identify specific compounds or functional groups. Pooled (summed for the entire spectrum) and crossed (resorted for each factor) analysis of variance (ANOVA) of the NIR data showed that the sample type (skim milk powder or non-fat dry milk), the supplier, the production site, the processing temperature (high, medium, or low temperature), and the day of analysis provided statistically significant sources of variation. In addition, the spectra correlated with residual fat, moisture, and protein contents.3 The results of the NIR study showed that a global reference standard consisted of variance from each of the sources listed above and that individual production sites were better able to detect variations in their
5 1 05 005 001 0005
5 1 05 005 001 0005
5 1 05 005 001 0005
5 1 05 005 001 0005
product or adulteration by employing a series of in-house reference materials. In this study, the 1H NMR data are analyzed by PCA, oneclass modeling, and pcANOVA to characterize the variance of the reference samples, i.e., to determine the fraction of variance from each experimental factor (type of milk powder, supplier, production site, processing temperature, and day of analysis). Newly prepared samples with systematically increasing concentrations of eight adulterants were analyzed to determine the level of detection using non-targeted 1H NMR. In addition, the impact of resolution on variance and detection of adulterants was determined using variable bin sizes.
■
MATERIALS AND METHODS
Standards and Solvents. Details were previously described.8 Briefly, ampules of 99.8% deuterium dimethyl disulfide (DMSO-d6) containing 0.05% (v/v) tetramethylsilane (TMS) were purchased from Cambridge Isotope Laboratories (Tewksbury, MA, U.S.A.). Sucrose, 8479
DOI: 10.1021/acs.jafc.8b00432 J. Agric. Food Chem. 2018, 66, 8478−8488
Article
Journal of Agricultural and Food Chemistry dicyandiamide, urea, melamine, maltodextrin, and ammonium sulfate were reagent-grade and commercially purchased from Sigma-Aldrich (St. Louis, MO, U.S.A.). Whey protein concentrate and soy protein isolate were obtained from industrial ingredient suppliers. Milk Powder Samples. Samples were acquired from commercial samples all over the world as previously described.2,3 The metadata is summarized in Table 1. Briefly, there were two sample types, skim milk powder (SMP) and non-fat dry milk (NFDM), three processing temperatures, high (H), medium (M), and low (L), 11 suppliers (national and international), and 15 production sites, and samples were analyzed over a period of 7 days.2 Milk Powder Solution Preparation. Samples were prepared for NMR analysis by weighing 6.0 ± 0.1 mg of milk powder (MP) and dissolving in 650 μL of DMSO-d6 in an Eppendorf tube (powders did not fully dissolve).8 This mixture was shaken at 2800 rpm on a touchsensitive miniRoto vortex mixer [Thermo Fisher Scientific, Waltham, MA, U.S.A. (formerly Fischer Scientific, Pittsburgh, PA, U.S.A.)] for 1 min. After ensuring that no dry powder remained, centrifugation at 2000g for 10 min was applied using a C1301 centrifuge (Labnet, Edison, NJ, U.S.A.), and the resulting clear, colorless supernatant liquid was decanted into a clean Eppendorf tube. Finally, 600 μL of each supernate was transferred via an autopipette into a high-quality, 5 mm NMR tube for analysis. Adulterant-Spiked Milk Powder Solution Preparation. As previously described,8 adulterant, synthetic mixtures were prepared using replicate solution preparations (per the Milk Powder Solution Preparation section) of sample S097 and two adulterant spiking solutions: a concentrated adulterant stock solution (CAS, 10 mg/mL) and a dilute adulterant stock solution (DAS, 0.1 mg/mL) (Table 2).8 The CAS was prepared similar to the milk powder samples (all powders were not fully dissolved) using 600 μL of DMSO-d6. The DAS was prepared by dissolving 6 μL of CAS in 594 μL of DMSO-d6. Adulterated solutions were prepared at 0.5, 1.0, and 5.0% (w/w) concentrations by sequentially spiking 3, 3, and 24 μL of CAS, respectively, into the same NMR tube containing 600 μL of MP S097 solution. The NMR tube was inverted after each spike to aid in mixing. Data collection was performed before the first spike (pre-spike) and after each sequential adulterant CAS solution spike (adulterant spiked). This cyclic data collection−spike procedure was carried out with all eight adulterants in eight non-adulterated S097 sample solution replicates for a total of 32 spectra. This same process was repeated using DAS, yielding a second set of eight pre-spike replicate spectra and spectra of eight adulterated S097 preparations each at 0.005, 0.01, and 0.5% (w/w). Sample solutions for the eight adulterants (referred to as 100% spike or adulterant standard) were prepared by combining 30 μL of CAS and 570 μL of DMSO-d6 in a NMR tube and mixing with tube inversion. Data Acquisition. Data were acquired with a Bruker AVANCE III 600 MHz NMR spectrometer (Bruker Biospin, Rheinstetten, Germany) equipped with a 5 mm TXI inverse probe with Z-gradient coils and a Bruker SampleXpress Lite sample changer. The instrumentation and analytical conditions have been previously described.8 Instrument performance was verified daily by performing 1 H line shape, 1H sensitivity, and 13C sensitivity tests, and temperature reference was completed in automation using the Assure-System Suitability Test routine of the spectrometer and standard samples. All NMR spectra were acquired at 298.0 K in an approximate 20 ppm spectral range between approximately −4 and 16 ppm. 1H NMR spectra were acquired in automation under the control of ICON-NMR software (Bruker Biospin, Rheinstetten, Germany) using a single 90° pulse experiment with 32 scans and a 10 s relaxation delay, requiring about 12 min per sample. Batch Sample Preparation and Data Collection Timeline. Batch sample preparation and NMR data collection were carried out in two studies (parts 1 and 2) as previously described.8 Preparation and analysis took place several months apart. The reference data and replicate sample S097 data were collected in part 1 (Table 3) and the spiked sample (S097 and adulterant standard data) data were collected in part 2 (Table 2). For part 1, each sample was prepared in triplicate on day 1 and 1H NMR spectra were acquired on each day of the 5 days
Table 3. Repeat Analyses of Sample S097 sample preparation S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097 S097
1 1 1 1 1 2 2 2 2 2 3 3 3 3 3 4 4 4 4 4 4
preparation date (month/day/year)
day of analysis
analysis time (h)
12/3/12 12/4/12 12/5/12 12/6/12 12/7/12 12/3/12 12/4/12 12/5/12 12/6/12 12/7/12 12/3/12 12/4/12 12/5/12 12/6/12 12/7/12 12/9/12 12/9/12 12/9/12 12/9/12 12/9/12 12/9/12
3 4 5 6 7 3 4 5 6 7 3 4 5 6 7 9 9 9 9 9 9
1 2 0.5 3 4 5
of batch analysis along with nine randomly selected authentic samples. 1 H NMR spectra of a separate S097 replicate on day 9 was analyzed continuously at six different times on a single day. In total, 66 data sets representing 46 commercial MP samples were collected in part 1. For part 2 (spiking study), the data were collected over an approximate 10 day period, randomly selecting each day one or more adulterant spiking set or adulterant standard sample to be prepared and analyzed. An adulterant spiking set consisted of S097 solutions (0.0%, w/w) and its corresponding three serially adulterant-spiked S097 solutions (either 0.005, 0.01, and 0.05 or 0.5, 1, and 5%, w/w), yielding a total of four spectra. In total, 72 spectra [9 spectra (i.e., 2 spiking sets plus the adulterant standard) × 8 adulterants] were collected in part 2. Data Processing. As previously described,8 each raw signal was multiplied by an exponential line-broadening factor of 0.3 Hz before Fourier transformation. Resulting NMR spectra were manually phasecorrected and referenced to TMS at a chemical shift (δ) of 0.0 ppm and saved in Bruker NMR data file format. For data analysis, the processed Bruker data files were converted to tab-separated value (TSV) text files, retaining 65 000 data points per sample. For visual evaluation, the TSV files were analyzed using MestReNova NMR software (version 11.0.2, Mestrelab Research, Santiago de Compostela, Spain). For the current study, the TSV files were transferred and combined into a single Excel file. Although all of the files were the same length, they did not have the same starting and ending points. Positive or negative shifts of 0−7 units (0.0−0.0021 ppm) were necessary to align the TMS peaks. Data Range. For all of the results presented in this study, intensities from 9.0 to 0.2 ppm were used. Solvent Peaks. For all of the results presented in this study, intensities between 2.47 and 2.56 ppm and between 3.30 and 3.40 ppm were set to 0.0 to eliminate the signals from DMSO and hydrogen-deuterium oxide (HDO). Data Analysis. The raw data consisted of 65 636 data points for each sample for the range from 16.0 to −3.0 ppm. The selected data range (9−0.2 ppm) resulted in files that were 66 samples × 28 175 data points. Binning of 100 and 333 points gave files of 66 × 295 and 66 × 86, respectively. Data were analyzed using PCA.20 SIMCA was applied using one-class modeling.20 Data for chemometric analysis were preprocessed by a square root transformation, normalization (∑X2 = 1.0), and mean centering. pcANOVA was applied as previously described.19 8480
DOI: 10.1021/acs.jafc.8b00432 J. Agric. Food Chem. 2018, 66, 8478−8488
Article
Journal of Agricultural and Food Chemistry
Figure 1. PCA scores plots for 1H NMR spectra for 46 samples from 9.0 to 0.2 ppm, minus the solvent peaks (2.47−2.56 and 3.30−3.40 ppm) with (A) original spectra unaligned with 0.0003 ppm resolution, (B) aligned spectra with 0.0003 ppm resolution, and (C) aligned spectra with 333 points per bin. Samples were processed at low (green asterisks), medium (blue squares), and high (red triangles) temperatures.
Figure 2. Illustration of the source of the first derivative shape for PCA loadings: (A) spectral profiles for unaligned peaks for whey at 5.74 ppm and (B) resultant PCA loadings.
■
function of magnetic shift) for PC1 showed a first derivative shape for every peak. Figure 2 shows an expanded view of the 1 H NMR spectrum around the peak at 5.74 ppm and illustrates how misalignment of the spectra can produce loadings with a first derivative structure. Because the peaks are not aligned, the first half of the leading peak gives a negative error and the last half of the trailing peak gives a positive error. Processing of spectra routinely references the chemical shift of the data to an internal standard, e.g., TMS. Processing of the data produced spectra containing 65 636 data points for each sample for the range from 16.0 to −3.0 ppm, at a resolution of 0.0003 ppm. Close examination of the spectra showed that the TMS peak was not always labeled as 0.0 (Figure 3A). Instead, the peaks ranged from −1 to +6 resolution units away from 0.0, corresponding to a shift of −0.0003 to +0.0018 ppm. Shifts of this magnitude are generally deemed inconsequential, especially if the spectra are binned to 0.02 ppm. Misalignment of the TMS peak with respect to 0.0 linearly affected the entire spectrum (Figure 3B). Consequently, alignment of all of the TMS peaks at 0.0 (Figure 3C) served to align the whole spectrum (Figure 3D). Figure 1B shows the PCA scores plot for the same data plotted in Figure 1A after alignment of the TMS peaks. Now the processing temperature is correlated with PC1. In addition,
RESULTS This research consisted of two parts. Part 1 determined the variance associated with each of the experimental factors (i.e., sample metadata): sample type (SMP or NFDM), processing temperature (high, medium, or low), supplier, production site, and repeat analyses of a selected sample (S097). Part 2 examined the ability to detect samples spiked with a series of concentrations for eight adulterants, ammonium sulfate (AS), dicyandiamide (DCD), maltodextrin (malt), melamine (mel), soy protein isolate (SPI), sucrose (sucr), urea, and whey, using a non-targeted approach. Sources of Variance. PCA. PCA and pcANOVA provide easy means of examining the similarity of samples and the variance associated with each of the experimental factors. Figure 1A shows the PCA scores plot obtained for the 1H NMR spectra of the 46 samples in Table 1. The samples are labeled with respect to their processing temperature and show a visual correlation with the second principal component (PC2) (i.e., the processing temperatures appear to separate vertically), which provided 22% of the total variance. None of the experimental factors (type, supplier, production site, or day of analysis) showed a correlation with the first principal component (PC1), which provided 65% of the variance. Close examination of the PCA loadings plot (loadings as a 8481
DOI: 10.1021/acs.jafc.8b00432 J. Agric. Food Chem. 2018, 66, 8478−8488
Article
Journal of Agricultural and Food Chemistry
Figure 3. Spectral profiles for (A) TMS unaligned near 0.00 ppm, (B) quartet at 4.18 ppm unaligned, (C) TMS aligned at 0.00 ppm, and (D) quartet at 4.18 ppm after alignment of TMS.
Table 4. pcANOVA for Aligned, Unbinned Data (Resolution = 0.0003 ppm) n total sum of squares SMP versus NFDM processing temperature supplier site cross factors analysis day preparation
mean residual mean residual mean residual mean residual mean residual mean residual mean residual
66 2 66 3 66 11 66 15 66 21 66 7 66 45 66
df
total variance
percent total variance (%)
mean variance
F value
p
1 65 2 64 10 55 14 51 20 140 6 59 44 21
0.4053 0.0251 0.3802 0.0680 0.3121 0.0583 0.2538 0.0900 0.1638 0.0601 0.1037 0.0189 0.0848 0.0744 0.0104
100 6
0.0251
50.68
