Food Targeting: Geographical Origin Determination of Hazelnuts

Subscriber access provided by UNIV OF NEWCASTLE

Article

Food Targeting: Geographical origin determination of hazelnuts (Corylus avellana) by LC-QqQ-MS/MS based targeted metabolomics application Sven Klockmann, Eva Reiner, Nicolas Cain, and Markus Fischer J. Agric. Food Chem., Just Accepted Manuscript • DOI: 10.1021/acs.jafc.6b05007 • Publication Date (Web): 09 Jan 2017 Downloaded from http://pubs.acs.org on January 10, 2017

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Journal of Agricultural and Food Chemistry is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 36

Journal of Agricultural and Food Chemistry

1

Food Targeting: Geographical Origin Determination of Hazelnuts (Corylus

2

avellana) by LC-QqQ-MS/MS Based Targeted Metabolomics Application

3 4

Sven Klockmann, Eva Reiner, Nicolas Cain, and Markus Fischer*

5 6

HAMBURG SCHOOL OF FOOD SCIENCE; Institute of Food Chemistry, University of

7

Hamburg, Grindelallee 117, 20146 Hamburg, Germany, *Corresponding author: Tel.: +49-

8

40-428384357; Fax: +49-40-428384342; E-Mail: [email protected]

9

1 ACS Paragon Plus Environment


Page 2 of 36

10

ABSTRACT

11

A targeted metabolomics LC-ESI-QqQ-MS application for geographical origin discrimination

12

based on 20 non-polar key metabolites was developed, validated according to accepted

13

guidelines and used for quantitation via stable isotope labeled internal standards in 202 raw

14

authentic hazelnut samples from six countries (Turkey, Italy, Georgia, Spain, France and

15

Germany) out of harvest years 2014 and 2015. Multivariate statistics were used for detection

16

of significant variations in metabolite levels between countries and moreover, a prediction

17

model using support vector machine classification (SVM) was calculated yielding 100%

18

training accuracy and 97% cross-validation accuracy which was subsequently applied to 55

19

hazelnut samples for confectionary industry gaining up to 80% correct classifications

20

compared to declared origin. The present method demonstrates the great suitability for

21

targeted metabolomics applications in geographical origin determination of hazelnuts and

22

their applicability in routine analytics.

23 24

KEYWORDS

25

Targeted Metabolomics, Triple quadrupole, Hazelnut, Corylus avellana, Geographical origin,

26

Chemometrics, Metabolic profiling, Lipids


Page 3 of 36


27

INTRODUCTION

28

While 64% of the hazelnut (Corylus avellana L.) world crop (800,000 tons p.a.) is provided

29

by Turkey, Italy represents the second largest producer (13%) followed by the United States,

30

Georgia, Azerbaijan, Spain, France, Iran and China (less than 5%).1 Only 10% are dedicated

31

to direct consumption (mostly as in-shell nuts) but 90% of the crop is processed by food

32

industry, represented by chocolate, confectionary and bakery industry (predominantly in their

33

shelled and roasted form). With regard to demands of manufacturing industries, the quality of

34

nuts essentially depends on its shape and size, but also its chemical composition, represented

35

by aroma profile or blanching character.2,3 There are numerous varieties cultivated in

36

commercial orchards, each with its own characteristics and demands. However, besides the

37

genotype, cultural techniques, postharvest management and especially the geographical

38

location (soil, sun exposure, rainfall, temperature, height above sea level, etc.) of the hazelnut

39

plant affects the yield as well as morphological, physical and chemical characteristics and

40

thus, the ‘quality’ of hazelnuts.4-6 Consequently, sales prices may vary depending on - despite

41

of the harvest year - the geographical origin of products with highest values of Italian nuts

42

gaining 5,207 USD/t (with shell) in 2014, whereas nuts from Turkey had a 18% lower value,

43

followed by USA (-24%), Georgia (-31%) and Azerbaijan (-49%).1 In general, high price

44

differences are a potential incentive to food fraud for profit enhancement. Since there are no

45

existing methods in routine analytics for origin authentication of hazelnuts, no information

46

about the magnitude of deliberate or accidental false declarations are currently available. To

47

overcome this circumstance, this paper presents a targeted metabolomics approach for the

48

differentiation of hazelnuts from distinct countries based on 20 previously identified marker

49

substances. In this previous study, these key metabolites were determined in a non-targeted

50

approach using UPLC-ESI-QTOF-MS analysis of hazelnut samples from harvest years 2014

51

and 2015. Different prediction models were created and evaluated showing best results using

52

support vector machine classification (SVM) in combination with soft independent modelling 3 ACS Paragon Plus Environment


Page 4 of 36

53

of class analogy (SIMCA). 100% of the training samples and 80% of the prediction samples

54

(hazelnuts from confectionary industry) could be predicted correct at no false positives.7

55

However, the following development of a targeted metabolomics application for absolute

56

quantitation purposes was conducted by LC-QqQ-MS instead. Triple quadrupole mass

57

spectrometry (QqQ-MS) provides greater linear dynamic range, higher precision, less matrix

58

interferences and better robustness compared to LC-QTOF-MS. Furthermore, it is one of the

59

most popular instruments for food-quality and -safety analysis and nowadays commonly

60

applied in most laboratories of industrial quality control and governmental food control,

61

because it offers high sensitivity, selectivity and specificity for identification and shows good

62

quantitation capabilities for the analysis of complex food samples when operated in multiple-

63

reaction monitoring (MRM) mode.8-10 This scanning technique has a unique capability for

64

simultaneous analysis of large number of compounds in complex mixtures and is able to

65

reduce chemical noise and increase selectivity and sensitivity by selection of specific mass

66

transitions from precursor to fragment ions.11,12 The obtained mini-fingerprints are

67

representative for each group of issue to be examined and thus, used for comparing and

68

differentiating them by use of commonly applied multivariate statistics such as analysis of

69

variances (ANOVA), principal component analysis (PCA), soft independent modeling of

70

class analogy (SIMCA), cluster analysis, linear discriminant analysis (LDA) or support vector

71

machine classification (SVM).6,13-15

72

Present studies concerning metabolomics based authentication of the geographical origin of

73

foods have been extensively reviewed by several authors, mostly dealing with NMR, IR,

74

stable isotope analysis or GC-MS as detection techniques.6,14,16,17 Furthermore, in this context,

75

current LC-MS approaches predominantly make use of non-targeted high resolution LC-ESI-

76

QTOF-MS applications while targeted approaches are still rare.13,18-20 Some studies used

77

metabolic profiling for investigating the geographic differences of certain substance classes,

78

like steviol glycosides in stevia and polyphenols in red wine or apple juices.21-23 Accordingly, 4 ACS Paragon Plus Environment

Page 5 of 36


79

the development of a target-oriented LC-ESI-QqQ-MS method for geographical origin

80

discrimination by only analyzing 20 previously identified key metabolites is unique. Besides

81

the already highlighted analytical benefits, the motivation for switching the analytical method

82

from non-targeted QTOF to targeted QqQ were the comparatively low costs and the widely

83

dissemination of such devices in modern laboratories, giving the great opportunity for a

84

consequent implementation of developed methods into routine analysis. Constantly growing

85

databases would ensure consideration of annual variation of hazelnut metabolome, caused by

86

mutable factors like climate, and improve statistical models by consecutive updating.

87

Otherwise, extraordinary phenomena such as extreme cold and rainy or hot and dry summers

88

may lead to misinterpretations caused by metabolic changes.

89 90

MATERIALS AND METHODS

91

Reagents and chemicals

92

Ultrapure water was obtained by purifying demineralized water in a Direct-Q® 3 UV-R

93

system (Merck Millipore, Darmstadt, Germany). LC-MS grade isopropanol, methanol, HPLC

94

grade chloroform, ammonium formiate ˃95%, formic acid >99% and butylated

95

hydroxytoluene (BHT) ≥99.8% were purchased from Carl Roth (Karlsruhe, Germany).

96

The reference standards 1,2-dilinoleoyl-sn-glycero-3-phosphocholine (PC(18:2/18:2)), 1-

97

palmitoyl-2-linoleoyl-sn-glycero-3-phosphocholine

(PC(16:0/18:2)),

98

glycero-3-phosphoethanolamine

and

99

(DG(16:0/18:1)) were acquired from Avanti Polar Lipids (Alabaster, AL, USA), 1,3-

(PE(18:2/18:2))

(DG(18:2/18:2)),

1,2-dilinoleoyl-sn-

1-palmitoyl-2-oleoyl-sn-glycerol

100

dilinoleoyl-rac-glycerol

1-palmitoyl-2-linoleoyl-rac-glycerol-3-

101

phosphoethanolamine (PE(16:0/18:2)) and 1-palmitoyl-2-oleoyl-3-linoleoyl-rac-glycerol

102

(TG(16:0/18:1/18:2)) from Cayman Chemical (Ann Arbor, MI, USA), 1,2-dioleoyl-sn-

103

glycerol (DG(18:1/18:1), 1,2-dioleoyl-sn-phosphoethanolamine (PE(18:1/18:1) and γ-

104

tocopherol from Sigma-Aldrich (Munich, Germany), 1,2-dioleoyl-3-palmitoyl-rac-glycerol 5 ACS Paragon Plus Environment


Page 6 of 36

105

(TG(16:0/18:1/18:1)) from Toronto Research Chemicals (Toronto, ON, Canada) and 1-

106

linoleoyl-2-hydoxy-sn-glycero-3-phosphocholine (PC(18:2/0:0)) from Echelon Biosciences

107

(Salt Lake City, UT, USA).

108

The stable isotope labeled internal standards (IS), 1-pentadecanoyl-2-oleoyl(d7)-sn-glycero-3-

109

phosphocholine

110

phosphoethanolamine

111

phosphocholine

112

(DG(15:0/18:1(d7))) and 1,3-dipentadecanoyl-2-oleyol(d7)-glycerol (TG(15:0/18:1(d7)/15:0))

113

were obtained from Avanti Polar Lipids (Alabaster, AL, USA).

(PC(15:0/18:1(d7))), (PE(15:0/18:1(d7))), (PC(18:1(d7)/0:0)),

1-pentadecanoyl-2-oleoyl(d7)-sn-glycero-31-oleoyl(d7)-2-hydoxy-sn-glycero-31-pentadecanoyl-2-oleyol(d7)-sn-glycerol

114 115

Standard solutions and calibration

116

For internal standard calibration all five stable isotope labeled lipids were pooled to one mix-

117

solution containing 302.48 µM PC(18:1d7/0:0), 272.00 µM DG(15:0/18:1d7), 229.87 µM

118

PE(15:0/18:1d7), 201.19 µM TG(15:0/18:1d7/15:0) and 212.36 µM PC(15:0/18:1d7).

119

Standard solutions of individual lipids were prepared by dissolving the respective compound

120

separately in methanol/chloroform to a stock solution of approx. 10 mM. Validation solutions

121

were created by mixing all stock solutions and diluting to the following concentrations in µM

122

with methanol: 0.02, 0.04, 0.06, 0.08, 0.1, 0.2, 0.4, 0.6, 0.8, 1, 2, 4, 6, 8, 10, 20, 40, 60, 80,

123

100 (for DG(18:1/18:1) and DG(18:2/18:2) additionally 200, 400, 600, 800, 1000), equally

124

performed with internal standard mix-solution. Validation was realized either in absence (base

125

calibration) or presence (matrix calibration) of a representative hazelnut extract as matrix

126

(50% of the resulting standard solution, respectively). 10 blank samples containing only pure

127

extraction solution were created for determination of the limit of detection (LOD).

128 129

Hazelnut samples


Page 7 of 36


130

Overall 202 authentic raw hazelnut samples of different varieties, origins and producers from

131

harvest years 2014 (107) and 2015 (95) were obtained for analyses. The samples were

132

harvested in the respective commercial relevant regions of each country, represented by 115

133

French (mainly Midi-Pyrénées and Aquitaine), 35 German (mainly in Bavaria), 22 Italian

134

(Piedmont, Campania and Lazio), 14 Turkish (Ordu, Akçakoca and Samsun), ten Georgian

135

(Guria, Samegrelo and Imereti) and six Spanish (Tarragona) samples. In addition, 50 samples

136

of raw hazelnut kernels for confectionary industry (non-authentic samples) were provided by

137

industrial partners, including one Spanish, one French, six Georgian, nine Italian and 33

138

Turkish samples. Furthermore, five samples of roasted and blanched hazelnut kernels that

139

would have been used for production of confectionary products were obtained (four Turkish

140

and one Italian). Each sample comprises either 1000 grams hazelnut kernels with skin (testa)

141

or 1500 grams unshelled hazelnuts.

142 143

Sample treatment

144

Harvest and postharvest processing was executed by suppliers under realistic conditions

145

following common commercial practices in drying, storage, and sometimes cracking

146

conditions as well as time of harvest. Authentic hazelnut samples were shipped as quickly as

147

possible under common storage conditions after the crop. For non-authentic samples no

148

information about postharvest storage time could be obtained. All hazelnut samples were

149

equally handled during all analytical processes being stored at -40 °C in the unprocessed state

150

and at -20 °C as lyophilized powder or extracts. Only hazelnut kernels with skin were used for

151

further procedure (except of roasted ones), those samples obtained in-shell were previously

152

cracked manually. After freezing with liquid nitrogen hazelnut kernels were ground in

153

combination with dry ice at a ratio of 1:1 using a Grindomix GM 300 knife mill equipped

154

with a stainless steel grinding container and a full metal knife (Retsch, Haan, Germany). A

155

representative aliquot of the grist was freeze-dried using a Gamma 1-20 freeze-dryer (Martin 7 ACS Paragon Plus Environment


Page 8 of 36

156

Christ Gefriertrocknungsanlagen, Osterode am Harz, Germany). 50 mg of each lyophilisate

157

was used for extraction adding 10 µL internal standard mixture, 990 µL extraction solvent

158

(chloroform/isopropanol 1/2 with 0.1% butylated hydroxytoluene) and two steel balls

159

followed by ball milling for 3 min at 3 m/s using a Bead Ruptor 24 equipped with a 1.5 mL

160

microtube carriage kit (Biolabproducts, Bebensee, Germany). The remaining suspension was

161

centrifuged for 10 min at 16,000 x g at 4 °C and the supernatant was used for LC-MS analysis

162

after filtration with a Rotilabo® PTFE syringe filter, 0.45 µm pore diameter (Carl Roth,

163

Karlsruhe, Germany).

164 165

HPLC-ESI-QqQ-MS/MS acquisition

166

Chromatographic separation was carried out by a 50 mm x 4.6 mm i.d., 2.7 µm, Poroshell 120

167

EC-C18 column (Agilent Technologies, Waldbronn, Germany) at 30 °C with a flow rate of

168

500 µL/min and elution solvents A water and B isopropanol, both containing 5 mM

169

ammonium formiate buffer at pH 3.5 using an Agilent 1260 Infinity Quaternary LC System

170

(Agilent Technologies, Waldbronn, Germany). The gradient elution started with 85% B for

171

2 min, linearly increased to 100% in 3 min and kept for 4 min, brought back to 85% in

172

0.5 min followed by 4 min of re-equilibration. Injection volume was set to 3 µL. For detection

173

a triple quadrupole-MS/MS API 2000 (Applied Biosystems, Darmstadt, Germany) equipped

174

with a turbo ion spray source was used with the following mass spectrometer settings in

175

positive ion mode: ion spray voltage = 5500 V; temperature = 450 °C; ion source

176

gas 1 = 30 psi; ion source gas 2 = 70 psi; curtain gas = 20 psi; collision gas = 5 psi; ion spray

177

probe vertical position = 3; ion spray probe horizontal position = 5. The MS acquisition was

178

divided into two periods (period 1 = 5.5 min; period 2 = 8.0 min) analyzing only compounds

179

eluting in the respective period to increase dwell-times at consistent cycle times for better

180

sensitivity and more data points. For each compound one mass transition for quantitation

181

purposes (quantifier) and one for additional identification (qualifier) was acquired. Dwell8 ACS Paragon Plus Environment

Page 9 of 36


182

time was set to 60 ms for quantifier in period 1, 100 ms for quantifier in period 2 and 20 ms

183

for all qualifiers. To optimize data acquisition parameters for multiple reaction monitoring

184

(MRM) ideal compound-dependent device voltages for each mass transition were determined

185

via automatic compound optimization using Analyst® Software (AB Sciex, version 1.6,

186

Foster City, CA, USA) injecting a 10 µM solution of each compound directly into the source

187

with a flow of 10 µL/min. For those metabolites no standard substance could be purchased the

188

acquisition parameters were extrapolated taking related structures as a basis. Various

189

transitions with different parameters were tested and quantifier and qualifier finally selected

190

according to highest signals achieved. To prevent bias due to instrumental drifts samples were

191

injected in a random order followed by a blank sample (pure extraction solvent) every 12

192

injections. Samples were acquired in triplicate analysis, each one in a separate batch.

193

Carryover effects were minimized by applying a needle wash step with methanol after each

194

injection.

195 196

Method validation

197

Linearity, lower limit of quantitation (LLOQ), precision and accuracy were assessed in

198

analogy to accepted guidelines.24 The limit of detection (LOD) was determined according to

199

DIN 32645.

200

Linearity was determined for all commercially available standards using a twenty-point (resp.

201

twenty-five-point for DG(18:1/18:1) and DG(18:2/18:2)) calibration curve (n = 5) with OLS

202

regression. Homoscedasticity was confirmed for all samples using Breusch-Pagan test and

203

linear range was calculated by means of Mandel’s fitting test.25,26 Furthermore, the LLOQ has

204

been set as the lowest standard on the calibration curve as it met the following conditions: The

205

analyte response was at least five times the response compared to a blank response and the

206

analyte peak is identifiable, discrete, and reproducible, and the back-calculated concentration



Page 10 of 36

207

should have precision that does not exceed 20% of the coefficient of variation CV and

208

accuracy within 20% of the nominal concentration.

209

Precision and accuracy were obtained determining the CV, resp. the back-calculated

210

concentration of every analyte by replicate analysis of calibration standards (n = 5) at three

211

concentration levels (smallest, highest and medium concentration of the determined linear

212

range). The mean value for precision and accuracy should be within 15% of the nominal value

213

except at LLOQ, where it should not deviate by more than 20%.

214

LOD was calculated based on the standard deviation of the response and the slope of the

215

calibration curve by replicate analysis of ten independent blank samples (pure extraction

216

solvent).

217

To determine the required dilution factor for measuring hazelnut extracts being in linear range

218

for each metabolite a dilution array was built by mixing equal aliquots of a randomly chosen

219

representative extract from each country and diluting the mixture with following dilution

220

factors: 1.18, 1.43, 1.82, 2.5, 4, 10, 11.76, 14.29, 18.18, 25, 40, 100, 117.65, 142.86, 181.82,

221

250, 400, 1000.

222 223

Data processing and chemometrics

224

Peak integration of LC-MS data was performed with Analyst® Software (AB Sciex, version

225

1.6, Foster City, CA, USA). Metabolite concentration was calculated based on its

226

corresponding internal standard, represented their respective substance class, adding a

227

correction factor for each metabolite in its linear ranges considering the individual signal

228

responses for each mass transition. For those metabolites no standard substance could be

229

obtained, an estimated correction factor was extrapolated by averaging the values of its

230

related substance class. Since no triacylglycerol marker substance standards were available

231

TG(16:0/18:1/18:1) and TG(16:0/18:1/18:2) were used additionally instead. Multivariate

232

statistics was assessed in The Unscrambler X 10.3 (Camo Software, Oslo, Norway). Prior to 10 ACS Paragon Plus Environment

Page 11 of 36


233

statistical process the data set was scaled applying interquartile range scaling using the range

234

between the 75th and 25th percentile for each metabolite separately. Scaling parameters were

235

calculated based on authentic sample data and afterwards applied to non-authentic sample

236

data. Classification models for prediction of unknown samples were applied using support

237

vector machine classification (SVM) applying 4th grade polynomial kernel type with

238

offset = 3, C = 1 and Gamma = 0.1 as optimal parameters. Optimal kernel function and

239

related parameters were identified by exercising grid search function via cross-validation. The

240

data of triplicate analysis of the 150 authentic hazelnut samples was used as training data for

241

model creation. To validate the model, cross-validation was applied by dividing the training

242

set into six segments. The model was then used for prediction of the 50 non-authentic

243

hazelnut samples assuming that one sample (three measurements) is only attributed to a

244

particular country if at least two of three classifications predict the same. Furthermore, linear

245

discriminant analysis based on PCA scores (PCA-LDA) using quadratic method at 17

246

components as optimal parameters and soft independent modeling of class analogy (SIMCA)

247

(α = 0.05) were tested. Two-sided T-Test models for each country pair (e.g. Italy vs. Turkey)

248

were calculated employing unscaled values to test the equality of two means. Furthermore,

249

one-way analysis of variance (ANOVA) was calculated in ProfileAnalysis 2.1 (Bruker

250

Daltronics, Bremen, Germany) to test the equality of three or more means.

251 252

RESULTS AND DISSCUSION

253

LC-MS acquisition and validation

254

The metabolites analyzed in the presented work were previously identified as marker

255

substances for geographical origin discrimination of hazelnuts via non-targeted UPLC-ESI-

256

QTOF-MS considering samples of five countries (Germany, France, Italy, Georgia, Turkey).7

257

Subsequently, a targeted HPLC-ESI-QqQ-MS method for quantitation of the previously

258

identified 20 marker substances by usage of internal standards was developed and validated 11 ACS Paragon Plus Environment


Page 12 of 36

259

(Table 1). For each substance class (phosphatidylcholines, phosphatidylethanolamines,

260

lysophosphatidylcholines, diacylglycerols and triacylglycerols) one representative deuterated

261

internal standard was used. With this method the great advantage over the already presented

262

non-targeted application is the possibility for determining absolute metabolite levels in mg/kg

263

hazelnut in only 13.5 min per sample by using a relatively cheap but robust high-throughput

264

HPLC-ESI-QqQ-MS application. Thus, the step from technically, financially and time

265

demanding high-resolution UPLC-ESI-QTOF-MS to this application is a reasonable

266

advancement and a great opportunity for routine analytics. Since origin discrimination

267

demands continual progression and improvement of statistical methods due to annual climatic

268

changes and perpetual developments in international trading, the comparability of results from

269

different acquisition times and laboratories is a crucial point. Thus, absolute quantitation via

270

internal standards using robust and commonly applied LC-ESI-QqQ-MS/MS provides

271

maximum comparability and the possibility for data fusion.

272

In contrast to UPLC-ESI-QTOF-MS acquisition, DG(16:0/16:1) could not be detected as its

273

sodium adduct. Instead, the ammonium adduct found during compound optimization was

274

used for quantitation, which is in compliance to the other di- and triacylglycerols the most

275

abundant ion. Validation of LC-MS method was performed for all commercially available

276

standards in accordance with guidelines for bioanalytical method validation of the FDA and

277

the German DIN 32645 either in absence (base calibration) or presence (matrix calibration) of

278

hazelnut matrix.24 Since none of the triacylglycerol marker substances could be purchased,

279

two structural similar triacylglycerols were used in method validation to demonstrate method

280

efficiency. In baseline calibration (Table 2) all compounds showed good linearity (R2 > 0.99)

281

in a broad dynamic linear range from lowest 0.03 µM to highest 106.46 µM. The limit of

282

detection (LOD) ranged from 0.19 nM to 978.49 nM. While phospholipids showed lowest

283

values, limits of detection for diacylglycerols were up to 22-fold higher and for

284

triacylglycerols up to 717-fold, expect of regarding at the internal standards, where all 12 ACS Paragon Plus Environment

Page 13 of 36


285

substance classes gained similar values. The obtained values for precision and accuracy at the

286

LLOQ were below 15% for 82% of the analytes, but all being less than 20%. 94% of the

287

values for expected concentrations ranged below 15% while none was above 20%.

288

To overcome matrix effects, linearity responses were studied in hazelnut matrix spiked with

289

mix-standard solution. The correlation coefficients of all the calibration curves showed good

290

linearity (R2 > 0.99) for 80% of the analytes and still acceptable linearity (0.98 < R2 < 0.99)

291

for 13%. Values for accuracy and precision at the LLOQ were below 15% for 80% of the

292

analytes, but all being less than 20%. 95% of the values for expected concentrations ranged

293

below 15% while none was above 20%, which demonstrated a minor influence of matrix

294

effects on the presented method.

295

For the remaining marker substances a dilution array of a representative mix of hazelnut

296

extracts from all six countries was used to estimate the required dilution factor for being in

297

linear range. Furthermore, this was used to determine the dilution factor for extracts being

298

used for quantitation of each metabolite. It turned out that all metabolites could be measured

299

in pure extracts except of TG(16:0/16:1/18:1), TG(14:0/16:0/18:1), TG(17:1/18:1/18:2) and

300

TG(18:2/18:2/18:3), which had to be quantitated in a 100-fold dilution. For those metabolites

301

that could be analyzed in both conditions the undiluted state was used. During first analyses

302

of hazelnut extracts it became apparent that the developed application was not sensitive

303

enough for quantitating γ-tocopherol since signal intensity was either below LOD or LLOQ.

304

Ensuing from LOD (81.4573 nM) and LLOQ (14.9995 µM), a γ-tocopherol content of

305

0.69 mg/kg should be theoretically detectable and 201.61 mg/kg quantifiable, respectively.

306

There are numerous publications about γ-tocopherol content in hazelnuts ranging from 2 to

307

240 mg/kg extracted hazelnut oil or not detectable, depending on variety, origin and analysis

308

method.27-32 Furthermore, values for mg/kg hazelnut range between 0.0038 and 20.8, or were

309

not detectable either.33-35 Although the development of a customized extraction procedure

310

including an enrichment step might overcome this problem, this was not contemplated as 13 ACS Paragon Plus Environment


Page 14 of 36

311

extraction protocol was designed for maximum reproducibility. Changing the composition of

312

the extraction solvent to lower chloroform rate at higher methanol content resulting in higher

313

polarity for better tocopherol extraction would impair efficiency for the less polar tri- and

314

diacylglycerols. This should not be pursued since signal intensity for DG(16:0/16:1) and

315

TG(15:0/16:0/18:1) already is near the lower limit. Hence, extraction was kept as simple and

316

fast as possible with the downside of losing γ-tocopherol and thus, no internal standard for

317

tocopherols needed to be applied during validation process.

318

Since detectors response for different mass transitions varied within substance classes

319

between internal standard and metabolites, a correction factor for each metabolite was

320

calculated to overcome this variation. For those metabolites no standard could be achieved, a

321

value was extrapolated by average substance class correction factor. For triacylglycerol

322

marker substances TG(16:0/18:1/18:1) and TG(16:0/18:1/18:2) were used for extrapolation

323

resulting in limited accuracy for calculated metabolite levels, though.

324 325

Method application

326

19 of the previously identified marker substances were analyzed in 202 authentic hazelnut

327

samples. In contrast to non-targeted acquisition, samples were out of six commercially

328

relevant countries now (Turkey, Italy, Georgia, Spain, France, and Germany), additionally

329

including Spain. Every sample was measured as triplicate. As expected, level distributions

330

were similar to previous findings with decreasing contents of lipids with at least one

331

polyunsaturated fatty acid side chain from Germany over France and Italy to Georgia, Turkey

332

and Spain and lipids with only monounsaturated fatty acids in reverse order. (Figure 1)

333

Between Georgia, Turkey and Spain the order partially changes depending on key metabolite.

334

Average metabolite levels ranged from 1.03 mg/kg (DG(16:0/16:1)) to 4350.25 mg/kg

335

(DG(18:1/18:1)) at confidence intervals (α = 0.05) between 1.22% and 21.45% of the average

336

value showing up to 6-fold changed between countries (PC(18:2/18:2)). (Table 3) Principal 14 ACS Paragon Plus Environment

Page 15 of 36


337

component analysis was calculated visualizing the relations of samples against each other

338

(Figure 2). The overall variance explained by PC1-PC3 already yields 80%. Samples lying

339

close together have a similar metabolic profile, which is the case for all countries though

340

overlapping partially. While German samples are only overlapping with some French ones,

341

Turkish, Georgian, Spanish and Italian samples are more mixed. Nevertheless, clusters for

342

each country are visible. Only Spanish samples are divided stronger and two Georgian

343

samples are further away from the rest but both not divided by harvest years and neither for

344

any other country. It can be concluded that the harvest year does not cause much bias to the

345

metabolome, at least for the present samples. T-test for comparison of means was calculated

346

for each country-pair and metabolite to evaluate the significance of differences between

347

countries. (Figure 3) P-values between 0.001 and 0.05 indicate evidence against the null

348

hypothesis (metabolite levels of two groups are equal) and thus, that these two groups are

349

significantly different, while p-values < 0.001 represent very strong evidence for a highly

350

significant difference. A p-value above 0.05 signifies no evidence for rejecting the null

351

hypothesis.36 The amounts of metabolites comprising significant differences for each country-

352

pair could be used to estimate the similarity of metabolite contents and thus, the

353

discriminative ability of the mini-fingerprints. Although some metabolite ranges show strong

354

overlapping, Turkey-Georgia, Turkey-France, Italy-Georgia and Italy-Spain obtain 10 to 14

355

highly significant metabolites and moreover, the number for the remaining country pairs is

356

higher than 17 at maximum two non-significant ones. Difficulties may occur especially

357

discriminating Georgia and Spain, involving eight metabolites with coinciding levels at only

358

seven highly significant. Nevertheless, seven metabolites with highly significant changes still

359

may be sufficient for efficient discrimination. Out of the 19 marker substances PE(18:2/18:2)

360

overall has the best discriminative abilities according to one-way analysis of variance

361

(ANOVA), which is comparing the means of all six countries simultaneously, owing the

362

lowest p-value followed by PC(18:2/18:2) and DG(18:2/18:2). Because of the broad range of 15 ACS Paragon Plus Environment


Page 16 of 36

363

metabolite contents a scaling step had to be applied to the data prior to multivariate statistics

364

to bring values to one scale which has also been done before PCA calculation. Here,

365

interquartile range scaling was utilized. Scaling parameters for each metabolite were

366

calculated based on authentic sample set and afterwards applied to non-authentic data. These

367

parameters can also be used for processing future measurements, so upcoming data can be

368

easily integrated into the present statistical models. With an appropriate scale data now can be

369

reasonably visualized without discriminating low values. (Figure 4) The bar charts can be

370

seen as a kind of barcodes or ‘mini-fingerprints’ being representative for each country. As

371

already appeared after calculating p-values in t-test, Georgia and Spain are very similar but

372

still differences can be overserved visually, while all other countries can be easily

373

differentiated by this chart. Because France’s metabolite levels are always located in the

374

middle of metabolite ranges, scaled values reckon around zero.

375 376

Sample prediction

377

For prediction of unknown samples SVM was used because of its great capabilities in

378

complex biological problems. SVM has some major advances over classical multivariate

379

approaches being able to separate even non-linear data by transformation of data into higher-

380

dimensional feature spaces and searching for optimal separating hyperplanes enabling

381

discrimination of complex cases and overlapping classes in original data.37,38 Several kernel

382

functions (sigmoid, radial basis, polynomial) give rise to classification with overlapping

383

groups mapping non-linear data into a linear separation case in feature space. In this case

384

polynomial function (with offset = 3, C = 1 and Gamma = 0.1 as optimal parameters) fits best

385

yielding 100% both for training accuracy as well as cross-validation accuracy. Cross-

386

validation was applied by multiple dividing training set into variable six segments, using 5/6

387

for model creation and subsequent prediction of the separated sixth. Even when using just two

388

segments, cross-validation accuracy still reached 97%, demonstrating the high robustness of 16 ACS Paragon Plus Environment

Page 17 of 36


389

the present model. Thus, the prediction model is able to differentiate all countries despite

390

overlapping metabolite ranges. This evinces the excellent suitability and strength of the

391

presented targeted metabolomics application for geographical origin discrimination of

392

hazelnuts.

393

In addition, 50 hazelnut samples for confectionary industry were analyzed in equal measure

394

and metabolite levels were calculated. Due to current market situation, mainly samples from

395

Turkey (33), Italy (nine) and Georgia (six) could be obtained and additionally one each from

396

France and Spain. Overlapping confidence intervals with authentic samples could be found in

397

77%. When only regarding countries with more than one confectionary sample, 93%

398

overlapping confidence intervals were obtained. By overlaying spider charts of scaled values

399

for authentic and confectionary samples these findings could be confirmed (Figure 5). Shapes

400

of spider diagrams reflect the similarities of the illustrated sample groups. While countries

401

among themselves have different shapes and can be easily differentiated by visual like bar

402

charts the contour of authentic and confectionary plots for each country closely resemble,

403

except of French ones. Nevertheless, the shape of its confectionary sample still coincides

404

most with the French’s authentic samples. Like in bar charts Georgia and Spain look very

405

similar, but characteristic differences in the markedness of spikes still can be perceived.

406

Employing confectionary samples on the SVM prediction model 80% were classified correct

407

at 20% wrong. One sample was defined as classified ‘correct’ if two or three single

408

measurements of triplicate analysis were predicted correct, ‘wrong’ if two or three predicted

409

the same but false country and ‘not classified’ if all three predictions were dissimilar.

410

Additionally, like in the non-targeted approach, PCA-LDA (98.4% training but only 24.0%

411

prediction accuracy) and SIMCA (multiple classifications, 0% false negative but only 25.3%

412

single correct attributions) were tested and the obtained results were again inferior to SVM.

413

Unlike previous findings, the combination of SVM with SIMCA as surveillance analysis

414

(because no sample was attributed false negative) did not improve the outcome dramatically 17 ACS Paragon Plus Environment


Page 18 of 36

415

since only one sample could be assigned ‘not classified’ instead of being false predicted.

416

Looking at the data of SVM, four out of six Georgian samples could not be classified

417

correctly, indicating a poor suitability for this country. Nevertheless, 93% of the determined

418

metabolite levels were inside the respective range for authentic Georgian samples and 84%

419

for all ten false predicted samples, showing that there is not much deviation in metabolite

420

profiles themselves but rather blurred borders. Because of the partly strong overlapping

421

metabolite levels of countries the margins become indistinct in certain cases and properly

422

prediction is not easy. Concerning the French sample, despite the already mentioned deviation

423

from authentic set which could be observed visually by comparing both spider charts, it was

424

predicted correct though. As the metabolic profile of French hazelnuts differs relative strong

425

to other countries a higher variation could be tolerated for still correct prediction unlike it

426

would be the case for closer resembling countries. However, there is the possibility of being

427

different from authentic samples due to either biological variance or economic handling

428

resulting in poor predictability. This indicates that further improvements of prediction models

429

still have to be evaluated. By increasing the sample amount used for model creation this

430

problem may be overcome, since statistical reliability could be improved as larger sets of

431

samples would take much better account of natural extremes. This thesis could be confirmed,

432

when using confectionary and authentic samples together for model creation in SVM,

433

reaching again both 100% for training accuracy and cross-validation accuracy, respectively.

434

However, since the declared geographical origin of samples for confectionary industry could

435

not be entirely verified, one cannot rule out the possibility that conscious or accidental

436

adulterations have been taken place distorting the results and furthermore, some samples may

437

remain undistinguishable with this method because of too strong overlapping scopes.

438

Especially samples from border regions remain challenging, probably having very similar

439

metabolic profiles. This becomes critical if areas of cultivation within one country are more

440

spatially divided than the proximity of one of them to other country’s ones. Thus, for some 18 ACS Paragon Plus Environment

Page 19 of 36


441

cases, a classification into (transnational) geographical areas would be more sensible but

442

usually does not reflect commercial reality of declaration. Fortunately, harvest areas for the

443

present issue are not directly linked to areas of other countries.

444

In addition, five confectionary samples with roasted and blanched hazelnut kernels were

445

equally analyzed to estimate the effect of roasting to the predictability of the model.

446

Surprisingly again 80% could be attributed correct while only one sample was predicted

447

wrong. The metabolic profiles of correct predicted samples predominantly coincide with

448

authentic ones, lying inside the respective ranges in 93% of the cases. Further investigations

449

still have to proof these preliminary findings but the results are demonstrating that the general

450

applicability is given and that the presented marker substances seem to remain unaffected by

451

temperature impact of roasting process as well as blanching. Even though the high

452

temperatures of roasting would have a supposedly great influence on hazelnut metabolome,

453

the results are in accordance to a study of Locatelli et al. (2015) who inferred from

454

chemometric evaluation that the fatty acid composition of hazelnuts is not influenced by

455

roasting and thus, geographical origin discrimination capacity remains unaffected.39 The same

456

holds true for Alasalvar et al. (2010) who reported minor influences on fatty acid profile of

457

different Turkish hazelnut cultivars.40 Therefore, it can be concluded that the lipid fraction of

458

hazelnuts is not very sensitive to the temperature impact of roasting process.

459

In comparison to the non-targeted metabolomics application this method offers the great

460

opportunity for being relatively easy implemented in other laboratories using commonly

461

applied HPLC-ESI-QqQ-MS instruments while still yielding good results in discriminative

462

abilities. It offers a robust and fast quantitation of 19 key metabolites that are suitable for

463

origin discrimination of hazelnuts from six countries. Ongoing analyses of further samples

464

from other countries and harvest years still have to be executed for constant improvement of

465

statistical methods for better prediction of raw as well as roasted hazelnut kernels. Data



Page 20 of 36

466

sharing and fusion could then be used to create a unique and comprehensive database

467

enabling greatest possible accuracy for prediction of unknown samples.

468 469

ABBREVIATIONS USED

470

MRM, multiple reaction monitoring; ANOVA, analysis of variances; SIMCA, soft

471

independent modelling of class analogy; PCA, principal component analysis; SVM, support

472

vector

473

phosphatidylcholine; PE, phosphatidylethanolamine; LOD, limit of detection; LLOQ, lower

474

limit of quantitation; FDA, Food and Drug Administration; Qnt, quantifier; Qal, qualifier;

475

IQR, interquartile range

machine

classification;

TG,

triacylglycerol;

DG,

diacylglycerol;

PC,

476 477

ACKNOWLEDGEMENTS

478

The authors are very grateful to SCA Unicoque, Erzeugerorganisation Deutscher

479

Haselnussanbauer UG, Schlüter&Maack GmbH, Amt für Ernährung, Landwirtschaft und

480

Forsten

481

AgroTeamConsulting, Institute of Biotechnology and Microbiology, University of Hamburg,

482

August Storck KG, Seeberger GmbH, Crisol de Frutos Secos, Azienda Agricola Cascina

483

Valcrosa, Basaran Entegre Gıda san. ve Tic. A.Ş, Alta Langa Azienda Agricola, Corilu

484

Societa Cooperativa Agricola, Coselva SCCL, Eganut LLC and Franken Genuss UG &

485

Co.KG for providing us with authentic hazelnut samples.

486

Furthermore we thank Ferrero OHG mbH, Heinrich Brüning GmbH, August Töpfer & Co.

487

(GmbH & Co.) KG, Lübecker Marzipan-Fabrik v. Minden & Bruhns GmbH & Co. KG, Carl

488

Wilhelm Clasen GmbH, Horst Walberg Trockenfrucht Import GmbH, Fratelli Caffa s.a.s.,

489

Ludwig Schokolade GmbH & Co. KG, Alfred Ritter GmbH & Co. KG, Kaufland

490

Omnichannel International GmbH & Co. KG, Stollwerck GmbH and Rapunzel Naturkost

491

GmbH for supporting us with raw hazelnut kernels for confectionary industry.

Fürth/Sortenversuchsanstalt

Gonnersdorf,

Stelma

SRL

Unipersonale,


Page 21 of 36


492

In addition, the authors thank René Bachmann and Thomas Hackl for excellent assistance in

493

sample preparation.

494 495 496

SUPPORTING INFORMATION

497

Detailed list with suppliers, provenance and cultivar information of all used authentic

498

hazelnut samples (Table S1) and confectionary samples (Table S2), metabolite concentrations

499

in validation solutions (Table S3), MRM acquisition parameters for each mass transition

500

(Table S4), scaling values and correction factors (Table S5), accuracy and precision for three

501

concentration levels in base calibration (Table S6) and matrix calibration (Table S8),

502

validation parameters for matrix calibration (Table S7), dilution array of a representative mix

503

of extract with regression equation, correlation coefficient and linear range (Table S9),

504

metabolite ranges of authentic samples for each country and metabolite (Table S10), p-values

505

for ANOVA calculations (Table S11), average and confidence intervals for confectionary

506

samples (Table S12), loadings plots of PCA calculations (Figure S1) and results of SVM

507

model predictions (Table S13) as well as equation for calculation of scaled values for each

508

metabolite (Equation S1).



Page 22 of 36

509

REFERENCES

510

(1)

511

2015; http://faostat3.fao.org/browse/Q/QC/E (accessed Dez 20, 2016).

512

(2)

513

Image analysis techniques for automated hazelnut peeling determination. Food Bioprocess

514

Technol. 2009, 3, 155-159.

515

(3)

Mehlenbacher, S. A. Hazelnuts (Corylus). Acta Hortic. 1991, 290, 791-838.

516

(4)

Locatelli, M.; Coïsson, J. D.; Travaglia, F.; Cereti, E.; Garino, C.; D’Andrea, M.;

517

Martelli, A.; Arlorio, M. Chemotype and genotype chemometrical evaluation applied to

518

authentication and traceability of “Tonda Gentile Trilobata” hazelnuts from Piedmont (Italy).

519

Food Chem. 2011, 129, 1865-1873.

520

(5)

521

plant metabolomics strategies: from analytical platforms to data acquisition and processing.

522

Nat. Prod. Rep. 2014, 31, 784-806.

523

(6)

524

authentication. Food Res. Int. 2014, 60, 95-107.

525

(7)

526

Metabolomic approaches for geographical origin discrimination of hazelnuts (Corylus

527

avellana) by UPLC-QTOF-MS. J. Agric. Food Chem. 2016, 64, 9253-9262.

528

(8)

529

mass spectrometry and related techniques. J. Chromatogr. A. 2012, 1259, 3-15.

530

(9)

531

spectrometry in food-safety and quality analysis. TrAC, Trends Anal. Chem. 2013, 52, 170-

532

185.

533

(10)

534

flight mass spectrometry for the determination of pesticide residues in comparison with an

FAOstat Food and Agriculture Organization of the United Nations, Agricultural data

Pallottino, F.; Menesatti, P.; Costa, C.; Paglia, G.; De Salvador, F. R.; Lolletti, D.

Ernst, M.; Silva, D. B.; Silva, R. R.; Vencio, R. Z.; Lopes, N. P. Mass spectrometry in

Cubero-Leon, E.; Peñalver, R.; Maquet, A. Review on metabolomics for food

Klockmann, S.; Reiner, E.; Bachmann, R.; Hackl, T.; Fischer, M. Food fingerprinting:

Holcapek, M.; Jirasko, R.; Lisa, M. Recent developments in liquid chromatography-

Wang, X.; Wang, S.; Cai, Z. The latest developments and applications of mass

Picó, Y.; Blasco, C.; Farré, M.; Barcelo, D. Analytical utility of quadrupole time-of-


Page 23 of 36


535

optimized column high-performance liquid chromatography/tandem mass spectrometry

536

method. J. AOAC Int. 2009, 92, 734-744.

537

(11)

538

Dossena, A.; Drahos, L.; Vekey, K. Applications of liquid chromatography-mass

539

spectrometry for food analysis. J. Chromatogr. A. 2012, 1259, 74-85.

540

(12)

541

reaction monitoring for quantitative biomarker analysis in proteomics and metabolomics. J.

542

Chromatogr. B: Anal. Technol. Biomed. Life Sci. 2009, 877, 1229-1239.

543

(13)

544

Chemical Industry Press: Beijing, 2014.

545

(14)

546

fingerprinting for authentication of food in official control. Food Res. Int. 2014, 60, 189-204.

547

(15)

548

BioSyst. 2012, 8, 470-481.

549

(16)

550

spectrometry for food safety, quality and traceability. TrAC, Trends Anal. Chem. 2013, 52,

551

74-87.

552

(17)

553

safety and quality of plant-derived food. Food Res. Int. 2013, 54, 1172-1183.

554

(18)

555

metabolome analysis of fresh tobacco leaves in different geographical regions using liquid

556

chromatography-mass spectrometry. Anal. Bioanal. Chem. 2015, 407, 5009-5020.

557

(19)

558

orange origin discrimination by ultra-high performance liquid chromatography coupled to

559

quadrupole time-of-flight mass spectrometry. Food Chem. 2014, 157, 84-93.

Di Stefano, V.; Avellone, G.; Bongiorno, D.; Cunsolo, V.; Muccilli, V.; Sforza, S.;

Kitteringham, N. R.; Jenkins, R. E.; Lane, C. S.; Elliott, V. L.; Park, B. K. Multiple

Qi, X.; Chen, X.; Wang, Y., Plant Metabolomics: Methods and applications.

Esslinger, S.; Riedl, J.; Fauhl-Hassek, C. Potential and limitations of non-targeted

Zhou, B.; Xiao, J. F.; Tuli, L.; Ressom, H. W. LC-MS-based metabolomics. Mol.

Castro-Puyana, M.; Herrero, M. Metabolomics approaches based on mass

Oms-Oliu, G.; Odriozola-Serrano, I.; Martín-Belloso, O. Metabolomics for assessing

Li, L.; Lu, X.; Zhao, J.; Zhang, J.; Zhao, Y.; Zhao, C.; Xu, G. Lipidome and

Diaz, R.; Pozo, O. J.; Sancho, J. V.; Hernandez, F. Metabolomic approaches for



Page 24 of 36

560

(20)

Bondia-Pons, I.; Savolainen, O.; Törrönen, R.; Martinez, J. A.; Poutanen, K.;

561

Hanhineva, K. Metabolic profiling of goji berry extracts for discrimination of geographical

562

origin by non-targeted liquid chromatography coupled to quadrupole time-of-flight mass

563

spectrometry. Food Res. Int. 2014, 63, 132-138.

564

(21)

565

according to variety and geographical origin based on polyphenolic profiles. J. Agric. Food

566

Chem. 2013, 61, 6949-6963.

567

(22)

568

MS/MS analysis of phenols for classification of red wine according to geographic origin,

569

grape variety and vintage. Food Chem. 2010, 122, 366-372.

570

(23)

571

M. Determination of six steviol glycosides of Stevia rebaudiana (Bertoni) from different

572

geographical origin by LC–ESI–MS/MS. Food Chem. 2013, 141, 745-753.

573

(24)

574

Health and Human Services, FDA, Center for Drug Evaluation and Research, Rockville, MD,

575

2001.

576

(25)

577

New York, 1964.

578

(26)

579

coefficient variation. Econometrica. 1979, 47, 1287-1294.

580

(27)

581

oils in hazelnuts (Corylus avellana L.) grown in New Zealand. J. Am. Oil Chem. Soc. 1997,

582

74, 755-759.

583

(28)

584

hazelnut oil composition with comparisons to other vegetable oils, particularly olive oil. Eur.

585

Food Res. Technol. 2003, 218, 13-19.

Guo, J.; Yue, T.; Yuan, Y.; Wang, Y. Chemometric classification of apple juices

Jaitz, L.; Siegl, K.; Eder, R.; Rak, G.; Abranko, L.; Koellensperger, G.; Hann, S. LC–

Montoro, P.; Molfetta, I.; Maldini, M.; Ceccarini, L.; Piacente, S.; Pizza, C.; Macchia,

U.S. FDA. Guidance for Industry: Bioanalytical method validation; US Department of

Mandel, J., The statistical analysis of experimental data. Dover Publications, INC.:

Breusch, T. S.; Pagan, A. R. A simple test for heteroscedasticity and random

Savage, G. P.; McNeil, D. L.; Dutta, P. C. Lipid composition and oxidative stability of

Benitez-Sánchez, P. L.; León-Camacho, M.; Aparicio, R. A comprehensive study of


Page 25 of 36


586

(29)

Alasalvar, C.; Amaral, J. S.; Satır, G.; Shahidi, F. Lipid characteristics and essential

587

minerals of native Turkish hazelnut varieties (Corylus avellana L.). Food Chem. 2009, 113,

588

919-925.

589

(30)

590

composition, tocols profile, and oxidative stability of some fully refined edible oils. Int. J.

591

Food Prop. 2015, 18, 2064-2076.

592

(31)

593

M. F. Determination of lipid and phenolic fraction in two hazelnut (Corylus avellana L.)

594

cultivars grown in Poland. Food Chem. 2015, 168, 615-622.

595

(32)

596

the most represented hazelnut varieties cultivated in Croatia. Ital. J. Food Sci. 2014, 26, 24-

597

30.

598

(33)

599

(Corylus avellana L.) varieties cultivated in Turkey. Food Chem. 2006, 99, 509-515.

600

(34)

601

and tocotrienol content of hazelnut cultivars grown in Portugal. J. Agric. Food Chem. 2006,

602

54, 1329-1336.

603

(35)

604

various tree nuts and peanuts. J. Agric. Food Chem. 2013, 61, 11636-11644.

605

(36)

606

and Purpose. Am. Stat. 2016, 70, 129-133.

607

(37)

608

classification in chemometrics. Crit. Rev. Anal. Chem. 2006, 36, 177-188.

609

(38)

610

dimension reduction and statistical analysis. TrAC, Trends Anal. Chem. 2011, 30, 827-841.

Ayyildiz, H. F.; Topkafa, M.; Kara, H.; Sherazi, S. T. H. Evaluation of fatty acid

Ciemniewska-Zytkiewicz, H.; Verardo, V.; Pasini, F.; Brys, J.; Koczon, P.; Caboni,

Vujević, P.; Petrović, M.; Vahčić, N.; Milinović, B.; Čmelik, Z. Lipids and minerals of

Köksal, A. Đ.; Artik, N.; Şimşek, A.; Güneş, N. Nutrient composition of hazelnut

Amaral, J. S.; Casal, S.; Alves, M. R.; Seabra, R. M.; Oliveira, B. P. P. Tocopherol

Esche, R.; Müller, L.; Engel, K.-H. Online LC-GC-based analysis of minor lipids in

Wasserstein, R. L.; Lazar, N. A. The ASA's Statement on p-Values: Context, Process,

Xu, Y.; Zomer, S.; Brereton, R. G. Support vector machines: A recent method for

Liland, K. H. Multivariate methods in metabolomics – from pre-processing to



Page 26 of 36

611

(39)

Locatelli, M.; Coisson, J. D.; Travaglia, F.; Bordiga, M.; Arlorio, M. Impact of

612

roasting on identification of hazelnut (Corylus avellana L.) origin: A chemometric approach.

613

J. Agric. Food Chem. 2015, 63, 7294-7303.

614

(40)

615

composition of Turkish hazelnut varieties (Corylus avellana L.). Int. J. Food Sci. Nutr. 2010,

616

61, 630-642.

Alasalvar, C.; Pelvan, E.; Topal, B. Effects of roasting on oil and fatty acid

617

618


Page 27 of 36


FIGURE CAPTIONS Figure 1: Metabolite levels of authentic hazelnut samples in mg/kg Figure 2: PCA scores plots of authentic hazelnut samples, A) PC1 vs. PC2 and B) PC1 vs. PC3 Figure 3: P-values of T-test calculations between country pairs for each metabolite Figure 4: Bar charts with confidence intervals of scaled values for authentic hazelnut samples Figure 5: Spider diagrams of authentic and confectionary hazelnut samples for each country displayed as scaled values



Page 28 of 36

TABLES Table 1: Chromatographic and mass spectrometric information of key metabolites for geographical origin discrimination of hazelnuts metabolite

sum formula

adduct ion

Rt [min]

precursor [Da]

Qnt [Da]

Qal [Da]

γ-tocopherol

C28H48O2

+H

2.60

417.646

151.1

69.1

PC(18:2/0:0)

C26H50NO7P

+H

1.19

520.353

184.2

104.2

PC(18:1d7/0:0)

C26H45D7NO7P

+H

1.28

529.358

184.2

104.2

DG(16:0/16:1)

C35H66O5

+NH4

3.49

584.525

311.4

313.4

DG(15:0/18:1d7)

C36H61D7O5

+NH4

3.81

605.431

570.6

299.3

DG(16:0/18:1)

C37H70O5

+NH4

4.26

612.607

339.4

313.5

DG(18:2/18:2)

C39H68O5

+NH4

3.38

634.520

377.3

95.2

DG(18:1/18:1)

C39H72O5

+NH4

4.43

638.640

399.5

83.2

TG(2:0/18:2/18:2)

C41H70O6

+NH4

4.12

676.551

379.4

599.5

TG(2:0/18:1/18:2)

C41H72O6

+NH4

4.73

678.570

381.3

379.3

PE(15:0/18:1d7)

C38H67D7NO8P

+H

2.44

711.571

570.6

57.1

PE(16:0/18:2)

C39H74NO8P

+H

2.41

716.581

575.6

81.2

PE(18:2/18:2)

C41H74NO8P

+H

2.26

740.647

599.5

95.2

PE(18:1/18:1)

C41H78NO8P

+H

2.82

744.470

603.6

69.0

PC(15:0/18:1d7)

C41H73D7NO8P

+H

2.43

753.592

184.2

86.2

PC(16:0/18:3)

C42H78NO8P

+H

2.13

756.550

184.2

86.1

PC(16:0/18:2)

C42H80NO8P

+H

2.41

758.635

184.3

86.2

PC(18:2/18:2)

C44H80NO8P

+H

2.25

782.596

184.2

86.1

PC(18:1/18:2)

C44H82NO8P

+H

2.48

784.596

184.2

86.1

TG(14:0/16:0/18:1)

C51H96O6

+NH4

8.19

822.755

549.5

577.5

TG(15:0/18:1d7/15:0)

C51H89D7O6

+NH4

8.15

829.782

570.6

523.5

TG(15:0/16:0/18:1)

C52H98O6

+NH4

6.76

836.771

563.5

537.5

TG(16:0/16:1/18:1)

C53H98O6

+NH4

8.23

848.770

575.5

81.1

TG(16:0/18:1/18:1)1

C55H102O6

+NH4

8.45

874.786

95.1

81.0

TG(16:0/18:1/18:2)1

C55H100O6

+NH4

8.26

876.801

81.0

95.2

TG(17:1/18:1/18:2)

C56H100O6

+NH4

8.20

886.787

587.5

589.5

TG(18:2/18:2/18:3)

C57H96O6

+NH4

7.76

894.756

597.5

599.5

Rt: retention time; Qnt: Quantifier; Qal: Qualifier; PC: Phosphatidylcholine; DG: Diacylglycerol; TG: Triacylglycerol; PE: Phosphatidylethanolamine 1 only used for method validation


Page 29 of 36


Table 2: Validation parameters for base calibration analyte

regression equation

R2

PC(18:2/0:0)

y = 2.61E+05 + 6.09E+04

0.9976 0.5773 – 9.6219

1.3634

DG(16:0/18:1)

y = 8.24E+04 – 1.78E+05

0.9999 3.3600 – 84.0000

3.5744

DG(18:2/18:2)

y = 8.98E+04 – 3.08E+04

0.9971 1.9449 – 58.3468

8.6852

DG(18:1/18:1)

y = 6.56E+04 – 4.21E+04

0.9961 2.3020 – 92.0800

30.2737

PE(16:0/18:2)

y = 9.99E+04 – 4.64E+04

0.9992 2.1487 – 21.4869

4.2465

PE(18:2/18:2)

y = 1.01E+05 + 3.21E+02

0.9967 0.0400 – 20.0028

2.4762

PE(18:1/18:1)

y = 7.93E+04 – 4.40E+03

0.9996 0.3978 – 59.6750

3.6945

PC(16:0/18:2)

y = 4.82E+05 + 5.18E+04

0.9989 0.3997 – 19.9872

1.6968

PC(18:2/18:2)

y = 4.68E+05 + 4.35E+04

0.9989 0.3996 – 19.9785

1.5106

γ-tocopherol

y = 1.37E+02 + 3.13E+03

0.9981 14.9995 – 74.9976 81.4573

TG(16:0/18:1/18:1)1

y = 5.85E+04 + 1.92E+05

0.9974 6.0187 - 20.0623

978.4892

1

TG(16:0/18:1/18:2)

y = 8.03E+04 + 1.63E+05

0.9963 4.2413 - 21.2063

733.9936

PC(18:1d7/0:0)

y = 4.13E+04 + 4.50E+03

0.9982 0.1184 - 23.6790

0.2000

DG(15:0/18:1d7)

y = 3.06E+04 - 2.64E+03

0.9979 0.1065 - 106.4625 0.4072

linear range [µM]

LOD [nM]

PE(15:0/18:1d7)

y = 3.28E+04 + 6.48E+03

0.9980 0.0900 - 17.9946

0.3735

PC(15:0/18:1d7)

y = 7.25E+04 - 2.45E+02

0.9986 0.0499 - 33.2476

0.1924

0.9982 0.0315 - 6.2999

0.3839

TG(15:0/18:1d7/15:0) y = 1.69E+05 - 2.15E+03 1

only used for method validation



Page 30 of 36

Table 3: Average metabolite levels for authentic hazelnut samples with confidence intervals (α = 0.05) in mg/kg PC(18:2/0:0) Germany 5.00 ± 0.28

DG(16:0/16:1) 1.03 ± 0.05

DG(16:0/18:1) 470.10 ± 25.02

DG(18:2/18:2) 1633.95 ± 73.10

DG(18:1/18:1) 2052.64 ± 86.24

France

3.86 ± 0.09

1.34 ± 0.03

690.28 ± 13.93

1236.34 ± 29.41

3044.35 ± 50.27

Georgia

2.26 ± 0.19

1.98 ± 0.10

917.22 ± 33.51

500.45 ± 36.56

3911.33 ± 108.52

Italy

2.48 ± 0.12

1.59 ± 0.07

854.77 ± 29.74

780.03 ± 45.66

3779.80 ± 73.22

Spain

1.87 ± 0.23

1.87 ± 0.27

992.95 ± 117.71

966.65 ± 134.04

4295.52 ± 234.07

Turkey

1.76 ± 0.07

1.78 ± 0.08

953.89 ± 48.02

535.37 ± 33.34

4350.25 ± 267.33

TG(2:0/18:1/18:2) 1295.90 ± 57.96

PE(16:0/18:2) 127.12 ± 5.44

PE(18:2/18:2) 143.63 ± 6.12

PE(18:1/18:1) 491.01 ± 18.34

TG(2:0/18:2/18:2) Germany 394.27 ± 23.51 France

220.57 ± 9.76

893.69 ± 29.03

92.87 ± 1.64

86.09 ± 2.45

602.82 ± 10.68

Georgia

78.99 ± 13.25

423.13 ± 55.02

43.22 ± 4.89

23.11 ± 3.25

621.24 ± 30.50

Italy

71.79 ± 9.10

398.00 ± 35.82

58.98 ± 2.53

39.88 ± 2.45

650.24 ± 20.02

Spain

154.89 ± 29.79

711.24 ± 110.87

53.48 ± 8.77

32.69 ± 5.06

569.65 ± 31.05

Turkey

103.22 ± 15.61

505.25 ± 53.06

41.14 ± 2.41

22.62 ± 2.65

693.27 ± 31.34

PC(16:0/18:2) 112.98 ± 4.33

PC(18:2/18:2) 177.01 ± 7.90

PC(18:1/18:2) 258.52 ± 5.78

TG(14:0/16:0/18:1) 99.02 ± 4.66

PC(16:0/18:3) Germany 4.95 ± 0.24 France

3.74 ± 0.08

85.90 ± 1.47

110.08 ± 2.97

230.97 ± 2.84

132.87 ± 3.86

Georgia

1.64 ± 0.20

49.87 ± 5.16

34.54 ± 5.09

129.44 ± 8.58

192.40 ± 18.76

Italy

2.20 ± 0.11

60.17 ± 2.55

54.46 ± 2.85

168.63 ± 4.33

182.41 ± 12.35

Spain

1.82 ± 0.39

53.85 ± 8.92

45.21 ± 8.58

144.35 ± 14.13

208.90 ± 24.71

Turkey

1.41 ± 0.14

40.24 ± 2.41

29.06 ± 3.29

120.27 ± 5.92

119.70 ± 7.08

TG(15:0/16:0/18:1) TG(16:0/16:1/18:1) TG(17:1/18:1/18:2) TG(18:2/18:2/18:3) γ-tocopherol Germany 18.62 ± 0.99 3658.28 ± 150.83 572.76 ± 26.86 1082.30 ± 64.11 < LLOQ France

20.90 ± 0.40

3505.22 ± 78.31

439.32 ± 10.00

730.32 ± 18.78

< LLOQ

Georgia

13.31 ± 0.66

3708.43 ± 434.26

292.88 ± 24.51

459.13 ± 34.98

< LLOQ

Italy

15.84 ± 0.98

3198.68 ± 160.06

262.15 ± 15.82

460.93 ± 30.57

< LLOQ

Spain

14.76 ± 1.28

4233.26 ± 538.53

338.35 ± 47.09

688.94 ± 104.53

< LLOQ

Turkey

15.86 ± 0.80

2533.94 ± 124.89

346.11 ± 25.45

692.57 ± 56.16

< LLOQ


Page 31 of 36


FIGURES Figure 1



Page 32 of 36

Figure 2


Page 33 of 36


Figure 3



Page 34 of 36

Figure 4


Page 35 of 36


Figure 5



Page 36 of 36

TABLE OF CONTENT


Food Targeting: Geographical Origin Determination of Hazelnuts

Recommend Documents