A rapid and non-destructive method for simultaneous determination of

Subscriber access provided by UNIV OF LOUISIANA

Food Safety and Toxicology

A rapid and non-destructive method for simultaneous determination of aflatoxigenic fungus and aflatoxin contamination on corn kernels Feifei Tao, Haibo Yao, Fengle Zhu, Zuzana Hruska, Yongliang Liu, Kanniah Rajasekaran, and Deepak Bhatnagar J. Agric. Food Chem., Just Accepted Manuscript • DOI: 10.1021/acs.jafc.9b01044 • Publication Date (Web): 15 Apr 2019 Downloaded from http://pubs.acs.org on April 16, 2019

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 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 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.

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 29

Journal of Agricultural and Food Chemistry

1

A rapid and non-destructive method for simultaneous determination of aflatoxigenic

2

fungus and aflatoxin contamination on corn kernels

3 4

Feifei Taoa, Haibo Yaoa, *, Fengle Zhua, Zuzana Hruskaa, Yongliang Liub, Kanniah Rajasekaranb, Deepak

5

Bhatnagarb

6 7 8 9

a Geosystems

Research Institute, Mississippi State University

Building 1021, Stennis Space Center, MS 39529, USA

10 11

b USDA-ARS,

Southern Regional Research Center, New Orleans, LA 70124, USA

12 13 14 15 16

*Contact

Information for Corresponding Author:

17

Haibo Yao

18

Geosystems Research Institute, Mississippi State University

19

1021 Balch Blvd., Stennis Space Center, MS 39529

20

E-mail: [email protected]

21

Phone: +1-228-688-3742

22 23 24 25 26 27

ACS Paragon Plus Environment


28

ABSTRACT: Conventional methods for detecting aflatoxigenic fungus and aflatoxin contamination are generally

29

time-consuming, sample-destructive and require skilled personnel to perform, making them impossible for large-

30

scale non-destructive screening detection, real-time and on-site analysis. Therefore, the potential of visible-near

31

infrared (Vis-NIR) spectroscopy over the 400-2500 nm spectral range was examined for determination of

32

aflatoxigenic fungus infection and the corresponding aflatoxin contamination on corn kernels, in a rapid and non-

33

destructive manner. The two A. flavus strains, AF13 and AF38 were used to represent the aflatoxigenic fungus and

34

non-aflatoxigenic fungus, respectively, for artificial inoculation on corn kernels. The partial least squares

35

discriminant analysis (PLS-DA) models based on different combinations of spectral range (I: 410-1070 nm; II:

36

1120-2470 nm), corn side (endosperm or germ side), spectral variable number (full spectra or selected variables),

37

modeling approach (two-step or one-step) and classification threshold (20 or 100 ppb) were developed and their

38

performance was compared. The first study focusing on detection of aflatoxigenic fungus-infected corn kernels

39

showed that, in classifying the “control+AF38-inoculated” and AF13-inoculated corn kernels, the full spectral PLS-

40

DA models using the preprocessed spectra over range II and one-step approach yielded more accurate prediction

41

results than using the spectra over range I and the two-step approach. The advantage of the full spectral PLS-DA

42

models established using one corn side than the other side was not consistent in the explored combination cases. The

43

best full spectral PLS-DA model obtained was obtained using the germ-side spectra over range II with the one-step

44

approach, which achieved an overall accuracy of 91.11%. The established CARS-PLSDA models performed better

45

than the corresponding full-spectral PLS-DA models, with the better model achieved an overall accuracy of 97.78%

46

in separating the AF13-inoculated corn kernels and the uninfected control and AF38-inoculated corn kernels. The

47

second study focusing on the detection of aflatoxin-contaminated corn kernels showed that, based on the aflatoxin

48

threshold of 20 ppb and 100 ppb, the best overall accuracy in classifying the aflatoxin-contaminated and healthy

49

corn kernels attained 86.67% and 84.44%, respectively, using the CARS-PLSDA models. The quantitative modeling

50

results using partial least squares regression (PLSR) obtained the correlation coefficient of prediction set (RP) of

51

0.91, indicating the possibility of using Vis-NIR spectroscopy to quantify aflatoxin concentration in aflatoxigenic

52

fungus-infected corn kernels.

53

KEYWORDS: aflatoxigenic fungus; aflatoxin; corn kernel; partial least squares discriminant analysis; competitive

54

adaptive reweighted sampling; partial least squares regression

55


Page 2 of 29

Page 3 of 29

56


INTRODUCTION

57

Aflatoxins are a group of highly toxic secondary metabolites produced by fungi of the genus Aspergillus,

58

predominantly A. flavus and A. parasiticus [1]. The dominant aflatoxins produced by A. flavus are aflatoxin B1

59

(AFB1) and B2 (AFB2), whereas A. parasiticus produces two additional aflatoxins, G1 (AFG1) and G2 (AFG2) [2].

60

Aflatoxin contamination can occur in a wide variety of agricultural commodities, including corn, rice, wheat,

61

peanuts, almond kernels, pistachio nuts, etc. Various methods have been developed and utilized to determine

62

aflatoxin contamination and fungal infection in foods. The available techniques for detecting aflatoxins include thin

63

layer chromatography (TLC), gas chromatography (GC), high-performance liquid chromatography (HPLC),

64

enzyme-linked immunosorbent assay (ELISA), fluorescence polarization assays, radio immunoassays among others

65

[1, 3, 4],

66

[1].

67

well-equipped laboratory. They are also expensive, time-consuming, destructive to the test samples, and most

68

importantly, subject to sampling error, making them impossible for large-scale non-destructive screening detection

69

or integration in an on-line sorting and production system.

70

and fungal infection in food is traditionally determined using microbiological methods in a laboratory setting

These methods may give highly accurate results in laboratories, however, most require skilled personnel and a

The optical-based methods have demonstrated great potential for real-time evaluation of food quality and safety

71

attributes, in a rapid and non-destructive manner [5, 6]. Considerable number of studies have been conducted using

72

fluorescence spectroscopy, visible-near infrared (Vis-NIR) spectroscopy, hyperspectral imaging (HSI) and Raman

73

spectroscopy to detect aflatoxin contamination and/or fungal infection in different varieties of foods [1, 7, 8], and their

74

potential have been indicated. However, in most of the studies reported, the samples infected with non-aflatoxigenic

75

and/or nontoxigenic fungi were not involved, which means that their effect on identifying the aflatoxigenic fungi-

76

infected and/or aflatoxin-contaminated commodities were generally neglected in statistical analysis of predictive

77

modeling. In real world, non-aflatoxigenic and nontoxigenic fungi coexist with the aflatoxigenic fungi, in either

78

crops, seeds, or soils [9-12]. Strains of A. flavus also vary greatly in aflatoxin production, with some producing

79

copious amounts and others none [9, 10, 13]. The relative distribution of aflatoxigenic versus non-aflatoxigenic isolates

80

is modulated by many factors including plant species present, soil composition, cropping history, crop management,

81

and environment conditions, including rain fall and temperature [10, 14]. Additionally, non-toxigenic strain of A. flavus

82

have been used as biological control agents on plants to control aflatoxin contamination nowadays [15-17].



83

As far as we know, only a few studies explored the feasibility of optical methods for differentiating

84

aflatoxigenic and non-aflatoxigenic fungi infection on agricultural commodities. Jin et al. [18] reported a study where

85

the HSI system under both halogen and ultraviolet (UV) illumination was employed to classify the toxigenic and

86

atoxigenic strains of A. flavus over the 400-1000 nm spectral range. One aflatoxigenic AF13 and three non-

87

aflatoxigenic AF2038, AF283 and AF38 were included in this study. The authors reported that under the halogen

88

light source, the average detection accuracies of 83% and 74% were obtained in identifying the toxigenic fungus

89

pixels and atoxigenic fungus pixels, respectively; and under the UV light source, the average detection accuracy

90

attained 67% and 85% correspondingly. Subsequently, based on fluorescence HSI, Yao et al. [19] detected aflatoxin

91

contamination in corn kernels which were field-inoculated with AF13 and AF38 fungal strains, separately. Using all

92

data from single kernels which included the non-treated controls, AF13- and AF38-inoculated samples, the overall

93

accuracies of 77.2% and 78.9% were obtained with the 20-ppb threshold, from the endosperm and germ sides,

94

respectively, and 91.7% and 94.4% correspondingly with the classification threshold of 100 ppb. The linear

95

discriminant analysis (LDA) models established in this study achieved high accuracies with all samples in

96

identifying the healthy corn kernels, while the accuracies in identifying the aflatoxin-contaminated corn kernels

97

were lower than 60%. Further endeavors are still needed to clarify the issue mentioned above.

98 99

Therefore, in the present study, the two A. flavus strains, AF13 and AF38 were used to represent the aflatoxigenic fungus and non-aflatoxigenic fungus, respectively, for artificial inoculation on single corn kernels. The

100

main objective of this study were: 1) to investigate the potential of using Vis-NIR spectroscopy to distinguish corn

101

kernels infected with aflatoxigenic fungus from the uninfected and non-aflatoxigenic fungus-infected kernels; 2) to

102

differentiate the aflatoxin-contaminated corn kernels from healthy kernels, as the presence of A. flavus in

103

commodities is not always associated with aflatoxin contamination; 3) to determine the characteristic wavelengths

104

using the competitive adaptive reweighted sampling (CARS) algorithm, for identifying corn kernels infected with

105

aflatoxigenic fungus or contaminated with aflatoxins, and establish the simplified classification models using the

106

determined characteristic wavelengths; 4) to quantify the aflatoxin concentration of aflatoxigenic fungus-infected

107

corn kernels.

108

MATERIALS AND METHODS


Page 4 of 29

Page 5 of 29


109

Sample preparation. Two strains of A. flavus, AF13 (aflatoxigenic) and AF38 (non-aflatoxigenic) were used as

110

inoculum separately for artificial laboratory inoculations. The fungal strains AF13 and AF38 were obtained from the

111

Food and Feed Safety Research Unit, Agricultural Research Service (ARS), USDA (New Orleans, LA). Both

112

isolates were cultured separately on potato dextrose agar (PDA) medium in plastic Petri dishes at 30 ̊C in a dark

113

incubator. After 5 days of growth, the conidia were harvested, and suspended in sterile distilled water at a dilution of

114

5×106 conidia/mL, determined with a hemocytometer.

115

Corn kernels (N78B-GT, Syngenta NK Brand Seeds, Laurinburg, NC, USA) were harvested in 2010 from the

116

ARS Field Station in Stoneville, MS, USA. The harvested kernels were dried to below 15% moisture content and

117

kept at room temperature (22 ̊C) before use. In all 180 corn kernels were randomly selected from a single year’s

118

field experiment and used with 3 treatments, namely, 60 kernels inoculated with the AF13 fungal strain, 60 kernels

119

inoculated with the AF38 strain, and 60 kernels inoculated with sterile distilled water as control. Before inoculation,

120

all corn kernels were surface sterilized in 70% ethanol and then rinsed in three changes of distilled water. Kernels in

121

each treatment group were inoculated by immersion and stirred for 1 min. Each group of kernels was incubated

122

separately in a humidity chamber using a plastic tray. Distilled water was added every other day to each incubation

123

chamber in order to maintain constant moisture. After 8 days, all samples were taken out from the incubator, and

124

transferred to individually labeled coin envelopes (1 kernel/envelop) and placed in a 60 °C oven for two days to

125

terminate fungal growth. For following spectroscopic and chemical analyses, all kernels were surface wiped to

126

remove all exterior signs of mold.

127

Vis-NIR spectroscopy and spectral acquisition. The Foss XDS rapid content analyzer (Foss NIRSystems Inc.,

128

Laurel, MD) which covers the spectral range of 400-2500 nm was utilized to scan the corn kernels in this study. This

129

is a dual-detector system including both the silicon (Si) and lead sulphide (PbS) detectors, which cover the spectral

130

ranges of 400-1100 and 1100-2500 nm, respectively. A sample holder made of optical-grade Spectralon diffuse

131

reflectance material (> 99% diffuse reflectance), with a 6×8 mm triangle hole in the middle was used for collecting

132

spectra from a constant area of each kernel. A total of 180 corn kernels from 3 groups were scanned in the

133

reflectance mode with the data recorded as the absorbance of log (1/R) at the spectral interval of 0.5 nm. Each kernel

134

was scanned from the endosperm and germ sides separately. The scanning of each side was performed by placing

135

the kernel side being tested over the triangle hole of the sample holder, facing the detector. The data acquisition rate

136

was 2 scans/s, and each side of the corn samples was scanned with 32 scans.



137

Chemical analysis. After spectroscopic scanning, each whole corn kernel was subjected to standard chemical

138

analysis for determination of its reference aflatoxin concentration. The chemical procedures for single kernel

139

analysis were adjusted from the AflaTest protocol (VICAM, Milford, Mass.), which is one of the USDA-approved

140

laboratory methods for aflatoxin determination. The detailed information for the chemical analysis method can be

141

found in Yao et al. [19].

142

Data analysis. In this study, both the qualitative and quantitative analyses were conducted, based on the standard

143

normal variate (SNV)-preprocessed absorbance spectra. Figure S1 demonstrates the data-processing procedures

144

employed for qualitative/classification analyses. The absorbance spectra collected from the endosperm and germ

145

sides of corn samples were processed separately to study the effect of corn sides on model performance. As the Vis-

146

NIR spectroscopy employed in this study is a dual-detector system, and the signal-to-noise ratio (SNR) at two ends

147

of each detector coverage range might be low, only the absorbance spectra over the spectral ranges of 410-1070 and

148

1120-2470 nm were used for the following data analysis. To simplify the descriptions, the symbols I and II were

149

used to represent the spectral ranges of 410-1070 and 1120-2470 nm, respectively. In addition, as in the U.S., the

150

maximum limits for total aflatoxins (AFB1, AFB2, AFG1, AFG2) are regulated by the U.S. Food and Drug

151

Administration (FDA) and are 20 ppb and 100 ppb for corn foods and corn feed products, respectively [20], both

152

thresholds were used to develop the classification models in this study. The detailed methods for model development

153

and evaluation were described in the following sections.

154

CARS algorithm for spectral variable selection. The variable selection procedure was conducted on the optimal full

155

spectral PLS-DA models, as the full spectra over the ranges I and II each includes a total of 1321 and 2701 spectral

156

variables and some data may be noise or contain irrelevant information that could worsen the results of the

157

calibration models. Therefore, in order to reduce the data dimensionality and simplify the complexity of

158

computation, the CARS algorithm based on the simple but effective principle ‘survival of the fittest’ was utilized to

159

select the optimal combinations of spectral variables in this study.

160

CARS works through selecting N subsets of variables by N sampling runs in an iterative manner and finally

161

chooses the subset with the lowest root mean square error of cross validation (RMSECV) or error rate value as the

162

optimal subset [21]. The optimal wavelengths refer to the wavelengths with large absolute coefficients in a multivariate

163

linear regression model, such as partial least squares (PLS). In each sampling run, CARS works in four steps: (1)


Page 6 of 29

Page 7 of 29


164

Monte Carlo for model sampling, which can be regarded as sampling in the model space combined with Monte Carlo

165

strategy; (2) employ exponentially decreasing function (EDF) to perform enforced wavelength selection; (3) adopt

166

adaptive reweighted sampling to realize a competitive selection of wavelengths; and (4) cross validation is utilized to

167

evaluate the subset [21]. In this study, the lowest error rate of 10-fold cross validation was used to evaluate the subset,

168

and the number (N) of Monte Carlo sampling runs was set to be 500.

169

Development and evaluation of classification models. The classification models were developed using the partial

170

least squares discriminant analysis (PLS-DA) method in this study. PLS-DA is a linear classification method that is

171

based on the well-known PLS regression. The detailed descriptions of this method are referenced elsewhere [22], and

172

therefore not shown here. The 10-fold cross validation method was conducted to determine the optimal number of

173

latent variables (LVs) of each model, and the LVs with the lowest error rate in cross validation were employed to

174

establish the PLS-DA classification model. Ten-fold cross validation is a common choice in k-fold cross validation

175

with k=10. The k-fold cross validation method is a popular procedure for estimating the performance of a

176

classification algorithm which randomly divides a data set into k disjoint folds with approximately equal size, and

177

each fold is in turn used to test the model induced from the other k−1 folds by a classification algorithm [23].

178

As two detection objectives are covered in this study, namely, a) identifying corn kernels infected with

179

aflatoxigenic fungus and b) identifying aflatoxin-contaminated corn kernels, their classification models were

180

developed and evaluated, separately (Figure S1). a) First, in separating the corn kernels infected with aflatoxigenic

181

fungus (AF13) from the uninfected and non-aflatoxigenic fungus-infected (AF38) corn kernels, both two-step and

182

one-step approaches were applied. As its name implies, the two-step approach logically resolves the detection

183

objective into two steps, based on the concept of decomposition in hierarchical classification. The first step of the

184

two-step method refers to classifying the fungus-infected (AF38- and AF13-inoculated) and uninfected corn kernels;

185

and the second step refers to classifying the corn kernels infected with aflatoxigenic fungus (AF13-inoculated) and

186

non-aflatoxigenic fungus (AF38-inoculated). Different from the two-step method, the one-step method refers to

187

establishing the classification models directly, in one step, indicating that when building the classification models,

188

the AF13-inoculated corn kernels were treated as one class, and all other corn kernels including the uninfected

189

control and AF38-inoculated kernels were treated as the other class. Therefore, in identifying the AF13-inoculated

190

corn kernels from all other kernels, a total of 2× 2× 2 (endosperm-side/germ-side spectra× spectral range I/II× one-

191

step/two-step method) types of full spectral PLS-DA models were established. Based on the optimal cases of the full



Page 8 of 29

192

spectral models which were developed from the endosperm and germ sides of corn kernels, separately, the CARS

193

algorithm was conducted to find informative spectral variables for identifying AF13 infection on corn kernels for

194

these cases. Using the determined spectral variables by CARS, the simplified CARS-PLSDA models were

195

developed correspondingly.

196

b) In identifying aflatoxin-contaminated corn kernels from healthy kernels, a total of 2× 2× 2 (endosperm-

197

side/germ-side spectra× spectral range I/II× 20-ppb threshold/100-ppb threshold) types of full spectral PLS-DA

198

models were established first. Then, based on the optimal cases of the full spectral models which were developed

199

from the endosperm and germ sides of corn kernels with the classification threshold of 20 and 100 ppb, the CARS

200

algorithm was conducted to select the informative spectral variables for aflatoxin contamination on corn kernels for

201

these cases. Using the determined spectral variables by CARS, the simplified CARS-PLSDA models were

202

established correspondingly.

203

Generally, in developing the classification models, three fourths of samples from each class were used to

204

establish the classification models, and the other one fourth of samples were used as independent predictions to

205

evaluate the performance of each model. Depending on the classification objective and threshold, the negative

206

samples (class 1) may represent the kernels from the “control+AF38-inoculated” group in objective (a), or kernels

207

with aflatoxin concentration≤ 20 ppb or ≤100 ppb in objective (b). The positive samples (class 2) may represent the

208

kernels from the AF13-inoculated group in objective (a), or kernels with aflatoxin concentration>20 ppb or >100

209

ppb in objective (b). The indices of class accuracy and overall accuracy, which are described in the following

210

equations, were calculated to evaluate the performance of each classification model. The higher the accuracy values,

211

the better predictive ability the classification model has. 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑎𝑠𝑠𝑖𝑔𝑛𝑚𝑒𝑛𝑡𝑠 𝑜𝑓 𝑒𝑎𝑐ℎ 𝑐𝑙𝑎𝑠𝑠 𝑇𝑜𝑡𝑎𝑙 𝑠𝑎𝑚𝑝𝑙𝑒 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑒𝑎𝑐ℎ 𝑐𝑙𝑎𝑠𝑠 𝑡𝑒𝑠𝑡𝑒𝑑

(1)

𝑆𝑢𝑚 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑎𝑠𝑠𝑖𝑔𝑛𝑚𝑒𝑛𝑡𝑠 𝑜𝑓 𝑎𝑙𝑙 𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑇𝑜𝑡𝑎𝑙 𝑠𝑎𝑚𝑝𝑙𝑒 𝑛𝑢𝑚𝑏𝑒𝑟

(2)

212

𝐶𝑙𝑎𝑠𝑠 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =

213

𝑂𝑣𝑒𝑟𝑎𝑙𝑙 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =

214

Lastly, in objective (a), the final overall accuracy of the two-step method in distinguishing the AF13-inoculated

215

kernels from the control and AF38-inoculated kernels is the multiplied result of overall accuracies obtained from the

216

above two individual steps. The final class accuracy for identifying the AF13- and AF38-inoculated kernels is the

217

multiplied result of the class accuracy obtained in identifying them in the second step and the class accuracy in


Page 9 of 29


218

identifying the fungus-infected kernels in the first step, separately. This is because they are based on the first-step

219

result in identifying the fungus-infected corn kernels. The final accuracy for predicting the uninfected control

220

kernels is equal to the accuracy in the first step since the control kernels are not involved in the second step.

221

Development and evaluation of quantitative models. To explore the feasibility of the Vis-NIR spectroscopy in

222

quantifying aflatoxin concentration of the corn kernels infected with aflatoxigenic fungus, the partial least squares

223

regression (PLSR) models were also established in this study. The SNV-preprocessed absorbance spectra were used

224

for establishing the PLSR models, and 2×2 (endosperm-side/germ-side spectra× spectral range I/II) cases were

225

covered in this section. Among the 60 AF13-inoculated corn kernels, 45 samples were employed to establish the

226

PLSR models, and the other 15 samples were used to evaluate the model performance. The optimal LVs number of

227

the PLSR model was determined by the leave-one-out cross validation (LOOCV) method, and the LVs with the

228

minimal standard error of cross validation (SECV) were used for model development. The LOOCV algorithm is a

229

particular case of k-fold cross validation where k is equal to the total number of samples trained. This method was

230

chosen for validating the PLSR models is because a relatively small data set was included in this process. The

231

performance of the established quantitative models was estimated by the indices of correlation coefficient (RP) and

232

root mean squared error (RMSEP) of prediction set. The higher RP and the lower RMSEP are, the better predictive

233

ability the model has.

234

RESULTS AND DISCUSSION

235

Chemical analysis results. According to the reference chemical data, the mean aflatoxin concentration of the 60

236

AF13-inoculated corn kernels was 599.03 ppb, indicating the aflatoxin-producing ability of the AF13 strain on corn

237

kernels. The maximum and minimum value was 3400.00 and 0.00 ppb, respectively, with a standard deviation (SD)

238

of 749.90 ppb (Table S1). The distribution plots of aflatoxin concentration of single corn kernels from the AF13-

239

inoculated group are shown in Figure 1(a). As observed, most of the corn kernels inoculated with the AF13 fungal

240

strain show aflatoxin contamination, based on both thresholds of 20 ppb and 100 ppb. Among the 60 AF13-

241

inoculated corn kernels, the aflatoxin concentration of 43 single kernels was over 20 ppb, and 36 single kernels

242

showed over-100 ppb aflatoxin concentration.

243 244

Figure 1(b) presents the distribution plots of aflatoxin concentration of the AF38-inoculated corn kernels. As can be observed, all aflatoxin concentrations of the AF13-inoculated corn kernels were low, with most of the



245

concentrations being 0.00 ppb or in the range of (0, 2] ppb. The maximum concentration of the AF38-inoculated

246

corn kernels was 11.00 ppb, which means that the corn kernels inoculated with the AF38 strain did not show over-

247

threshold aflatoxin contamination, based on either the 20-ppb or the 100-ppb threshold. For the control group, 57

248

among the 60 kernels showed 0.00-ppb aflatoxin concentration, with only 3 kernels containing very low amounts of

249

aflatoxins, i.e. 0.01, 0.53 and 3.00 ppb. The reference data of the control group also indicated that the corn kernels

250

used in the present study were not originally contaminated by aflatoxins.

251

Characterization of Vis-NIR spectra. The mean spectra of original absorbance of control, AF38-inoculated and

252

AF13-inoculated corn kernels were plotted from the endosperm and germ side separately and are presented in Figure

253

2. As shown in Figures 2(a) and 2(b), the mean absorbance spectra of the two fungus-inoculated groups are more

254

analogous among all three mean spectra, particularly the spectra acquired from the endosperm side. Obvious spectral

255

differences could be observed between the mean spectra of the control and the fungus-inoculated kernels, either

256

from the endosperm or germ side. More specifically, the mean absorbance from the endosperm side of the fungus-

257

infected corn kernels is higher than the control kernels before 780 nm, and lower over almost the whole NIR range.

258

For the spectra acquired from the germ side, the mean absorbance of the fungus-inoculated corn kernels showed

259

higher intensities than the control kernels over the whole spectral range I, and lower over almost the whole spectral

260

range II. The phenomenon observed here is in accordance with the study reported by Pearson & Wicklow [24].

261

Pearson & Wicklow [24] studied the potential of using Vis-NIR spectroscopy over 500-1700 nm to identify the corn

262

kernels infected by eight different fungi, which included A. flavus, A. niger, Acremonium zeae, Diplodia maydis,

263

Fusarium graminearum, Fusarium verticillioides, Penicillium spp. and Trichoderma viride. The authors found that

264

for all fungus-infected corn kernels with extensive discoloration, except Acremonium zeae, their absorbance is

265

higher than that of undamaged kernels below 700 nm and lower between 900 and 1700 nm. Differences in

266

absorbance spectra between the fungus-infected and sound corn kernels can possibly be explained by the scattering

267

and absorbance characteristics caused by fungal growth and metabolic activities in the kernel. As fungal infections

268

generally cause discoloration of kernels, this would cause higher visible wavelength absorbance [24]. A fungus-

269

infected kernel would also scatter more light than a sound, vitreous kernel, since the invasion of the fungus tends to

270

cause the kernel endosperm to become porous [25, 26]. This scattering would cause less NIR radiation to be absorbed

271

in the reflectance mode. Powdery substances with refractive indices different from that of air, such as those in the

272

air-endosperm interface of infected kernels, cause more light to be reflected [27], as opposed to the more crystalline-


Page 10 of 29

Page 11 of 29


273

like property of normal kernels. Compared to the endosperm-side spectra, the differences in the mean spectra of the

274

AF13-inoculated and AF38-inoculated kernels are more obvious from the germ side, particularly over the spectral

275

range II. However, as the spectral variance among the kernels of each group is great, the differences existing among

276

their group mean spectra is not appropriate to be used as classification criterion for predicting each individual

277

kernel. Therefore, the chemometric technique of PLS-DA was applied for classifications in the following sections.

278

Correlation analysis between absorbance and aflatoxin produced in AF13-inoculated kernels. To better

279

understand the linear relationship between absorbance and aflatoxin concentration of the AF13-inoculated corn

280

kernels, the Pearson correlation coefficient (R) was calculated between them at each wavelength. The correlation

281

coefficients were calculated using the original endosperm-side and germ-side spectra, separately. The obtained

282

correlation coefficients were plotted over the whole spectral range and were shown in Figure 3. Overall, the

283

correlation coefficient profiles calculated using the endosperm-side and germ-side spectra demonstrated analogous

284

features, such as, both profiles showing positive correlation coefficients at short wavebands, and negative

285

correlations at long wavebands. This could be explained by the absorbance differences existing between the control

286

and AF13-inoculated corn kernels (Figure 2), as the amount of aflatoxin produced is directly associated with the

287

growth of AF13 and its metabolic activities. As observed in Figure 2, generally, the mean spectral absorbance of the

288

AF13-inoculated corn kernels is higher than that of the control kernels over short wavebands, and lower over long

289

wavebands; therefore, it is reasonable to infer the positive correlation between the aflatoxin amount and spectral

290

absorbance over short wavebands, and negative correlation over long wavebands. Additionally, both correlation

291

coefficients calculated using either endosperm-side or germ-side spectra, varied greatly over short wavebands, and

292

showed small variations over long wavebands.

293

In detail, the correlation coefficients between the endosperm-side spectral absorbance and aflatoxin

294

concentration are positive before 790 nm. The correlation coefficients calculated using the germ-side spectra are all

295

positive till around 903 nm. Based on the endosperm-side spectra, the maximum positive correlation coefficient of

296

0.60, appeared around 592 nm, and then the coefficient value decreased rapidly. The strongest negative correlation

297

appeared nearby 1415 nm, with the coefficient value of -0.58. Based on the germ-side spectra, the greatest positive

298

correlation coefficient of 0.53, locating nearby 558 nm, was a little smaller, compared to the 0.60 calculated using

299

the endosperm-side spectra. As indicated in Figure 3, after 1132 nm, the negative correlations between the germ-side

300

spectral absorbance and aflatoxin concentration are all slightly stronger than those calculated using the endosperm-



301

side spectra. The strongest negative correlation of -0.64, calculated using the germ-side spectra, appeared around

302

1404 nm, which is nearby the location (1415 nm) of the strongest negative correlation calculated using the

303

endosperm-side spectra.

304

Modeling results in identifying corn kernels infected with aflatoxigenic fungus. Full spectral PLS-DA models.

305

Modeling results with two-step method. In the first step of the two-step method, namely, classifying the uninfected

306

control and fungus-inoculated corn kernels, 135 corn kernels which included 45 uninfected control corn kernels

307

(class 1 in this case) and 90 fungus-inoculated kernels (class 2 in this case, including 45 AF13-inoculated corn

308

kernels and 45 AF38-inoculated kernels) were used to establish the PLS-DA models. The other 45 corn kernels,

309

namely, 15 control kernels and 30 fungus-inoculated kernels (15 AF13-inoculated kernels and 15 AF38-inoculated

310

kernels) were used as prediction samples to evaluate the model performance. In the second step, namely, classifying

311

the AF38- and AF13-inoculated corn kernels, 90 corn kernels which included 45 AF13-inoculated and 45 AF38-

312

inoculated corn kernels were used to develop the classification models, and the other 30 kernels, including 15 AF13-

313

inoculated and 15 AF38-inoculated kernels were employed to evaluate the model performance.

314

The full spectral PLS-DA modeling results for each step of the two-step method are shown in Table S2. For

315

the first step, all four PLS-DA models obtained over 91.00% overall accuracies for the prediction set. Among the

316

four models, the PLS-DA model established using the germ-side spectra over the spectral range II achieved the best

317

overall accuracy of 100.0%. In comparisons, the germ-side spectra performed better than the endosperm-side

318

spectra, and the models established using the spectral range II achieved better prediction results than the spectral

319

range I in classifying the uninfected control and fungus-inoculated corn kernels. Generally, the classification models

320

established performed better in identifying the fungus-inoculated corn kernels than the uninfected control kernels.

321

In the second step of the two-step method, namely, classifying the AF38- and AF13-inoculated corn

322

kernels, the classification models established did not perform as well as in the first step. However, a few overall

323

accuracies obtained for the prediction set were still acceptable. The best overall accuracy of 86.67% was obtained

324

using the endosperm-side spectra over the spectral range II. Using this model, the prediction accuracy in identifying

325

the AF13-inoculated corn kernels achieved 100.00%. In addition, the classification model established using the

326

endosperm-side spectra over range I attained an overall accuracy of 80.00% for the prediction set. However, the

327

classification models established using the germ-side spectra did not perform as well as the endosperm-side spectra


Page 12 of 29

Page 13 of 29


328

in classifying the AF38- and AF13-inoculated corn kernels, as the better overall accuracy obtained using the germ-

329

side spectra only achieved 73.33%. By comparisons, it was observed that the absorbance spectra over range II

330

performed better than range I in classifying the AF38- and AF13-inoculated corn kernels, which was in accordance

331

with the phenomenon observed in the first step. While, different with the prediction results obtained in the first step,

332

the models established with the endosperm-side spectra yielded more accurate prediction results than the models

333

developed using the germ-side spectra, in classifying the AF38- and AF13-inoculated corn kernels.

334

Table S3 shows the final full spectral PLS-DA modeling results using the two-step method. As can be

335

observed, the best prediction result was obtained with an overall accuracy of 82.82% for the prediction set, using the

336

endosperm-side spectra over range II. Using this model, the prediction accuracy in identifying the AF13-inoculated

337

corn kernels attained 100.00%. While, the final prediction results obtained using other 3 combinations of spectral

338

range and corn side did not perform as well. That might be due to the inconsistency of corn side effect on the model

339

performance of the two individual steps. In other words, during the two-step classifications, the germ-side spectra

340

performed better in classifying the uninfected control and fungus-inoculated corn kernels (step 1), while the

341

endosperm-side spectra yielded more accurate prediction results in classifying the AF38- and AF13-inoculated corn

342

kernels (step 2). However, it was consistent that the models developed over the spectral range II were more accurate

343

than those established over range I.

344

Modeling results with one-step method. Using the one-step method, 135 corn kernels which included 90

345

aflatoxigenic fungus-negative kernels (class 1 in this case, referring to 45 control kernels and 45 AF38-inoculated

346

kernels) and 45 aflatoxigenic fungus-positive corn kernels (class 2 in this case, referring to AF13-inoculated kernels)

347

were used to establish the PLS-DA models. The other 45 corn kernels, namely, 30 aflatoxigenic fungus-negative

348

kernels (15 control kernels and 15 AF38-inoculated kernels) and 15 aflatoxigenic fungus-positive kernels (AF13-

349

inoculated kernels) were used as prediction samples to evaluate the model performance. Table S4 shows the

350

determined LVs number and the one-step PLS-DA modeling results in classifying the “control+AF38-inoculated”

351

and AF13-inoculated corn kernels, using the full spectra collected from the endosperm and germ side, separately.

352

Among all the 4 PLS-DA models established, the best overall accuracy of 91.11% was achieved for the prediction

353

set, using the endosperm-side spectra over the spectral range II. Compared to the prediction results obtained using

354

the two-step method, the overall accuracies obtained using the one-step method are all higher. This may be due to

355

that when using the two-step method, the calculation of overall accuracy takes the accuracies obtained in both steps



356

into account, while the one-step method does not separate the uninfected control and AF38-inoculated corn kernels.

357

That means the classification error in separating the uninfected control and AF38-inoculated corn kernels was not

358

included in the one-step method, but was included in the two-step method.

359

The prediction results of the models established using the endosperm-side spectra over ranges I and II were

360

also acceptable, which achieved the overall accuracy of 84.44% and 86.67%, respectively. Particularly, the model

361

developed based on the endosperm-side spectra over range II yielded a prediction accuracy of 100.00% in

362

identifying the AF13-inoculated corn kernels from the control and AF38-inoculated corn kernels. Overall, the

363

prediction results using the one-step method were more accurate using the spectra over range II than range I, which

364

was in accordance with the final prediction results obtained with the two-step method. However, the advantage of

365

the classification models established using one corn side than the other side was not consistent over both spectral

366

ranges. This might be due to the aforementioned-inconsistency of corn side effect on the predictive ability of the

367

classification models.

368

Variable selection results by CARS and CARS-PLSDA modeling results. Based on the optimal full spectral PLS-

369

DA models, the CARS algorithm was performed to select the key wavelengths in discriminating the AF13-

370

inoculated corn kernels from the uninfected control and AF38-inoculated corn kernels, from the endosperm and

371

germ side separately. As shown in the above-mentioned comparisons, the one-step PLS-DA classification models

372

established over the spectral range II performed better than over range I with the one-step method and all two-step

373

classification cases, for both endosperm-side and germ-side spectra, the CARS algorithm was executed for these 2

374

combinations. As an example, Figure S2 presents the process of variable selection by CARS in one-step classifying

375

the “control+AF38-inoculated” and AF13-inoculated corn kernels, based on the endosperm-side spectra over range

376

II. As shown in Figure S2(a), with increasing number of sampling runs, the number of sampled spectral variables in

377

the variable subset decreased first fast at the first stage of EDF and then slowly at the second stage of EDF, namely,

378

the “fast selection” and “refined selection” stages. The changes of the error rate of 10-fold cross validation from all

379

spectral variable subsets in CARS calculations were plotted in Figure S2(b). As can be observed, with the increasing

380

number of sampling runs, the error rate of 10-fold cross validation first fluctuated in a gentle way and showed a

381

declining trend between the sampling run of 1-321, because of elimination of irrelevant spectral variables. Then,

382

with the increasing number of sampling run between 322 and 500, the error rate of 10-fold cross validation increased

383

after too many variables were eliminated, which may have included important spectral variables. The CARS


Page 14 of 29

Page 15 of 29


384

calculations showed the lowest global error rate of 10-fold cross validation when the number of sampling run was

385

321, represented by the vertical dashed lines in Figure S2. Figure S2(c) shows the regression coefficient path of each

386

wavelength variable in the execution of CARS with the number of sampling runs set to 500. Each curve in Figure

387

S2(c) reflects the changing of regression coefficient of one spectral variable at different sampling runs. As can be

388

observed, at the beginning of each sampling run, the absolute regression coefficient values of each spectral variable

389

were extremely low. With the increasing number of sampling runs, the values of some variables started growing,

390

while the rest of variables dropped to 0, which were removed by CARS because of their incompetence. The

391

regression coefficients of each spectral variable at the sampling run of 321 were extracted and shown in Figure 4(a).

392

Based on the regression coefficients, a total of 27 spectral variables were obtained in the determined optimal

393

variable subset by CARS.

394

By performing similar CARS procedures, the key wavelengths in classifying the AF13-inoculated corn

395

kernels and the uninfected control and AF38-inoculated corn kernels were identified based on the germ-side spectra

396

over range II. The regression coefficients of each spectral variable at the sampling run of 338 were extracted and

397

shown in Figure 4(b), due to where the lowest global error rate of 10-fold cross validation was achieved. Based on

398

the regression coefficients, a total of 21 spectral variables were determined by the CARS algorithm. By comparing

399

the key wavelengths determined by CARS from the endosperm and germ sides over range II, it was observed that

400

the determined wavelengths from both corn sides owned some common wavelengths, which focused around 1816,

401

2036, 2085, 2132, 2180, 2256, 2275 and 2320 nm. This indicates that even though differences exist between the

402

endosperm and germ sides of corn kernels in physical structure and chemical composition, the changes caused by

403

the infection and growth of the AF13 fungus from the endosperm and germ sides of corn kernels still possessed

404

some common features, and these changes can be sensed by the Vis-NIR spectroscopic method to discriminate the

405

AF13 infection from the uninfected control and AF38-infected corn kernels. The common wavelengths determined

406

over the spectral range II could be assigned as: 1) 2036 nm: the combination band region of CONH2(H); 2) 2085

407

nm: the combination band region of CONH2(H) and ROH; 3) 2132 nm: the combination band region of RNH2; 4)

408

2180 nm: the combination band region of RNH2 and CC; 5) 2256 and 2275 nm: the combination band region of CH2

409

and CH3; 6) 2320 nm: the combination band region of CH, CH2 and CH3. The wavelengths associated with

410

CONH2(H) and RNH2 bonds may reflect the changes of protein in corn kernels caused by the growth of AF13 and

411

its metabolic activities. The other determined wavelengths relating to the ROH, CC, CH, CH2 and CH3 bonds are



412

most probably due to the caused changes of carbohydrate and oil in corn kernels by the infection of AF13. In

413

addition, the common key wavelengths determined from the endosperm and germ sides of corn kernels around 1816,

414

2085, 2132 and 2180 nm were found to be in accordance with the interesting wavelengths reported on differentiating

415

the total fungi-infected corn and the uninfected corn [28].

416

Using the key wavelengths determined by CARS in the above-mentioned two cases, the simplified PLS-

417

DA models were established separately. Table 1 presents the obtained CARS-PLSDA results in classifying the

418

“control+AF38-inoculated” and AF13-inoculated corn kernels. Using the 27 and 21 selected wavelengths by CARS

419

from the endosperm and germ sides, the overall accuracy for the prediction set achieved 88.89% and 97.78%,

420

respectively, which are both higher than those obtained with the corresponding full-spectral models. This indicates

421

the usefulness of selecting the most informative spectral variables in improving the model performance.

422

Furthermore, the data dimensionality employed and the complexity of computation in establishing the classification

423

models could be greatly reduced by using the selected wavelengths. For instance, compared to the full spectral PLS-

424

DA models which were established using 2701 spectral variables over range II, the employed spectral variables were

425

decreased at least 99.00% in the simplified CARS-PLSDA models.

426

Modeling results in classifying aflatoxin-contaminated and healthy corn kernels. Full spectral PLS-DA

427

models. The second objective of this study focuses on aflatoxin contamination detection. Based on the aflatoxin

428

threshold of 20 ppb, 43 corn kernels from the AF13-inoculated group were determined as aflatoxin-positive and all

429

other kernels, which included 60 control, 60 AF38-inoculated and 17 AF13-inoculated corn kernels were regarded

430

as aflatoxin-negative (healthy). According to the approximate ratio of 3:1 of the sample number contained in the

431

calibration and prediction sets, 135 corn kernels which included 103 aflatoxin-negative kernels and 32 aflatoxin-

432

positive kernels were used as calibration samples, and the other 45 kernels including 34 aflatoxin-negative and 11

433

aflatoxin-positive kernels, were used as prediction samples to evaluate the established models when taking 20 ppb as

434

the classification threshold. When using the 100-ppb aflatoxin threshold, 36 corn kernels from the AF13-inoculated

435

group were determined as aflatoxin-positive kernels, and all other kernels, which included 60 control, 60 AF38-

436

inoculated and 24 AF13-inoculated corn kernels were regarded as aflatoxin-negative. Therefore, the calibration set

437

included a total of 135 corn kernels, namely, 108 aflatoxin-negative and 27 aflatoxin-positive kernels, and the

438

prediction set included a total of 45 corn kernels, namely, 36 aflatoxin-negative and 9 aflatoxin-positive kernels,

439

when taking 100 ppb as the classification threshold.


Page 16 of 29

Page 17 of 29


440

Based on the classification thresholds of 20 ppb and 100 ppb in aflatoxin concentration of each corn kernel,

441

the full spectral PLS-DA models were established separately. Table S5 presents the obtained results for all 8 cases.

442

Based on the classification threshold of 20 ppb, the best overall accuracy of 82.22% was obtained, which was

443

comparable to the best prediction result of 80.00% achieved with the 100-ppb threshold. In addition, both the best

444

full-spectral PLS-DA models established with the classification thresholds of 20 ppb and 100 ppb, were achieved

445

using the germ-side spectra over range II, which combination also performed the best with full spectra in one-step

446

classifying the “control+AF38-inoculated” and AF13-inoculated corn kernels. The full spectral PLS-DA model

447

established using the endosperm-side spectra over range I also yielded an overall accuracy of 80.00%, with the

448

classification threshold of 20 ppb. The other full spectral models did not perform so well in classifying the aflatoxin-

449

contaminated and healthy corn kernels, as the overall accuracies obtained for the prediction set were all lower than

450

80.00%.

451

Variable selection results by CARS and CARS-PLSDA modeling results. Based on the optimal combinations for

452

the full spectral models, the CARS procedure was conducted and the informative wavelengths were selected for

453

each classification threshold and corn side. In detail, for the classification threshold of 20 ppb, the CARS algorithm

454

was calculated for the two cases of “endosperm side+ spectral range I” and “germ side+ spectral range II” and

455

calculated for the cases of “endosperm side+ spectral range II” and “germ side+ spectral range II” for the 100-ppb

456

threshold. Using the 20-ppb threshold, a total of 61 and 65 spectral variables were selected by CARS for the

457

combination of “endosperm side+ spectral range I” and “germ side+ spectral range II”, respectively; and 62 and 52

458

variables were determined using the endosperm-side and germ-side spectra over range II separately, with the

459

classification threshold of 100 ppb.

460

Using the selected variables, the simplified CARS-PLSDA models were established separately with the

461

classification threshold of 20 ppb and 100 ppb, to separate the aflatoxin-contaminated and healthy corn kernels. The

462

classification results of the established CARS-PLSDA models are shown in Table 2. Based on the classification

463

threshold of 20 ppb, the better CARS-PLSDA model achieved an overall accuracy of 86.67%, using the selected

464

variables from the endosperm side over range I. When taking 100 ppb as the classification threshold, the superior

465

CARS-PLSDA model was obtained using the selected variables from the germ side over range II, which achieved an

466

overall accuracy of 84.44%. The other 2 CARS-PLSDA models also yielded acceptable prediction results in

467

classifying the aflatoxin-contaminated and healthy corn kernels, achieving the overall accuracies of 82.22% and



468

80.00%. Overall, compared to the corresponding full spectral PLS-DA models, the established CARS-PLSDA

469

models all obtained improved or equivalent overall accuracies.

470

PLSR modeling results in quantifying aflatoxin concentration of AF13-inoculated corn kernels. To investigate

471

the feasibility of using the Vis-NIR spectroscopy to quantify aflatoxin concentration of corn kernels infected with

472

aflatoxigenic fungus, the PLSR model was established for each case, separately. As mentioned above, 45 kernels

473

were used to establish the PLSR model and the other 15 kernels were used as prediction samples for evaluating the

474

established model performance. The statistical indices including the maximum, minimum, mean and SD values for

475

the calibration and prediction sets are shown in Table S1. Table S6 presents the PLSR modeling results obtained for

476

the calibration and prediction sets separately, for each case. Based on the endosperm-side spectra, the better PLSR

477

model in quantifying aflatoxin concentration was obtained over range II, which achieved a RP and RMSEP of 0.91

478

and 284.27 ppb, respectively. The superior PLSR model using the germ-side spectra was achieved over range I, with

479

RP and RMSEP of 0.87 and 377.96 ppb. Overall, the obtained RPs were acceptable with a few combination cases,

480

however, the RMSEPs were all great (≥ 280 ppb). The results obtained here using the Vis-NIR spectroscopy were

481

similar to those reported by Chu et al. [29], where short-wave infrared hyperspectral imaging over the spectral range

482

of 1000-2500 nm was applied to quantify the AFB1 concentration of single corn kernels from the germ side. The

483

authors obtained the determination coefficient of validation set (RV2) of 0.70, while the root mean square error of

484

validation set (RMSEV) was 524.4 ppb. The great RMSEP or RMSEV might be due to the wide ranges of reference

485

aflatoxin concentration in both studies, which is 0-3400 ppb in this study, and 0-3800 ppb in Chu et al.’s work [29].

486

The inaccurately predicted and meanwhile great data values might have strong negative effect on the obtained

487

RMSEPs or RMSEVs. However, it needs to be noted that for the PLSR quantification and the aforementioned-

488

classifications in separating healthy and aflatoxin-contaminated corn kernels, the accuracy of the reference lab

489

method used may be one factor affecting the prediction results of the established models. For practical reasons, the

490

reference chemical test was only performed once for each corn kernel. However, over the past few years, our lab has

491

been participating in a proficiency study, every 6 months, which tests the reliability of our instrument and

492

methodology and compares it to numerous labs across the world. A Cochran test of variance and a Grubbs test of the

493

mean recorded by each lab is conducted to identify outliers. A Z-value is assigned to the results of each lab, which

494

indicates deviation from the assigned mean. The Z range is between -3 and 3, where any value outside this range is

495

excluded. Therefore, the closer the Z value is to zero, the more reliable is the test. The specific test that was the


Page 18 of 29

Page 19 of 29


496

closest to this experiment had Z equaled to -0.38. This indicates our method, compared to all the other quantification

497

methods, is very reliable. Additionally, the instrument is calibrated weekly with standards provided by the

498

manufacturer (Mycotoxin Calibration Standards, VICAM: the green, red and yellow readings for 0.2 g equivalent

499

used in the current study were -2, 110, and 54 +/- 5, respectively, which indicates the instrument is calibrated

500

properly) to ensure the accuracy of the chemical analysis. Nevertheless, the quantitative modeling results using

501

PLSR, indicates the possibility of using Vis-NIR spectroscopy to quantify aflatoxin concentration in corn kernels

502

infected with aflatoxigenic fungus.

503

As mentioned previously, the two strains of AF13 and AF38 were chosen in this study as examples of

504

reliable and consistent aflatoxin producing (AF13) and non-toxin (AF38) producing fungi. Future studies may

505

incorporate additional strains of aflatoxigenic and non-aflatoxigenic fungi as well as diverse corn cultivars. Further

506

exploration with more advanced spectral preprocessing methods and modeling approaches will also be conducted in

507

the future. With continuous developments of optical hardware and chemometric techniques, the proposed approach

508

may become a rapid and non-invasive screening procedure to prevent aflatoxins from entering the food and feed

509

chains.

510

Supporting Information Available

511

Figure of data-processing routines for classifications; Figure of calculations of CARS algorithm in selecting optimal

512

wavelengths; Figure of scatter plots of reference versus predicted aflatoxin concentration in AF13-inoculated corn

513

kernels; Table of descriptive statistics of aflatoxin concentration of single AF13-inoculated kernels by the reference

514

method; Table of PLS-DA modeling results for each step of the two-step method using the full spectra; Table of

515

final PLS-DA modeling results of the two-step method using the full spectra; Table of one-step PLS-DA modeling

516

results in classifying the “control+AF38-inoculated” and AF13-inoculated corn kernels using the full spectra; Table

517

of PLS-DA modeling results in classifying healthy and aflatoxin-contaminated corn kernels using the full spectra;

518

Table of PLSR results for quantifying aflatoxin concentration in AF13-inoculated corn kernels.

519

ACKNOWLEDGEMENTS

520

This publication is a contribution of the Mississippi Agricultural and Forestry Experiment Station. The authors

521

gratefully acknowledge the financial support of the USDA cooperative agreement No. 58-6435-3-024, the U. S.

522

Agency for International Development via the Peanut Mycotoxin Innovation Laboratory at University of Georgia

523

(Subaward No. RC710-059/4942206), and Mississippi State University Special Research Initiative Program. This



524

material is also based upon work that is supported by The National Institute of Food and Agriculture, U.S.

525

Department of Agriculture, Hatch multistate project under accession number 1008884. Lastly, we would like to

526

thank Mr. Russell Kincaid for field corn sample preparation and Ms. Dawn Darlington for aflatoxin chemical

527

analysis.

528

REFERENCES

529

(1) Tao, F.; Yao, H.; Hruska, Z.; Burger, L. W.; Rajasekaran, K.; Bhatnagar, D. Recent development of optical

530

methods in rapid and non-destructive detection of aflatoxin and fungal contamination in agricultural products.

531

Trends Anal. Chem. 2018, 100, 65-81.

532

(2) Payne, G. A. Chapter 9: Process of contamination by aflatoxin-producing fungi and their impact on crops. In:

533

Shinha, K. K.; Bhatnagar, D. Mycotoxins in Agriculture and Food Safety, CRC Press. 1998. pp. 279.

534

(3) Wang, W.; Ni, X.; Lawrence, K. C.; Yoon, S.-C.; Heitschmidt, G. W.; Feldner, P. Feasibility of detecting

535

Aflatoxin B1 in single maize kernels using hyperspectral imaging. J. Food Eng. 2015, 166, 182-192.

536 537 538 539 540 541 542 543 544 545 546 547

(4) Liu, J.-M.; Wang, Z.-H.; Ma, H.; Wang, S. Probing and quantifying the food-borne pathogens and toxins: From in vitro to in vivo. J. Agric. Food Chem., 2018, 66 (5), 1061-1066. (5) Huang, H.; Yu, H.; Xu, H.; Ying, Y.; Near infrared spectroscopy for on/in-line monitoring of quality in foods and beverages: A review. J. Food Eng. 2008, 87, 303-313. (6) Wang, W.; Paliwal, J. Near-infrared spectroscopy and imaging in food quality and safety. Sensing and Sens Instrum. Food Qual. Saf. 2007, 1(4), 193-207. (7) Wu, Q.; Xie, L.; Xu, H. Determination of toxigenic fungi and aflatoxins in nuts and dried fruits using imaging and spectroscopic techniques. Food Chem. 2018, 252, 228-242. (8) Lee, K.-M.; Herrman, T. J.; Bisrat, Y.; Murray, S. C. Feasibility of surface-enhanced Raman spectroscopy for rapid detection of aflatoxins in maize. J. Agric. Food Chem. 2014, 62(19), 4466-4474. (9) Abbas, H. K.; Zablotowicz, R. M.; Locke, M. A. Spatial variability of Aspergillus flavus soil populations under different crops and corn grain colonization and aflatoxins. Canadian Journal of Botany 2004, 82, 1768-1775.

548

(10) Abbas, H. K.; Zablotowicz, R. M.; Weaver, M. A.; Horn, B. W.; Xie, W.; Shier, W. T. Comparison of cultural

549

and analytical methods for determination of aflatoxin production by Mississippi Delta Aspergillus isolates.

550

Canadian Journal of Microbiology 2004, 50, 193-199.


Page 20 of 29

Page 21 of 29

551


(11) Abbas, H. K.; Weaver, M. A.; Zablotowicz, R. M.; Horn, B. W.; Shier, W. T. Relationships between aflatoxin

552

production, sclerotia formation and source among Mississippi Delta Aspergillus isolates. Eur J Plant Pathol.

553

2005, 112, 283-287.

554

(12) Pildain, M. B.; Vaamonde, G.; Cabral, D. Analysis of population structure of Aspergillus flavus from peanut

555

based on vegetative compatibility, geographic origin, mycotoxin and sclerotia production. Int J Food Microbiol.

556

2004, 93, 31-40.

557 558 559 560 561 562 563

(13) Horn, B. W. Ecology and population biology of aflatoxigenic fungi in soil. Journal of Toxicology: Toxin Reviews 2003, 22, 351-379. (14) Zablotowicz, R. M.; Abbas, H. K.; Locke, M. A. Population ecology of Aspergillus flavus associated with Mississippi Delta soils. Food Addit Contam. 2007, 24, 1102-1108. (15) Abbas, H. K.; Zablotowicz, R. M.; Bruns, H. A.; Abel, C. A. Biocontrol of aflatoxin in corn by inoculation with non-aflatoxigenic Aspergillus flavus isolates. Biocontrol Sci Technol. 2006, 16, 437-449. (16) Abbas, H. K.; Zablotowicz, R. M.; Bruns, H. A.; Abel., C. A. Development of non-toxigenic strains of

564

Aspergillus flavus for control of aflatoxin in Maize. In: Burton, E. N.; Williams, P. V. eds. Crop Protection,

565

Research Advances. New York, NY: Nova Sciences Publishers, 2008, pp. 181-192.

566

(17) Abbas, H. K.; Accinelli, C.; Shier, W. T. Biological control of aflatoxin contamination in U.S. crops and the use

567

of bioplastic formulations of Aspergillus flavus biocontrol strains to optimize application strategies. J. Agric.

568

Food Chem. 2017, 65 (33), 7081-7087.

569 570 571

(18) Jin, J.; Tang, L.; Hruska, Z.; Yao, H. Classification of toxigenic and atoxigenic strains of Aspergillus flavus with hyperspectral imaging. Comput. Electron. Agric. 2009, 69, 158-164. (19) Yao, H.; Hruska, Z.; Kincaid, R.; Brown, R. L.; Bhatnagar, D.; Cleveland, T. E. Detecting maize inoculated

572

with toxigenic and atoxigenic fungal strains with fluorescence hyperspectral imagery. Biosyst. Eng. 2013, 115,

573

125-135.

574

(20) Zhu, F.; Yao, H.; Hruska, Z.; Kincaid, R.; Brown, R.; Bhatnagar, D.; Cleveland, T. Integration of fluorescence

575

and reflectance visible near-infrared (VNIR) hyperspectral images for detection of aflatoxins in corn kernels.

576

Trans. ASABE 2016, 59, 785-794.

577 578

(21) Li, H.; Liang, Y.; Xu, Q.; Cao, D. Key wavelengths screening using competitive adaptive reweighted sampling method for multivariate calibration. Analytica Chimica Acta 2009, 648, 77-84.



579 580 581 582 583 584 585 586

(22) Ballabio, D.; Consonni, V. Classification tools in chemistry. Part 1: linear models. PLS-DA. Anal. Methods 2013, 5, 3790-3798. (23) Wong, T.-T. Performance evaluation of classification algorithms by k-fold and leave-one-out cross validation. Pattern Recognit. 2015, 48, 2839-2846. (24) Pearson, T. C.; Wicklow, D. T. Detection of corn kernels infected by fungi. Trans. ASABE 2006, 49, 12351245. (25) Hesseltine, C. W.; Shotwell, O. L. New methods for rapid detection of aflatoxin. Pure and Appl. Chem. 1973, 35(3), 259-266.

587

(26) Lillehoj, E. B.; Kwolek, W. F.; Peterson, R. E.; Shotwell, O. L.; Hesseltine, C. W. Aflatoxin contamination,

588

fluorescence, and insect damage in corn infected with Aspergillus flavus before harvest. Cereal Chem. 1976,

589

53(4), 505-512.

590

(27) Birth, G. S.; Hecht, H. G. The physics of near-infrared reflectance. In Near-Infrared Technology in the

591

Agriculture and Food Industries, 1-15. Williams, P.; Norris, K. eds. St. Paul, Minn.: American Association of

592

Cereal Chemists. 1987.

593

(28) Berardo, N.; Pisacane, V.; Battilani, P.; Scandolara, A.; Pietri, A.; Marocco, A. Rapid detection of kernel rots

594

and mycotoxins in maize by near-infrared reflectance spectroscopy. J. Agric. Food Chem. 2005, 53(21), 8128-

595

8134.

596

(29) Chu, X.; Wang, W.; Yoon, S.-C.; Ni, X.; Heitschmidt, G. W. Detection of aflatoxin B1 (AFB1) in individual

597

maize kernels using short wave infrared (SWIR) hyperspectral imaging. Biosyst. Eng. 2017, 157, 13-23.

598 599 600 601 602 603 604 605 606


Page 22 of 29

Page 23 of 29


607 608 609

Figure Captions

610

Figure 1. Distribution plots of aflatoxin concentration of single corn kernels from the (a) AF13-inoculated group,

611

(b) AF38-inoculated group.

612

Figure 2. The mean spectra of original absorbance of control, AF38 and AF13-inoculated samples from the (a)

613

endosperm and (b) germ sides of corn kernels.

614

Figure 3. Correlation coefficients between original absorbance and aflatoxin concentration of AF13-inoculated

615

corn kernels.

616

Figure 4. Regression coefficients of selected wavelengths by CARS in one-step classifying the “control+AF38-

617

inoculated” and AF13-inoculated corn kernels: (a) 27 variables based on endosperm-side spectra over range II, (b)

618

21 variables based on germ-side spectra over range II.



Page 24 of 29

(a)

(b)

Figure 1. Distribution plots of aflatoxin concentration of single corn kernels from the (a) AF13-inoculated group, (b) AF38-inoculated group.


Page 25 of 29


(a)

(b)

Figure 2. The mean spectra of original absorbance of control, AF38 and AF13-inoculated samples from the (a) endosperm and (b) germ sides of corn kernels.



Figure 3. Correlation coefficients between original absorbance and aflatoxin concentration of AF13-inoculated corn kernels.


Page 26 of 29

Page 27 of 29


(a)

(b)

Figure 4. Regression coefficients of selected wavelengths by CARS in one-step classifying the “control+AF38inoculated” and AF13-inoculated corn kernels: (a) 27 variables based on endosperm-side spectra over range II, (b) 21 variables based on germ-side spectra over range II.



Table 1. One-step CARS-PLSDA modeling results in classifying the “control+AF38-inoculated” and AF13-inoculated corn kernels. Corn side

Spectral range II

Data set

LVs number 11

Correct sample Correct sample Accuracy for Accuracy for Overall number for class 1 number for class 2 class 1 (%) class 2 (%) accuracy Endosperm Calibration 90 45 100.00 100.00 100.00 Prediction 28 12 93.33 80.00 88.89 Germ II Calibration 12 90 43 100.00 95.56 98.52 Prediction 29 15 96.67 100.00 97.78 Note: Class 1 refers to the negative group, representing the control and AF38-inoculated corn kernels; Class 2 refers to the positive group, representing the AF13inoculated corn kernels.

Table 2. CARS-PLSDA modeling results in classifying healthy and aflatoxin-contaminated corn kernels. Data set LVs Correct sample Correct sample Accuracy for Accuracy for Overall number number for class 1 number for class 2 class 1 (%) class 2 (%) accuracy (%) Classification threshold: 20 ppb Endosperm I Calibration 18 99 31 96.12 96.88 96.30 Prediction 31 8 91.18 72.73 86.67 Germ II Calibration 18 99 32 96.12 100.00 97.04 Prediction 29 8 85.29 72.73 82.22 Classification threshold: 100 ppb Endosperm II Calibration 16 107 27 99.07 100.00 99.26 Prediction 32 4 88.89 44.44 80.00 Germ II Calibration 15 107 27 99.07 100.00 99.26 Prediction 32 6 88.89 66.67 84.44 Note: Class 1 refers to the negative group, representing corn kernels with aflatoxin≤ 20/100 ppb; Class 2 refers to the positive group, representing corn kernels with aflatoxin>20/100 ppb. Corn side

Spectral range


Page 28 of 29

Page 29 of 29


TOC Graphic


A rapid and non-destructive method for simultaneous determination of

Recommend Documents