Combining E-Nose and Lateral Flow Immunoassays (LFIAs) for Rapid Occurrence/Co-Occurrence Aflatoxin and Fumonisin Detection in Maize

The aim of this study was to evaluate the potential use of an e-nose in combination with lateral flow immunoassays for rapid aflatoxin and fumonisin occurrence/co-occurrence detection in maize samples. For this purpose, 161 samples of corn have been used. Below the regulatory limits, single-contaminated, and co-contaminated samples were classified according to the detection ranges established for commercial lateral flow immunoassays (LFIAs) for mycotoxin determination. Correspondence between methods was evaluated by discriminant function analysis (DFA) procedures using IBM SPSS Statistics 22. Stepwise variable selection was done to select the e-nose sensors for classifying samples by DFA. The overall leave-out-one cross-validated percentage of samples correctly classified by the eight-variate DFA model for aflatoxin was 81%. The overall leave-out-one cross-validated percentage of samples correctly classified by the seven-variate DFA model for fumonisin was 85%. The overall leave-out-one cross-validated percentage of samples correctly classified by the nine-variate DFA model for the three classes of contamination (below the regulatory limits, single-contaminated, co-contaminated) was 65%. Therefore, even though an exhaustive evaluation will require a larger dataset to perform a validation procedure, an electronic nose (e-nose) seems to be a promising rapid/screening method to detect contamination by aflatoxin, fumonisin, or both in maize kernel stocks.


Introduction
Maize (Zea mays L.) is one of the most important food and feed commodities among cereal crops. However, its diffusion, popularity, and, most of all, safety for consumption are threatened by mycotoxin contamination, each sample was assigned to one of three classes: below the regulatory limits, single-contaminated, or co-contaminated. Specifically, of the 161 samples analyzed, 118 were below the regulatory limits, 20 were single-contaminated, and 23 were co-contaminated. Among the regular samples (samples below the regulatory limits), the recorded levels for both AF and FM were below the limits for foods established by the EU. Considering single-contaminated samples, thirteen were contaminated by AF, while seven were contaminated by FM.
Toxins 2018, x, x FOR PEER REVIEW 3 of 10 contamination, each sample was assigned to one of three classes: below the regulatory limits, singlecontaminated, or co-contaminated. Specifically, of the 161 samples analyzed, 118 were below the regulatory limits, 20 were single-contaminated, and 23 were co-contaminated. Among the regular samples (samples below the regulatory limits), the recorded levels for both AF and FM were below the limits for foods established by the EU. Considering single-contaminated samples, thirteen were contaminated by AF, while seven were contaminated by FM. Figure 1. Distribution of maize kernel samples according to the presence of aflatoxin (AF) and fumonisin (FM) determined by LFIA kit (Envirologix™, Portland, ME, USA). Below the regulatory limits (AF < 5 ppb, and FM < 4 ppm); single contaminated, above the regulatory limit for 1 mycotoxin (AF > 5 ppb, or FM > 4 ppm); or co-contaminated, above the regulatory limit for 2 mycotoxins (AF > 5 ppb, and FM > 4 ppm). N, number.

Electronic Nose
Data from the e-nose were analyzed using DFA, and the performance of the models in predicting maize kernels below the regulatory limits, single contaminated, and co-contaminated with AF and/or FM is summarized in Table 1.
Stepwise variable selection was done to select the e-nose sensors for classifying samples by DFA. The discriminant function used to identify aflatoxin contaminated or below the regulatory limits samples included several e-nose sensors: W1C aromatic, W5S broad-range, W3C Aromatic, W1S Broad-methane, W1W Sulphur-organic, W2S broad-alcohol, W2W Sulph-clor, W3S methane-aliph.
The overall leave-out-one cross-validated percentage of samples correctly classified by the eight variate DFA models for AF was 84.1%. In the case of below the regulatory limits samples, the percentage of samples correctly classified was 89.2%, while in the case of contaminated samples, it was 68.9%.
The discriminant function used to identify FM contaminated or below the regulatory limits samples included seven e-nose sensors, W1C aromatic, W5S broad-range, W5C Arom-aliph, W1W Sulphur-organic, W2S broad-alcohol, W2W Sulph-clor, W3S methane-aliph. The overall leave-outone cross-validated percentage of samples correctly classified by the seven-variate DFA model for fumonisin was 85.4%. In the case of below the regulatory limits samples, the percentage of samples correctly classified was 87.2%, while in the case of contaminated samples, it was 82.0%. . Below the regulatory limits (AF < 5 ppb, and FM < 4 ppm); single contaminated, above the regulatory limit for 1 mycotoxin (AF > 5 ppb, or FM > 4 ppm); or co-contaminated, above the regulatory limit for 2 mycotoxins (AF > 5 ppb, and FM > 4 ppm). N, number.

Electronic Nose
Data from the e-nose were analyzed using DFA, and the performance of the models in predicting maize kernels below the regulatory limits, single contaminated, and co-contaminated with AF and/or FM is summarized in Table 1.
Stepwise variable selection was done to select the e-nose sensors for classifying samples by DFA. The discriminant function used to identify aflatoxin contaminated or below the regulatory limits samples included several e-nose sensors: W1C aromatic, W5S broad-range, W3C Aromatic, W1S Broad-methane, W1W Sulphur-organic, W2S broad-alcohol, W2W Sulph-clor, W3S methane-aliph.
The overall leave-out-one cross-validated percentage of samples correctly classified by the eight variate DFA models for AF was 84.1%. In the case of below the regulatory limits samples, the percentage of samples correctly classified was 89.2%, while in the case of contaminated samples, it was 68.9%.
The discriminant function used to identify FM contaminated or below the regulatory limits samples included seven e-nose sensors, W1C aromatic, W5S broad-range, W5C Arom-aliph, W1W Sulphur-organic, W2S broad-alcohol, W2W Sulph-clor, W3S methane-aliph. The overall leave-out-one cross-validated percentage of samples correctly classified by the seven-variate DFA model for fumonisin was 85.4%. In the case of below the regulatory limits samples, the percentage of samples correctly classified was 87.2%, while in the case of contaminated samples, it was 82.0%. A further step in the study was to test the potential of the e-nose in detecting co-contaminated samples. For this purpose, the discriminant function used to identify below the regulatory limits, single-contaminated, and co-contaminated samples included almost all the e-nose sensors (nine of ten), namely: W1C aromatic, W5S broad-range, W3C aromatic, W6S Hydrogen, W1S broad-methane, W1W Sulphur-organic, W2S Broad-alcohol, W2W Sulph-clor, and W3S Methane-aliph. The overall leave-out-one cross-validated percentage of samples correctly classified, by the nine-variate DFA model for the three classes of contamination (below the regulatory limits, single-contaminated, co-contaminated), was 65.2%. In the case of regular samples (below the regulatory limits samples), the percentage correctly classified was 65.4%, while it dropped to 61.0% and 67.4% for single-contaminated and co-contaminated samples, respectively. The same results are shown in the discriminant analysis scatterplot reported in Figure 2. In general, co-contaminated samples (red crosses) tended to group together, confirming results obtained for cross-validation, which, for this class of samples, reached 67% of samples correctly classified. By contrast, regular and single-contaminated samples, shown in the same scatterplot as green circles and yellow triangles, respectively, were more dispersed, indicating higher variability in the VOCs' profile.
Toxins 2018, x, x FOR PEER REVIEW 5 of 10 A further step in the study was to test the potential of the e-nose in detecting co-contaminated samples. For this purpose, the discriminant function used to identify below the regulatory limits, single-contaminated, and co-contaminated samples included almost all the e-nose sensors (nine of ten), namely: W1C aromatic, W5S broad-range, W3C aromatic, W6S Hydrogen, W1S broad-methane, W1W Sulphur-organic, W2S Broad-alcohol, W2W Sulph-clor, and W3S Methane-aliph. The overall leave-out-one cross-validated percentage of samples correctly classified, by the nine-variate DFA model for the three classes of contamination (below the regulatory limits, single-contaminated, cocontaminated), was 65.2%. In the case of regular samples (below the regulatory limits samples), the percentage correctly classified was 65.4%, while it dropped to 61.0% and 67.4% for singlecontaminated and co-contaminated samples, respectively. The same results are shown in the discriminant analysis scatterplot reported in Figure 2. In general, co-contaminated samples (red crosses) tended to group together, confirming results obtained for cross-validation, which, for this class of samples, reached 67% of samples correctly classified. By contrast, regular and singlecontaminated samples, shown in the same scatterplot as green circles and yellow triangles, respectively, were more dispersed, indicating higher variability in the VOCs' profile. Figure 2. Discriminant analysis scatterplot relative to classification of: 0 = below the regulatory limits (AF < 5 ppb, FM < 4 ppm), green circles (O); 1 = single contaminated, above the regulatory limit for 1 mycotoxin (AF > 5 ppb, or FM > 4 ppm), yellow triangles (Δ); 2 = co-contaminated, above the regulatory limit for 2 mycotoxins (AF > 5ppb, and FM > 4 ppm), red crosses (×). DF, discriminant function.

Discussion
The discriminant function used to identify AF contaminated or below the regulatory limits samples included eight e-nose sensors (W1C aromatic, W5S broad-range, W2S broad-alcohol, and W3S methane-aliph, etc.), that indicated a quite complicated odor profile. It is known that e-nose sensors are nonspecific but can be highly sensitive, responding to a range of different compounds. The conductivity of the polymer changes when molecules are absorbed at the sensor surface. The sensors respond strongly to the presence of alcohols, ketones, fatty acids, and esters, but have a reduced response to fully oxidized materials such as CO2, NO2, and H2O. Thus, the present results are in line with other studies reporting that the production of mycotoxins by particular mold strains is generally associated with the production of volatile substances such as alcohols, aldehydes, ketones, and esters [17]. Moreover, as previously reported by Cheli and co-workers [1], the present work confirms the significant contribution of e-nose metal-oxide-semiconductor (MOS) sensors that are able to detect nitrogen oxides and ozone, which were related to the AF contamination in maize A further step in the study was to test the potential of the e-nose in detecting co-contaminated samples. For this purpose, the discriminant function used to identify below the regulatory limits, single-contaminated, and co-contaminated samples included almost all the e-nose sensors (nine of ten), namely: W1C aromatic, W5S broad-range, W3C aromatic, W6S Hydrogen, W1S broad-methane, W1W Sulphur-organic, W2S Broad-alcohol, W2W Sulph-clor, and W3S Methane-aliph. The overall leave-out-one cross-validated percentage of samples correctly classified, by the nine-variate DFA model for the three classes of contamination (below the regulatory limits, single-contaminated, cocontaminated), was 65.2%. In the case of regular samples (below the regulatory limits samples), the percentage correctly classified was 65.4%, while it dropped to 61.0% and 67.4% for singlecontaminated and co-contaminated samples, respectively. The same results are shown in the discriminant analysis scatterplot reported in Figure 2. In general, co-contaminated samples (red crosses) tended to group together, confirming results obtained for cross-validation, which, for this class of samples, reached 67% of samples correctly classified. By contrast, regular and singlecontaminated samples, shown in the same scatterplot as green circles and yellow triangles, respectively, were more dispersed, indicating higher variability in the VOCs' profile.

Discussion
The discriminant function used to identify AF contaminated or below the regulatory limits samples included eight e-nose sensors (W1C aromatic, W5S broad-range, W2S broad-alcohol, and W3S methane-aliph, etc.), that indicated a quite complicated odor profile. It is known that e-nose sensors are nonspecific but can be highly sensitive, responding to a range of different compounds. The conductivity of the polymer changes when molecules are absorbed at the sensor surface. The sensors respond strongly to the presence of alcohols, ketones, fatty acids, and esters, but have a reduced response to fully oxidized materials such as CO2, NO2, and H2O. Thus, the present results are in line with other studies reporting that the production of mycotoxins by particular mold strains is generally associated with the production of volatile substances such as alcohols, aldehydes, ketones, and esters [17]. Moreover, as previously reported by Cheli and co-workers [1], the present work confirms the significant contribution of e-nose metal-oxide-semiconductor (MOS) sensors that are able to detect nitrogen oxides and ozone, which were related to the AF contamination in maize samples quantified by ELISA.
); 1 = single contaminated, above the regulatory limit for 1 mycotoxin (AF > 5 ppb, or FM > 4 ppm), yellow triangles ( A further step in the study was to test the potential of the e-nose in detecting co-contam samples. For this purpose, the discriminant function used to identify below the regulatory single-contaminated, and co-contaminated samples included almost all the e-nose sensors (n ten), namely: W1C aromatic, W5S broad-range, W3C aromatic, W6S Hydrogen, W1S broad-me W1W Sulphur-organic, W2S Broad-alcohol, W2W Sulph-clor, and W3S Methane-aliph. The o leave-out-one cross-validated percentage of samples correctly classified, by the nine-variat model for the three classes of contamination (below the regulatory limits, single-contaminat contaminated), was 65.2%. In the case of regular samples (below the regulatory limits sample percentage correctly classified was 65.4%, while it dropped to 61.0% and 67.4% for contaminated and co-contaminated samples, respectively. The same results are shown discriminant analysis scatterplot reported in Figure 2. In general, co-contaminated sample crosses) tended to group together, confirming results obtained for cross-validation, which, f class of samples, reached 67% of samples correctly classified. By contrast, regular and contaminated samples, shown in the same scatterplot as green circles and yellow tria respectively, were more dispersed, indicating higher variability in the VOCs' profile. Figure 2. Discriminant analysis scatterplot relative to classification of: 0 = below the regulatory limi (AF < 5 ppb, FM < 4 ppm), green circles (O); 1 = single contaminated, above the regulatory limit for mycotoxin (AF > 5 ppb, or FM > 4 ppm), yellow triangles (Δ); 2 = co-contaminated, above the regulato limit for 2 mycotoxins (AF > 5ppb, and FM > 4 ppm), red crosses (×). DF, discriminant function.

Discussion
The discriminant function used to identify AF contaminated or below the regulatory samples included eight e-nose sensors (W1C aromatic, W5S broad-range, W2S broad-alcoho W3S methane-aliph, etc.), that indicated a quite complicated odor profile. It is known that sensors are nonspecific but can be highly sensitive, responding to a range of different compo The conductivity of the polymer changes when molecules are absorbed at the sensor surfac sensors respond strongly to the presence of alcohols, ketones, fatty acids, and esters, but h reduced response to fully oxidized materials such as CO2, NO2, and H2O. Thus, the present are in line with other studies reporting that the production of mycotoxins by particular mold is generally associated with the production of volatile substances such as alcohols, alde ketones, and esters [17]. Moreover, as previously reported by Cheli and co-workers [1], the p work confirms the significant contribution of e-nose metal-oxide-semiconductor (MOS) senso ); 2 = co-contaminated, above the regulatory limit for 2 mycotoxins (AF > 5 ppb, and FM > 4 ppm), red crosses ( W1W Sulphur-organic, W2S Broad-alcohol, W2W Sulph-clor, and W3S Methan leave-out-one cross-validated percentage of samples correctly classified, by th model for the three classes of contamination (below the regulatory limits, singl contaminated), was 65.2%. In the case of regular samples (below the regulatory percentage correctly classified was 65.4%, while it dropped to 61.0% and contaminated and co-contaminated samples, respectively. The same results discriminant analysis scatterplot reported in Figure 2. In general, co-contami crosses) tended to group together, confirming results obtained for cross-valida class of samples, reached 67% of samples correctly classified. By contrast, contaminated samples, shown in the same scatterplot as green circles an respectively, were more dispersed, indicating higher variability in the VOCs' pr

Discussion
The discriminant function used to identify AF contaminated or below th samples included eight e-nose sensors (W1C aromatic, W5S broad-range, W2S W3S methane-aliph, etc.), that indicated a quite complicated odor profile. It is sensors are nonspecific but can be highly sensitive, responding to a range of di The conductivity of the polymer changes when molecules are absorbed at the sensors respond strongly to the presence of alcohols, ketones, fatty acids, and reduced response to fully oxidized materials such as CO2, NO2, and H2O. Thus are in line with other studies reporting that the production of mycotoxins by par is generally associated with the production of volatile substances such as a ). DF, discriminant function.

Discussion
The discriminant function used to identify AF contaminated or below the regulatory limits samples included eight e-nose sensors (W1C aromatic, W5S broad-range, W2S broad-alcohol, and W3S methane-aliph, etc.), that indicated a quite complicated odor profile. It is known that e-nose sensors are nonspecific but can be highly sensitive, responding to a range of different compounds. The conductivity of the polymer changes when molecules are absorbed at the sensor surface. The sensors respond strongly to the presence of alcohols, ketones, fatty acids, and esters, but have a reduced response to fully oxidized materials such as CO 2 , NO 2 , and H 2 O. Thus, the present results are in line with other studies reporting that the production of mycotoxins by particular mold strains is generally associated with the production of volatile substances such as alcohols, aldehydes, ketones, and esters [17]. Moreover, as previously reported by Cheli and co-workers [1], the present work confirms the significant contribution of e-nose metal-oxide-semiconductor (MOS) sensors that are able to detect nitrogen oxides and ozone, which were related to the AF contamination in maize samples quantified by ELISA.
With regard to the eight-variate DFA model for AF, figures obtained in the present study, when compared with the literature [1], seem to indicate that the e-nose in the present study was less effective at distinguishing between contaminated and below the regulatory limits samples. However, in making the comparison, it must be kept in mind that Cheli and co-workers [1] used highly contaminated samples. In that preliminary study, the range of AF contamination was 6-100 ppb, with more than 50% of the samples having >20 ppb of AF.
Moving to the evaluation of the potential use of e-noses for rapid FM detection in maize, the present study indicates that the percentage of samples correctly classified by the seven-variate DFA model for FM was 71%. Although only a few studies have been conducted in this field [15], Keshri and Magan [19] reported that, using an e-nose system, it is possible to differentiate between mycotoxigenic and non-mycotoxigenic strains of such spoilage fungi (e.g., Fusarium spp.) based on their volatile production patterns. However, some of the mycotoxigenic strains used in that study were also able to produce toxins, other than FM (i.e., zearalenone and trichothecene). Similarly, Gobbi and co-workers [20] reported that the e-nose could correctly recognize high FM content (>1000 ppm) and low FM content (<1.6 ppm) in maize cultures (contamination obtained by in vitro incubation of coarsely cracked maize kernels with fumonisin-producing fungi) and provide a fair quantitative estimation. However, in these studies [20], the fungal contamination/spoilage was obtained in vitro from maize cultures inoculated with selected Fusarium strains, and sterilized maize cultures were used as reference material/control. By contrast, samples used in the present study were collected from different stockpiles in Italy, which implies that both the fungal strains and the level of contamination were not standardized. Accordingly, in the present work, a higher variability in the odor profile compared to the samples experimentally contaminated in previous studies [20] could be assumed.
Combining the results obtained from the overall leave-out-one cross-validation, it can be suggested that the predictive accuracy of the model was limited. In fact, although below the co-contaminated samples tended to group together, reaching 67% of samples correctly classified, both single-contaminated and regular (below the regulatory limit) samples were more dispersed, indicating a higher variability of the VOCs' profile. For these classes, the percentages of samples correctly classified were 65% and 61% for regular (below the regulatory limits) and single-contaminated samples, respectively. These figures indicate that more than one third of the samples were wrongly classified, making the method still unsuitable for the purpose of mycotoxin detection. For comparison, Olsson et al. [15] investigated the possibility of using fungal volatile metabolites as indicators of two mycotoxins (ochratoxin A and deoxynivalenol) in barley, using both e-noses and gas chromatography combined with mass spectrometry (GC-MS). In that study, the authors reported that the e-nose misclassified less than 20% of samples (seven of 37 samples) in the case of ochratoxin A, while the deoxynivalenol level could be estimated using a partial least square (PLS) model constructed using the sensor signals from the e-nose. However, in this case, even if several samples were contaminated by both ochratoxin A and deoxynivalenol, detection of co-contamination by ochratoxin A and deoxynivalenol, was not the aim of the study.
Therefore, the present findings seem to indicate that complex chemical patterns of volatile components prevented a complete characterization of the different volatile organic compounds present in regular, single-contaminated and co-contaminated maize, limiting their correct classification to a range values of 60-67%.

Conclusions
Combining the present results, it can be concluded that e-noses have some potential for detection of selected mycotoxins like aflatoxin and fumonisin in maize. Surprisingly, the e-nose was more effective in detecting co-contaminated samples, reaching 67% of samples correctly classified, while the same value drops to 65% and 61% for regular and single-contaminated samples, respectively.
In the cereal industry, detection of mycotoxin contamination and co-contamination represents a major analytical challenge, and systematic, economical, straightforward cereal tests for rapid and accurate diagnosis of maize safety are needed. At the industrial level, the main question is the choice of the best analytical method for practically enabling rapid decision-making regarding the acceptance or rejection of lots of cereal and ensuring safety standards. Although, for an exhaustive evaluation of e-nose potential for mycotoxins detection other larger datasets are needed, the e-nose seems to be a promising rapid/screening method to detect mycotoxin contamination in maize kernel stocks. Its potential seems improved by its combination with commercially available rapid kit assays for mycotoxin detection. Both are characterized by limited sample preparation, real-time analysis, rapid detection of mold contamination at an early phase, evaluation of co-contamination, easy automation to create models for use as quality control tools, and possible integration into production processes.

Samples Collection and Analysis
For this study, 161 samples of maize (Zea mays L.), collected from different stockpiles in Italy between October 2016 and November 2017, were used. Representative samples were obtained according to European Commission Regulation no. 152/2009 [21]. From each sample, two aliquots were subsampled randomly and analyzed with two rapid/screening methods for detection of total aflatoxin and fumonisin contamination or co-contamination. The methods used were: (i) LFIA; and (ii) e-nose.

Lateral Flow Immuno Assay
The first aliquot of each maize sample was immediately analyzed for determination of total aflatoxin (AF) and fumonisin (FM) by commercial LFIA kit (Envirologix™), made up of a lateral flow strip with antibody immobilized on a test zone, the same principle as ELISA tests [12]. Analyses were performed immediately after receiving samples from stockpiles. According to the protocol, quantification of AF and FM was performed in parallel on two representative sub-aliquots. Briefly, each maize sample was finely ground and weighed. After that, according to the protocol for both aflatoxin and fumonisin detection, maize samples were rehydrated in distilled water, and the supernatant was extracted and an aliquot was mixed in a buffer and incubated for 2 min. Finally, a colorimetric strip was added for 4 min or 5 min (for aflatoxin or fumonisin, respectively) and then read by QuickScan Envirologix TM . The detailed protocol is reported in Table 2. For both AF and FM detection, the base range protocol was adopted (AF detection from 2.7 ppb to 30 ppb; FM detection from 1.5 ppm to 7 ppm). Each sample was analyzed in duplicate. Results obtained for AF and FM were used in classifying each sample as follows: (i) Samples below the regulatory limits; (ii) single-contaminated; or (iii) co-contaminated samples. Further details about the grouping (meet regulatory limit, above regulatory limit for 1 mycotoxin, and above regulatory limit for 2 mycotoxins) of the different samples according to the LFIA results are reported in the statistical analysis section.

Electronic Nose Analysis
The second aliquots of maize samples were stored at −18 • C in vacuum-sealed conditions prior to e-nose analysis, in order to prevent the development of further odors and off-odors that could affect the reliability of the results. For the e-nose analysis, 10 g of each sample of maize kernels was placed into an airtight 20 mL glass vial and sealed with a chlorobutyl/PTFE magnetic cap. The e-nose's pipe was connected to the vials with a needle: Between the needle and the pipe, a filter was inserted in order to block the dust that was present in some samples and could ruin the sensors. One more needle was inserted into each vial in order to avoid creating negative pressure inside the vial.
The analysis was performed on a portable electronic nose 3 (PEN 3) model e-nose from Airsense Analytics GmbH (Schwerin, Germany). The sensor array consisted of 10 metal-oxide-semiconductor (MOS) chemical sensors made of a ceramic substrate heated by a wire resistor and coated with a metal oxide semiconducting film. At operating conditions, interactions between the volatiles from the head space and the sensor's surface induced changes in the conductance of the semiconductor. Thus, the ratio G/G0 (in which G and G0 represent the resistance of a sensor when detecting a gas and when inhaling clean air, respectively) was recorded by the e-nose dedicated software. The characteristics of the e-nose sensor array are listed in Table 3. Preliminary trials were performed to evaluate the sensitivity of e-nose sensors using experimentally contaminated maize kernels [1].
After an equilibration period (2 h to 3 h) at room temperature, the gas in the head space of each vial was analyzed by the e-nose with a flow of 400 mL/min. During the analysis (60 s), each sensor's signal was recorded when the sensor's response curve showed a stabilized conductance for at least 10 s. During this time (measurement time), data from the raw sensor signals for each maize sample were recorded (with a 1 s interval). Thus, each sample was analyzed in duplicate and run three times (i.e., 6 observation per samples). The model has been fitted with the mean of 6 observations for each second (N of measurements) per sample (i.e., the sample odor profile).
Ten different descriptors, representing each sensor of the e-nose, were used to detect mycotoxin contamination. After acquisition, all sensor signal measurements were collected in Excel files and used for dataset assembly.

Statistical Analysis
Based on the levels of AF and FM measured by LFIA, each sample was assigned to one of three classes: (i) below the regulatory limits (AF < 5 ppb, and FM < 4 ppm); (ii) single-contaminated, above the regulatory limit for 1 mycotoxin (AF > 5 ppb, or FM > 4 ppm); or (iii) co-contaminated, above the regulatory limit for 2 mycotoxins (AF > 5ppb, and FM > 4 ppm).
Discriminant function analysis (DFA) was done on the e-nose data. The analysis was done on the original dataset with 9660 measurements (161 samples × 10 MOS × 2 replicates × 3 runs). All analysis was done with IBM SPSS 22.0 (IBM Corp., Armonk, NY, USA) predictive analytics software. In DFA, the aim is basically to build up a predictive model for group membership. The model is composed of a discriminant function based on linear combinations of predictor variables. In addition, those predictor variables provide the best discrimination between groups. The stepwise variable selection was conducted to ensure getting the most significant variables to carry out DFA. This method shows the variables that give the best results for DFA, by regarding the Wilk's test results. Leave-out-one-sample cross-validation was also done.   Funding: This study was funded by ATPr&d S.r.l.

Conflicts of Interest:
The authors declare no conflict of interest.