- Email: [email protected]
A multivariate regression model for detection of fumonisins content in maize from near infrared spectra
Food Chemistry 141 (2013) 4289–4294
Contents lists available at SciVerse ScienceDirect
Food Chemistry journal homepage: www.elsevier.com/locate/foodchem
Analytical Methods
A multivariate regression model for detection of fumonisins content in maize from near infrared spectra Della Riccia Giacomo, Del Zotto Stefania ⇑ Dept. of Mathematics and Computer Science – Research Center Norbert Wiener, University of Udine, via delle Scienze 206, 33100 Udine, Italy
a r t i c l e
i n f o
Article history: Received 27 November 2012 Received in revised form 3 May 2013 Accepted 3 July 2013 Available online 12 July 2013 Keywords: Fumonisins NIR spectroscopy Multivariate regression Statistical model PLS inference Calibration Full-cross validation
a b s t r a c t Fumonisins are mycotoxins produced by Fusarium species that commonly live in maize. Whereas fungi damage plants, fumonisins cause disease both to cattle breedings and human beings. Law limits set fumonisins tolerable daily intake with respect to several maize based feed and food. Chemical techniques assure the most reliable and accurate measurements, but they are expensive and time consuming. A method based on Near Infrared spectroscopy and multivariate statistical regression is described as a simpler, cheaper and faster alternative. We apply Partial Least Squares with full cross validation. Two models are described, having high correlation of calibration (0.995, 0.998) and of validation (0.908, 0.909), respectively. Description of observed phenomenon is accurate and overfitting is avoided. Screening of contaminated maize with respect to European legal limit of 4 mg kg1 should be assured. Ó 2013 Elsevier Ltd. All rights reserved.
1. Introduction Fumonisins, first isolated in 1988 (Gelderblom et al., 1988), are mycotoxins synthesized as secondary metabolites by sixteen fungi species primarily belonging to Fusarium genus. The main producers of fumonisins are Fusarium verticillioides (Saccardo) Nirenberg 1976 (synonym: Fusarium moniliforme J. Sheldon 1904; teleomorph Gibberella moniliformis Wineland) and Fusarium Proliferatum (Matsushima) Nirenberg 1976, both belonging to the Liseola section. Twenty-eight different fumonisins have been identified to date and they are clustered into four groups (A, B, C and P series) based on structural similarities. Occurrence of fumonisin B1 (FB1) and fumonisin B2 (FB2) are the most frequent. Fusarium species are among the most common natural contaminant of maize plants (Zea mays) worldwide (Munkvold and Desjardins, 1997) and compared with several cereals as sorghum, wheat and barley, maize has the highest fumonisins production. In particular FB1 usually accounts for about 70% of the total fumonisins content found in naturally infected maize (Rheeder et al., 2002). Fusarium have been historically recognized as the cause of different potentially serious disease of maize-fed livestock. Fumonisins, indeed, bring on acute or chronic primary mycotoxicosis since they inhibit the biosynthesis of complex ⇑ Corresponding author. Tel.: +39 0432 558429; fax: +39 0432 558499. E-mail addresses: [email protected] (D.R. Giacomo), [email protected] (D.Z. Stefania). 0308-8146/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.foodchem.2013.07.021
sphingolipids, endogen antitumor agents. Moreover fumonisins mycotoxicosis can involve secondary maize-fed livestock disease, so that fumonisins can be considered carcinogens and immunosuppressive causing sometimes animals’ death. For instance, FB1 is directly related to equine leukoencephalomalacia (Marasas et al., 1988) and pig pulmonary edema (Harrison et al., 1990). In several animal species, as rats and mice, it is hepatoxic and nefrotoxic (Gelderblom et al., 1991), then it is linked to hypercholesterolemia, immunological alteration, renal and liver toxicity. Fumonisins are characterized by high thermostability, since their molecular structure can be destroyed only after thermic exposure greater than 220 °C. Moreover fumonisins are not metabolized by animals so that, due to carry over, their occurrences could be found in cattle products like milk, cheese, meat or eggs. As a consequence, human contamination has two kind of sources: consumption of directly contaminated maize-based dietary staples or of indirectly toxic food produced by infected animals. A possible association between human esophageal cancer and a diet based on maize with high levels of fumonisins in South Africa has been suggested, although at present no definitive conclusion can be made. International Agency for Research on Cancer states that there is inadequate evidence in humans for carcinogenicity of fumonisins so that it defines FB1 belonging to ‘Group 2B carcinogens’, i.e. possibly carcinogenic to humans (IARC, 2002). Because of maize wide geographical distribution, Fusarium frequent occurrence in maize, overall high levels of fumonisins production and its association with known animal mycotoxicosis it
4290
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
is necessary to minimize risk related to consumption of fumonisins contaminated maize-based products both by maize-fed livestock and by human beings. Considering that, EU Scientific Committee established a Tolerable Daily Intake (TDI) for fumonisins with respect to several maize products intended to both animals or humans. Commission Regulation (EC) No. 1126/2007 (Commission Regulation, 2007) amending (EC) No. 1881/2006 (Commission Regulation, 2006b) sets maximum levels for certain contaminants in several foodstuffs as regards Fusarium toxins in maize and maize products, whereas Commission Recommendation 2006/576/EC (Commission Recommendation, 2006) reports law limits for cattle feed. Aim of this research is to explore Near Infrared (NIR) spectroscopy capabilities to detect fumonisins in maize. NIR spectroscopy, compared to traditional techniques based on chemical analysis like HPLC, appears as a cheaper, faster and simpler alternative. Chemometrics is an essential part of NIR spectroscopy, so that statistical inference based on multivariate regression methods (like Partial Least Squares calibration and full cross validation) can assure knowledge of fumonisins level in maize from its NIR spectrum. A preliminary study was suggested in Gaspardo et al. (2012). However, even if those results seemed to be promising, they showed two drawbacks. First of all, full cross validation values are poor, moreover, a specific analysis of wavelengths was not taken into account. As a consequence a further examination is done in order to (i) improve full cross validation outcomes; (ii) do a more precise survey of wavelengths that enter into the model; (iii) extract as much information as possible from observations. This paper presents a complete analysis in which two statistical models are developed with data measured from maize meal amounts. They describe relationship between fumonisins content and NIR spectra and they allow screening of maize to comply with European Regulation, a good classification ability and a rather accurate fumonisins prediction, since validation results are promising. The above models are relatively simple because fumonisins contents are expressed in terms of about twenty principal components, only. 1.1. Reference analytical methods Chemical analysis procedures are nowadays the most widely accepted reference methods for determining fumonisins in maize. Any general analytical method usually consist of several phases. After sampling, mycotoxin of interest should be extracted from matrix. The extraction is usually followed by a clean-up step where unwanted substances, which may interfere with the detection of the analyte, are removed from the extract. A final separation step will complete the procedure (Krska et al., 2007). Fusarium mycotoxins constitute a very heterogeneous group of compounds and separate procedures are usually utilized for quantification of individual toxins; however, protocols for simultaneous determination of toxins belonging to different groups are also available. Currently, quantitative methods of analysis for most Fusarium toxins are high-performance liquid chromatography (HPLC), gas chromatography (GC), enzyme-linked immunosorbent assay (ELISA), thin layer chromatography (TLC). Existing methods are considered time consuming (the more recent techniques require at least 30 min to process each mixture) because of extensive procedures which every phase consists of. Moreover expensive solvents and reagents are needed and only trained staff can do such complex chemical analyses. Finally, these methods are frequently destructive and involve large quantities of grain.
composition of the biological material being analyzed (Williams and Norris, 2001). NIR spectra can be collected either from the reflectance of element or the transmittance through substance. Both measurements provide an alternative technology for simultaneous determination of multiple constituents of biological material and they are commonly used to predict composition of bulk whole grain in maize (Orman and Schumann, 1991). Reflectance or transmittance visible and NIR spectroscopy has been used to detect relative composition of protein, oil and starch in maize kernel (Baye et al., 2006) and several single-kernel maize attributes such as oil content (Orman and Schumann, 1992) or levels of transgenic traits (Kramer et al., 2000). Automated systems for detecting attributes of single seeds have been described by Dowell et al. (1998) and Pearson (1999). Furthermore, NIR measurements are also applied to identify different mycotoxins in many commodities as DON in single wheat kernels (Dowell and Ram, 1999); aflatoxin in single whole corn kernels (Pearson et al., 2001); fumonisin in single corn kernels (Dowell et al., 2002); fumonisin FB1 in maize meal (Berardo et al., 2005); aflatoxin B1, ochratoxin A and total aflatoxins in spices as red paprika (Hernández-Hierro et al., 2008); aflatoxin B1 in maize and barley (Fernández-Ibañez et al., 2009); aflatoxin B1 in red chili powder (Tripathi and Mishra, 2009), just to name a few. When NIR spectra are measured from maize meal, attention should be taken during previous sample preparation phase. As in every measurement process, a standard procedure should be followed in order to assure reproducible sampling and to obtain spectra that can be compared and analyzed together avoiding noise in the following statistical model caused by some variables that are not taken into account. In particular, same granularity of maize meal amounts should be guaranteed during milling process and NIR should analyze every time the same quantity of maize meal, characterized by a homogeneous and compact surface in contact with the glass of the dish. Measurements of NIR spectra are faster than traditional techniques because materials can be screened rapidly (less than 1 min) and minimum preparation, as described above, is required with respect to elaborated and time consuming procedures of chemical methods. Moreover, NIR spectroscopy needs no reagents during analyses, so it is cheaper than analytical methods, too. Finally, NIR spectroscopy is a nondestructive technology since it preserves maize after the measurement for further analysis. 2. Materials and methods Maize amounts are collected in Friuli Venezia Giulia Region (Italy) from 4 drying and storage stations and 18 dairy farms located in 5 different areas previously detected on the basis of latitude, pedoclimatic conditions, agronomic features and farms concentration, such that representative data would be available. Maize just harvested are collected during September and October 2008 from drying and storage stations. Maize harvested and stored within farm silos for at least 8 months are taken three times (May, July and October) for two consecutive years (2008 and 2009) from dairy farms. Sampling is carried out according to EC Regulation No. 401/2006 (Commission Regulation, 2006a) and after drying at 60 °C over night maize grains are ground in a laboratory mill for 2 min at high speed, mixed accurately to ensure homogeneity and stored in a cool, dry place until analysis. 2.1. HPLC measurements
1.2. NIR spectroscopy Organic molecules have specific absorption patterns in the NIR region of electromagnetic spectrum that report chemical
In this research FB1 and FB2 concentration in maize is determined according to the AOAC Official Method 2001.04 (Visconti et al., 2001). FumoniTest Wide Bore immunoaffinity columns and
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
ortho-phthaldialdehyde solution are purchased from Vicam (Watertown, MA); HPLC-grade solvents and sodium dihydrogen phosphate are from Sigma–Aldrich (Chemie GbH, Steinhiem, Germany); fumonisins from Alexis Biochemicals, Axxora (Deutschland GmbH) is used; a Nova-Pack C18 (3.9 150 mm, 4 lm particle size, Waters, Ireland) is employed. HPLC system (Varian Analytical Instruments, USA) is equipped with a Prostar 363 fluorescence detector set at 335 nm excitation and 440 nm emission. Twenty grams of maize were analyzed each time. Fumonisins sum is used in the following statistical analysis and is denoted by [FB1 + FB2]. Values are in ‘ppm’ scale, that means ‘parts per million’ and is equivalent to mg kg1. 2.2. NIR measurements One hundred grams of maize meal are scanned with a Perkin Elmer FT-NIR spectrometer (Perkin Elmer Italia S.p.A.) using the integrating sphere with an adapter Sample Cup Spinner for rotation. Scanning time is approximately 50 s. Software Spectrum v.5.3 (Perkin Elmer Italia S.p.A.) installed in a personal computer interfaced to spectrometer records FT-NIR spectrum as a result of the following procedure. An interferogram, a component of the detector signal modulated as a function of optical path difference, is firstly stored. Successively the interferogram is converted to a frequency-domain single-beam spectrum. A reference reflectance spectrum, that had been taken before element’s scanning, is finally subtracted from the single-beam spectrum to give a spectrum collected in diffuse reflectance mode. This procedure is repeated until fifty spectra are done and automatically averaged into one absorbance spectrum. Absorbance is defined as the logarithm of the inverse of reflectance. Values are collected at 2 nm interval for wavelengths between 650 nm and 2500 nm (926 variables). The final spectral data are exported into standard JCAMP data format for the following statistical analysis. 2.3. Statistical inference: PLS regression Multivariate linear regression is applied in order to explain and quantify relationship between fumonisins contamination level (response or dependent variable) and maize NIR spectra (926 predictors or independent variables). With spectral data, advanced techniques are required because of predictors high cardinality and collinearity. Usually these methods first build new variables, called principal components (PC) as orthogonal linear combination of original predictors, then they specify a simple model between response and PC. Namely Partial Least Squares (PLS), introduced by Wold (1975), belong to this kind of regression methods and their form is a modification of NIPALS algorithm (Wold, 1966) for computing PC. PLS produce a sequence of models and estimate the best one through validation. In chemometrics literature, PLS have been heavily promoted as an alternative to ordinary least squares (OLS) in the poorly conditioned or ill-conditioned problems encountered here, see Martens and Naes, 2001 or Esbensen, 2004 for further details. With PLS analysis authors would define models that have a twofold purposes: (i) to assure screening of maize amounts with respect to threshold of 4 mg kg1 in order to comply European Regulation for human consumption of maize based products; (ii) to provide a tool for classification and rather accurate predictions into the whole scale of fumonisins contamination level. Both these aims together with properties that every statistical models should comply suggest several criteria that have to be followed as guide lines during the process of model definition. First of all final models should separate correctly and without false positives elements of dataset with respect to EU defined threshold of 4 mg kg1. Secondly, indexes of models goodness and their predictive ability should be satisfactory. In particular correlation
4291
between measured and predicted values is taken into account. Thirdly, residual variance should be decreasing and final models should not overcome a reasonable number of PC. In this research, all chemometrics analysis are developed with software ‘The Unscrambler v.9.6’ (CAMO Software AS, Oslo, Norway). Firstly, some preliminaries descriptive analysis are done both with graphic tools, as histogram or spectra visualization, and numerical results like main statistics of every variable (mean, min, max, standard deviation, number of missing data, etc.). Then, multivariate linear regression is applied choosing Partial Least Squares (PLS) with full cross validation. Every regression is applied to centered data. After each computation, outcome model is reviewed analyzing graphical and numerical outputs that allow to describe model properties and to compare results with criteria defined in this section.
3. Results and discussion The dataset contains information on maize amounts about their NIR spectra and their fumonisins contamination level, measured by HPLC technique. Every NIR spectrum consists of absorbance values for each wavelength from 650 nm to 2500 nm at intervals of 2 nm, recorded in 926 predictors. The sum of fumonisin FB1 and FB2 levels varies from 0.357 mg kg1 to 11.845 mg kg1 with a mean value equal to 3.353 mg kg1. Positive skewness and histogram confirm that statistically there are less highly contaminated elements than those with low value of fumonisins contamination. In particular there are only 59 elements with [FB1 + FB2] greater than 4 mg kg1, whereas the remaining 66% elements do not exceed this threshold. Regarding regression analysis, several PLS models with full cross validation have been tried in order to improve results presented in Gaspardo et al. (2012). At the end, the dataset reduces to a calibration set with 128 elements and the residual variance curve suggests a number of PC equal to 17. Correlation of calibration and validation are equal to 0.995 and 0.908, and values of SEC, RMSEC, SEP and RMSEP are respectively 0.232, 0.231, 0.933 and 0.929. Since both correlation values are rather high, we can say that our model: (i) describes with acceptable accuracy the relation between fumonisins content and NIR spectra avoiding overfitting, (ii) would have a good prediction ability for future unknown maize amounts that could be defined as members of the population to which dataset belongs. In order to evaluate screening and classification ability of the model, lines corresponding to target thresholds are added in calibration plot (Fig. 1). They allow to visualize easily if there are false positives and/or false negatives (by positive we mean contaminated above 4 mg kg1), if estimated values agree with measured one or if some element is not fitted in the right group. First of all, focusing on values around 4 mg kg1, we can observe that no element is misclassified and that legal limit is satisfied, without false positives and/or false negatives. So, the developed model assures a good screening ability. Some other interesting intervals can also be drawn. Identified groups are [0;1.3), [1.3;2.5), [2.5;3), [3;4), [4;5) and values higher than 5 mg kg1. No point is outside highlighted areas, since every calibrated contamination level is sufficiently accurate. As a consequence, the model can be also used as a classifier. We then applied the model to predict the remaining 45 elements that did not enter in the calibration set. Analysis of prediction plot (Fig. 2) allows to evaluate their behaviour and legal limit equal to 4 mg kg1 is still an interesting guide line. For example, there are 14 elements well classified by the model, even though not accurately predicted. Indeed, if measured value of these elements are lower (higher) than legal limit, their predicted value is
4292
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
Fig. 3. Second model: plot of calibration and validation residual variance versus number of PC.
Fig. 1. First model: calibration plot. Threshold equal to EU limit of 4 mg kg1 is highlighted as some other classes.
0.144, 0.143, 0.893 and 0.890. Calibration plot is analyzed in Fig. 4 where some interesting thresholds are drawn. They identify groups ½0; 1Þ; ½1; 2Þ; ½2; 2:5Þ; ½2:5; 3Þ; ½3; 3:5Þ; ½3:5; 4Þ; ½4; 5Þ; ½5; 6Þ; ½6; 7Þ and values greater than 7 mg kg1. While screening ability at 4 mg kg1 is still assured, accuracy is improved with respect to the previous model and more groups are highlighted. In particular, law limits equal to 1 mg kg1 and 2 mg kg1 are now in evidence, they are two other maximum levels established in Commission Regulation (EC) No. 1126/2007; we refer the reader to that document for more details. The remaining 40 elements are drawn in Fig. 5 where threshold equal to law limit of 4 mg kg1 is also depicted. There are 11 points correctly classified even if not accurately predicted. Then, there are 22 false positives plus 7 false negatives, that can be defined as outliers, in the sense that we assume that they do not belong to the model as said before. Eventually these outliers will be recovered by using future data. Comparing these two models, we can observe that (i) both are better than the model presented in Gaspardo et al. (2012) since all results are improved, screening ability is assured and classification can also be done; (ii) the last overcomes the former. Indeed, correlation values are increased then, a better screening ability and a greater accuracy are obtained. But which of the two models has a larger predictability? The answer is left for future research and in particular the possibility of having a bigger external test set could help considerably.
Fig. 2. First model: prediction plot. Threshold equal to 4 mg kg1 is highlighted.
lower (higher) than this threshold, too. There are 26 elements that can be considered false positives, because their prediction overcomes 4 mg kg1 even if their measured value is lower than legal limit. The remaining 5 elements are false negatives, i.e. they are contaminated maize amounts predicted as safe. These extreme predictions suggest that corresponding elements can be defined as outliers, i.e. statistical units that do not belong to the model. For the time being, they could be ignored and any comment about them should be postponed. After a more careful choice of elements that enter into the calibration set, another promising model was obtained. This model has 21 PC and is fitted on 133 elements. The curve of residual variance is a monotonically decreasing function (Fig. 3). Correlations of calibration and validation are very high, equal to 0:998 and 0:909, whereas SEC, RMSEC, SEP and RMSEP are respectively
Fig. 4. Second model: calibration plot. Threshold equal to EU limit of 4 mg kg1 is highlighted as some other classes.
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
4293
limits must be abode; fumonisins presence in maize should be periodically monitored in order to know as soon as possible if lots are contaminated. This provided information allows identification of maize with significant fumonisins levels so that infected batches can be segregated before they are commingled with maize characterized by low or negligible fumonisins content. Then, fumonisins contaminated maize can be safely removed from batches intended for food or feed use and assigned for other purpose, as energy production. Since traditional fumonisins detection procedure, like HPLC, are accurate but expensive and time consuming, in this research a method based on NIR spectra and multivariate regression is evaluated. The results obtained so far suggest that NIR might be a suitable alternative. However, further studies are desired in order to confirm its good properties, and eventually to develop more efficient models. Acknowledgements
Fig. 5. Second model: prediction plot. Threshold equal to 4 mg kg1 is highlighted.
As to the number of variables to be used in the definition of a model, all frequencies are considered as starting point, then in order to improve the results we checked whether some variables should be removed. We excluded the 52 variables belonging to the visible spectrum (650–752 nm), whose signal brings too much noise and makes extraction of significant information difficult. This is probably due to NIR instrument that works also with visible spectrum even if it should be considered an extreme range where spectra reproducibility, accuracy and precision cannot be assured. At the end, the last model computes the PC using 632 variables. We can state that excluded wavelengths are redundant and bring on noisy information. However, such aspects should be investigated more deeply and we leave to a chemist the choice of which range of NIR wavelengths is more appropriate. In conclusion, maize is a dietary staple both for cattle breedings and human beings in many countries. Fusarium development in maize is a very complex phenomenon that depends on environmental patterns (relations and interactions between plant, soil and weather properties). Fusarium species damage maize plant then, when suitable conditions appear, produce fumonisins as secondary metabolites. Factors that cause fumonisins infection are not so clear. Fumonisins role in fungi life is also not always known. Moreover, Fusarium presence does not always involve fumonisins occurrence and at the same time Fusarium absence does not imply lack of fumonisins contamination, since fumonisins can be observed without visible symptoms of Fusarium instances. Furthermore, both Fusarium and fumonisins can occur in every phase of the maize’s chain from sowing to storage, from field to mills. Actually, Fusarium and fumonisins contamination can be controlled, reduced but not completely removed (Munkvold and Desjardins, 1997) applying some reasonably achieved agricultural and manufacturing practices. Thus, suitable new strategies should be adopted in order to deal successfully with such problematic and likely dangerous situations. 4. Conclusion Taking into account that (i) Fusarium and fumonisins infection of maize causes both serious health consequences and economic loss due to decreased quality and quantity of batches; (ii) legal
This research is part of ‘Micosafe’, a project of University of Udine proposed and developed by Department of Agricultural and Environmental Sciences in cooperation with Department of Mathematics and Computer Science, International Research Center for the Mountain (CirMont) and Breeders Association FVG. We thank prof. Bruno Stefanon, prof. Giuseppe Firrao, doc. Brigitta Gaspardo, doc. Emanuela Torelli and Sirio Cividino. The project was supported with funding from Natural, Agricultural and Forestry Resources Office of Friuli Venezia Giulia Region. Aim of this project is to monitor in Friuli Venezia Giulia during three years, from January 2008 to December 2010, contamination rate of some mycotoxins in the whole supply chain, from maize grain and silage to unifeed and milk. Another goal is to find new faster and cheaper methods for mycotoxin analysis, comparing them with traditional analytical techniques. Finally ‘Micosafe’ wants to set some guidelines for mycotoxin prevention as regard both agricultural practices and cattle breeding management. We also would like to thank the anonymous reviewers for their excellent remarks; the modifications they suggested will certainly contribute to the clarity of the paper. References Baye, T. M., Pearson, T. C., & Settles, A. M. (2006). Development of a calibration to predict maize seed composition using single kernel near infrared spectroscopy. Journal of Cereal Science, 43, 236–243. Berardo, N., Pisacane, V., Battilani, P., Scandolara, A., Pietri, A., & Marocco, A. (2005). Rapid detection of kernel rots and mycotoxins in maize by Near-Infrared Reflectance Spectroscopy. Journal of Agricultural and Food Chemistry, 53, 8128–8134. Commission Recommendation (2006). Commission Recommendation of 17 August 2006 on the presence of deoxynivalenol, zearalenone, ochratoxin A, T-2 and HT2 and fumonisins in products intended for animal feeding (Text with EEA relevance) (2006/576/EC). Official Journal of the European Union, L229/7-9 (23.8.2006). Commission Regulation (2006). Commission Regulation (EC) No. 401/2006 of 23 February 2006 laying down the methods of sampling and analysis for the official control of the levels of mycotoxins in foodstuffs (Text with EEA relevance). Official Journal of the European Union, L70/12-34 (9.3.2006). Commission Regulation (2006). Commission Regulation (EC) No. 1881/2006 of 19 December 2006 setting maximum levels for certain contaminants in foodstuffs (Text with EEA relevance). Official Journal of the European Union, L364/5-24 (20.12.2006). Commission Regulation (2007). Commission Regulation (EC) No. 1126/2007 of 28 September 2007 amending Regulation (EC) No. 1881/2006 setting maximum levels for certain contaminants in foodstuffs as regards Fusarium toxins in maize and maize products. Official Journal of the European Union, L255/14-17 (29.9.2007). Dowell, F. E., Throne, J. E., & Baker, J. E. (1998). Automated nondestructive detection of internal insect infestation of wheat kernels by using Near-Infrared Reflectance Spectroscopy. Journal of Economic Entomology, 91, 899–904. Dowell, F. E., Pearson, T. C., Maghirang, E. B., Xie, F., & Wicklow, D. T. (2002). Reflectance and Transmittance Spectroscopy applied to detecting Fumonisin in single corn kernels infected with Fusarium verticillioides. Cereal Chemistry, 79(2), 222–226.
4294
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
Dowell, F. E., Ram, M. S., & Seitz, L. M. (1999). Predicting scab, vomitoxin, and ergosterol in single wheat kernels using Near-Infrared Spectroscopy. Cereal Chemistry, 76(4), 573–576. Esbensen, K. H. (2004). Multivariate data analysis – In practice (5th ed.). Aalborg University, Esbjerg: CAMO Process AS. Fernández-Ibañez, V., Soldado, A., Martínez-Fernández, A., & de la Roza-Delgado, B. (2009). Application of near infrared spectroscopy for rapid detection of aflatoxin B1 in maize and barley as analytical quality assessment. Food Chemistry, 113, 629–634. Gaspardo, B., Del Zotto, S., Torelli, E., Cividino, S. R., Firrao, G., Della Riccia, G., & Stefanon, B. (2012). A rapid method for detection of fumonisins B1 and B2 in corn meal using Fourier transform near infrared (FT-NIR) spectroscopy implemented with integrating sphere. Food Chemistry, 135, 1608–1612. Gelderblom, W. C. A., Jaskiewicz, K., Marasas, W. F. O., Thiel, P. G., Horak, R. M., Vleggaar, R., & Kriek, N. P. J. (1988). Fumonisins-novel mycotoxins with cancerpromoting activity produced by Fusarium moniliforme. Applied and Environmental Microbiology, 54(7), 1806–1811. Gelderblom, W. C. A., Kriek, N. P. J., Marasas, W. F. O., & Thiel, P. G. (1991). Toxicity and carcinogenicity of the Fusarium moniliforme metabolite, fumonisn B1 , in rats. Carcinogenesis, 12(7), 1247–1251. Harrison, L. R., Colvin, B. M., Greene, J. T., Newman, L. E., & Cole, J. R. (1990). Pulmonary Edema and hydrothorax in swine produced by fumonisin B1, a toxic metabolite of Fusarium moniliforme. Journal of Veterinary Diagnostic Investigation, 2, 217–221. Hernández-Hierro, J. M., García-Villanova, R. J., & Gonzáles-Martín, I. (2008). Potential of near infrared spectroscopy for the analysis of mycotoxins applied to naturally contaminated red paprika found in the Spanish market. Analytica Chimica Acta, 622, 189–194. International Agency for Research on Cancer Fumonisin B1 (2002). IARC Summaries, & Evaluations, 82, 301. Kramer, K. J., Morgan, T. D., Throne, J. E., Dowell, F. E., Bailey, M., & Howard, J. A. (2000). Transgenic avidin maize is resistant to storage insect pests. Nature Biotechnology, 18, 670–674. Krska, R., Welzig, E., & Boudra, H. (2007). Analysis of Fusarium toxins in feed. Animal Feed Science and Technology, 137, 241–264. Marasas, W. F. O., Kellerman, T. S., Gelderblom, W. C. A., Coetzer, J. A. W., Thiel, P. G., & van der Lugt, J. J. (1988). Leukoencephalomalacia in a horse induced by
fumonisin B1 isolated from Fusarium moniliforme. Onderstepoort Journal of Veterinary Research, 55, 197–203. Martens, H., & Naes, T. (2001). Multivariate calibration. Chichester, UK: Wiley. Munkvold, G. P., & Desjardins, A. E. (1997). Fumonisins in maize: Can we reduce their occurrence? Plant Disease, 81, 556–565. Orman, B. A., & Schumann, R. A. (1991). Comparison of near-infrared spectroscopy calibration methods for the prediction of protein, oil, and starch in maize grain. Journal of Agricultural and Food Chemistry, 39(5), 883–886. Orman, B. A., & Schumann, R. A. (1992). Nondestructive single-kernel oil determination of maize by near-infrared transmission spectroscopy. Journal of the American Oil Chemists’ Society, 69(10), 1036–1038. Pearson, T. C. (1999). Use of near infrared light transmission to automatically detect almonds with concealed damage. Lebensm. Wiss. Technology, 32, 73–78. Pearson, T. C., Wicklow, D. T., Maghirang, E. B., Xie, F., & Dowell, F. E. (2001). Detecting aflatoxin in single corn kernels by Transmittance and Reflectance Spectroscopy. Transactions of the American Society of Agricultural Engineers, 44(5), 1247–1254. Rheeder, J. P., Marasas, W. F. O., & Vismer, H. F. (2002). Production of Fumonisin analogs by fusarium species. Applied and Environmental Microbiology, 68(5), 2101–2105. Tripathi, S., & Mishra, H. N. (2009). A rapid FT-NIR method for estimation of aflatoxin B1 in red chili powder. Food Control, 20, 840–846. Visconti, A., Solfrizzo, M., & De Girolamo, A. (2001). Determination of Fumonisins B1 and B2 in corn and Corn Flakes by Liquid Chromatography with Immunoaffinity Column Cleanup: Collaborative Study. Journal of AOAC, 84, 1828–1837. Williams, P., & Norris, K. (2001). Near-infrared technology in the agricultural and food industries (2nd ed.). St Paul, MN: American Association of Cereal Chemists Inc.. Wold, H. (1966). Estimation of principal component and related models by iterative least squares. In P. R. Krishnaiah (Ed.), Multivariate Analysis (pp. 391–420). New York: Academic Press. Wold, H. (1975). Path models with latent variables: The non-linear iterative partial least squares (NIPALS) approach. In H. M. Blalock et al. (Eds.), Quantitative sociology: Intentional perspective on mathematical and statistical modeling (pp. 307–357). Academic Press.
Contents lists available at SciVerse ScienceDirect
Food Chemistry journal homepage: www.elsevier.com/locate/foodchem
Analytical Methods
A multivariate regression model for detection of fumonisins content in maize from near infrared spectra Della Riccia Giacomo, Del Zotto Stefania ⇑ Dept. of Mathematics and Computer Science – Research Center Norbert Wiener, University of Udine, via delle Scienze 206, 33100 Udine, Italy
a r t i c l e
i n f o
Article history: Received 27 November 2012 Received in revised form 3 May 2013 Accepted 3 July 2013 Available online 12 July 2013 Keywords: Fumonisins NIR spectroscopy Multivariate regression Statistical model PLS inference Calibration Full-cross validation
a b s t r a c t Fumonisins are mycotoxins produced by Fusarium species that commonly live in maize. Whereas fungi damage plants, fumonisins cause disease both to cattle breedings and human beings. Law limits set fumonisins tolerable daily intake with respect to several maize based feed and food. Chemical techniques assure the most reliable and accurate measurements, but they are expensive and time consuming. A method based on Near Infrared spectroscopy and multivariate statistical regression is described as a simpler, cheaper and faster alternative. We apply Partial Least Squares with full cross validation. Two models are described, having high correlation of calibration (0.995, 0.998) and of validation (0.908, 0.909), respectively. Description of observed phenomenon is accurate and overfitting is avoided. Screening of contaminated maize with respect to European legal limit of 4 mg kg1 should be assured. Ó 2013 Elsevier Ltd. All rights reserved.
1. Introduction Fumonisins, first isolated in 1988 (Gelderblom et al., 1988), are mycotoxins synthesized as secondary metabolites by sixteen fungi species primarily belonging to Fusarium genus. The main producers of fumonisins are Fusarium verticillioides (Saccardo) Nirenberg 1976 (synonym: Fusarium moniliforme J. Sheldon 1904; teleomorph Gibberella moniliformis Wineland) and Fusarium Proliferatum (Matsushima) Nirenberg 1976, both belonging to the Liseola section. Twenty-eight different fumonisins have been identified to date and they are clustered into four groups (A, B, C and P series) based on structural similarities. Occurrence of fumonisin B1 (FB1) and fumonisin B2 (FB2) are the most frequent. Fusarium species are among the most common natural contaminant of maize plants (Zea mays) worldwide (Munkvold and Desjardins, 1997) and compared with several cereals as sorghum, wheat and barley, maize has the highest fumonisins production. In particular FB1 usually accounts for about 70% of the total fumonisins content found in naturally infected maize (Rheeder et al., 2002). Fusarium have been historically recognized as the cause of different potentially serious disease of maize-fed livestock. Fumonisins, indeed, bring on acute or chronic primary mycotoxicosis since they inhibit the biosynthesis of complex ⇑ Corresponding author. Tel.: +39 0432 558429; fax: +39 0432 558499. E-mail addresses: [email protected] (D.R. Giacomo), [email protected] (D.Z. Stefania). 0308-8146/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.foodchem.2013.07.021
sphingolipids, endogen antitumor agents. Moreover fumonisins mycotoxicosis can involve secondary maize-fed livestock disease, so that fumonisins can be considered carcinogens and immunosuppressive causing sometimes animals’ death. For instance, FB1 is directly related to equine leukoencephalomalacia (Marasas et al., 1988) and pig pulmonary edema (Harrison et al., 1990). In several animal species, as rats and mice, it is hepatoxic and nefrotoxic (Gelderblom et al., 1991), then it is linked to hypercholesterolemia, immunological alteration, renal and liver toxicity. Fumonisins are characterized by high thermostability, since their molecular structure can be destroyed only after thermic exposure greater than 220 °C. Moreover fumonisins are not metabolized by animals so that, due to carry over, their occurrences could be found in cattle products like milk, cheese, meat or eggs. As a consequence, human contamination has two kind of sources: consumption of directly contaminated maize-based dietary staples or of indirectly toxic food produced by infected animals. A possible association between human esophageal cancer and a diet based on maize with high levels of fumonisins in South Africa has been suggested, although at present no definitive conclusion can be made. International Agency for Research on Cancer states that there is inadequate evidence in humans for carcinogenicity of fumonisins so that it defines FB1 belonging to ‘Group 2B carcinogens’, i.e. possibly carcinogenic to humans (IARC, 2002). Because of maize wide geographical distribution, Fusarium frequent occurrence in maize, overall high levels of fumonisins production and its association with known animal mycotoxicosis it
4290
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
is necessary to minimize risk related to consumption of fumonisins contaminated maize-based products both by maize-fed livestock and by human beings. Considering that, EU Scientific Committee established a Tolerable Daily Intake (TDI) for fumonisins with respect to several maize products intended to both animals or humans. Commission Regulation (EC) No. 1126/2007 (Commission Regulation, 2007) amending (EC) No. 1881/2006 (Commission Regulation, 2006b) sets maximum levels for certain contaminants in several foodstuffs as regards Fusarium toxins in maize and maize products, whereas Commission Recommendation 2006/576/EC (Commission Recommendation, 2006) reports law limits for cattle feed. Aim of this research is to explore Near Infrared (NIR) spectroscopy capabilities to detect fumonisins in maize. NIR spectroscopy, compared to traditional techniques based on chemical analysis like HPLC, appears as a cheaper, faster and simpler alternative. Chemometrics is an essential part of NIR spectroscopy, so that statistical inference based on multivariate regression methods (like Partial Least Squares calibration and full cross validation) can assure knowledge of fumonisins level in maize from its NIR spectrum. A preliminary study was suggested in Gaspardo et al. (2012). However, even if those results seemed to be promising, they showed two drawbacks. First of all, full cross validation values are poor, moreover, a specific analysis of wavelengths was not taken into account. As a consequence a further examination is done in order to (i) improve full cross validation outcomes; (ii) do a more precise survey of wavelengths that enter into the model; (iii) extract as much information as possible from observations. This paper presents a complete analysis in which two statistical models are developed with data measured from maize meal amounts. They describe relationship between fumonisins content and NIR spectra and they allow screening of maize to comply with European Regulation, a good classification ability and a rather accurate fumonisins prediction, since validation results are promising. The above models are relatively simple because fumonisins contents are expressed in terms of about twenty principal components, only. 1.1. Reference analytical methods Chemical analysis procedures are nowadays the most widely accepted reference methods for determining fumonisins in maize. Any general analytical method usually consist of several phases. After sampling, mycotoxin of interest should be extracted from matrix. The extraction is usually followed by a clean-up step where unwanted substances, which may interfere with the detection of the analyte, are removed from the extract. A final separation step will complete the procedure (Krska et al., 2007). Fusarium mycotoxins constitute a very heterogeneous group of compounds and separate procedures are usually utilized for quantification of individual toxins; however, protocols for simultaneous determination of toxins belonging to different groups are also available. Currently, quantitative methods of analysis for most Fusarium toxins are high-performance liquid chromatography (HPLC), gas chromatography (GC), enzyme-linked immunosorbent assay (ELISA), thin layer chromatography (TLC). Existing methods are considered time consuming (the more recent techniques require at least 30 min to process each mixture) because of extensive procedures which every phase consists of. Moreover expensive solvents and reagents are needed and only trained staff can do such complex chemical analyses. Finally, these methods are frequently destructive and involve large quantities of grain.
composition of the biological material being analyzed (Williams and Norris, 2001). NIR spectra can be collected either from the reflectance of element or the transmittance through substance. Both measurements provide an alternative technology for simultaneous determination of multiple constituents of biological material and they are commonly used to predict composition of bulk whole grain in maize (Orman and Schumann, 1991). Reflectance or transmittance visible and NIR spectroscopy has been used to detect relative composition of protein, oil and starch in maize kernel (Baye et al., 2006) and several single-kernel maize attributes such as oil content (Orman and Schumann, 1992) or levels of transgenic traits (Kramer et al., 2000). Automated systems for detecting attributes of single seeds have been described by Dowell et al. (1998) and Pearson (1999). Furthermore, NIR measurements are also applied to identify different mycotoxins in many commodities as DON in single wheat kernels (Dowell and Ram, 1999); aflatoxin in single whole corn kernels (Pearson et al., 2001); fumonisin in single corn kernels (Dowell et al., 2002); fumonisin FB1 in maize meal (Berardo et al., 2005); aflatoxin B1, ochratoxin A and total aflatoxins in spices as red paprika (Hernández-Hierro et al., 2008); aflatoxin B1 in maize and barley (Fernández-Ibañez et al., 2009); aflatoxin B1 in red chili powder (Tripathi and Mishra, 2009), just to name a few. When NIR spectra are measured from maize meal, attention should be taken during previous sample preparation phase. As in every measurement process, a standard procedure should be followed in order to assure reproducible sampling and to obtain spectra that can be compared and analyzed together avoiding noise in the following statistical model caused by some variables that are not taken into account. In particular, same granularity of maize meal amounts should be guaranteed during milling process and NIR should analyze every time the same quantity of maize meal, characterized by a homogeneous and compact surface in contact with the glass of the dish. Measurements of NIR spectra are faster than traditional techniques because materials can be screened rapidly (less than 1 min) and minimum preparation, as described above, is required with respect to elaborated and time consuming procedures of chemical methods. Moreover, NIR spectroscopy needs no reagents during analyses, so it is cheaper than analytical methods, too. Finally, NIR spectroscopy is a nondestructive technology since it preserves maize after the measurement for further analysis. 2. Materials and methods Maize amounts are collected in Friuli Venezia Giulia Region (Italy) from 4 drying and storage stations and 18 dairy farms located in 5 different areas previously detected on the basis of latitude, pedoclimatic conditions, agronomic features and farms concentration, such that representative data would be available. Maize just harvested are collected during September and October 2008 from drying and storage stations. Maize harvested and stored within farm silos for at least 8 months are taken three times (May, July and October) for two consecutive years (2008 and 2009) from dairy farms. Sampling is carried out according to EC Regulation No. 401/2006 (Commission Regulation, 2006a) and after drying at 60 °C over night maize grains are ground in a laboratory mill for 2 min at high speed, mixed accurately to ensure homogeneity and stored in a cool, dry place until analysis. 2.1. HPLC measurements
1.2. NIR spectroscopy Organic molecules have specific absorption patterns in the NIR region of electromagnetic spectrum that report chemical
In this research FB1 and FB2 concentration in maize is determined according to the AOAC Official Method 2001.04 (Visconti et al., 2001). FumoniTest Wide Bore immunoaffinity columns and
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
ortho-phthaldialdehyde solution are purchased from Vicam (Watertown, MA); HPLC-grade solvents and sodium dihydrogen phosphate are from Sigma–Aldrich (Chemie GbH, Steinhiem, Germany); fumonisins from Alexis Biochemicals, Axxora (Deutschland GmbH) is used; a Nova-Pack C18 (3.9 150 mm, 4 lm particle size, Waters, Ireland) is employed. HPLC system (Varian Analytical Instruments, USA) is equipped with a Prostar 363 fluorescence detector set at 335 nm excitation and 440 nm emission. Twenty grams of maize were analyzed each time. Fumonisins sum is used in the following statistical analysis and is denoted by [FB1 + FB2]. Values are in ‘ppm’ scale, that means ‘parts per million’ and is equivalent to mg kg1. 2.2. NIR measurements One hundred grams of maize meal are scanned with a Perkin Elmer FT-NIR spectrometer (Perkin Elmer Italia S.p.A.) using the integrating sphere with an adapter Sample Cup Spinner for rotation. Scanning time is approximately 50 s. Software Spectrum v.5.3 (Perkin Elmer Italia S.p.A.) installed in a personal computer interfaced to spectrometer records FT-NIR spectrum as a result of the following procedure. An interferogram, a component of the detector signal modulated as a function of optical path difference, is firstly stored. Successively the interferogram is converted to a frequency-domain single-beam spectrum. A reference reflectance spectrum, that had been taken before element’s scanning, is finally subtracted from the single-beam spectrum to give a spectrum collected in diffuse reflectance mode. This procedure is repeated until fifty spectra are done and automatically averaged into one absorbance spectrum. Absorbance is defined as the logarithm of the inverse of reflectance. Values are collected at 2 nm interval for wavelengths between 650 nm and 2500 nm (926 variables). The final spectral data are exported into standard JCAMP data format for the following statistical analysis. 2.3. Statistical inference: PLS regression Multivariate linear regression is applied in order to explain and quantify relationship between fumonisins contamination level (response or dependent variable) and maize NIR spectra (926 predictors or independent variables). With spectral data, advanced techniques are required because of predictors high cardinality and collinearity. Usually these methods first build new variables, called principal components (PC) as orthogonal linear combination of original predictors, then they specify a simple model between response and PC. Namely Partial Least Squares (PLS), introduced by Wold (1975), belong to this kind of regression methods and their form is a modification of NIPALS algorithm (Wold, 1966) for computing PC. PLS produce a sequence of models and estimate the best one through validation. In chemometrics literature, PLS have been heavily promoted as an alternative to ordinary least squares (OLS) in the poorly conditioned or ill-conditioned problems encountered here, see Martens and Naes, 2001 or Esbensen, 2004 for further details. With PLS analysis authors would define models that have a twofold purposes: (i) to assure screening of maize amounts with respect to threshold of 4 mg kg1 in order to comply European Regulation for human consumption of maize based products; (ii) to provide a tool for classification and rather accurate predictions into the whole scale of fumonisins contamination level. Both these aims together with properties that every statistical models should comply suggest several criteria that have to be followed as guide lines during the process of model definition. First of all final models should separate correctly and without false positives elements of dataset with respect to EU defined threshold of 4 mg kg1. Secondly, indexes of models goodness and their predictive ability should be satisfactory. In particular correlation
4291
between measured and predicted values is taken into account. Thirdly, residual variance should be decreasing and final models should not overcome a reasonable number of PC. In this research, all chemometrics analysis are developed with software ‘The Unscrambler v.9.6’ (CAMO Software AS, Oslo, Norway). Firstly, some preliminaries descriptive analysis are done both with graphic tools, as histogram or spectra visualization, and numerical results like main statistics of every variable (mean, min, max, standard deviation, number of missing data, etc.). Then, multivariate linear regression is applied choosing Partial Least Squares (PLS) with full cross validation. Every regression is applied to centered data. After each computation, outcome model is reviewed analyzing graphical and numerical outputs that allow to describe model properties and to compare results with criteria defined in this section.
3. Results and discussion The dataset contains information on maize amounts about their NIR spectra and their fumonisins contamination level, measured by HPLC technique. Every NIR spectrum consists of absorbance values for each wavelength from 650 nm to 2500 nm at intervals of 2 nm, recorded in 926 predictors. The sum of fumonisin FB1 and FB2 levels varies from 0.357 mg kg1 to 11.845 mg kg1 with a mean value equal to 3.353 mg kg1. Positive skewness and histogram confirm that statistically there are less highly contaminated elements than those with low value of fumonisins contamination. In particular there are only 59 elements with [FB1 + FB2] greater than 4 mg kg1, whereas the remaining 66% elements do not exceed this threshold. Regarding regression analysis, several PLS models with full cross validation have been tried in order to improve results presented in Gaspardo et al. (2012). At the end, the dataset reduces to a calibration set with 128 elements and the residual variance curve suggests a number of PC equal to 17. Correlation of calibration and validation are equal to 0.995 and 0.908, and values of SEC, RMSEC, SEP and RMSEP are respectively 0.232, 0.231, 0.933 and 0.929. Since both correlation values are rather high, we can say that our model: (i) describes with acceptable accuracy the relation between fumonisins content and NIR spectra avoiding overfitting, (ii) would have a good prediction ability for future unknown maize amounts that could be defined as members of the population to which dataset belongs. In order to evaluate screening and classification ability of the model, lines corresponding to target thresholds are added in calibration plot (Fig. 1). They allow to visualize easily if there are false positives and/or false negatives (by positive we mean contaminated above 4 mg kg1), if estimated values agree with measured one or if some element is not fitted in the right group. First of all, focusing on values around 4 mg kg1, we can observe that no element is misclassified and that legal limit is satisfied, without false positives and/or false negatives. So, the developed model assures a good screening ability. Some other interesting intervals can also be drawn. Identified groups are [0;1.3), [1.3;2.5), [2.5;3), [3;4), [4;5) and values higher than 5 mg kg1. No point is outside highlighted areas, since every calibrated contamination level is sufficiently accurate. As a consequence, the model can be also used as a classifier. We then applied the model to predict the remaining 45 elements that did not enter in the calibration set. Analysis of prediction plot (Fig. 2) allows to evaluate their behaviour and legal limit equal to 4 mg kg1 is still an interesting guide line. For example, there are 14 elements well classified by the model, even though not accurately predicted. Indeed, if measured value of these elements are lower (higher) than legal limit, their predicted value is
4292
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
Fig. 3. Second model: plot of calibration and validation residual variance versus number of PC.
Fig. 1. First model: calibration plot. Threshold equal to EU limit of 4 mg kg1 is highlighted as some other classes.
0.144, 0.143, 0.893 and 0.890. Calibration plot is analyzed in Fig. 4 where some interesting thresholds are drawn. They identify groups ½0; 1Þ; ½1; 2Þ; ½2; 2:5Þ; ½2:5; 3Þ; ½3; 3:5Þ; ½3:5; 4Þ; ½4; 5Þ; ½5; 6Þ; ½6; 7Þ and values greater than 7 mg kg1. While screening ability at 4 mg kg1 is still assured, accuracy is improved with respect to the previous model and more groups are highlighted. In particular, law limits equal to 1 mg kg1 and 2 mg kg1 are now in evidence, they are two other maximum levels established in Commission Regulation (EC) No. 1126/2007; we refer the reader to that document for more details. The remaining 40 elements are drawn in Fig. 5 where threshold equal to law limit of 4 mg kg1 is also depicted. There are 11 points correctly classified even if not accurately predicted. Then, there are 22 false positives plus 7 false negatives, that can be defined as outliers, in the sense that we assume that they do not belong to the model as said before. Eventually these outliers will be recovered by using future data. Comparing these two models, we can observe that (i) both are better than the model presented in Gaspardo et al. (2012) since all results are improved, screening ability is assured and classification can also be done; (ii) the last overcomes the former. Indeed, correlation values are increased then, a better screening ability and a greater accuracy are obtained. But which of the two models has a larger predictability? The answer is left for future research and in particular the possibility of having a bigger external test set could help considerably.
Fig. 2. First model: prediction plot. Threshold equal to 4 mg kg1 is highlighted.
lower (higher) than this threshold, too. There are 26 elements that can be considered false positives, because their prediction overcomes 4 mg kg1 even if their measured value is lower than legal limit. The remaining 5 elements are false negatives, i.e. they are contaminated maize amounts predicted as safe. These extreme predictions suggest that corresponding elements can be defined as outliers, i.e. statistical units that do not belong to the model. For the time being, they could be ignored and any comment about them should be postponed. After a more careful choice of elements that enter into the calibration set, another promising model was obtained. This model has 21 PC and is fitted on 133 elements. The curve of residual variance is a monotonically decreasing function (Fig. 3). Correlations of calibration and validation are very high, equal to 0:998 and 0:909, whereas SEC, RMSEC, SEP and RMSEP are respectively
Fig. 4. Second model: calibration plot. Threshold equal to EU limit of 4 mg kg1 is highlighted as some other classes.
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
4293
limits must be abode; fumonisins presence in maize should be periodically monitored in order to know as soon as possible if lots are contaminated. This provided information allows identification of maize with significant fumonisins levels so that infected batches can be segregated before they are commingled with maize characterized by low or negligible fumonisins content. Then, fumonisins contaminated maize can be safely removed from batches intended for food or feed use and assigned for other purpose, as energy production. Since traditional fumonisins detection procedure, like HPLC, are accurate but expensive and time consuming, in this research a method based on NIR spectra and multivariate regression is evaluated. The results obtained so far suggest that NIR might be a suitable alternative. However, further studies are desired in order to confirm its good properties, and eventually to develop more efficient models. Acknowledgements
Fig. 5. Second model: prediction plot. Threshold equal to 4 mg kg1 is highlighted.
As to the number of variables to be used in the definition of a model, all frequencies are considered as starting point, then in order to improve the results we checked whether some variables should be removed. We excluded the 52 variables belonging to the visible spectrum (650–752 nm), whose signal brings too much noise and makes extraction of significant information difficult. This is probably due to NIR instrument that works also with visible spectrum even if it should be considered an extreme range where spectra reproducibility, accuracy and precision cannot be assured. At the end, the last model computes the PC using 632 variables. We can state that excluded wavelengths are redundant and bring on noisy information. However, such aspects should be investigated more deeply and we leave to a chemist the choice of which range of NIR wavelengths is more appropriate. In conclusion, maize is a dietary staple both for cattle breedings and human beings in many countries. Fusarium development in maize is a very complex phenomenon that depends on environmental patterns (relations and interactions between plant, soil and weather properties). Fusarium species damage maize plant then, when suitable conditions appear, produce fumonisins as secondary metabolites. Factors that cause fumonisins infection are not so clear. Fumonisins role in fungi life is also not always known. Moreover, Fusarium presence does not always involve fumonisins occurrence and at the same time Fusarium absence does not imply lack of fumonisins contamination, since fumonisins can be observed without visible symptoms of Fusarium instances. Furthermore, both Fusarium and fumonisins can occur in every phase of the maize’s chain from sowing to storage, from field to mills. Actually, Fusarium and fumonisins contamination can be controlled, reduced but not completely removed (Munkvold and Desjardins, 1997) applying some reasonably achieved agricultural and manufacturing practices. Thus, suitable new strategies should be adopted in order to deal successfully with such problematic and likely dangerous situations. 4. Conclusion Taking into account that (i) Fusarium and fumonisins infection of maize causes both serious health consequences and economic loss due to decreased quality and quantity of batches; (ii) legal
This research is part of ‘Micosafe’, a project of University of Udine proposed and developed by Department of Agricultural and Environmental Sciences in cooperation with Department of Mathematics and Computer Science, International Research Center for the Mountain (CirMont) and Breeders Association FVG. We thank prof. Bruno Stefanon, prof. Giuseppe Firrao, doc. Brigitta Gaspardo, doc. Emanuela Torelli and Sirio Cividino. The project was supported with funding from Natural, Agricultural and Forestry Resources Office of Friuli Venezia Giulia Region. Aim of this project is to monitor in Friuli Venezia Giulia during three years, from January 2008 to December 2010, contamination rate of some mycotoxins in the whole supply chain, from maize grain and silage to unifeed and milk. Another goal is to find new faster and cheaper methods for mycotoxin analysis, comparing them with traditional analytical techniques. Finally ‘Micosafe’ wants to set some guidelines for mycotoxin prevention as regard both agricultural practices and cattle breeding management. We also would like to thank the anonymous reviewers for their excellent remarks; the modifications they suggested will certainly contribute to the clarity of the paper. References Baye, T. M., Pearson, T. C., & Settles, A. M. (2006). Development of a calibration to predict maize seed composition using single kernel near infrared spectroscopy. Journal of Cereal Science, 43, 236–243. Berardo, N., Pisacane, V., Battilani, P., Scandolara, A., Pietri, A., & Marocco, A. (2005). Rapid detection of kernel rots and mycotoxins in maize by Near-Infrared Reflectance Spectroscopy. Journal of Agricultural and Food Chemistry, 53, 8128–8134. Commission Recommendation (2006). Commission Recommendation of 17 August 2006 on the presence of deoxynivalenol, zearalenone, ochratoxin A, T-2 and HT2 and fumonisins in products intended for animal feeding (Text with EEA relevance) (2006/576/EC). Official Journal of the European Union, L229/7-9 (23.8.2006). Commission Regulation (2006). Commission Regulation (EC) No. 401/2006 of 23 February 2006 laying down the methods of sampling and analysis for the official control of the levels of mycotoxins in foodstuffs (Text with EEA relevance). Official Journal of the European Union, L70/12-34 (9.3.2006). Commission Regulation (2006). Commission Regulation (EC) No. 1881/2006 of 19 December 2006 setting maximum levels for certain contaminants in foodstuffs (Text with EEA relevance). Official Journal of the European Union, L364/5-24 (20.12.2006). Commission Regulation (2007). Commission Regulation (EC) No. 1126/2007 of 28 September 2007 amending Regulation (EC) No. 1881/2006 setting maximum levels for certain contaminants in foodstuffs as regards Fusarium toxins in maize and maize products. Official Journal of the European Union, L255/14-17 (29.9.2007). Dowell, F. E., Throne, J. E., & Baker, J. E. (1998). Automated nondestructive detection of internal insect infestation of wheat kernels by using Near-Infrared Reflectance Spectroscopy. Journal of Economic Entomology, 91, 899–904. Dowell, F. E., Pearson, T. C., Maghirang, E. B., Xie, F., & Wicklow, D. T. (2002). Reflectance and Transmittance Spectroscopy applied to detecting Fumonisin in single corn kernels infected with Fusarium verticillioides. Cereal Chemistry, 79(2), 222–226.
4294
D.R. Giacomo, D.Z. Stefania / Food Chemistry 141 (2013) 4289–4294
Dowell, F. E., Ram, M. S., & Seitz, L. M. (1999). Predicting scab, vomitoxin, and ergosterol in single wheat kernels using Near-Infrared Spectroscopy. Cereal Chemistry, 76(4), 573–576. Esbensen, K. H. (2004). Multivariate data analysis – In practice (5th ed.). Aalborg University, Esbjerg: CAMO Process AS. Fernández-Ibañez, V., Soldado, A., Martínez-Fernández, A., & de la Roza-Delgado, B. (2009). Application of near infrared spectroscopy for rapid detection of aflatoxin B1 in maize and barley as analytical quality assessment. Food Chemistry, 113, 629–634. Gaspardo, B., Del Zotto, S., Torelli, E., Cividino, S. R., Firrao, G., Della Riccia, G., & Stefanon, B. (2012). A rapid method for detection of fumonisins B1 and B2 in corn meal using Fourier transform near infrared (FT-NIR) spectroscopy implemented with integrating sphere. Food Chemistry, 135, 1608–1612. Gelderblom, W. C. A., Jaskiewicz, K., Marasas, W. F. O., Thiel, P. G., Horak, R. M., Vleggaar, R., & Kriek, N. P. J. (1988). Fumonisins-novel mycotoxins with cancerpromoting activity produced by Fusarium moniliforme. Applied and Environmental Microbiology, 54(7), 1806–1811. Gelderblom, W. C. A., Kriek, N. P. J., Marasas, W. F. O., & Thiel, P. G. (1991). Toxicity and carcinogenicity of the Fusarium moniliforme metabolite, fumonisn B1 , in rats. Carcinogenesis, 12(7), 1247–1251. Harrison, L. R., Colvin, B. M., Greene, J. T., Newman, L. E., & Cole, J. R. (1990). Pulmonary Edema and hydrothorax in swine produced by fumonisin B1, a toxic metabolite of Fusarium moniliforme. Journal of Veterinary Diagnostic Investigation, 2, 217–221. Hernández-Hierro, J. M., García-Villanova, R. J., & Gonzáles-Martín, I. (2008). Potential of near infrared spectroscopy for the analysis of mycotoxins applied to naturally contaminated red paprika found in the Spanish market. Analytica Chimica Acta, 622, 189–194. International Agency for Research on Cancer Fumonisin B1 (2002). IARC Summaries, & Evaluations, 82, 301. Kramer, K. J., Morgan, T. D., Throne, J. E., Dowell, F. E., Bailey, M., & Howard, J. A. (2000). Transgenic avidin maize is resistant to storage insect pests. Nature Biotechnology, 18, 670–674. Krska, R., Welzig, E., & Boudra, H. (2007). Analysis of Fusarium toxins in feed. Animal Feed Science and Technology, 137, 241–264. Marasas, W. F. O., Kellerman, T. S., Gelderblom, W. C. A., Coetzer, J. A. W., Thiel, P. G., & van der Lugt, J. J. (1988). Leukoencephalomalacia in a horse induced by
fumonisin B1 isolated from Fusarium moniliforme. Onderstepoort Journal of Veterinary Research, 55, 197–203. Martens, H., & Naes, T. (2001). Multivariate calibration. Chichester, UK: Wiley. Munkvold, G. P., & Desjardins, A. E. (1997). Fumonisins in maize: Can we reduce their occurrence? Plant Disease, 81, 556–565. Orman, B. A., & Schumann, R. A. (1991). Comparison of near-infrared spectroscopy calibration methods for the prediction of protein, oil, and starch in maize grain. Journal of Agricultural and Food Chemistry, 39(5), 883–886. Orman, B. A., & Schumann, R. A. (1992). Nondestructive single-kernel oil determination of maize by near-infrared transmission spectroscopy. Journal of the American Oil Chemists’ Society, 69(10), 1036–1038. Pearson, T. C. (1999). Use of near infrared light transmission to automatically detect almonds with concealed damage. Lebensm. Wiss. Technology, 32, 73–78. Pearson, T. C., Wicklow, D. T., Maghirang, E. B., Xie, F., & Dowell, F. E. (2001). Detecting aflatoxin in single corn kernels by Transmittance and Reflectance Spectroscopy. Transactions of the American Society of Agricultural Engineers, 44(5), 1247–1254. Rheeder, J. P., Marasas, W. F. O., & Vismer, H. F. (2002). Production of Fumonisin analogs by fusarium species. Applied and Environmental Microbiology, 68(5), 2101–2105. Tripathi, S., & Mishra, H. N. (2009). A rapid FT-NIR method for estimation of aflatoxin B1 in red chili powder. Food Control, 20, 840–846. Visconti, A., Solfrizzo, M., & De Girolamo, A. (2001). Determination of Fumonisins B1 and B2 in corn and Corn Flakes by Liquid Chromatography with Immunoaffinity Column Cleanup: Collaborative Study. Journal of AOAC, 84, 1828–1837. Williams, P., & Norris, K. (2001). Near-infrared technology in the agricultural and food industries (2nd ed.). St Paul, MN: American Association of Cereal Chemists Inc.. Wold, H. (1966). Estimation of principal component and related models by iterative least squares. In P. R. Krishnaiah (Ed.), Multivariate Analysis (pp. 391–420). New York: Academic Press. Wold, H. (1975). Path models with latent variables: The non-linear iterative partial least squares (NIPALS) approach. In H. M. Blalock et al. (Eds.), Quantitative sociology: Intentional perspective on mathematical and statistical modeling (pp. 307–357). Academic Press.