Predição de micotoxinas em cereais e subprodutos via espectroscopia de reflectância no infravermelho próximo

Abstract

Cereal grain quality can be altered by the presence of fungi, which may produce mycotoxins that have the potential to cause harm to human and animal health. It is thus essential to monitor the levels of such substances in products intended for consumption. Despite being efficient, current analytical methodologies are time consuming and require the use of varied reagents and instruments, so the Industry demands the development of fast and reliable methods to expedite decision making. In this setting, the present work employs near infrared reflectance spectroscopy (NIRS) in association with chemometric methods for quantification and classification to build multivariate models for predicting mycotoxins. Four studies were performed in the following matrices: corn; corn distiller’s dried grains with solubles (DDGS); and wheat flour. The analyzed mycotoxins were aflatoxin B1 (AFB1), fumonisin B1 (FB1) + fumonisin B2 (FB2) (total fumonisins, FBs), deoxynivalenol (DON) and zearalenone (ZEA). Spectral data were processed through partial least squares and the number of principal components of the models was determined by cross-validation. Liquid chromatography coupled to tandem mass spectrometry was used as the reference methodology. The first study developed prediction curves for FBs and ZEA in corn. Correlation coefficient (R), determination coefficient and residual prediction deviation (RPD) for FBs and ZEA were, respectively: 0.809 and 0.991; 0.899 and 0.984; and 3.33 and 2.71. The second study assessed mycotoxicological prevalence and chemical composition (water activity, crude protein, ether extract, starch and apparent metabolisable energy in poultry) in 8,854 spectra of corn originating from Argentina, Bolivia, Brazil (stratified per regions), Colombia and Peru in 2020. FBs showed the greatest prevalence in South American as well as in Brazilian samples: 91.6% and 92.6%, respectively. Crude protein ranged from 6.7% in Colombia to 8.4% in Bolivia in relation to the mean (7.4%). The chemical composition of the samples from the Southeast region of Brazil presented the largest positive variability in relation to the means. The third study was conducted in DDGS and elaborated prediction curves for FB1 and FB2. One hundred ninety samples were used to build the models, being 132 for calibration and 58 for external validation. The results of R and RPD for FB1 and FB2 were, respectively: 0.90 and 0.88; and 2.16 and 2.06. The fourth study evaluated DONcontaminated wheat flour samples using partial least-squares discriminant analysis (PLS-DA) and principal components-linear discriminant analysis (PC-LDA). The samples were classified according to the maximum tolerated limit (MTL) for DON in Brazil, 750 μg.kg-¹, and two groups were established for the calibration set: category A (≤ 450 μg kg-¹), non-contaminated or below the MTL; and category B (> 450 μg kg-¹), contaminated or above the MTL. Validation samples analyzed via PLS-DA showed correct classification rates between 85 and 87.5%; for PC-LDA, the hit rate was over 85%. Both methods presented a 10-15% error. The results achieved through these studies evidence the potential of the alternative technology NIRS to be used in the Industry, providing agility to the analytical process of the ingredients. Therefore, decisions can be made assertively and thus ensure food quality and safety.