Empirical vs. mechanistic models for deoxynivolenol prediction in wheat.

Fusarium head blight (FHB) is among the major causes of reduced quality in wheat and its products and the main concern regards related mycotoxins along the cereal food-feed chain; Deoxynivalenol (DON) and its acetylatede derivatives 3-and 15-acetyl deoxiynivalenol, zearalenone, T-2 and HT-2 toxins are the most important. European Commission established legal limits in cereal products, with the maxim level of DON in unprocessed soft and durum wheat at 1250 µg/kg and 1750 µg/kg, respectively. Legal limits were also defined for zearalenone, while only recommendations were delivered for T2 and HT2 toxins. The use of models to predict the outcome of the DON contamination at harvest is desirable to enhance and trigger management opportunities. Moreover, accurate predictions of DON occurrence from Fusarium spp. fungi are important because its presence is not strictly related with visible symptoms. Mycotoxin predictive models have potential to support rationale and sustainable management strategies and thus contribute to safe food production. To date, several models have been developed to predict DON contamination in wheat. Based on the approach followed for their development, mathematical models can be distinguished into mechanistic and empirical. The former are drawn considering the cause-effect relationship among variables, while the latter describe the relation between the driving factors of the phenomena and are developed by statistical analyses of data collected in field. One mechanistic model for the prediction of DON in wheat has been published/developed in Italy, while several empiric models were developed in different countries worldwide and, a neural network. The empirical model recently developed in The Netherlands, with the aim to predict in wheat at harvest with output information destined to different groups of end-users, was selected in this study. This study aimed to compare the predictive performance/approach of the mechanistic and empirical models, find out pros and cons in the two approaches and consider possible advantages in their joint use. Results showed good prediction of both modeling approaches for independent wheat fields (sampled in Italy and The Netherlands). Both models predicted correctly DON content in wheat kernels below the legal limit in around 90% of the samples. Comparing the output of the models in the cross validation, samples correctly predicted as not contaminated by one model are confirmed frequently also by the other; therefore, predictions are in agreement and the reliability of predictions is consequently maximised. The joint application of two models give different advantages in practice. Data input are diverse and predictions are available both during the growing season and at harvest, when outputs can be compared. Output data in agreement reduces the uncertainty of predictions and those divers underline the uncertainty. Besides, both the idea of powerfulness of mechanistic models because based on underlying processes driving patterns and the notion that the variables interact the way they do and the assumption that the relationship extends past the measured values are both exploited with this approach. Research supported by EC KBBE-2007-222690-2 MYORED. Marco Camardo Leggieri carried out this work within the PhD School “Agrisystem” of the Università Cattolica del Sacro Cuore (Italy).