Risk-oriented Prediction of Deoxynivalenol and T-2 Toxin Contamination in Malian Cottonseed Products using Fungal and Storage Indicators
Mamadou Kanadjigui
CERFITEX, Ségou, Mali.
Hamssatou Aliou Maïga
Agence Nationale de la Sécurité Sanitaire des Aliments (ANSSA), Institute National de la Santé Publique (INSP), and Technology of Bamako (USTTB), Bamako, Mali.
Kangaye Amadou Diallo
LaboREM-Biotech, Faculty of Sciences and Techniques (FST), University of Sciences, Techniques and Technology of Bamako (USTTB), Bamako, Mali and African Center of Excellence in Bioinformatics (ACE-B), University of Sciences, Techniques and Technology of Bamako (USTTB), Bamako, Mali.
Awa Soronfé Doumbia
CERFITEX, Ségou, Mali.
Zoumana Traore
CERFITEX, Ségou, Mali.
Amadou Hamadoun Babana *
LaboREM-Biotech, Faculty of Sciences and Techniques (FST), University of Sciences, Techniques and Technology of Bamako (USTTB), Bamako, Mali.
*Author to whom correspondence should be addressed.
Abstract
Aims: To determine whether fungal loads and storage-environment measurements can predict deoxynivalenol and T-2 toxin concentrations in Malian cottonseed and cottonseed cake, and to compare statistical and machine-learning algorithms for prioritising lots for confirmatory testing within a feed-safety surveillance framework.
Study Design: A cross-sectional analytical study with internal validation of prediction models.
Place and Duration of Study: Cotton products were sampled in Koutiala, Sikasso and Ségou, Mali, from November to September 2024.
Methodology: Eighty-one independent composite samples were analysed in triplicate, and analytical replicates were averaged at the sample level before modelling. Temperature, relative humidity, Fusarium spp., Aspergillus flavus and Aspergillus niger formed the primary predictor set. Linear regression, Elastic Net, Random Forest and Extreme Gradient Boosting were compared using five-fold outer cross-validation repeated three times, three-fold inner-loop tuning and group-aware splitting of identical analytical profiles. Performance was assessed using root mean squared error, mean absolute error and the coefficient of determination.
Results: Random Forest performed best for both toxins. With five predictors, mean root mean squared error, mean absolute error and R² were 10.52 µg/kg, 7.97 µg/kg and 0.933 for deoxynivalenol, and 6.74 µg/kg, 5.45 µg/kg and 0.913 for T-2 toxin, respectively. Fungal variables alone retained similar performance, whereas temperature and relative humidity alone were less predictive. Adding total Aspergillus counts provided no material improvement. After 15 identical profiles were excluded, R² remained 0.934 for deoxynivalenol and 0.916 for T-2 toxin.
Conclusion: Fungal loads contained most of the predictive information in this dataset, and Random Forest captured relationships that linear models represented less effectively. The findings support a preliminary feed-safety surveillance approach in which higher-risk cottonseed lots are prioritised for confirmatory toxin analysis. The models do not replace laboratory measurement, establish regulatory compliance or constitute a formal exposure assessment. Prospective external validation is required before operational use.
Keywords: Cottonseed, cottonseed cake, deoxynivalenol, T-2 toxin, Fusarium, feed safety, mycotoxin surveillance, Mali