The transfer learning model TabPFN achieved the highest overall performance in predicting mycotoxin contamination in Irish oat crops within a multi-response prediction framework. Contamination from mycotoxins presents a major ongoing threat to cereal crop quality, food safety, and agricultural productivity. Reliable and accurate prediction of mycotoxin levels is designed to support proactive intervention strategies and reduce substantial economic losses across the agricultural industry.
To address this multi-response prediction challenge, researchers compiled and analyzed a dedicated dataset of oat samples collected across Ireland. The dataset incorporated a diverse combination of environmental, agronomic, and geographical predictors to capture the varying conditions associated with crop development. Using these diverse variables, the study investigated how advanced neural networks and specialized transfer learning architectures compare when modeling complex contamination outcomes.
The study evaluated five distinct machine learning modeling approaches across the oat dataset. These methods consisted of a baseline multilayer perceptron (MLP), an MLP trained with a pre-training phase, and three separate tabular transfer learning architectures: TabPFN, TabNet, and FT-Transformer. The research team evaluated each modeling approach across both regression and classification metrics, calculating and reporting performance outcomes on average as well as separately for each individual toxin.
Performance assessments demonstrated that the transfer learning model TabPFN delivered the strongest results across both regression and classification evaluations. In regression testing, TabPFN achieved a root mean squared error (RMSE) of 0.65 and an R-squared (R2) value of 0.86. In classification testing, the model recorded an area under the receiver operating characteristic curve (AUC) of 0.98 alongside an F1 score of 0.92. Across the evaluated metrics, the baseline multilayer perceptron followed closely behind TabPFN in overall predictive effectiveness.
In addition to evaluating model accuracy, the researchers conducted a permutation-based variable importance analysis to identify the most influential predictors across both prediction tasks. The analysis revealed that weather history patterns recorded during the 90-day pre-harvest period served as the most important predictors of mycotoxin contamination. Alongside these 90-day pre-harvest weather patterns, grain moisture content emerged as a critical predictor determining the presence and severity of contamination in the harvest.
Subject of Research: Prediction of mycotoxin contamination in Irish oats using machine learning
Article Title: Predicting Mycotoxin Contamination in Irish Oats Using Deep and Transfer Learning
Article References: Inglis, A., Doohan, F., Natarajan, S., McNulty, B., Elliott, C., Nugent, A., Meneely, J., Greer, B., Kildea, S., Bucur, D., Danaher, M., Rocco, M. D., Black, L., Gauley, A., McKenna, N., & Parnell, A. (2026). Predicting Mycotoxin Contamination in Irish Oats Using Deep and Transfer Learning. npj Science of Food. https://doi.org/10.1038/s41538-026-01208-5
Image Credits: AI Generated
DOI: 10.1038/s41538-026-01208-5
Keywords: mycotoxin contamination, Irish oats, transfer learning, TabPFN, deep learning, multilayer perceptron, TabNet, FT-Transformer, food safety, crop quality, agricultural productivity, grain moisture, weather history, multi-response prediction, variable importance analysis, environmental predictors
News Source: Alan Morgan. (November 19, 2026). Transfer Learning Model Accurately Predicts Mycotoxin Contamination in Irish Oat Crops. Scienmag.



