Aike Potze 'Prediction of GxExM with multimodal deep learning'

Prediction of GxExM with multimodal deep learning
Aike Potze (1), Fred van Eeuwijk (2), Ioannis N. Athanasiadis (1)
1 Artificial Intelligence Group, Wageningen University and Research. 2 Biometris, Wageningen University and Research.

Predicting yield for future genotypes and environments is a long-standing challenge in agriculture. Deep neural networks have emerged as a promising approach to this task, due to their flexibility as universal function approximators in learning nonlinear functions that can represent complex biological processes. Surprisingly, recent intercomparison studies have found deep neural networks underperforming in comparison to classical linear methods and machine learning methods, even for continent-scale multi-environment trials (Washburn et al., 2024). We show that a simple approach, combining a linear mixed model and deep neural networks, results in nonlinear models that predict yield effectively in the next year of a multi-environment trial. We employ a linear mixed model to decompose plot yield into genetic, environmental and interaction effects, and use these effects as targets during the training of deep neural networks. Our method improves prediction of both future genetic main effect of untested hybrids and environment main effect of future trials, as well as reaching a new state-of-the-art performance in yield prediction at the plot level on an open benchmark dataset.

References:
Washburn, J. D., Varela, J. I., Xavier, A., Chen, Q., Ertl, D., Gage, J. L., Holland, J. B., Lima, D. C., Romay, M. C., Lopez-Cruz, M., de Los Campos, G., Barber, W., Zimmer, C., Trucillo Silva, I., Rocha, F., Rincent, R., Ali, B., Hu, H., Runcie, D. E., … de Leon, N. (2024). Global Genotype by environment prediction competition reveals that diverse modeling strategies can deliver satisfactory maize yield estimates. bioRxiv, 2024.09. 13.612969.