gogna@ipk-gatersleben.de1, reif@ipk-gatersleben.de1, zhao@ipk-gatersleben.de1

Breeding programs in Europe often prioritize average genotype performance in target environments, potentially overlooking the late-stage selection of environment-optimized candidates. The latter translates into suboptimal cultivar performance at the farm level, especially under environment diversification exacerbated by climate change.

To address this challenge, we propose leveraging BigData and Artificial Intelligence (AI) for genomic predictions. In this study, we compiled data on winter wheat grain yield from 13,285 genotypes across 98,175 yield plots in diverse 117 central European environments from 2010 to 2022. With integrated genomic data we found that as the size of the training dataset increased, convolutional neural networks (CNN) outperformed traditional genomic best linear unbiased predictions (GBLUP) in predicting average genotype performance.

We then expanded our prediction models to account for genotype-environment interactions (GxE) by additionally including enviromic data. In doing so, we observed a notable improvement of up to 22% in predicting environment-specific performance of new hybrids within a network of testing environments. Lastly, to better understand environment factors driving GxE, we conducted analyses on a core set comprising 500 genetically diverse lines in the studied environments.

Using machine learning, we successfully identified pivotal markers for environment classification in Central Europe.

Our results suggest that BigData not only facilitates absorption of AI but also offers new avenues for widening the genetic bottleneck often encountered when progressing candidates from early limited-environment to late stage multi-environment evaluations.