Using self‐organizing maps, genomic estimated breeding values, and environmental covariates in predictive breeding to improve performance and stability in maize ( Zea mays L.)

Abstract To develop genotypes, breeders test and select genotypes across sites and years representative of their target population of environments (TPE). This is not cost‐effective at early breeding stages when evaluating thousands of genotypes. We demonstrate the use of self‐organizing maps (SOMs), genomic estimated breeding values (GEBV), and recursive partitioning to identify environmental groups (EGs), environmental covariates differentiating among EGs, and hybrids stable across EGs. Genotypes were predicted based on 310 site‐specific estimation sets trained on as many sites in the Argentine maize TPE between 2018 and 2023. Analyses showed that the first cropping season in the Center and South was associated with drought and heat stress during grain fill; the second cropping season was likely to experience drought during the vegetative phase. The second season in the North showed higher temperatures and higher vapor pressure deficit (VPD). SOM grouped sites with similar genotypic ranking for grain yield in repeatable EGs irrespective of geographical zone, season, or year for lines and hybrids. Drought, heat, and high VPD at different phenological stages explained differences in yield among EGs. Lines and hybrids stable across EGs were identified using SOM. We infer that these hybrids would show good adaptation and stability across conditions used for model training. Using site‐specific GEBVs in an analogous way to phenotypic data allowed simulating the performance of lines and hybrids at early breeding stages beyond solely relying on phenotypic evaluations. Since SOM predominantly focuses on yield, it is critical to use SOM in combination with other methods and information available.