Genomic Selection Using Factor-Analytic Regression to Predict Hybrid Performance in Untested Environments by Leveraging Environmental Data

Genotype-by-environment interaction (G×E) remains one of the major challenges to optimizing plant breeding programs and achieving consistent genetic gains. Most approaches characterize G×E only within the environments already sampled by the trial network, limiting recommendations to a fraction of the target population of environments (TPE) that breeding programs actually serve. To advance our understanding of methods that better incorporate GxE into breeding decisions, we implemented the Geographic Information Systems–Factor Analytic (GIS-FA) framework, which integrates factor analytic (FA) models with environmental covariates and genomic kernels to predict hybrid performance in untested environments. Using a multi-environment trial dataset of 117 maize hybrids evaluated for grain yield across 41 environments in Brazil (2010-2014), each defined as a location-year combination, and genotyped with 27,250 SNP markers, we tested this framework under three cross-validation schemes, simulating prediction of an untested environment, an untested location, and an untested year. Unlike traditional G×E models, GIS-FA links environmental descriptors directly to the latent factor loadings estimated by FA models. These loadings are then extrapolated to new environments using partial least squares regression, decoupling environmental characterization from the FA structure and allowing many environmental covariates to be incorporated regardless of trial network size. By incorporating a nonlinear Gaussian genomic kernel into the FA framework, we hypothesized that we could potentially capture additive and non-additive genetic effects. Indeed, the predictive accuracy improved by up to 9.7% relative to the environment-only model, with the largest gain observed when predicting performance in a completely untested year. Pairwise genetic correlations among tested environments further revealed a heterogeneous G×E structure with substantial crossover interaction, reinforcing the value of spatially targeted, rather than broadly generalized, hybrid recommendations. Beyond predictive accuracy, GIS-FA generated performance maps, "which-won-where" maps, and pairwise comparison maps across the Brazilian territory. The results highlight the practical advantages of GIS-FA for tropical maize breeding programs: (i) robust prediction of hybrid performance beyond experimental networks; (ii) strategic genotype recommendation in the target population of environments (TPEs); and (iii) improved accuracy of site-specific predictions through the addition of genomic information.