Genomic selection (GS) has successfully been used in plant breeding to improve selection efficiency and reduce breeding time and cost. However, there is not a clear strategy on how to incorporate genotype × environment interaction (GEI) to GS models. Increased prediction accuracy could be achieved using mixed models to exploit GEI by borrowing information from other environments. The objective of this work was to compare strategies to exploit GEI in GS using mixed models. Specifically, we compared strategies to predict new genotypes by borrowing information from other environments modeling the correlation matrix across environments and to design sets of environments aiming for low GEI to predict genomic performance in new environments. We evaluated 1477 advanced wheat (Triticum aestivum L.) lines for yield in 35 location–year combinations genotyped with genotyping‐by‐sequencing (GBS). Mixed models were used to obtain either overall or by‐environment predictions for different sets of environments. Overall accuracy was high (0.5). Borrowing information from relatives evaluated in multiple environments and modeling the correlation matrix across environments was the best strategy to predict new genotypes. On the other hand, the best strategy for predicting the performance of genotypes in new environments was either to predict across locations for single years or to predict within defined mega‐environments (MEs) for any year or location. In summary, higher predictive ability was obtained by characterizing and by modeling GEI in the GS context.

The efficient use of testing resources is one of the key factors for successful plant breeding programs. Controlling micro‐ and macro‐environmental variability is an effective way of improving the testing efficiency and the selection of superior genotypes. Common experimental designs in genotypic testing usually use replicated or augmented experiments at each location, but they are balanced across locations. Some studies suggest that the increase in population size even at the expense of balanced experiments might be beneficial if genotype × environment interaction (GEI) is modeled. The objective of this study was to compare strategies for micro and macro‐environmental variability control that include GEI information to optimize resource allocation in multi‐environment trials (METs). Six experimental designs combined with four spatial correction models were compared for efficiency under three experimental sizes using simulations under a real yield variability map. Additionally, six resource allocation strategies were evaluated in terms of accuracy and the expected response to selection. The α‐lattice (ALPHA) experimental design was the best one at controlling micro‐environmental variability. The moderate mega‐environmental design (MED) strategy had the largest response to selection. This strategy uses historical mega‐environments (MEs) to unbalance genotypic testing within MEs while modeling GEI. The MED was the best resource allocation strategy and could potentially increase selection response up to 43% in breeding programs when genotypes are evaluated in METs.

Controlling spatial variation in agricultural field trials is the most important step to compare treatments efficiently and accurately. Spatial variability can be controlled at the experimental design level with the assignment of treatments to experimental units and at the modeling level with the use of spatial corrections and other modeling strategies. The goal of this study was to compare the efficiency of methods used to control spatial variation in a wide range of scenarios using a simulation approach based on real wheat data. Specifically, classic and spatial experimental designs with and without a two-dimensional autoregressive spatial correction were evaluated in scenarios that include differing experimental unit sizes, experiment sizes, relationships among genotypes, genotype by environment interaction levels, and trait heritabilities. Fully replicated designs outperformed partially and unreplicated designs in terms of accuracy; the alpha-lattice incomplete block design was best in all scenarios of the medium-sized experiments. However, in terms of response to selection, partially replicated experiments that evaluate large population sizes were superior in most scenarios. The AR1 $$\times $$× AR1 spatial correction had little benefit in most scenarios except for the medium-sized experiments with the largest experimental unit size and low GE. Overall, the results from this study provide a guide to researchers designing and analyzing large field experiments.Supplementary materials accompanying this paper appear online.