Genomic selection allowing for marker‐by‐environment interaction

Abstract Genomic selection has been routinely implemented in plant breeding in two stages. The first stage usually omits the marker information and estimates adjusted means of genotypes across environments. The second stage uses the adjusted means to predict genomic breeding values. However, if the effects of markers vary substantially between different environments, it may be important to account for this variation for varieties adapted to different environments. Using two maize data sets, we investigated whether modelling the marker‐by‐environment interaction can improve the predictive ability of genomic selection relative to modelling genotype‐by‐environment interaction alone. Modelling the marker‐by‐environment interaction did not substantially increase the predictive ability relative to modelling only the genotype‐by‐environment interaction for the two tested data sets. Thus, genomic selection, carried out in a stagewise fashion, such that the marker information is omitted until the last stage of the process, may suffice for most practical purposes. Moreover, predictive ability did not reduce substantially even when the number of markers with consistent effects across environments used for genomic prediction was reduced to about 50.