ABSTRACT
Genomic prediction is a useful tool to accelerate genetic gain in selection using DNA marker information. However, this technology usually relies on models that are not designed to accommodate population heterogeneity, which results from differences in marker signals across genetic backgrounds. Previous studies have proposed to cope with population heterogeneity using diverse approaches: (i) either ignoring it, therefore relying on the robustness of standard approaches; (ii) reducing it, by selecting homogenous subsets of individuals in the sample; or (iii) modelling it by using interactive models. In this study we assessed all three possible approaches, applying existing and novel procedures for each of them. All procedures developed are based on deterministic optimizations, can account for heteroscedasticity, and are applicable in contexts of admixed populations. In a case study on a diverse switchgrass sample, we compared the procedures to a control where predictions rely on homogeneous subsamples. Ignoring heterogeneity was often not detrimental, and sometimes beneficial, to prediction accuracy, compared to the control. Reducing heterogeneity did not result in further increases in accuracy. However, in scenarios of limited subsample sizes, a novel procedure, which accounted for redundancy within subsamples, outperformed the existing procedure, which only considered relationships to selection candidates. Modelling heterogeneity resulted in substantial increases in accuracy, in the cases where accounting for population heterogeneity yielded a highly significant improvement in fit. Our study exemplifies advantages and limits of the various approaches that are promising in various contexts of population heterogeneity, e.g. prediction based on historical datasets or dynamic breeding.