Genomic Prediction of Breeding Values when Modeling Genotype × Environment Interaction using Pedigree and Dense Molecular Markers

TLDR

Genomic selection is a key tool in plant and animal breeding, and multienvironment models can improve prediction accuracy by borrowing information across environments and traits. This study develops multienvironment genomic selection models and evaluates their predictive accuracy against models that omit pedigree or marker data. A statistical framework integrating pedigree and dense marker information was applied to wheat multienvironment trials, addressing two prediction scenarios: CV1 for untested genotypes and CV2 for genotypes missing data in some environments. Models that combined pedigree and marker data outperformed pedigree‑only and simple linear mixed models, demonstrating that multienvironment genomic selection substantially enhances trial evaluation.

Abstract

ABSTRACT Genomic selection (GS) has become an important aid in plant and animal breeding. Multienvironment (multitrait) models allow borrowing of information across environments (traits), which could enhance prediction accuracy. This study presents multienvironment (multitrait) models for GS and compares the predictive accuracy of these models with: (i) multienvironment analysis without pedigree and marker information, and (ii) multienvironment pedigree or/and marker‐based models. A statistical framework for incorporating pedigree and molecular marker information in models for multienvironment data is described and applied to data that originate from wheat ( Triticum aestivum L.) multienvironment trials. Two prediction problems relevant to plant breeders are considered: (CV1) predicting the performance of untested genotypes (“newly” developed lines), and (CV2) predicting the performance of genotypes that have been evaluated in some environments but not in others. Results confirmed the superiority of models using both marker and pedigree information over those based on pedigree information only. Models with pedigree and/or markers had better predictive accuracy than simple linear mixed models that do not include either of these two sources of information. We concluded that the evaluation of such trials can benefit greatly from using multienvironment GS models.