Multienvironment and Multitrait Genomic Selection Models in Unbalanced Early‐Generation Wheat Yield Trials

Brian P Ward,Priyanka Tyagi,Carl A Griffey,Clay H Sneller,Frederic L Kolb,Gina Brown-Guedira,David A Van Sanford

doi:10.2135/cropsci2018.03.0189

Abstract

The majority of studies evaluating genomic selection (GS) for plant breeding have used single‐trait, single‐site models that ignore genotype × environment interaction (GEI) effects. However, such studies do not accurately reflect the complexities of many applied breeding programs, and previous papers have found that models that incorporate GEI effects and multiple traits can increase the accuracy of genomic estimated breeding values (GEBVs). This study's goal was to test GS methods for prediction in scenarios that simulate early‐generation yield testing by correcting for field spatial variation, and fitting multienvironment and multitrait models on data for 14 traits of varying heritability evaluated in unbalanced designs across four environments. Corrections for spatial variation increased across‐environment trait heritability by 25%, on average, but had little effect on model predictive ability. Results between all models were generally equivalent when predicting the performance of newly introduced genotypes. However, models incorporating GEI information and multiple traits increased prediction accuracy by up to 9.6% for low‐heritability traits when phenotypic data were sparsely collected across environments. The results suggest that GS models using multiple traits and incorporating GEI effects may best be suited to predicting line performance in new environments when phenotypic data have already been collected across a subset of the total testing environments.

Highlights

The majority of studies evaluating genomic selection (GS) for plant breeding have used single-trait, single-site models that ignore genotypeenvironment interaction (GEI) effects
Genomic selection (GS) is a method of marker-based selection that uses a large number of markers spread throughout the genome, such that every quantitative trait locus (QTL) affecting a trait is assumed to be in high linkage disequilibrium (LD) with at least one marker
The genomic relationship matrix (GRM) calculated from the genotypic data showed several groups of highly interrelated genotypes present within the Virginia and Illinois germplasm (Supplemental Fig. S1), and a principal component analysis of the Single nucleotide polymorphism (SNP) matrix showed that Virginia genotypes formed a large cluster, with a second, overlapping cluster forming from a combination of Kentucky and Illinois genotypes (Supplemental Fig. S2)

Summary

Introduction

The majority of studies evaluating genomic selection (GS) for plant breeding have used single-trait, single-site models that ignore genotypeenvironment interaction (GEI) effects. Jarquín et al (2014) introduced a reaction-norm model modeling GEI as functions of markers and environmental covariates Their results demonstrated that the incorporation of GEI effects into the model could increase both across- and within-environment predictive ability, depending on the specific model specification and cross-validation procedure. Zhang et al (2015) studied GS across multiple environments using a biparental maize population and found that complex traits such as grain yield benefitted the most from incorporation of GEI effects, whereas simpler traits such as anthesis date demonstrated more modest gains in predictive ability. They used a strategy of performing GS prediction within megaenvironments to increase the ability to predict genotype performance in new environments

Objectives

Methods

Results

Discussion

Conclusion