Modeling the impact of resource allocation decisions on genomic prediction using maize multi‐environment data

Abstract

AbstractIn a hybrid maize (Zea mays L.) breeding program that utilizes genomic selection, resource allocation used in phenotypic data acquisition must be balanced between population size, number of environments, and the number of testers used for generating hybrids. Plant breeders evaluate newly developed inbred lines using multi‐environment trials to account for genotype‐by‐environment interaction effects. The replication of hybrids across environments in these trials impacts the training data accuracy for developing genomic prediction models. This study examined the impact of resource allocation scenarios on genomic prediction accuracy using a multi‐environment trial dataset generated using inbred lines crossed to multiple testers. A total of 369 Stiff Stalk double haploid lines from a synthetic mapping population were testcrossed to three non‐Stiff Stalk inbred lines as testers, PHZ51, PHK76, and PHP02, and evaluated across 34 environments by the Genomes to Fields Initiative in 2020 and 2021. Resource allocation scenarios significantly impacted site‐specific genomic prediction accuracy for unobserved hybrids in unobserved environments. A training set with three to five environments that had the highest quality data produced similar prediction accuracy as data from 10 random environments for both observed and unobserved hybrids, indicating that strong prediction models can be built with a limited set of environments for both grain yield and plant height. We found that resource‐efficient prediction models that use data from one tester and three to five environments can effectively conduct selection of untested hybrids and in untested environments. Public research programs are often limited in testing capacity, and this study provides support for genomic selection in resource‐limited breeding programs.