Phenomic vs Genomic Prediction - A Comparison of Prediction Accuracies for Grain Yield in Hard Winter Wheat Lines

Abstract

Common bread wheat (Triticum aestivum L.) is a key component of global diets, but the genetic improvement of wheat is not keeping pace with the growing demands of the worlds population. To increase efficiency and reduce costs, breeding programs are rapidly adopting the use of unoccupied aerial vehicles (UAVs) to conduct high-throughput spectral analyses. This study examined the effectiveness of multispectral indices in predicting grain yield compared to genomic prediction. Multispectral data were collected on advanced generation yield nursery trials during the 2019-2021 growing seasons in the Colorado State University Wheat Breeding Program. Genome-wide genotyping was performed on these advanced generations and all plots were harvested to measure grain yield. Two methods were used to predict grain yield: genomic estimated breeding values (GEBVs) generated by a genomic best linear unbiased prediction (gBLUP) model and phenomic phenotypic estimates (PPEs) using only spectral indices via multiple linear regression (MLR), k-nearest neighbors (KNN), and random forest (RF) models. In cross-validation, PPEs produced by MLR, KNN, and RF models had higher prediction accuracy ([r]: 0.41 [≤][r] [≤] 0.48) than GEBVs produced by gBLUP ([r] = 0.35). In leave-one-year-out forward validation using only multispectral data for 2020 and 2021, PPEs from MLR and KNN models had higher prediction accuracy of grain yield than GEBVs of those same lines. These findings suggest that a limited number of spectra may produce PPEs that are more accurate than or equivalently accurate as GEBVs derived from gBLUP, and this method should be evaluated in earlier development material where sequencing is not feasible.