Near Infrared Reflectance Spectroscopy Phenomic and Genomic Prediction of Maize Agronomic and Composition Traits Across Environments

Abstract

For nearly two decades, genomic selection has supported efforts to increase genetic gains in plant and animal improvement programs. However, novel phenomic strategies helping to predict complex traits in maize have proven beneficial when integrated into across– and within-environment genomic prediction models. One phenomic data modality is near infrared spectroscopy (NIRS), which records reflectance values of biological samples (e.g., maize kernels) based on chemical composition. Predictions of seven maize agronomic traits and three kernel composition traits across two years (2011-2012) and two management conditions (water stressed and well-watered) were conducted using combinations of NIRS and genomic data within four different cross-validation prediction scenarios. In aggregate, models incorporating NIRS data alongside genomic data improved predictive ability over models using only genomic data in 5 of 28 trait/cross-validation scenarios for across-environment prediction and 15 of 28 trait/environment scenarios for within-environment prediction, while the model with NIRS data alone had the highest prediction ability in only 1 of 28 scenarios for within-environment prediction. Potential causes of the surprisingly lower phenomic than genomic prediction power in this study are discussed, including sample size, sample homogenization, and low G×E. A genome-wide association study (GWAS) implicated known (i.e.,MADS69,ZCN8, sh1, wx1, du1) and unknown candidate genes linked to plant height and flowering-related agronomic traits as well as compositional traits such as kernel protein and starch content. This study demonstrated that including NIRS with genomic markers is a viable method to predict multiple complex traits with improved predictive ability and elucidate underlying biological causes.Key messageGenomic and NIRS data from a maize diversity panel were used for prediction of agronomic and kernel composition traits while uncovering candidate genes for kernel protein and starch content.

Near‐infrared reflectance spectroscopy phenomic prediction can perform similarly to genomic prediction of maize agronomic traits across environments

“…Using the caret package in R (Kuhn, 2008), the lasso algorithm was run with fivefold three‐replication CV using the maximum available hybrids in each environment to determine the most important NIRS bands (cm −1 ) associated with GY and kernel composition traits. Model performance was assessed using R 2 (DeSalvio, 2023). The lasso algorithm nominated important NIRS bands linked to GY and KW in each environment.…”

“…Here, $$\sigma _j^2$$, $$\sigma _{jk}^2$$, and $$\sigma _{{{\mathrm{\varepsilon }}_{jklm}}}^2$$ are the variance components of the overall effects of pedigree, the interaction between pedigree and environment, and the error term, respectively, n is the number of replicates, and k is the number of environments. Pedigree effects (BLUPs) for GY, KW, and each NIRS band are shown in the Supporting Information (DeSalvio, 2023).…”

“…Five genomic and phenomic best linear unbiased prediction (GBLUP/PBLUP) models, including main and interaction effects between genomic/phenomic markers and environments modeled with different covariance structures, were compared in this study (Table 1). The data preparation and matrix multiplication to obtain each kernel are described in the Supporting Information repository (DeSalvio, 2023) and in the following section. In the series of equations in Table 1, Y represents a $$982 \times 1$$ vector of BLUPs from Equation (1) for GY or KW across all four environments.…”

“…An annotated script of the R code used in this research can be accessed via GitHub (DeSalvio, 2023) [https://github.com/ajdesalvio/Maize-NIRS-GBS.git]. All files needed to reproduce the results are provided in the GitHub repository for users to access.…”