Environment-specific genomic prediction ability in maize using environmental covariates depends on environmental similarity to training data

Abstract

Technology advances have made possible the collection of a wealth of genomic, environmental, and phenotypic data for use in plant breeding. Incorporation of environmental data into environment-specific genomic prediction (GP) is hindered in part because of inherently high data dimensionality. Computationally efficient approaches to combining genomic and environmental information may facilitate extension of GP models to new environments and germplasm, and better understanding of genotype-by-environment (G × E) interactions. Using genomic, yield trial, and environmental data on 1,918 unique hybrids evaluated in 59 environments from the maize Genomes to Fields project, we determined that a set of 10,153 SNP dominance coefficients and a 5-day temporal window size for summarizing environmental variables were optimal for GP using only genetic and environmental main effects. Adding marker-by-environment variable interactions required dimension reduction, and we found that reducing dimensionality of the genetic data while keeping the full set of environmental covariates was best for environment-specific GP of grain yield, leading to an increase in prediction ability of 2.7% to achieve a prediction ability of 80% across environments when data were masked at random. We then measured how prediction ability within environments was affected under stratified training-testing sets to approximate scenarios commonly encountered by plant breeders, finding that incorporation of marker-by-environment effects improved prediction ability in cases where training and test sets shared environments, but did not improve prediction in new untested environments. The environmental similarity between training and testing sets had a greater impact on the efficacy of prediction than genetic similarity between training and test sets.

Outlook for Implementation of Genomics-Based Selection in Public Cotton Breeding Programs

“…Directly accounting for G × E is promising for genomic predictions in cotton, although the biggest challenge is accumulating the amount of data necessary to accurately capture these high-level interactions (thousands of markers by many environments, plus the shared effect of a marker across environments). Rogers and Holland (2022) [ 30 ] found that such high-complexity models improve the prediction accuracy only when environments overlap between the training and testing set, further demonstrating our inability to capture genotype × environments based on the measurable environmental variables.…”

Envirome-Wide Associations Enhance Multi-Year Genome-Based Prediction of Historical Wheat Breeding Data

“…In this sense, a good envirotyping protocol must be designed and conducted in order to avoid misspecification of those similarities, while differentiating the effect of different time-scales (e.g., time-windows during crop lifetime, phenology) according to the crop phenology. For example, there is no meaning in using monthly temperature values to model the growing conditions faced by an annual crop such as maize, due to the fact that during 30-day intervals, the crop goes through a wide number of development stages, and for each one of them, the crop will not face the same environment due to the differential needs and sensibility to inputs and stress factors, as evidenced by Rogers and Holland (2022). On the other hand, the use of high-resolution data, such as an hour-scale, may not aggregate value and moreover, may even be a source of noise that will limit the accuracy of the reaction-norm models (Jarquin et al, 2020).…”

Yield Prediction Through Integration of Genetic, Environment, and Management Data Through Deep Learning

“…2021; Rogers et al . 2021; Rogers and Holland 2021), machine learning (Westhues et al . 2021), physiological crop growth models (Technow et al .…”

“…Within agriculture, diverse methods have been applied to the task of predicting phenotype ranging from classical statistics (Jarquin et al 2021;Rogers and Holland 2021), machine learning (Westhues et al 2021), physiological crop growth models (Technow et al 2015), to combinations of these and other methods (Messina et al 2018;Shahhosseini et al 2021). Each model contains limitations such as lacking the capacity to model complex non-linear responses (linear models) or interactions between factors, interpretability within a biological framework (machine learning models), or dependence on costly, low throughput data for calibration (crop growth models).…”

“…DNNs require an abundance of data for training. We begin by detailing a workflow that incorporates a wider number of years from the Genomes to Fields Initiative in than previous studies Washburn et al 2021;Rogers and Holland 2021), while also limiting the effect of errant and absent measurements. Improving on past studies, we propose a new approach to model optimization whereby the model is broken into sub-modules for each data type and interactions between them, then each submodule is consecutively optimized, using a bayesian optimization procedure to find a suitable structure based on the data itself.…”