Leveraging biological insight and environmental variation to improve phenotypic prediction: Integrating crop growth models (CGM) with whole genome prediction (WGP)

Abstract

A successful strategy for prediction of crop yield that accounts for the effects of genotype, environment and their interactions with management will create many opportunities for enhancing the productivity of agricultural systems. Crop growth models (CGMs) have a history of application for crop management decision support. Recently, whole genome prediction (WGP) methodologies have been developed and applied in breeding to enable prediction of crop traits for new genotypes, and thus increased the size of plant breeding programs without expanding expensive field testing. The integration of a CGM into the algorithm for WGP, referred to as CGM-WGP, has opened up the potential for prediction of G × E × M interactions for breeding and product placement applications. The main objectives of this study were to extend CGM-WGP methodology to train models using data from multiple environments, and to evaluate, using both synthetic and experimental data from a maize drought breeding program, whether CGM-WGP methodology can enable improved phenotypic prediction when G × E interactions are an important determinant of performance. The CGM-WGP methodology was improved by 1) reformulating the model as a Bayesian generalized linear hierarchical model, and 2) by sampling the posterior distribution using a Metropolis-within-Gibbs sampling algorithm. The increased efficiency of the algorithm enabled the use of multiple environments and larger populations than those used in previous studies. Synthetic datasets included three environments and an empirical dataset included two environments contrasting for drought stress pattern and intensity. The empirical dataset included four double haploid populations expressing different levels of G × E interaction. Collectively, the prediction accuracy results for the empirical study indicate there were realized advantages in prediction accuracy for yield, in both the water limited and the not water limited environments, from the modeling of the G × E interactions by the CGM-WGP methodology relative to the reference method BayesA. Similarly, the difference in CGM-WGP accuracy relative to BayesA increased with decreasing similarity between the environment types utilized for training and evaluating the predictions. The synthesis provided in this work that encompasses crop physiology and modeling, quantitative genetics, genomic prediction and breeding, should stimulate a cross disciplinary dialogue towards building the next generation of prediction methodologies.