Nonlinear kernels, dominance, and envirotyping data increase the accuracy of genome-based prediction in multi-environment trials

Abstract

Modern whole-genome prediction (WGP) frameworks that focus on multi-environment trials (MET) integrate large-scale genomics, phenomics, and envirotyping data. However, the more complex the statistical model, the longer the computational processing times, which do not always result in accuracy gains. We investigated the use of new kernel methods and modeling structures involving genomics and nongenomic sources of variation in two MET maize data sets. Five WGP models were considered, advancing in complexity from a main-effect additive model (A) to more complex structures, including dominance deviations (D), genotype × environment interaction (AE and DE), and the reaction-norm model using environmental covariables (W) and their interaction with A and D (AW + DW). A combination of those models built with three different kernel methods, Gaussian kernel (GK), Deep kernel (DK), and the benchmark genomic best linear-unbiased predictor (GBLUP/GB), was tested under three prediction scenarios: newly developed hybrids (CV1), sparse MET conditions (CV2), and new environments (CV0). GK and DK outperformed GB in prediction accuracy and reduction of computation time (~up to 20%) under all model-kernel scenarios. GK was more efficient in capturing the variation due to A + AE and D + DE effects and translated it into accuracy gains (~up to 85% compared with GB). DK provided more consistent predictions, even for more complex structures such as W + AW + DW. Our results suggest that DK and GK are more efficient in translating model complexity into accuracy, and more suitable for including dominance and reaction-norm effects in a biologically accurate and faster way.

Fig. 1. Partition of the variance components related to the different genetics, environmental, and residual sources of variation.

Each panel is based on the combination of the five WGP models (vertical titles) built with the three different kernel methods (x axis) over HEL and USP data sets (horizontal titles).

Fig. 2. Resolution of the genomic-enabled models and kernel methods in predicting genotypes in novel environments.

a Predictive ability of specific hybrids (each row) involving (q − 1) tested environments plus a one novel environment for sets HEL and USP, respectively; b typology of predictive abilities for sets HEL and USP, respectively. Predictive ability values are represented from warm colors (red, worst results) to cold colors (blue and purple, better results).

Fig. 3. Joint accuracy trends of best combinations of kernel method and model in predicting novel environments (CV0) for both maize data sets (HEL and USP).

On the X axis, the environments were ordered from less predictable (S4) to higher predictable (S3). Environments with the acronym S denote sites (from 1 to 5, in the HEL set) and with E denoting environments (from 1 to 8, in the USP set).