Evaluation of a Computational Diagnostic for Epistasis in Plant Breeding Populations

Previously we reported the inability of genomic prediction methods based on linear models to accurately predict trait values composed of an epistatic genetic architecture. We also reported non-parametric genomic prediction methods applied to the same data produced reasonably accurate predictions. The difference led us to propose analyses by paired parametric and non-parametric methods to the same data could be used as a diagnostic for epistatic genetic architectures in typical plant breeding populations. The suggested computational diagnostic was based on evaluation of 14 genomic prediction methods applied to eight sets of simulated conditions consisting of three factors, each with two levels. Because the potential set of factors that might affect accuracies of genomic predictions is unknown, there is a need for a systematic approach to identify combinations of factors that impact estimates of accuracy. Herein we propose the application of response surface methods to systematically identify conditions that maximize the difference between estimated accuracies of genomic prediction methods. The results indicate that genetic architecture and repeatability at their upper boundaries for complete epistasis and repeatability have the greatest influence on the differences between parametric and non-parametric estimated prediction accuracies. Further, the surface is very steep in the vicinity of the boundary conditions, indicating that the proposed diagnostic is of limited value for discovery of epistatic genetic architectures.