Evaluation of RR-BLUP Genomic Selection Models that Incorporate Peak Genome-Wide Association Study Signals in Maize and Sorghum

Certain agronomic crop traits are complex and thus governed by many small-effect loci. Statistical models typically used in a genome-wide association study (GWAS) and genomic selection (GS) quantify these signals by assessing genomic marker contributions in linkage disequilibrium (LD) with these loci to trait variation. These models have been used in separate quantitative genetics contexts until recently, when, in published studies, the predictive ability of GS models that include peak associated markers from a GWAS as fixed-effect covariates was assessed. Previous work suggests that such models could be useful for predicting traits controlled by several large-effect and many small-effect genes. We expand this work by evaluating simulated traits from diversity panels in maize (Zea mays L.) and sorghum [Sorghum bicolor (L.) Moench] using ridge-regression best linear unbiased prediction (RR-BLUP) models that include fixed-effect covariates tagging peak GWAS signals. The ability of such covariates to increase GS prediction accuracy in the RR-BLUP model under a wide variety of genetic architectures and genomic backgrounds was quantified. Of the 216 genetic architectures that we simulated, we identified 60 where the addition of fixed-effect covariates boosted prediction accuracy. However, for the majority of the simulated data, no increase or a decrease in prediction accuracy was observed. We also noted several instances where the inclusion of fixed-effect covariates increased both the variability of prediction accuracies and the bias of the genomic estimated breeding values. We therefore recommend that the performance of such a GS model be explored on a trait-by-trait basis prior to its implementation into a breeding program.