Genomic Selection in the Presence of Strong Subpopulation Structure.

In genome-wide association studies (GWAS), subpopulation structure is known to potentially cause spurious long-distance or even unlinked associations between marker alleles and the phenotype. In GWAS, therefore, the concern is with avoiding false discoveries due to structure. In genomic selection (GS), the concern shifts to being able to maintain predictive ability despite a structured training dataset. Rare spurious associations will not be an important cause of loss of predictive ability. Rather, consistency of linkage disequilibrium across sub-populations, both in strength of correlation and phase across many (all) loci will be critical. If linkage disequilibrium is not consistent, allelic effects estimated in one sub-population will not be predictive for another sub-population. On the other hand, if sub-populations with differing LD relationships are combined to train a model, marker alleles may not be strongly predictive in either sub-population. The correlation of the LD measure r has been evaluated in barley where it had decayed substantially already at a map distance of 1 cM. An unpublished study in oat shows higher correlations of r than in barley. Accurately predicting breeding values of lines from different sub-populations becomes important in plant breeding when lines from outside of the breeding program are introduced to increase genetic diversity of the breeding program.  In this study, we report on marker density and training population requirements for GS to effectively incorporate favorable alleles from such exotic lines.