Dissecting Genome-Wide Association Signals: Lessons From Flavonoid Pigmentation Traits in Sorghum.

Genome-wide association studies (GWAS) are a powerful and widely-applicable approach to dissect the genetic basis of natural variation. In practice, however, the effects of complex genetic architecture and structured populations on the accuracy and precision of GWAS signals remain poorly understood, though several strategies have be been developed to address these confounding factors. Flavonoid pigmentation traits are a useful system to evaluate mapping strategies, as they are genetically well-characterized in several plants species and show abundant natural variation in grain and vegetative phenotypes. Here we dissect the genetic control of flavonoid pigmentation in the cereal grass sorghum (Sorghum bicolor) using high-resolution genotyping-by-sequencing (GBS) SNP markers and a variety of genome-wide mapping approaches. Studying the grain tannin phenotype, we find that naive General Linear Models are not able to precisely map tan1-a allele, with either a small panel (n=142) or larger association panel (n=350), and that synthetic associations limit the mapping of the Tannin1 locus to Mb-resolution. In contrast, a Mixed Linear Model (MLM) that account for population structure correctly identify tan1-a loss-of-function allele with gene-level resolution, but only with the larger association panel. Stepwise approaches, including Multi-Locus Mixed Models, effectively account for genetic heterogeneity and epistasis (complementary dominance) but remain sensitive to synthetic associations caused by allelic heterogeneity. In addition, we find that the tan1-a allele can be mapped with gene resolution in a biparental recombinant inbred line family (n=263) using GBS markers, but lower precision in the mapping of vegetative pigmentation traits suggested that consistent gene-level resolution will likely require multiple RILs in a nested association mapping design. Our findings highlight that with a combination of experimental strategies, encompassing population design, statistical modelling, and large sample sizes, it is possible to achieve gene-level mapping of complex traits in structured populations.