Genome-wide association mapping including phenotypes from relatives without genotypes in a single-step (ssGWAS) for 6-week body weight in broiler chickens

Huiyu Wang, Ignacy Misztal, Ignacio Aguilar, Andres Legarra, Rohan L. Fernando, Zulma Vitezica, Ron Okimoto, Terry Wing, Rachel Hawken, William M. Muir

The purpose of this study was to compare results obtained from various methodologies for genome-wide association studies, when applied to real data, in terms of number and commonality of regions identified and their genetic variance explained, computational speed, and possible pitfalls in interpretations of results. Methodologies include: two iteratively reweighted single-step genomic BLUP procedures (ssGWAS1 and ssGWAS2), a single-marker model (CGWAS), and BayesB. The ssGWAS methods utilize genomic breeding values (GEBVs) based on combined pedigree, genomic and phenotypic information, while CGWAS and BayesB only utilize phenotypes from genotyped animals or pseudo-phenotypes. In this study, ssGWAS was performed by converting GEBVs to SNP marker effects. Unequal variances for markers were incorporated for calculating weights into a new genomic relationship matrix. SNP weights were refined iteratively. The data was body weight at 6 weeks on 274,776 broiler chickens, of which 4553 were genotyped using a 60 k SNP chip. Comparison of genomic regions was based on genetic variances explained by local SNP regions (20 SNPs). After 3 iterations, the noise was greatly reduced for ssGWAS1 and results are similar to that of CGWAS, with 4 out of the top 10 regions in common. In contrast, for BayesB, the plot was dominated by a single region explaining 23.1% of the genetic variance. This same region was found by ssGWAS1 with the same rank, but the amount of genetic variation attributed to the region was only 3%. These findings emphasize the need for caution when comparing and interpreting results from various methods, and highlight that detected associations, and strength of association, strongly depends on methodologies and details of implementations. BayesB appears to overly shrink regions to zero, while overestimating the amount of genetic variation attributed to the remaining SNP effects. The real world is most likely a compromise between methods and remains to be determined.