|FRAGOMENI, BRENO - University Of Connecticut|
|LOURENCO, D - University Of Connecticut|
|LEGARRA, ANDRES - French National Institute For Agricultural Research|
|MISZTAL, IGNACY - University Of Georgia|
Submitted to: Journal of Dairy Science
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/16/2019
Publication Date: 11/1/2019
Citation: Fragomeni, B.O., Lourenco, D.A.L., Legarra, A., Van Raden, P.M., Misztal, I. 2019. Alternative SNP weighting for single-step genomic best linear unbiased predictor evaluation of stature in US Holsteins in the presence of selected sequence variants. Journal of Dairy Science. 102(11):10012–10019. https://doi.org/10.3168/jds.2019-16262.
Interpretive Summary: The purpose of this study was to evaluate the impact of including extra sequence variants to the SNP panel used in genomic evaluation in the US Holstein population. Additionally, we evaluated the impact of different SNP weighting approaches in single step GBLUP and GBLUP. A total of 17k extra variants were included in the regular 50k SNP chip, and the model allowed markers to have different variances. Improvements were observed in GBLUP, however, impacts of SNP weighting and extra variants did not affect ssGBLUP accuracies substantially.
Technical Abstract: Causal variants inferred from sequence data analysis are expected to increase accuracy of genomic selection. In this work we evaluate the extra accuracy using selected sequence variants, for US Holstein data and the trait stature, by three prediction methods: GBLUP using de-regressed proofs assuming either homozygous (HOMVAR – ignoring the number of equivalent daughters per bull) or heterozygous (HETVAR) residual variances, and by single-step GBLUP (ssGBLUP) using actual phenotypes. Phenotypic data included 3,999,631 records for stature on 3,027,304 Holstein cows. Genotypes on 54,087 SNP markers (54k) were available for 26,877 bulls. Addition of 16,648 selected sequence variants were available, for a total of 70,735 (70k) markers. In all methods, SNPs in the genomic relationship matrix (GRM) were unweighted or weighted iteratively, with weights derived either by quadratic functions of SNP effects or Nonlinear A. Adjusted reliabilities of predictions were obtained by cross validation. With unweighted GRM derived from 54k markers, the reliabilities (*100) were 68.8 for GBLUP HOMVAR, 72.4 for GBLUP HETVAR, and 75.3 for ssGBLUP. With unweighted GRM derived from 70k markers, the reliabilities were 69.5, 73.4 and 76.0, respectively. Weighting by Nonlinear A changed reliabilities to 70.9, 73.3, and 75.9, respectively. Addition of selected sequence variants increased accuracy very little. Weighting by quadratic functions reduced reliabilities. Weighting by Nonlinear A increased accuracies in GBLUP HOMVAR but had only a small effect in ssGBLUP. Reliabilities by DGV extracted from ssGBLUP using unweighted GRM with 54k were higher than reliabilities by any GBLUP. Thus, ssGBLUP seems to capture more information than GBLUP and there is less room for extra accuracy. Improvements with weighting may partly due to deficiencies in the model such as incorrect modeling of residuals or imperfect pseudo-observations.