Improvement of genomic prediction by integrating additional single nucleotide polymorphisms selected from imputed whole genome sequencing data

Aoxing Liu,Guosheng Su,Yachun Wang,Mogens Sandø Lund,Emre Karaman,Gert Pedersen Aamand,Didier Boichard,Ulrik Sander Nielsen,Sebastien Fritz

doi:10.1038/s41437-019-0246-7

Abstract

The availability of whole genome sequencing (WGS) data enables the discovery of causative single nucleotide polymorphisms (SNPs) or SNPs in high linkage disequilibrium with causative SNPs. This study investigated effects of integrating SNPs selected from imputed WGS data into the data of 54K chip on genomic prediction in Danish Jersey. The WGS SNPs, mainly including peaks of quantitative trait loci, structure variants, regulatory regions of genes, and SNPs within genes with strong effects predicted with variant effect predictor, were selected in previous analyses for dairy breeds in Denmark–Finland–Sweden (DFS) and France (FRA). Animals genotyped with 54K chip, standard LD chip, and customized LD chip which covered selected WGS SNPs and SNPs in the standard LD chip, were imputed to 54K together with DFS and FRA SNPs. Genomic best linear unbiased prediction (GBLUP) and Bayesian four-distribution mixture models considering 54K and selected WGS SNPs as one (a one-component model) or two separate genetic components (a two-component model) were used to predict breeding values. For milk production traits and mastitis, both DFS (0.025) and FRA (0.029) sets of additional WGS SNPs improved reliabilities, and inclusions of all selected WGS SNPs generally achieved highest improvements of reliabilities (0.034). A Bayesian four-distribution model yielded higher reliabilities than a GBLUP model for milk and protein, but extra gains in reliabilities from using selected WGS SNPs were smaller for a Bayesian four-distribution model than a GBLUP model. Generally, no significant difference was observed between one-component and two-component models, except for using GBLUP models for milk.

Highlights

Genomic prediction has been widely applied in dairy cattle breeding (Hayes et al 2009)
DKUSCOW Danish and US bulls and Danish cows as the reference population a54K: single nucleotide polymorphisms (SNPs) in the 54K chip b54K + DFS + FRA: SNPs in 54K chip together with whole genome sequencing (WGS) SNPs selected by analysis of data from major dairy breeds in Denmark–Finland–Sweden and France cG1: one-component Genomic best linear unbiased prediction (GBLUP) model; G2: two-component GBLUP model; B1: one-component Bayesian four-distribution mixture model; B2: two-component Bayesian four-distribution mixture model dLetters in the right lower position were for comparisons among models using the same reference population and SNP scenario
C54K: SNPs in 54K chip; 54K + DFS: SNPs in 54K chip together with WGS SNPs selected by analysis of data from major dairy breeds in Denmark–Finland–Sweden; 54K + FRA: SNPs in 54K chip together with WGS SNPs selected by analysis of data from major dairy breeds in France; 54K +DFS + FRA: SNPs in 54K chip together with WGS SNPs selected by analysis of data from major dairy breeds in Denmark–Finland–Sweden and France dLetters in the left upper positions were for comparisons among reference populations using the same SNP scenario and model; letters in the right lower position were for comparisons among SNP scenarios using the same reference populations and model

Summary

Introduction

Genomic prediction has been widely applied in dairy cattle breeding (Hayes et al 2009). To achieve reliable prediction for breeding values of candidate animals, a reference population consisted of a large number of individuals with both phenotypes and genotypes is required (Karaman et al 2016). Assembling such a sufficiently large reference population, may not be possible for traits that are hard to measure, such as feed intake (Berry et al 2014), or for breeds that are numerically small, such as Danish Jersey (Lund et al 2016). To improve reliabilities of genomic prediction, especially for a numerically small breed, many approaches have been investigated (Brøndum et al 2015; Lund et al 2016; van den Berg et al 2016b). The inclusion of a large number of noncausative SNPs may bring only noise to genomic prediction (Pérez-Enciso et al 2015)

Methods

Results

Discussion

Conclusion