Use and optimization of different sources of information for genomic prediction

Abstract

Abstract Background Molecular data is now commonly used to predict breeding values (BV). Various methods to calculate genomic relationship matrices (GRM) have been developed, with some studies proposing regression of coefficients back to the reference matrix of pedigree-based relationship coefficients (A). The objective was to compare the utility of two GRM: a matrix based on linkage analysis (LA) and anchored to the pedigree, i.e. $${\mathbf{G}}_{{{\mathbf{LA}}}} ,$$ G LA , and a matrix based on linkage disequilibrium (LD), i.e. $${\mathbf{G}}_{{{\mathbf{LD}}}}$$ G LD , using genomic and phenotypic data collected on 5416 broiler chickens. Furthermore, the effects of regressing the coefficients of $${\mathbf{G}}_{{{\mathbf{LD}}}}$$ G LD back to A (LDA) and to $${\mathbf{G}}_{{{\mathbf{LA}}}}$$ G LA (LDLA) were evaluated, using a range of weighting factors. The performance of the matrices and their composite products was assessed by the fit of the models to the data, and the empirical accuracy and bias of the BV that they predicted. The sensitivity to marker choice was examined by using two chips of equal density but including different single nucleotide polymorphisms (SNPs). Results The likelihood of models using GRM and composite matrices exceeded the likelihood of models based on pedigree alone and was highest with intermediate weighting factors for both the LDA and LDLA approaches. For these data, empirical accuracies were not strongly affected by the weighting factors, although they were highest when different sources of information were combined. The optimum weighting factors depended on the type of matrices used, as well as on the choice of SNPs from which the GRM were constructed. Prediction bias was strongly affected by the chip used and less by the form of the GRM. Conclusions Our findings provide an empirical comparison of the efficacy of pedigree and genomic predictions in broiler chickens and examine the effects of fitting GRM with coefficients regressed back to a reference anchored to the pedigree, either A or $${\mathbf{G}}_{{{\mathbf{LA}}}}$$ G LA . For the analysed dataset, the best results were obtained when $${\mathbf{G}}_{{{\mathbf{LD}}}}$$ G LD was combined with relationships in A or $${\mathbf{G}}_{{{\mathbf{LA}}}}$$ G LA , with optimum weighting factors that depended on the choice of SNPs used. The optimum weighting factor for broiler body weight differed from weighting factors that were based on the density of SNPs and theoretically derived using generalised assumptions.