Efficient computation of base generation allele frequencies

Abstract

Several aspects of genomic prediction require use of allele frequencies that ideally reflect the base generation of the available pedigree. This includes computation of model-based reliabilities of direct genomic values (DGV) in the context of multi-step genomic evaluations, computation of genomic relationships to be used in single-step GBLUP, and computation of relationships among metafounders. In many cases, the allele frequencies computed from the currently genotyped population are used instead, motivated by the observation that computation of base generation allele frequencies is time consuming. Our aim was to compare the efficiency and accuracy of different methods to compute base generation allele frequencies. The first method employed the gene content method, by running a BLUP on the SNP genotypes and considering a heritability of 0.99. Either a univariate BLUP run for each SNP, or a multiple-trait BLUP run for several SNPs was performed, considering zero genetic correlations among the SNPs. The second method employed a general least squares estimator that is equivalent to the first method, albeit that it does not consider a residual variance. First analyses on simulated data without selection, missing genotypes or genotype errors in the data showed that the second method is superior in both accuracy and efficiency, but only if the inverse of the A matrix was computed using imputation on the fly. The implementation of the second method required less than two minutes to compute base generation allele frequencies for 1 670 SNPs based on 100 078 genotyped animals, and a total pedigree of 325 266 animals. Subsequent analyses with datasets simulating selection, missing genotypes and genotyping errors, that are closer to data used in practice, supported the results that the second method is more efficient and accurate.