Effects of allele frequency estimation on genomic predictions and inbreeding coefficients.

Genetic calculations often require estimating allele frequencies, which differ across time due to selection and drift. Data were 50,000 simulated markers and 39,985 actual markers for 2391 genotyped Holstein bulls. Gene content of relatives and gene frequencies in the base (founder) population were estimated using pedigrees and a linear model. Ancestors born since 1950 were included, for a total of 22,088 animals. Because pedigrees were very complete, only one unknown-parent group was used. Convergence to 5 digits of accuracy required about 1000 iterations. Total time was 2 processor days and proportional to number of animals times markers, but actual clock time was reduced by processing loci on separate chromosomes in parallel. Simple allele frequencies were obtained from only the known genotypes. True base frequencies were correlated with estimated base frequencies by 0.98 versus 0.94 with simple frequencies. Genomic predictions and inbreeding coefficients were computed in four ways, using true or estimated base frequencies, simple frequencies, or an “estimate” of .5 for each marker. When allele frequencies estimates were used instead of 0.5 to assign mixed model coefficients, solutions converged more slowly but predictions were more accurate. From simulated data, realized reliabilities for young bulls were 62.8% using either true or estimated base frequencies, 62.6% using simple frequencies, and 62.0% using frequencies set to 0.5. Pedigree and genomic inbreeding coefficients were correlated by 0.73 using true base frequencies, 0.67 using estimated base frequencies, 0.12 using simple frequencies, and 0.72 when frequencies were set to 0.5. Genomic inbreeding coefficients were biased downward by 7% to 9% using either frequency estimate, upward by 31% using 0.5, but were reasonable when true frequencies were used. Frequency estimation had small effects on genomic predictions but large effects on genomic inbreeding coefficients in both simulated and real data.