COMMENTARY COMMENTARY
Holsteins are the genomic selection poster cows Jeremy F. Taylora,1, Kristen H. Taylorb, and Jared E. Deckera
Genomic selection (GS) is the process by which the genetic improvement of plants or animals is accomplished using the genomic prediction (GP) of additive genetic merits [known as genomic estimated breeding values (GEBVs)] of selection candidates. Alternative statistical models for GP were first described in 2001 by Meuwissen et al. (1), who used simulation to evaluate the performance of linear mixed models and Bayesian mixture models for the prediction of marker effects and GEBVs. This work was truly visionary, because it was not until January 2008 that the first high-density genotyping chip for an agricultural species, the Illumina BovineSNP50 (2), became publically available, allowing the generation of datasets that would enable GP and facilitate the deployment of GS (Fig. 1). In PNAS, Garc´ıa-Ruiz et al. (3) characterize the impact of 7 y of implementation of GS on a national breeding program and remarkably demonstrate that rates of annual genetic improvement in US Holstein dairy cows have increased from 50% to 100% for moderately heritable yield traits and from 300% to 400% for lowly heritable fitness traits. These increases in response to selection come with little evidence of any increase in rates of inbreeding that can lead to reductions in population fitness. Moreover, the rate of adoption of GS within the US dairy industry has been astounding. Although Garc´ıa-Ruiz et al. (3) analyzed data from over 25 million US Holstein cows born since 1975 and 316,485 bulls born since 1950, almost 1.2 million of these animals have now been chip-genotyped (Fig. 1), representing an industry investment of at least $50 million. Garc´ıa-Ruiz et al. (3) demonstrate that GS is the most important technology adopted by the US dairy industry since artificial insemination (AI) 75 y ago and that this industry has become the GS poster child as agriculture attempts to feed a growing population with increasing constraints on land and water availability and greenhouse gas emissions. The statistical methodologies underlying GP are now reasonably mature and build on Henderson’s mixed linear model equations for the best linear unbiased prediction (BLUP) of additive genetic merit using pedigree and phenotype data (4). When all animals are genotyped, BLUPs of additive genetic merits are obtained by replacing the pedigree relationship matrix
with the genomic relationship matrix in Henderson’s mixed model equations, and the predicted additive genetic merits for genotyped animals that do not have phenotypes (generally young selection candidates) can be simply obtained by including them in the model with null phenotypes (5). This formulation of the prediction problem (equation 13 in ref. 5) reveals that the reliability (precision of estimation) of GEBVs for animals without phenotype data depends on the extent of their relatedness to the animals with phenotypes and the number of animals with phenotypes. Therefore, the optimal design for a GS program requires (i) the ability to capture phenotypic data recurrently on genotyped animals to increase the reference population size and (ii) inclusion of parents and grandparents of selection candidates in the reference population to maximize the relatedness of reference individuals to selection candidates. As it turns out, GS was a technology waiting to be invented for the US dairy industry. More than 70% of all US dairy cows are bred by AI, and because nearly all of the produced female calves have historically been retained as herd replacements, selection differentials and generation intervals for the sires of bulls and sires of cows pathways have contributed the most to selection response (3). Because milk yield is expressed only in females, bulls were historically progeny-tested to evaluate their ability to produce high-milk-yielding daughters. This process was very expensive, costing about $500,000 per selected AI bull; however, these bulls achieved reliabilities exceeding 80% for their estimated merits but were 7 y old when their semen was released (3). By 2009, a year after the release of the BovineSNP50 chip, 3,576 Holstein bulls with highly reliable pedigreeestimated additive genetic merits had been genotyped and GEBVs with reliabilities of 50% were estimated for the young bulls awaiting progeny test results. This accomplishment was equivalent to the reliability of a progeny-test estimate based on 11 daughters (6). By continuing to collect phenotypes on the daughters of all tested bulls and maintaining a single generation between relatives in the reference population and the selection candidates, the only limit to the GEBV reliabilities was the size of the reference population.
a
Division of Animal Sciences, University of Missouri, Columbia, MO 65211; and bDepartment of Pathology and Anatomical Sciences, University of Missouri, Columbia, MO 65211 Author contributions: J.F.T., K.H.T., and J.E.D. wrote the paper. The authors declare no conflict of interest. See companion article on page E3995. 1 To whom correspondence should be addressed. Email: [email protected].
7690–7692 | PNAS | July 12, 2016 | vol. 113 | no. 28
www.pnas.org/cgi/doi/10.1073/pnas.1608144113
Fig. 1. Annual numbers of citations of Meuwissen et al. (1), rate of genetic improvement in milk production from Garc´ıa-Ruiz et al. (3), and numbers of Holstein cows chip-genotyped by December of each year from the Council for Dairy Cattle Breeding database (https://www.cdcb.us/Genotype/cur_density.html).
By 2011, GEBV reliabilities for yield traits had reached 75%, only slightly less than for 7-y-old progeny tested bulls, allowing the semen of 2-y-old bulls to be marketed (7). This finding changed the structure of the national breeding program. Once breeders began to trust the GEBV reliabilities, progeny testing began to phase out and generation intervals on the sire of bull and sire of cow pathways began to decrease (3). The reliability of GEBVs for young females was the same as for young males but was substantially increased over their pedigree-based estimates, allowing reductions in generation interval, particularly on the dam of bull pathway. Moreover, selection differentials increased by simply genotyping more young animals. Reducing the cost of progeny testing allowed investments in genotyping larger numbers of young bulls, but the key to improving the selection differential in the dams of sons was to reduce the cost of genotyping. This was accomplished using low-density genotyping chips, initially with 2,900 variants, to allow the widespread genotyping of females (7). By correcting paternity errors, which can be as high as 10% (8), and by generating GEBVs for young cows, it became possible for a cow to mother a bull destined to become an AI sire before she ever lactated. Although young bulls must reach 9–12 mo of age before they produce fertile sperm, young females are born with all of their oocytes, and the biological boundaries to generation interval are currently being explored. Holstein calves have already been born from oocytes aspirated from 8-mo-old females that were fertilized in vitro by the sperm of 10-mo-old bulls. This process is known as velogenetics, first described in 1991 (9), but it too had to await the development of low-cost, high-density genotyping chips and GS before becoming practicable. Finally, because one-half of the additive genetic variation in a population is expressed among full-sibling progeny, further increases in selection differentials may be possible by superovulating
Taylor et al.
dams of bulls and obtaining GEBVs for the harvested embryos. Male embryos that achieve the highest GEBVs can be transplanted into recipient cows to produce the next generation of bulls. This procedure requires the amplification of DNA recovered from only a few cells, which results in slight reductions in the fidelity of chip genotypes (10) and the viability of transferred embryos but increases the selection differential by applying selection to parental gametes. The development of genotyping chips and GS have been enabling technologies for the US dairy industry, and we anticipate that they will enable the superimposition of new reproductive technologies that will further increase genetic improvement. Using species-specific genotyping assays (11), GP models have been evaluated in several other plant and animal species. The reliabilities (called accuracies in other industries) of GEBVs can be almost as high as achieved within the US dairy industry for beef cattle (12), pigs (13), and chickens (14). Although the limited use of AI within the US beef cattle industry has limited the impact of GS on genetic improvement relative to Holsteins, more than 200,000 registered Angus, the predominant US beef breed, have been chipgenotyped (15). Of particular importance, the impact of GS within the US dairy industry has encouraged public investment by the US Department of Agriculture’s National Institute for Food and Agriculture in competitive research grants to develop GEBVs for traits such as feed efficiency (www.swinefeedefficiency.com, www.beefefficiency. org, and www.dairy-efficiency.org/?q=node/4) and disease resistance (www.brdcomplex.org), which have previously been unavailable. Genetic improvement of these traits will have far-reaching effects on the profitability and sustainability of US livestock production. Although a comprehensive analysis of the impacts of GS on the US beef industry has not yet been performed, the corporate ownership of the international pig and chicken breeding programs may preclude
PNAS | July 12, 2016 | vol. 113 | no. 28 | 7691
these analyses from publication for the foreseeable future. This fact makes the work of Garc´ıa-Ruiz et al. (3) even more valuable, because the recent increases in the efficiency and productivity of the US dairy industry clearly justify the public investment in the technology. Although cattle are an old species with a variant-rich genome, their recent history of domestication and breed formation has led to populations such as Holsteins with an effective population size of only about 100 animals (many variants on few haplotypes). This situation contrasts strongly with humans, a relatively young species with less genomic variation but with an effective population size of
10,000–20,000 individuals (fewer variants but more haplotypes). These demographic differences have an impact on the nature of the allele frequency spectra, with humans possessing much more rare variation than cows. Consequently, the application of GP in human medicine to predict complex traits, such as the risk of disease (16), may require much higher marker densities (possibly detected by nextgeneration sequencing) and reference population sizes than are currently used in livestock (17). However, the impacts of GS may extend far beyond the potential for clinical utility to, for example, the optimization of patient selection for clinical trials, potentially saving millions.
1 Meuwissen THE, Hayes BJ, Goddard ME (2001) Prediction of total genetic value using genome-wide dense marker maps. Genetics 157(4):1819–1829. 2 Matukumalli LK, et al. (2009) Development and characterization of a high density SNP genotyping assay for cattle. PLoS One 4(4):e5350. 3 Garc´ıa-Ruiz A, et al. (2016) Changes in genetic selection differentials and generation intervals in US Holstein dairy cattle as a result of genomic selection. Proc Natl Acad Sci USA 113:E3995–E4004. 4 Henderson CR (1973) Sire evaluation and genetic trends. J Anim Sci 1973:10–41. 5 Taylor JF (2014) Implementation and accuracy of genomic selection. Aquaculture 420–421(Suppl 1):S8–S14. 6 VanRaden PM, et al. (2009) Invited review: Reliability of genomic predictions for North American Holstein bulls. J Dairy Sci 92(1):16–24. 7 Wiggans GR, Vanraden PM, Cooper TA (2011) The genomic evaluation system in the United States: Past, present, future. J Dairy Sci 94(6):3202–3211. 8 Visscher PM, Woolliams JA, Smith D, Williams JL (2002) Estimation of pedigree errors in the UK dairy population using microsatellite markers and the impact on selection. J Dairy Sci 85(9):2368–2375. 9 Georges M, Massey JM (1991) Velogenetics, or the synergistic use of marker assisted selection and germ-line manipulation. Theriogenology 35(1):151–159. 10 Shojaei Saadi HA, et al. (2014) Impact of whole-genome amplification on the reliability of pre-transfer cattle embryo breeding value estimates. BMC Genomics 15:889. 11 Eggen A (2012) The development and application of genomic selection as a new breeding paradigm. Anim Front 2(1):10–15. 12 Saatchi M, et al. (2011) Accuracies of genomic breeding values in American Angus beef cattle using K-means clustering for cross-validation. Genet Sel Evol 43:40. 13 Veroneze R, et al. (2015) Accuracy of genome-enabled prediction exploring purebred and crossbred pig populations. J Anim Sci 93(10):4684–4691. 14 Weng Z, et al. (2016) Effects of number of training generations on genomic prediction for various traits in a layer chicken population. Genet Sel Evol 48:22. 15 The Angus Report (June 6, 2016) Top News. Available at https://youtube.com/watch?v=Eyvl-5-0lV8&feature=youtu.be. Accessed June 8, 2016. 16 Wray NR, et al. (2013) Pitfalls of predicting complex traits from SNPs. Nat Rev Genet 14(7):507–515. 17 MacLeod IM, Hayes BJ, Goddard ME (2014) The effects of demography and long-term selection on the accuracy of genomic prediction with sequence data. Genetics 198(4):1671–1684.
7692 | www.pnas.org/cgi/doi/10.1073/pnas.1608144113
Taylor et al.
Holsteins are the genomic selection poster cows Jeremy F. Taylora,1, Kristen H. Taylorb, and Jared E. Deckera
Genomic selection (GS) is the process by which the genetic improvement of plants or animals is accomplished using the genomic prediction (GP) of additive genetic merits [known as genomic estimated breeding values (GEBVs)] of selection candidates. Alternative statistical models for GP were first described in 2001 by Meuwissen et al. (1), who used simulation to evaluate the performance of linear mixed models and Bayesian mixture models for the prediction of marker effects and GEBVs. This work was truly visionary, because it was not until January 2008 that the first high-density genotyping chip for an agricultural species, the Illumina BovineSNP50 (2), became publically available, allowing the generation of datasets that would enable GP and facilitate the deployment of GS (Fig. 1). In PNAS, Garc´ıa-Ruiz et al. (3) characterize the impact of 7 y of implementation of GS on a national breeding program and remarkably demonstrate that rates of annual genetic improvement in US Holstein dairy cows have increased from 50% to 100% for moderately heritable yield traits and from 300% to 400% for lowly heritable fitness traits. These increases in response to selection come with little evidence of any increase in rates of inbreeding that can lead to reductions in population fitness. Moreover, the rate of adoption of GS within the US dairy industry has been astounding. Although Garc´ıa-Ruiz et al. (3) analyzed data from over 25 million US Holstein cows born since 1975 and 316,485 bulls born since 1950, almost 1.2 million of these animals have now been chip-genotyped (Fig. 1), representing an industry investment of at least $50 million. Garc´ıa-Ruiz et al. (3) demonstrate that GS is the most important technology adopted by the US dairy industry since artificial insemination (AI) 75 y ago and that this industry has become the GS poster child as agriculture attempts to feed a growing population with increasing constraints on land and water availability and greenhouse gas emissions. The statistical methodologies underlying GP are now reasonably mature and build on Henderson’s mixed linear model equations for the best linear unbiased prediction (BLUP) of additive genetic merit using pedigree and phenotype data (4). When all animals are genotyped, BLUPs of additive genetic merits are obtained by replacing the pedigree relationship matrix
with the genomic relationship matrix in Henderson’s mixed model equations, and the predicted additive genetic merits for genotyped animals that do not have phenotypes (generally young selection candidates) can be simply obtained by including them in the model with null phenotypes (5). This formulation of the prediction problem (equation 13 in ref. 5) reveals that the reliability (precision of estimation) of GEBVs for animals without phenotype data depends on the extent of their relatedness to the animals with phenotypes and the number of animals with phenotypes. Therefore, the optimal design for a GS program requires (i) the ability to capture phenotypic data recurrently on genotyped animals to increase the reference population size and (ii) inclusion of parents and grandparents of selection candidates in the reference population to maximize the relatedness of reference individuals to selection candidates. As it turns out, GS was a technology waiting to be invented for the US dairy industry. More than 70% of all US dairy cows are bred by AI, and because nearly all of the produced female calves have historically been retained as herd replacements, selection differentials and generation intervals for the sires of bulls and sires of cows pathways have contributed the most to selection response (3). Because milk yield is expressed only in females, bulls were historically progeny-tested to evaluate their ability to produce high-milk-yielding daughters. This process was very expensive, costing about $500,000 per selected AI bull; however, these bulls achieved reliabilities exceeding 80% for their estimated merits but were 7 y old when their semen was released (3). By 2009, a year after the release of the BovineSNP50 chip, 3,576 Holstein bulls with highly reliable pedigreeestimated additive genetic merits had been genotyped and GEBVs with reliabilities of 50% were estimated for the young bulls awaiting progeny test results. This accomplishment was equivalent to the reliability of a progeny-test estimate based on 11 daughters (6). By continuing to collect phenotypes on the daughters of all tested bulls and maintaining a single generation between relatives in the reference population and the selection candidates, the only limit to the GEBV reliabilities was the size of the reference population.
a
Division of Animal Sciences, University of Missouri, Columbia, MO 65211; and bDepartment of Pathology and Anatomical Sciences, University of Missouri, Columbia, MO 65211 Author contributions: J.F.T., K.H.T., and J.E.D. wrote the paper. The authors declare no conflict of interest. See companion article on page E3995. 1 To whom correspondence should be addressed. Email: [email protected].
7690–7692 | PNAS | July 12, 2016 | vol. 113 | no. 28
www.pnas.org/cgi/doi/10.1073/pnas.1608144113
Fig. 1. Annual numbers of citations of Meuwissen et al. (1), rate of genetic improvement in milk production from Garc´ıa-Ruiz et al. (3), and numbers of Holstein cows chip-genotyped by December of each year from the Council for Dairy Cattle Breeding database (https://www.cdcb.us/Genotype/cur_density.html).
By 2011, GEBV reliabilities for yield traits had reached 75%, only slightly less than for 7-y-old progeny tested bulls, allowing the semen of 2-y-old bulls to be marketed (7). This finding changed the structure of the national breeding program. Once breeders began to trust the GEBV reliabilities, progeny testing began to phase out and generation intervals on the sire of bull and sire of cow pathways began to decrease (3). The reliability of GEBVs for young females was the same as for young males but was substantially increased over their pedigree-based estimates, allowing reductions in generation interval, particularly on the dam of bull pathway. Moreover, selection differentials increased by simply genotyping more young animals. Reducing the cost of progeny testing allowed investments in genotyping larger numbers of young bulls, but the key to improving the selection differential in the dams of sons was to reduce the cost of genotyping. This was accomplished using low-density genotyping chips, initially with 2,900 variants, to allow the widespread genotyping of females (7). By correcting paternity errors, which can be as high as 10% (8), and by generating GEBVs for young cows, it became possible for a cow to mother a bull destined to become an AI sire before she ever lactated. Although young bulls must reach 9–12 mo of age before they produce fertile sperm, young females are born with all of their oocytes, and the biological boundaries to generation interval are currently being explored. Holstein calves have already been born from oocytes aspirated from 8-mo-old females that were fertilized in vitro by the sperm of 10-mo-old bulls. This process is known as velogenetics, first described in 1991 (9), but it too had to await the development of low-cost, high-density genotyping chips and GS before becoming practicable. Finally, because one-half of the additive genetic variation in a population is expressed among full-sibling progeny, further increases in selection differentials may be possible by superovulating
Taylor et al.
dams of bulls and obtaining GEBVs for the harvested embryos. Male embryos that achieve the highest GEBVs can be transplanted into recipient cows to produce the next generation of bulls. This procedure requires the amplification of DNA recovered from only a few cells, which results in slight reductions in the fidelity of chip genotypes (10) and the viability of transferred embryos but increases the selection differential by applying selection to parental gametes. The development of genotyping chips and GS have been enabling technologies for the US dairy industry, and we anticipate that they will enable the superimposition of new reproductive technologies that will further increase genetic improvement. Using species-specific genotyping assays (11), GP models have been evaluated in several other plant and animal species. The reliabilities (called accuracies in other industries) of GEBVs can be almost as high as achieved within the US dairy industry for beef cattle (12), pigs (13), and chickens (14). Although the limited use of AI within the US beef cattle industry has limited the impact of GS on genetic improvement relative to Holsteins, more than 200,000 registered Angus, the predominant US beef breed, have been chipgenotyped (15). Of particular importance, the impact of GS within the US dairy industry has encouraged public investment by the US Department of Agriculture’s National Institute for Food and Agriculture in competitive research grants to develop GEBVs for traits such as feed efficiency (www.swinefeedefficiency.com, www.beefefficiency. org, and www.dairy-efficiency.org/?q=node/4) and disease resistance (www.brdcomplex.org), which have previously been unavailable. Genetic improvement of these traits will have far-reaching effects on the profitability and sustainability of US livestock production. Although a comprehensive analysis of the impacts of GS on the US beef industry has not yet been performed, the corporate ownership of the international pig and chicken breeding programs may preclude
PNAS | July 12, 2016 | vol. 113 | no. 28 | 7691
these analyses from publication for the foreseeable future. This fact makes the work of Garc´ıa-Ruiz et al. (3) even more valuable, because the recent increases in the efficiency and productivity of the US dairy industry clearly justify the public investment in the technology. Although cattle are an old species with a variant-rich genome, their recent history of domestication and breed formation has led to populations such as Holsteins with an effective population size of only about 100 animals (many variants on few haplotypes). This situation contrasts strongly with humans, a relatively young species with less genomic variation but with an effective population size of
10,000–20,000 individuals (fewer variants but more haplotypes). These demographic differences have an impact on the nature of the allele frequency spectra, with humans possessing much more rare variation than cows. Consequently, the application of GP in human medicine to predict complex traits, such as the risk of disease (16), may require much higher marker densities (possibly detected by nextgeneration sequencing) and reference population sizes than are currently used in livestock (17). However, the impacts of GS may extend far beyond the potential for clinical utility to, for example, the optimization of patient selection for clinical trials, potentially saving millions.
1 Meuwissen THE, Hayes BJ, Goddard ME (2001) Prediction of total genetic value using genome-wide dense marker maps. Genetics 157(4):1819–1829. 2 Matukumalli LK, et al. (2009) Development and characterization of a high density SNP genotyping assay for cattle. PLoS One 4(4):e5350. 3 Garc´ıa-Ruiz A, et al. (2016) Changes in genetic selection differentials and generation intervals in US Holstein dairy cattle as a result of genomic selection. Proc Natl Acad Sci USA 113:E3995–E4004. 4 Henderson CR (1973) Sire evaluation and genetic trends. J Anim Sci 1973:10–41. 5 Taylor JF (2014) Implementation and accuracy of genomic selection. Aquaculture 420–421(Suppl 1):S8–S14. 6 VanRaden PM, et al. (2009) Invited review: Reliability of genomic predictions for North American Holstein bulls. J Dairy Sci 92(1):16–24. 7 Wiggans GR, Vanraden PM, Cooper TA (2011) The genomic evaluation system in the United States: Past, present, future. J Dairy Sci 94(6):3202–3211. 8 Visscher PM, Woolliams JA, Smith D, Williams JL (2002) Estimation of pedigree errors in the UK dairy population using microsatellite markers and the impact on selection. J Dairy Sci 85(9):2368–2375. 9 Georges M, Massey JM (1991) Velogenetics, or the synergistic use of marker assisted selection and germ-line manipulation. Theriogenology 35(1):151–159. 10 Shojaei Saadi HA, et al. (2014) Impact of whole-genome amplification on the reliability of pre-transfer cattle embryo breeding value estimates. BMC Genomics 15:889. 11 Eggen A (2012) The development and application of genomic selection as a new breeding paradigm. Anim Front 2(1):10–15. 12 Saatchi M, et al. (2011) Accuracies of genomic breeding values in American Angus beef cattle using K-means clustering for cross-validation. Genet Sel Evol 43:40. 13 Veroneze R, et al. (2015) Accuracy of genome-enabled prediction exploring purebred and crossbred pig populations. J Anim Sci 93(10):4684–4691. 14 Weng Z, et al. (2016) Effects of number of training generations on genomic prediction for various traits in a layer chicken population. Genet Sel Evol 48:22. 15 The Angus Report (June 6, 2016) Top News. Available at https://youtube.com/watch?v=Eyvl-5-0lV8&feature=youtu.be. Accessed June 8, 2016. 16 Wray NR, et al. (2013) Pitfalls of predicting complex traits from SNPs. Nat Rev Genet 14(7):507–515. 17 MacLeod IM, Hayes BJ, Goddard ME (2014) The effects of demography and long-term selection on the accuracy of genomic prediction with sequence data. Genetics 198(4):1671–1684.
7692 | www.pnas.org/cgi/doi/10.1073/pnas.1608144113
Taylor et al.
