Precision of genetic parameters and breeding values estimated in marker assisted BLUP genetic evaluation
© Neuner et al; licensee BioMed Central Ltd. 2009
Received: 24 February 2009
Accepted: 04 March 2009
Published: 04 March 2009
In practical implementations of marker-assisted selection economic and logistic restrictions frequently lead to incomplete genotypic data for the animals of interest. This may result in bias and larger standard errors of the estimated parameters and, as a consequence, reduce the benefits of applying marker-assisted selection. Our study examines the impact of the following factors: phenotypic information, depth of pedigree, and missing genotypes in the application of marker-assisted selection. Stochastic simulations were conducted to generate a typical dairy cattle population. Genetic parameters and breeding values were estimated using a two-step approach. First, pre-corrected phenotypes (daughter yield deviations (DYD) for bulls, yield deviations (YD) for cows) were calculated in polygenic animal models for the entire population. These estimated phenotypes were then used in marker assisted BLUP (MA-BLUP) evaluations where only the genotyped animals and their close relatives were included.
Models using YD of cows (bull dams) in addition to DYD of bulls resulted in much smaller standard errors for the estimated variance components. The bias in DYD models was larger than in models including YD. Depth of pedigree had the strongest impact on the standard errors of all the estimated variance components. As expected, estimation of variance components was less precise with larger proportions of animals without genotypes in the pedigree. Accuracies of MA-BLUP breeding values for young bull candidates were strongly affected by the inclusion of cow information, but only marginally influenced by pedigree depth and proportions of genotyped animals.
Advances in molecular genetics have led to the identification of several genes and of genetic markers linked or associated with genes that affect traits of interest in livestock (QTL). Once QTL are detected, the aim of animal breeders is to integrate linked markers for QTL into the breeding program, in so-called marker assisted selection schemes (MAS). An overview about different possibilities to apply marker- and gene-assisted selection is given by Dekkers . While the highest benefits are expected from gene-assisted selection using direct markers , in many cases, practical applications of MAS have to rely on anonymous markers that are often assumed to be in population-wide linkage equilibrium with the QTL.
The statistical model for using marker information in BLUP (best linear unbiased prediction) (MA-BLUP) genetic evaluations was developed by Fernando and Grossman . The MA-BLUP methodology allows the simultaneous estimation of QTL and polygenic effects. The QTL effect is accounted for in the mixed model as an extra random effect with the covariance structure proportional to the IBD (identity by descent) matrix at the QTL position given the linked markers .
Components of an integrated system to apply MAS for routine evaluations are given by Dekkers . In nearly all implementations, it will be necessary to take three decisions: i) how many animals are to be included in the MA-BLUP model, ii) which phenotypes should be used, and iii) how much effort is justified in order to completely genotype ancestors of the current young bull generation.
The aim of our study is to examine these three questions with respect to bias and standard errors of estimated variance components and accuracies of MA-BLUP breeding values by means of simulation.
A stochastic simulation model was applied to generate a data set that was then analyzed with five different models. Each simulation cycle consisted of two phases: data generation and analysis of the simulated data sets. The number of replicates for each variant was 100.
Simulated characteristics of the cow population and of the breeding program
Milking cows, nb
Cows in lactation, %
Bull dams, nb
Age at first calving, months
Intergestation period, months
AI bull population
Age at birth of first progeny, months
Age at first breeding value estimation, months
Service life as proven bull, months
Maximum age, months
Bulls sampled, nb/year
Sires for insemination service per year, nb
Sires of bulls used per year, nb
Daughter records per bull sampled, nb
A single-trait model for 305-day milk yield with a heritability of 0.36 and an additive genetic variance of 260,100 kg2 was chosen. Genetic parameters were in agreement with the actual first lactation parameters of German Fleckvieh . The overall breeding value of each animal was the sum of a 'residual polygenic breeding value' and a 'QTL breeding value'. A single biallelic QTL with an allele frequency of 0.5 was assumed and the QTL was bracketed by two marker loci located 3 cM and 2 cM apart, each with 10 alleles but different allelic distributions. Allele frequencies for the marker 3 cM apart from the QTL were 40, 19, 15, 12, 7, 2, 2, 1, 1 and 1% (polymorphic information content, PIC = 0.732), and for the marker 2 cM apart were 60, 20, 8, 4, 2, 2, 1, 1, 1 and 1% (PIC = 0.555), respectively. The Haldane mapping function  was assumed to simulate meiosis. The multipoint polymorphism content  for the simulated QTL position was 0.753.
All calculations assumed a QTL accounting for 20% of the overall additive genetic variance of the trait investigated, without dominance effects at the QTL.
Analysis of simulated data sets
In routine genetic evaluations of dairy cattle all pedigreed animals are included. However, when applying MAS, only a small fraction of animals might be genotyped at genetic markers. Since only the genotyped animals provide information for the estimation of QTL variance components and breeding values in MA-BLUP models, the 'two-step approach' as described by Liu et al. , Druet et al.  and Bennewitz et al.  was used in this study.
MA-BLUP using a two-step approach
In the first step, a classical polygenic animal model (AM) evaluation, assuming the true variance components are known, was conducted for the entire population in order to estimate daughter yield deviations (DYD) for bulls and yield deviations (YD) for cows, respectively . The pedigree contained about 260,000 animals of which 251,000 were cows with phenotypic records. To estimate AM-based breeding values, DYD, and YD, the package MiX99  was used.
The second evaluation step was applied only to a subset of genotyped animals from the population (MA-BLUP pedigree). Usually the pedigree used for MA-BLUP evaluations contains only young bull candidates, young bulls currently used, waiting bulls and their parents and one or more generations of ancestors. For the current study it was assumed that complete marker information was theoretically available for all animals in the MA-BLUP pedigree. Phenotypic observations in step 2 were either DYD of bulls alone, or DYD together with YD of cows. The different amount of information available for DYD was accounted for by applying weighting factors to DYD. YD were not weighted, because each cow had only one record in the current study (Druet 2006; personal communication) and fixed effects that may have an impact on the accuracy of the estimation of YD, e.g. herd effects, were not assumed. When using DYD and YD together in one model for MA-BLUP evaluations, it is necessary to consider that these two information types represent different amounts of genetic and residual variance. Neuner et al.  have provided a detailed description of this issue. In order to account for these differences weighting of information was necessary. The weighting factors EDC (effective daughter contributions, Fikse and Banos ) and γ  were applied to twice the DYD in DYD-YD evaluations.
An MA-BLUP model equivalent to that of Fernando and Grossman  was used for the estimation of MA-BLUP breeding values:
y i = μ + u i + v i + e i (1)
Pedigree for MA-BLUP
Main characteristics of the short and the deep pedigree and the related phenotypic data that are used for MA-BLUP evaluations
Animals with records (DYD, YD) in the MA-BLUP pedigree, nb
Bulls with record (DYD) in the MA-BLUP pedigree, nb
Cows with record (YD) in the MA-BLUP pedigree, nb
Waiting bulls without records in the MA-BLUP pedigree, nb
Young bull candidates without records in the MA-BLUP pedigree, nb
Schemes with missing genotypes
For random QTL models according to Fernando and Grossman  an IBD matrix reflects the covariance structure of QTL gametic effects in the pedigree. George et al.  have reported an extensive overview on algorithms to calculate IBD matrices even for complex pedigrees and incomplete marker information. They concluded that simulation-based algorithms like the multiple-site segregation sampler LOKI  are efficient tools to calculate IBD matrices for complex pedigree structures.
Proportions of genotypes that are assumed to be missing for moderate and extensive gaps in the genotyping structure
Scenario 1, moderate
Scenario 2, extensive
In total, the proportion of missing genotypes is about 41% for scenario 1 and 61% for scenario 2 for both pedigree depths. All IBD matrices applied for MA-BLUP evaluations were genotypic relationship matrices for the QTL and calculated using the package LOKI .
Genetic parameters for MA-BLUP models and MA-BLUP EBV were estimated with the ASREML package  using a MA-BLUP model equivalent to that of Fernando and Grossman  and assuming the QTL position is known from the mapping experiment.
Parameter combinations applied
Simulated and estimated parameters for the estimation of variance components when evaluation models were only based on daughter yield deviations (DYD) for bulls or DYD for bulls in combination with yield deviations (YD) for cows
Parameters considered for the estimation of variance components were the bias of estimated variance components and their standard errors.
The estimates deviation from the simulated parameters was used to check for bias due to the pedigree depth and/or missing genotypes. Standard errors of the estimates were used to assess the precision of estimates between different models. In order to assess the fit of the genetic model, the likelihood ratio test (LRT = -2 ln(L0(no QTL present)-L1(QTL present)) was calculated. L0 and L1 represent the likelihood values at the REML solutions of evaluations where no QTL was assumed to be segregating (no QTL present), and where a QTL was assumed to be segregating (QTL present), respectively.
To examine the impact of the different models on the estimation of MA-BLUP breeding values, the correlation of true and estimated breeding values was calculated for each group of animals (bulls, cows, young bull candidates).
Results presented are the averages of 100 replicates for a 20% simulated proportion of genetic variance explained by the QTL. Even if QTL variances close to zero were estimated in single replicates as an effect of a weak data structure, especially in DYD models for short pedigrees, these estimates were not excluded from the analysis in all investigated situations. This was necessary, as it was our interest, to elaborate and characterize the range of estimated parameters when a QTL is known to segregate with a fixed variance. In contrast, studies that aim at a possible gain due to MAS  allow for the exclusion of replicates that will not give benefits for the breeding program, because structure of the data and relevant parameters indicate that MAS will not improve selection.
Variance component estimation
Results for the estimation of variance components for the investigated combinations of phenotypic information, pedigree depth and missing genotypes are summarized in Table 4. The first two lines of table 4 concern the question: which amount of phenotypic information should be used? In DYD models, the additive genetic variance () was underestimated while the genetic variance explained by the QTL () and the ratio of and were overestimated. The estimates in DYD-YD models were closer to the simulated parameters. The standard errors of the estimated variance components were lower in DYD-YD models.
Results for the impact of missing genotypes are shown in lines three to five of Table 4. Missing genotypes do not lead to biased estimates for variance components. However, they increase the standard error for whereas the standard errors for all other variance components are hardly affected (see Figure 2).
Accuracy of MA-BLUP EBV
Accuracies of estimated breeding values in MA-BLUP evaluation models that were only based on daughter yield deviations (DYD) for bulls or on DYD for bulls and yield deviations for cows
Young bull candidates
Accuracies in AM evaluations were 0.950 for proven bulls, 0.760 for cows and 0.560 for young bulls. Average MA-BLUP accuracies for all categories of animals are hardly affected by the model, pedigree depth or missing genotypes, but strongly affected by the inclusion of YD. The increase in pedigree depth causes a slight improvement of accuracy for young bulls, because the gametic effects can be estimated more accurately. However, this slight increase is gradually lost as the amount of missing genotypes increases.
Questions examined in this study came up during the implementation of MAS for Simmental cattle in Germany and Austria. Our main interest was to find out how to set up an appropriate MA-BLUP system given that it was already decided to use MAS, and in view of the fact that not all relevant animals can be genotyped. In this study, we have investigated the impact of three factors on the estimation of variance components for marker-assisted selection under realistic assumptions: phenotypic information, depth of the pedigree and missing genotypes.
Design and simulation
Early studies about properties of MAS are in most cases based on nucleus breeding programs [17, 18], but this assumption does not hold for the majority of cattle breeding programs in Europe. Therefore, in this study, MAS was integrated in an existing breeding program using a two-step approach as is the case for practical applications in France and Germany [19, 7].
Genetic parameters, information content of genetic markers and variance explained by the QTL in the simulation were in agreement with recent results found in German Simmental cattle (unpublished data). Allele frequencies of 0.5 for alternative QTL alleles were assumed to reduce the risk of loosing alternative alleles due to random drift along the simulation process. As mentioned by Rönnegård and Carlborg  simulations have shown that random QTL models are capable of giving unbiased estimates even when the QTL is biallelic. Furthermore, the assumptions of a biallelic QTL and balanced allele frequencies reflect findings for QTL in cattle. A well-known example for a biallelic QTL is the K232A substitution in DGAT1. Estimated allele frequencies for alleles at K232A were 0.548 and 0.452 in German Holsteins . Another biallelic QTL is reported for the bovine prolactin receptor in Finnish Ayrshire  with allele frequencies of 0.45/0.55 for snp 6. Guillaume et al.  have reported the proportions of genetic variance explained by single QTL used for the French MAS-program, which range from 5 to 40% for individual QTL. Hence, a QTL explaining 20% of the genetic variance in this study is in the scope of findings for MAS programs in reality. In contrast to practical applications of MAS using multiple QTL [23, 7] this study was restricted to a single QTL. Our intention was to examine the influence of different structures of information (pedigree depth, completeness of genotyping) on the estimation of QTL effects. We did not want to examine the effect of several QTL on the accuracy of genetic parameters or on the benefits of MAS. We think that our conclusions will also hold for situations with more QTL, because, in practice, we observe that animals are either genotyped for all QTL or not genotyped at all.
Choice of phenotypic information
According to our results, the choice of phenotypic information in MA-BLUP models is important. We have shown (Neuner et al., ) that the two-step approach intrinsically causes a loss of information because not all relatives contribute to the MA-BLUP breeding value. In consequence, some proportion of the QTL information is required to compensate this loss of information. The inclusion of yield deviations for dams reduces this loss of information. As we have shown here, yield deviations also improve the precision for the estimated variance components and deviations from the simulated parameters become smaller. Neuner et al.  have already discussed the topic concerning the 'choice of weighting factors in MA-BLUP models'. According to the results, the weighting factors daughter equivalents and effective daughter contributions do not introduce bias. One challenge is to combine correctly DYD and YD in one model as these two types of information contain different amounts of genetic and residual variance .
The advantage of a more extensive pedigree of genotyped animals for MA-BLUP is obvious. The more animals have complete information, the smaller is the observed standard error of the estimated parameters. Similar to the effect of having more offspring for progeny tested bulls, a deeper pedigree implies more data and more informative matings for the estimation of QTL effects of parents and grandparents in MA-BLUP models. Our results show that using a deep pedigree with many gaps is still preferable over a short but complete pedigree. The reason for this is that a deeper pedigree improves the estimation of polygenic and residual variances as compared to shorter pedigrees.
The effect of a more parsimonious pedigree has also been shown by George et al. . By altering the number of offspring per mating from 1.8 to 14.3 offspring per mating, the number of progeny per parent providing information to estimate genetic parameters and MA-BLUP EBV was higher. Even if the approach of George et al.  was different from the one in our study, similar effects of a larger pedigree were observed: more accurate estimates and increased power.
The third factor investigated was the effect of missing genotypes in MA-BLUP evaluations. If marker information was complete and could be used to infer the transmission of QTL alleles, then the IBD matrix would only contain 1s and 0s. At the other extreme, if no marker information was available, the IBD matrix would become identical to the numerator relationship matrix, i.e. all covariance elements with the parents will be equal to 0.5, signifying equal probability of inheriting either allele from a parent. In the end, this would result in identical estimators for residual polygenic and QTL variance. Several approaches exist to deal with the problem that non-genotyped animals do not contribute information for QTL models . A well-known approach is the multiple-site segregation sampler LOKI  that was used in this study. We found that missing genotypes did not lead to biased variance components. In contrast to our results, George et al.  have reported that, if more genotypes are missing, the QTL variance is overestimated, the residual polygenic variance is underestimated and bias is increased. The main reason for these contradictory results could be that the structure of pedigree and missing data in our study allowed a much better reconstruction of missing genotypes by LOKI. In the sheep pedigree of George et al.  the number of male progeny and grand progeny was smaller than that in the cattle breeding program of this research. Thus, less descendants are available to contribute information for the reconstruction of their ancestors' genotypes. Furthermore, the amount of phenotypic information per sire is very different in both studies. Compared to George et al. , both the better ability to reconstruct missing genotypes and the higher amount of phenotypic information for MA-BLUP result in unbiased estimates in our study.
Effects on accuracy of MA-BLUP
A QTL explaining 20% of the additive genetic variance was intentionally chosen, because such a QTL is in the order of magnitude that we observed in Simmental cattle and because it shows nicely that MA-BLUP is not necessarily more accurate than conventional BLUP. Our results show that the accuracy of MA-BLUP breeding values is little affected by the pedigree depth and hardly affected by missing genotypes. However, it is strongly affected by the inclusion of YD in the MA-BLUP system . Guillaume et al.  have reported empirical results for the accuracies of EBV for young bulls in the French MAS program. The results of our study are in line with their findings for milk yield. When 40% of the genetic variance were marked with four QTL, they found an increase in accuracy for EBV of young bulls in MA-BLUP models of 0.030 compared to the AM model. In our study, we observed accuracies that were 0.006 higher for one QTL explaining 20% of the genetic variance. The main reason for this small increase was the loss of information due to the two-step approach. As mentioned by Neuner et al.  the loss has to be compensated before additional gain can arise. With respect to accuracy of EBV in MA-BLUP models, Villanueva et al.  have investigated the benefit of increased pedigree and marker information. They simulated four additional generations of random selection in order to extend their data set. The increased amount of marker genotype information significantly increased the accuracy of the estimation of the QTL effects from 0.54 to 0.65. Parameters in their study were 0.25 for the heritability and 0.24 for the ratio of genetic variance explained by the QTL. Spelman  also concluded that if more animals are genotyped in each generation and if more generations of genotypic information are used for MAS, an increase in accuracy of the estimation of QTL effects and therefore in MAS superiority are obtained.
At present many research programs are conducted in order to implement genomic breeding value estimations as described by Meuwissen et al. . Although first implementations exist  it is not sure that MAS schemes will be abandoned. Improvements in existing MAS programs will directly improve the selection until genomic selection will be applied. We also expect that QTL mapping based on the variance component approach  will continue in the future. Our results can be helpful in optimizing the choice of individuals to be genotyped.
In genomic selection projects, which animals to genotype and which source of phenotypic information to use will also have to be decided. Comparable to our study, we expect it will be necessary to rank animals selected for genotyping by their importance in the pedigree and their impact on further generations. For the phenotypic information, applied appropriate weighting and combination of DYD and YD will also be necessary for a correct modeling of genetic and residual variances. As soon as dams will be genotyped for genomic selection, considering their genotypes and YD for parameter estimation will help to improve the accuracy of estimated QTL effects in the same way as in our current study.
The main conclusion of this study is that phenotypic information of cows and an increased depth of the pedigree have an important impact on the precision of genetic parameters estimated in MA-BLUP models. Furthermore, deep pedigrees with many missing genotypes provide more accurate estimates than short pedigrees with complete genotyping. While the estimation of variance components is considerably improved by a deep pedigree with no missing genotypes, the accuracy of MA-BLUP genetic evaluation is hardly affected by these factors. As a consequence, if exact QTL parameters are available from other sources, the same accuracy of MA-BLUP can also be achieved with a short pedigree and many missing genotypes.
The authors gratefully acknowledge financial support from of the German Federal Ministry of Education and Research (projects FUGATO MAS.-Net, grant no. 0313390F and FUGATO+ BRAIN grant no. 0315136C) and of the Förderverein Biotechnologieforschung, Bonn.
- Dekkers JCM: Commercial application of marker- and gene-assisted selection in livestock: Strategies and lessons. J Anim Sci. 2004, 82 (E-Suppl): E313-328.PubMedGoogle Scholar
- Fernando RL, Grossman M: Marker assisted selection using best linear unbiased prediction. Genet Sel Evol. 1989, 21: 467-477. 10.1051/gse:19890407.PubMed CentralView ArticleGoogle Scholar
- Neuner S, Emmerling R, Thaller G, Götz KU: Strategies for estimating genetic parameters in marker-assisted best linear unbiased prediction models in dairy cattle. J Dairy Sci. 2008, 91: 4344-4354. 10.3168/jds.2008-1058.View ArticlePubMedGoogle Scholar
- Interbull: Description of national genetic evaluation system and trend validation for production traits. 2007, [http://www-interbull.slu.se/national_ges_info2/framesida-ges.htm]Google Scholar
- Haldane JBS: The combination of linkage values, and the calculation of distances between the loci of linked factors. J Genet. 1919, 8: 299-309. 10.1007/BF02983270.View ArticleGoogle Scholar
- Rijsdijk FR, Sham PC: Improved estimation of multipoint IBD sharing and multipoint information content by linear regression. Behaviour Genet. 2002, 32: 211-220. 10.1023/A:1016073113019.View ArticleGoogle Scholar
- Liu Z, Reinhardt F, Szyda J, Thomsen H, Reents R: A marker assisted genetic evaluation system for dairy cattle using a random QTL model. Interbull Bull. 2004, 32: 170-174.Google Scholar
- Druet T, Fritz S, Boichard D, Colleau JJ: Estimation of genetic parameters for quantitative trait loci for dairy traits in the French Holstein population. J Dairy Sci. 2006, 89: 4070-4076.View ArticlePubMedGoogle Scholar
- Bennewitz J, Reinsch N, Reinhardt F, Liu Z, Kalm E: Top down preselection using marker-assisted estimates of breeding values in dairy cattle. J Anim Breed Genet. 2004, 121: 307-318. 10.1111/j.1439-0388.2004.00467.x.View ArticleGoogle Scholar
- VanRaden PM, Wiggans GR: Derivation, calculation, and use of national animal model information. J Dairy Sci. 1991, 74: 2737-2746.View ArticlePubMedGoogle Scholar
- Vuori K, Strandén I, Lidauer M, Mäntysaari EA: MiX99 – Effective solver for large and complex linear mixed models. Proceedings of the 8th World Congress on Genetics Applied to Livestock Production: 13–18. 2006, August ; Belo HorizonteGoogle Scholar
- Fikse WF, Banos G: Weighting factors of sire daughter information in international genetic evaluations. J Dairy Sci. 2001, 84: 1759-1767.View ArticlePubMedGoogle Scholar
- George AW, Visscher PM, Haley CS: Mapping quantitative trait loci in complex pedigrees: a two-step variance component approach. Genetics. 2000, 156: 2081-2092.PubMed CentralPubMedGoogle Scholar
- Heath SC: Markov chain Monte Carlo segregation and linkage analysis for oligogenic models. Am J Hum Genet. 1997, 61: 748-760. 10.1086/515506.PubMed CentralView ArticlePubMedGoogle Scholar
- Gilmour AR, Thompson R, Cullis BR: Average information REML: an efficient algorithm for variance parameter estimation in linear mixed models. Biometrics. 1995, 51: 1440-1450. 10.2307/2533274.View ArticleGoogle Scholar
- Villanueva B, Pong-Wong R, Woolliams JA: Marker assisted selection with optimised contributions of the candidates to selection. Genet Sel Evol. 2002, 34: 679-703. 10.1051/gse:2002031.PubMed CentralView ArticlePubMedGoogle Scholar
- Meuwissen THE, Goddard ME: The use of marker haplotypes in animal breeding schemes. Genet Sel Evol. 1996, 28: 161-176. 10.1051/gse:19960203.PubMed CentralView ArticleGoogle Scholar
- Ruane J, Colleau JJ: Marker-assisted selection for a sex-limited character in a nucleus breeding population. J Dairy Sci. 1996, 79: 1666-1678.View ArticlePubMedGoogle Scholar
- Druet T, Fritz S, Colleau JJ, Gautier M, Eggen A, Rossignol MN, Boscher MY, Malafosse A, Boichard D: Genetic markers in breeding programs. Proceedings of the 26th European Holstein and Red Holstein Conference:. 2005, [http://www.whff.info/pdf/26ehc_prague/druet.pdf]May ; PragueGoogle Scholar
- Rönnegård L, Carlborg Ö: Separation of base allele and sampling term effects gives new insights in variance component QTL analysis. BMC Genet. 2007, 8: 1-14. 10.1186/1471-2156-8-1.PubMed CentralView ArticlePubMedGoogle Scholar
- Thaller G, Krämer W, Winter A, Kaupe B, Erhardt G, Fries R: Effects of DGAT1 variants on milk production traits in German cattle breeds. J Anim Sci. 2003, 81: 1911-1918.PubMedGoogle Scholar
- Viitala S, Szyda J, Blott S, Schulman N, Lidauer M, Maki-Tanila A, Georges M, Vilkki J: The role of the bovine growth hormone receptor and prolactin receptor genes in milk, fat and protein production in Finnish Ayrshire dairy cattle. Genetics. 2006, 173: 2151-2164. 10.1534/genetics.105.046730.PubMed CentralView ArticlePubMedGoogle Scholar
- Guillaume F, Fritz S, Boichard D, Druet T: Correlations of marker-assisted breeding values with progeny-test breeding values for eight hundred ninety-nine French Holstein bulls. J Dairy Sci. 2008, 91: 2520-2522. 10.3168/jds.2007-0829.View ArticlePubMedGoogle Scholar
- Spelman RJ: Major factors in marker-assisted selection genetic response in dairy cattle populations. Proceedings of the 6th World Congress on Genetics Applied to Livestock Production: 11–16 January 1998; Armidale. 1998, 26: 365-368.Google Scholar
- Meuwissen THE, Hayes BJ, Goddard ME: Prediction of total genetic value using genome-wide dense marker maps. Genetics. 2001, 157: 1819-1829.PubMed CentralPubMedGoogle Scholar
- VanRaden PM, Van Tassell CP, Wiggans GR, Sonstegard TS, Schnabel RD, Schenkel F: Genomic data and cooperation result in faster progress. Interbull Bull. 2008, 38:Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.