Genomic prediction when some animals are not genotyped
 Ole F Christensen^{1}Email author and
 Mogens S Lund^{1}
DOI: 10.1186/12979686422
© Christensen and Lund; licensee BioMed Central Ltd. 2010
Received: 28 September 2009
Accepted: 27 January 2010
Published: 27 January 2010
Abstract
Background
The use of genomic selection in breeding programs may increase the rate of genetic improvement, reduce the generation time, and provide higher accuracy of estimated breeding values (EBVs). A number of different methods have been developed for genomic prediction of breeding values, but many of them assume that all animals have been genotyped. In practice, not all animals are genotyped, and the methods have to be adapted to this situation.
Results
In this paper we provide an extension of a linear mixed model method for genomic prediction to the situation with nongenotyped animals. The model specifies that a breeding value is the sum of a genomic and a polygenic genetic random effect, where genomic genetic random effects are correlated with a genomic relationship matrix constructed from markers and the polygenic genetic random effects are correlated with the usual relationship matrix. The extension of the model to nongenotyped animals is made by using the pedigree to derive an extension of the genomic relationship matrix to nongenotyped animals. As a result, in the extended model the estimated breeding values are obtained by blending the information used to compute traditional EBVs and the information used to compute purely genomic EBVs. Parameters in the model are estimated using average information REML and estimated breeding values are best linear unbiased predictions (BLUPs). The method is illustrated using a simulated data set.
Conclusions
The extension of the method to nongenotyped animals presented in this paper makes it possible to integrate all the genomic, pedigree and phenotype information into a onestep procedure for genomic prediction. Such a onestep procedure results in more accurate estimated breeding values and has the potential to become the standard tool for genomic prediction of breeding values in future practical evaluations in pig and cattle breeding.
Background
Genomic selection [1] has become the new paradigm in animal breeding programs using markerassisted selection. It may increase the rate of genetic improvement, reduce the generation time, and provide higher accuracy of estimated breeding values (EBVs). Genomic prediction of breeding values can be based on a linear mixed model using matrix computations or a nonlinear mixture type of model using Markov chain Monte Carlo (McMC) procedures. In this paper we provide a natural extension of a linear mixed model to the situation with nongenotyped animals.
as in VanRaden [3], but other types of genomic relationship matrices are discussed in the discussion section. In VanRaden [3] it is assumed that all animals are genotyped, which is unlikely to be a common scenario. In particular, in pig breeding it is probable that only boars or other selection candidates are genotyped, and in cattle breeding, traits being recorded for millions of animals it is very unlikely that all will be genotyped. We present an extension of matrix (1) in the situation where not all animals are genotyped. The approach presented here combines the relationship matrix (1) with a model for the markers. By marginalisation of the markers of nongenotyped animals a natural extension of (1) is obtained. The resulting extension of the genomic relationship matrix is the same as the one derived in Legarra et al. [5], but the details in the derivation are somewhat different and the derivation therefore sheds more light on this result.
To capture genetic variation not associated to the markers in a given SNPpanel, the model can also contain a polygenic genetic effect with the usual pedigree derived additive relationship matrix, as considered by [4, 6] among others. The extension of the genomic relationship matrix to nongenotyped animals together with the addition of the polygenic effect provide a natural onestep procedure to blend the information from relatives and the genomic information into a combined genomically enhanced breeding value (GEBV). Genomic prediction with both a polygenic effect and with incomplete genotyping has been considered by a number of authors. Using a joint model for phenotypes and markers and using Bayesian inference, a general solution to sample missing markers in each McMC iteration has been suggested [4, 7]. However, with a large number of SNP markers and many animals without genotypes such a solution seems computationally unfeasible in practice. In Gianola et al. [7] bivariate models are suggested, where the two traits are the traits of the genotyped and nongenotyped animals, respectively, and the genetic effect for a genotyped animal is the sum of a polygenic effect and a genomic effect whereas the genetic effect for a nongenotyped animal is just a polygenic effect (correlated with the polygenic effect of the genotyped animals). Since the model does not contain a genomic genetic effect for the nongenotyped animals, the phenotypic information from nongenotyped animals closely related to a given genotyped animal does not propagate properly into the estimate of the genomic genetic effect for this animal. Alternatively, the approach by Baruch and Weller [8] involves several steps, where first, expected genotypes are computed for nongenotyped animals, then marker effects are estimated (using expected genotypes for nongenotyped animals), phenotypes are adjusted by known or expected marker effects, and finally polygenic EBVs are computed from adjusted phenotypes. Although somewhat similar in idea to the approach taken here, the approach in [8] does not propagate any uncertainty from one step in the procedure to the next step, and the effects are not estimated simultaneously.
Methods
with s = ∑_{ j }2ρ_{ j }(1  ρ_{ j }).
where y is phenotype, X and Z are incidence matrices, β denotes fixed effects, e is error, is the polygenic genetic effect, and is the genomic genetic effect. Here A is the usual pedigree derived additive relationship matrix, and G*(m^{ obs }) is the extension of (2) to be derived in the following section.
In the following sections, first, we derive the extension of the marker based relationship matrix, G*(m^{ obs }), and second, we study the variancecovariance matrix of the combined genetic effect g + a. Then procedures for parameter estimation using AIREML, and breeding value estimation are presented. Finally, a simulation data set is described.
Genomic relationship matrix with a relationship of markers
Gengler et al. [9] suggested that missing genotypes could be modelled using the usual mixed model methodology with relationship matrix A. We now combine that idea with the genomic relationship matrix on the form (1). For simplicity, the derivation is made for the form (2), but it is straightforward to generalise to (1) also.
where M is the gene content matrix. We assume that E[M_{ j }] = 1p_{ j }, Var(M_{ j }) = v_{ j }A, with A the usual relationship matrix, v_{ j }= 2ρ_{ j }(1  ρ_{ j }), and s = ∑_{ j }v_{ j }. The covariances of M_{ j }, and M_{ j' }for two different loci j ≠ j' are on the form Cov(M_{ j }, M_{ j' }) = v_{j,j'}A where the v_{j,j'}s are unspecified since they are cancelling in the derivations that follow.
When all animals have been genotyped, G*(m^{ obs }) = G(m^{ obs }), and when no animals have been genotyped, G*(m^{ obs }) = A, which makes the extension in (4) rather elegant. We assume that the distribution of [gm^{ obs }] is multivariate normal, which for the nongenotyped animals is not strictly true, but an approximation.
Considering the terms in (6), because of the low dimension of G(m^{ obs }) and A_{11} a direct inversion of these matrices should be possible for practical computations, and A^{1} is a sparse matrix which can be computed directly without constructing A itself and using standard techniques. To compute A_{11} there might be cases where most of the A matrix has to be computed, potentially causing a memory storage problem.
Alternatively, A_{11} = ((A^{1})^{1})_{11} may be computed using the formula (5) on A^{1} and using sparse matrix computation. The formula (6) requires that G(m^{ obs }) is invertible which may not actually be the case. In the next section this problem is automatically solved by combining the genomic genetic effect g with the polygenic effect a.
where A_{22}  A_{21} A_{12} is easily obtained from A^{1}, and the determinant can be computed using sparse matrix computation.
The combined genetic effect
The parameter w is interpreted as the relative weight on the polygenic effect, and it may be estimated from data as shown in the next section or be chosen to equal a small value.
and here G_{ w }is necessarily invertible when w > 0 (even when G(m^{ obs }) is singular).
Variance component estimation
where . We will not enter into details, but just note that the sparse structure of the left hand side matrix in (9) is the cornerstone for the fast computation of the AImatrix used in the numerical maximisation of the REML likelihood. Considering the terms in this matrix, then Z^{T}Z is a sparse matrix, and from (4) we see that has some sparse structure, although is a dense matrix. Depending on the proportion of animals genotyped it may in some cases not be necessarily advantageous to compute the AImatrix using (9), but instead an AIREML algorithm based on the inverse phenotypic variancecovariance matrix, , could be used, see [12]. Here, we assume that the majority of animals are not genotyped and use the sparse structure of G*(m^{ obs })^{1} for AIREML based on the mixed model equations.
The AIREML method based on the mixed model equations is implemented in software DMU [13] and requires input in the form of the vector of phenotypes, the nonzero entries of and the logdeterminant log(det( )) = log(det(G_{ w })) + log(det(A_{22}  A_{21} A_{12})). For a given w the software provides estimates of and , values of the REML loglikelihood at the maximum and (when required) BLUE solution and BLUP solution . Here, the parameter w is estimated by using a grid of values, i.e. w = 0.01, 0.03, ..., 0.19, and computing the REML loglikelihood for each value. The resulting profile likelihood curve, log , has a peak at the estimate , and a measure of the associated uncertainty is the interval {wlog > log  3.84} where 3.84 is the 95% quantile of a χ^{2}(1)distribution.
Breeding value estimation
Here we consider estimation (prediction) of breeding values. For animals included in the parameter estimation (animals with phenotypes, and some additional animals whose markers provide information about the unknown markers for nongenotyped animals with phenotypes), the GEBVs are the solution vector to (9) with the parameter values being the estimated ones from the previous section. The software DMU provides these GEBVs and their precision.
where , Z_{ all }and now contain all animals. Again software DMU provides these GEBVs and their precision.
For a scenario with a large number of genotyped animals whose marker information does not provide information for the parameter estimation, Appendix B presents a method for breeding value estimation where only part of the needs to be computed.
A simulated data set
The simulated data set is inspired by a pig nucleus breeding program, but is formulated in a simplified form. We assume, 10 chromosomes each 160 cM long, and a panel of p = 5000 equidistant SNP markers is used. It is assumed that 500 QTLs affect the phenotype, and the size of these effects is simulated from a Gamma(5.4, 0.42)distribution. First, a base population consisting of 150 boars and 1500 sows is generated by assuming random mating for 50 generations in a population with an effective population size of 100. Then the following mating and selection scheme is followed for five generations. In each generation, 150 boars are mated with 1500 sows to produce 15000 offspring (half of them males). For the next generation, the 150 boars with the highest value of their own phenotype are selected, and 1500 sows are selected randomly. It is assumed that family records are available for all five generations, phenotypes of all boars available for all five generations (35000 records), and the selected boars in the last three generations are genotyped (450 animals). In addition, to estimate the allele frequencies required for the method, the 150 boars in the base population are genotyped (and the allele frequencies used are the estimated frequencies from these 150 boars). For prediction, it is assumed that 300 selection candidates (without phenotypes) for generation 6 are genotyped.
where y_{ EBV }is the vector of traditional EBVs, and with G_{ w }= 0.99G(m^{ obs }) + 0.01A_{11}.
For the onestep method, the genotypes of the selection candidates provide information about the genotypes of their (nongenotyped) mothers and hence information about other nongenotyped animals further back in the pedigree. Therefore they also provide some information about the genotypes of the boars without offspring, and since these boars have phenotypes but not genotypes then the selection candidates should be included in the parameter estimation. However, to investigate how important it is to include these animals, a second analysis (onestep2) is also performed where they are not included. Finally, to investigate the importance of obtaining the allele frequencies in the base population, the scenario where the boars in the base population have not been genotyped is also studied. The use of three different allele frequencies are compared: 1) true allele frequencies (obtained from the 150 boars in the base population), 2) estimated allele frequencies for boars used in generation 3, 3) allele frequencies estimated using the approach by Gengler et al. [9].
Results
Results from model with = 0.01.
Method 

 Cor. true BV 

onestep  4.16  16.22  0.6598 
ped  5.03  15.80  0.3537 
twostep  7.56  0.069  0.5869 
onestep2  5.98  15.58  0.6596 
Column four in Table 1 shows the result obtained when ignoring the genotypes of the 300 selection candidates in the parameter estimation (onestep2). Even though the parameter estimates are somewhat different betweeen onestep and onestep2, only a minor difference in the correlation between GEBVs and the true breeding values is seen. Hence, for this data set this specific computational shortcut performs well. Finally, the results from the analyses where the boars in the base population are not genotyped show that the choice of allele frequencies is very important for parameter estimation. When using the true allele frequencies, ≈ 0 is obtained, whereas when using allele frequencies estimated from the observed genotypes, = 1 is obtained for both methods estimating the allele frequencies. Since = 1 corresponds to the usual animal model, no further results from this comparison are shown here. We conclude that for this data set the parameter estimation is sensitive to the allele frequencies used in the onestep method.
Discussion
For genomic prediction an extension of a linear mixed model to nongenotyped animals has been derived here. The extension of the method makes it possible to integrate in an optimal way the genomic, pedigree and phenotype information into a onestep procedure for breeding value estimation. Due to the simplicity of the method, the fact that it extends the traditional breeding value estimation method in a natural way, and the possibilities of handling large populations, such a onestep procedure has the potential to become the standard tool for genomic prediction of breeding values in practical pig or cattle evaluations in the future. The practical implementation of the approach uses an existing software DMU, and therefore the approach can be easily extended to other types of models implemented by that software, in particular multivariate analysis and generalised linear mixed models.
For such a onestep procedure to become the standard tool for computing GEBVs in practical pig or cattle evaluations, some technical issues of the method need further development. First, computing times necessary for the construction and the inversion of G(m^{ obs }) are proportional to p and , respectively. These computations seem to be the computational bottlenecks for the method, and for a very large number of genotyped animals the method may not be feasible. Further research on efficient computation of G(m^{ obs })^{1} seems necessary. Second, some computational shortcuts in the method could be imagined, as illustrated in our results by the good performance of the onestep method even when the marker information from selection candidates is ignored in the parameter estimation. Investigations by extensive simulation studies may reveal the benefits of other potential shortcuts. Third, the allele frequencies in the base population are considered known, or at least easily accessible. As illustrated in the results, the parameter estimation seems to be sensitive to the choice of these allele frequencies in a scenario with selection and where the base population itself has not been genotyped. To investigate whether the problems may be related to the strong selection on phenotype for the simulation data set, this analysis was repeated for a simulation with boars selected randomly. Here more sensible parameter estimates were obtained in the sense that ≈ 0 when allele frequencies were estimated from observed genotypes. For practical dairy cattle evaluations, Misztal et al. [16] investigated the use of a number of different allele frequencies and obtained the best results by using ρ_{ j }= 1/2 for all j but replacing s = 2∑_{ j }ρ_{ j }(1  ρ_{ j }) = p/2 with a another scaling s which in practice was larger than p/2. Of course, whether that result is due to selection in this real data set is not known. Further research on the effect of selection and on how to handle appropriately the issue with allele frequencies is needed.
An assumption behind the genomic relationship matrix (2) is that all regions of the genome are equally important for the trait of interest. It is possible to instead use G(m) ∝ (m  p)h(m  p)^{ T }where h is a diagonal matrix with known weights h_{ jj }= with b_{ j }s being estimated SNP effects (estimated using for example a nonlinear mixture type of model as in [1]). However, incorporating uncertainty on such estimated SNP effects into the method seems less straightforward.
with correlation parameter ϕ, corresponds to the method in [4] in it's dual formulation as a linear mixed model. For this choice of markerbased relationship matrix, the derivation of K*(m^{ obs }) = Var [gm^{ obs }] is also possible, but as shown in Appendix C the form of the result differs from (4) in a number of ways. The implication is that using (4) and (6) with a marker based relationship matrix defined by (11) is possible, but lacks theoretical justification.
Appendix A
Here the mean and variances of the conditional distribution [g  m^{ obs }] (with M^{ miss }marginalised out) are derived using formulas for conditional expectations, variances and covariances.
when Var(M) = V ⊗ A.
where G*(m^{ obs }) is defined in (4). However, this is not strictly correct for a nongenotyped animal i where g_{ i } X ~N (0, X) with X here being a random variable with distribution [∑_{ j }(M_{ ij } p_{ j })^{2}m^{ obs }]. This conditional distribution will never lead to a marginal normal distribution for g_{ i }(the only exception is when X is a constant). The normal distribution of gm^{ obs }is therefore only an approximation.
Appendix B
In some scenarios the number of genotyped animals not included in the parameter estimation may be large, for example if phenotypes are expensive to obtain and therefore only observed on a small subset of the population. To reduce the computational burden of creating the whole (m^{obs,other}) for all animals, a procedure is presented where only a part of this matrix needs to be computed.
This shows that the GEBVs of such genotyped animals only depend on . It also shows that only a part of the full genomic relationship matrix for genotyped animals is necessary to compute, since G_{w,33}= (1  w)G(m^{ other }) + wA_{33} does not enter into (12).
where (A_{ all })^{1} is sparse and is computed directly, and and are dummy variables.
Appendix C
to nongenotyped animals.
with and where these expectations and variances can be computed from the conditional expectations and variances given in Appendix A. The form exp(v^{2}/(1 + τ^{2}))/ with the variance τ^{2} occurring in two places, implies that that the elements in K*(m^{ obs }) cannot be expressed in matrix form as in (4) but are on a more complicated form.
Declarations
Acknowledgements
The work was part of the project "Svineavl, Genomisk selektion" funded by the Danish Ministry of Food, Agriculture and Fisheries, and Danish Pig Production. Guosheng Su is acknowledged for help in relation to the generation of the simulation study, and Per Madsen is acknowledged for his unselfish work on creating and maintaining the software DMU. A reviewer is thanked for his suggestions on how to improve the presentation.
Authors’ Affiliations
References
 Meuwissen THE, Hayes BJ, Goddard ME: Prediction of total genetic value using genomewide dense marker maps. Genetics. 2001, 157: 18191829.PubMed CentralPubMedGoogle Scholar
 VanRaden PM: Efficient methods to compute genomic predictions. Interbull Bull. 2007, 37: 111114.Google Scholar
 VanRaden PM: Efficient methods to compute genomic predictions. J Dairy Sci. 2008, 91: 44144423. 10.3168/jds.20070980.View ArticlePubMedGoogle Scholar
 Gianola D, van Kamm BCHM: Reproducing kernel Hilbert spaces regression methods for genomic prediction of quantitative traits. Genetics. 2008, 178: 22892303. 10.1534/genetics.107.084285.PubMed CentralView ArticlePubMedGoogle Scholar
 Legarra A, Aguilar I, Misztal I: A relationship matrix including full pedigree and genomic information. J Dairy Sci. 2009, 92: 46564663. 10.3168/jds.20092061.View ArticlePubMedGoogle Scholar
 Calus MPL, Veerkamp RF: Accuracy of breeding values when using and ignoring the polygenic effect in genomic breeding value estimation with a marker density of one SNP per cM. J Anim Breed Genet. 2007, 124: 362368.View ArticlePubMedGoogle Scholar
 Gianola D, Fernando RL, Stella A: Genomicassisted prediction of genetic value with semiparametric procedures. Genetics. 2006, 173: 17611776. 10.1534/genetics.105.049510.PubMed CentralView ArticlePubMedGoogle Scholar
 Baruch E, Weller JI: Incorporation of genotype effects into animal model evaluations when only a small fraction of the population has been genotyped. Animal. 2009, 3: 1623. 10.1017/S1751731108003339.View ArticlePubMedGoogle Scholar
 Gengler N, Mayeres P, Szydlowski M: A simple method to approximate gene content in large pedigree populations: application to the myostation gene in dualpurpose Belgian Blue cattle. Animal. 2007, 1: 2128. 10.1017/S1751731107392628.View ArticlePubMedGoogle Scholar
 Gilmour AR, Thompson R, Cullis BR: Average information REML: an efficient algorithm for parameter estimation in linear mixed models. Biometrics. 1995, 51: 14401450. 10.2307/2533274.View ArticleGoogle Scholar
 Johnson DL, Thompson R: Restricted maximum likelihood estimation of variance components for univariate animal models using sparse matrix techniques and average information. J Dairy Sci. 1995, 78: 449456.View ArticleGoogle Scholar
 Lee SH, Werf van der JHJ: An efficient variance component approach implementing an average REML suitable for combined LD and linkage mapping with a general pedigree. Genet Sel Evol. 1995, 38: 2543. 10.1186/1297968638125.View ArticleGoogle Scholar
 Madsen P, Jensen J: A users guide to DMU, version 6, release 4.7. Manual, Faculty of agricultural science, University of Aarhus. 2008Google Scholar
 VanRaden PM, Van Tassel CP, Wiggans GR, Sonstegard TS, Schnabel RD, Taylor JF, Schenkel FS: Invited review: reliability of genomic predictions for North American Holstein bulls. J Dairy Sci. 2009, 92: 1624. 10.3168/jds.20081514.View ArticlePubMedGoogle Scholar
 Su G, Guldbrandtsen B, Gregersen VR, Lund MS: Preliminary investigation on reliability of genomic estimated breeding values in the Danish Holstein population. J Dairy Sci. 2010,Google Scholar
 Misztal I, Legarra A, Aguilar I: Computing procedures for genetic evaluation including phenotypic, full pedigree and genomic information. Proceedings of the annual meeting EAAP: 2427 August 2009; Barcelona, Spain. 2009Google Scholar
Copyright
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.