Design of reference populations for genomic selection in crossbreeding programs
© van Grevenhof and van der Werf; licensee BioMed Central. 2015
Received: 20 May 2014
Accepted: 30 January 2015
Published: 7 March 2015
In crossbreeding programs, genomic selection offers the opportunity to make efficient use of information on crossbred (CB) individuals in the selection of purebred (PB) candidates. In such programs, reference populations often contain genotyped PB animals, although the breeding objective is usually more focused on CB performance. The question is what would be the benefit of including a larger proportion of CB individuals in the reference population.
In a deterministic simulation study, we evaluated the benefit of including various proportions of CB animals in a reference population for genomic selection of PB animals in a crossbreeding program. We used a pig breeding scheme with selection for a moderately heritable trait and a size of 6000 for the reference population.
Applying genomic selection to improve the performance of CB individuals, with a genetic correlation between PB and CB performance (rPC) of 0.7, selection accuracy of PB candidates increased from 0.49 to 0.52 if the reference population consisted of PB individuals, it increased to 0.55 if the reference population consisted of the same number of CB individuals, and to 0.60 if the size of the CB reference population was twice that of the reference population for each PB line. The advantage of using CB rather than PB individuals increased linearly with the proportion of CB individuals in the reference population. This advantage disappeared quickly if rPC was higher or if the breeding objective put some emphasis on PB performance. The benefit of adding CB individuals to an existing PB reference population was limited for high rPC.
Using CB rather than PB individuals in a reference population for genomic selection can provide substantial advantages, but only when correlations between PB and CB performances are not high and PB performance is not part of the breeding objective.
Traits expressed in purebred (PB) and crossbred (CB) individuals are genetically not the same . PB and CB performances can be considered as two genetically correlated traits; the correlation between PB and CB performances (rPC) is affected by the extent of non-additive genetic effects (particularly dominance) and the genetic distance between lines or breeds crossed. This correlation also often reflects genotype by environment interactions due to nucleus PB and commercial CB animals being exposed to different environments. Wei and van der Werf  proposed to consider PB and CB performances as two correlated traits and to use a multi-trait model for their genetic evaluation. The breeding objective usually focuses on the performance of CB animals.
In practice, it is often difficult to use CB information since both performance and pedigree records can be difficult to obtain on CB animals. Moreover, selection for CB performance is either on half-sibs, which does not exploit variation within the half-sib family, or on progeny information, which would lengthen the generation interval if it was used. Bijma and van Arendonk  showed that extensive use of sib information on CB individuals can lead to increased rates of inbreeding.
Genomic selection could benefit crossbreeding programs since it allows using information on CB animals that is available at an early age and the method uses within-family variation [4,5]. Dekkers  proposed to use marker information that was calibrated based on the performance of commercial CB animals. He found a significant increase in the rates of genetic gain compared to using only PB phenotypic information, or combined PB and CB information, whereas the rate of inbreeding decreased.
Genomic selection uses marker genotypes and phenotypes in a reference population to predict breeding values of selection candidates that have been genotyped . The effectiveness of genomic selection will depend on the size and composition of the reference population used for genomic predictions [8,9]. In crossbreeding programs, the breeding objective often includes CB performance, or in some cases, both PB and CB performance. Therefore, it seems reasonable to recommend the use of performance and genotypic data on CB individuals for genomic selection in crossbreeding programs. However, although large amounts of phenotypic as well as genotypic information on PB animals are usually already available, collecting CB information might be difficult, expensive and time-consuming. Thus, it is relevant to evaluate the benefit of using CB information in a reference population compared to that of PB information.
Dekkers  found that genetic gains were substantially less when genomic prediction was based on PB phenotypic data compared to CB phenotypic data. However, this assessment was based on comparisons for which the molecular breeding value (mbv) was combined with PB phenotypic information only, and the mbv was based on either PB or CB individuals with phenotypes, but not on both. Also, accuracies of mbv based on PB or CB individuals were assumed to be the same. However, if the linkage phase between a marker and a quantitative trait locus (QTL) is not consistent across breeds or selection lines, then a CB animal will have only one haplotype that is potentially informative to predict the mbv of a PB selection candidate whereas a PB animal has two relevant haplotypes. For a reference population of same size, phenotyping the CB individuals would require the estimation of twice the number of chromosome segments, therefore contributing less information per animal to a particular PB line. Nevertheless, the information provided is for the more relevant trait (CB performance).
This study aimed at assessing more generally the benefit of including CB information in the reference population of a crossbreeding program using genomic selection. The efficiency of investing in CB information to enable genomic selection of PB animals for CB performance was explored by varying the proportion of PB and CB animals in a reference population, or by adding CB individuals to an existing PB reference population. The size of the reference population, the size of the correlation between PB and CB performance and the emphasis on PB performance in the breeding objective were varied. Additionally, two examples of hard-to measure traits, relevant to pig breeding schemes, were considered.
Selection index methodology provides a suitable framework to predict the accuracy of estimated breeding values for various breeding program scenarios, including the use of genomic information . Deterministic simulation was used to predict the accuracy of estimated breeding values of selection candidates in a two-way crossing system. PB animals were selected for an index that included varying amounts of records and genotypes on CB and PB individuals. In scenarios with genomic selection, the make-up of the reference population was varied in terms of size, and proportion of CB individuals.
We assumed a pig breeding nucleus with two PB lines that included 500 breeding females and 25 females mated to one male. We assumed that in each full-sib family, two males and two females were measured to become selection candidates for nucleus replacement. Nucleus replacements were selected on an index measured at a fixed time of selection.
For the base situation, we considered a trait that could be measured on both sexes before selection started. The trait heritability was 0.25. When selection started, phenotypic information was available on own performance, sire and dam performance, three full-sib records and 40 half-sib records of PB family members. To be able to compare the results of our study with those reported in , these parameters were set to the same values as in . PB animals were mated to produce CB animals. The breeding goal was to improve the performance of CB animals, although, in some breeding schemes, PB performance also had some economic value. PB individuals had 10 CB half-sibs that were phenotypically measured at the time of their selection (no CB progeny). PB animals could also be genotyped before selection and the reliability of their genomically estimated breeding values depended on size and composition of the reference population. We assumed a reference population of varying size and with a varying proportion of CB animals.
Base parameters used in the simulations
Phenotypic standard deviation, SD
Common environment among full-sibs, c2
Economic value, PB_EV / CB_EV
Effective population size, Ne
Reference population size, nP
Purebred-Crossbred correlation, rPC
We also studied two single-trait examples relevant to pig breeding schemes that can be considered as ‘hard-to-measure traits’ and which therefore are expected to benefit more from genomic selection. We considered two sow traits that had different heritabilities and that could be measured within the breeding program. These traits were number of piglets born alive in first parity (PBA1) and length of productive life (LPL). PBA1 phenotypes were available for young sows and the heritability was 0.12 . We evaluated scenarios based on accuracy of male selection. Males were selected based on information on the dam, two full-sib records, 20 half-sib records and five half-sib CB records. For genomic selection, we assumed that a total of 4000 records were available in the reference population of which 1000 were PB records and 3000 were individuals with a varying proportion of CB animals (pCB).
LPL is a measure of longevity that is only available on CB sows (because PB dams are kept only for a short time to limit generation interval) and late in life. Without genomic selection, this trait cannot be selected for in breeding programs. Genomic selection creates the potential to select on CB performance via genomic breeding values. LPL was assumed to have a heritability of 0.06 in CB individuals . For genomic selection, we assumed a reference population of 6000 individuals, with pCB ranging from 0 to 1. We assumed that, for the purpose of creating a reference population, some PB sows will be kept longer and measured for the trait, without actually being used as nucleus dams. The size of this reference population may seem unrealistically large, but such a size is required in order to obtain a reasonable genomic prediction accuracy for such a lowly heritable trait. For both traits, rPC and EV_PB were varied.
Accuracy of molecular breeding values
The accuracy r ag of the estimated breeding values based on genomic information (g) was derived from the size and composition of total reference population. We considered two reference populations, one for PB and one for CB of size (1-pCB) · nP and pCB · nP, respectively, nP is the size of the combined reference population. We used the formula of Daetwyler et al.  to predict r ag of both g PB and g CB : rag = h2/(h2 + λ), where h2 is the trait heritability and λ = ng/N where N is size of reference population (N = (1-pCB) · nP for purebreds and N = pCB · nP for crossbreds), ng refers to the effective number of chromosome segments (independent loci). As suggested by Hayes et al. , ng can be approximated as 2NeL, where Ne is the effective population size and L is the genome length in Morgans. For the pure line breeding population, we assumed an effective size of Ne = 100. We assumed that in the CB reference population, the number of chromosome segments to be estimated is twice that in the PB reference population, since, at each segment of the genome, the haplotypes originate from two different breeds or lines. Note that this is equal to assuming that each CB individual is half as informative as a PB individual (as 2ng/N = ng/ (N/2)).
Following Dekkers [6,10], multi-trait selection indices were derived to predict the accuracy of the selection index. We combined phenotypic information with genomic predictions of breeding value, i.e. the molecular breeding value (g), from both PB and CB information sources to estimate breeding values for PB (a PB) and CB (a CB ) performances. For the selection index, the molecular breeding value g was considered as a separate trait, which is correlated to the breeding value it predicts, with the correlation between a i and g i (r aigi ) equal to the accuracy of the molecular breeding value; the heritability of g was equal to 0.999. Hence, the multi-trait model used for the selection index contained four traits: a PB, a CB, g PB and g CB. The correlation between a PB and a CB was equal to rPC and the correlation between g PB and g CB was assumed to be equal to r aPBgPB.r aCBgCB.rPC, which are similar to those in  but we assumed that all the additive genetic variance was captured by markers, which in turn is an assumption consistent with Daetwyler et al. . The selection index was optimized for the breeding objective with various degrees of emphasis on a PB and a CB and no emphasis on any of the g i. The accuracy of the selection index was used as a criterion to compare the performance of different scenarios.
We varied the size of the total reference population (nP), the degree of emphasis put on PB performance in the breeding objective (PB_EV), and the correlation between PB and CB performances (rPC). In all comparisons, we plotted the index accuracy versus pCB.
Reference population structure
Correlation between PB and CB performances
Effect of CB individuals in a reference population
Genomic selection can be very valuable in crossbreeding programs since it allows efficient selection on CB performance. In general, a larger proportion of CB animals in the reference population will increase the selection accuracy when the breeding goal is focussed on CB performance. However, it might be difficult and expensive to collect phenotypes and genotypes on CB individuals, whereas most breeding programs have routine phenotyping measurement and genotyping of nucleus animals in the pure lines. We found that the effect of replacing PB with CB animals in the reference population was highly positive but only when the correlation between PB and CB performance was low (rPC < 0.7) and the breeding objective emphasis was mainly focused on improving CB performance. In our example, with an rPC of 0.7 and a breeding objective that focused only on CB performance, the index accuracy increased from 0.526 to 0.554 when using a CB instead of a PB reference population. With an rPC of 0.9, the index accuracy decreased slightly from 0.635 to 0.626 for a CB compared to a PB reference population. Therefore, the additional gain from using CB rather than PB individuals in the reference population decreased rapidly with higher values of rPC and also when emphasis was put on PB performance in the breeding objective. The increase in index accuracy from including genomic information was also relatively low when selection of PB animals was already based on phenotypic information from selection candidates and their relatives (including CB).
Dekkers  concluded that marker-based selection in crossbreeding programs was much more effective when molecular prediction was based on commercial CB data rather than on PB data, which contradicts our findings that the additional gain from using CB animals in the reference set is in many cases only modest. Dekkers  used an rPC of 0.7 and a breeding objective that targeted CB performance only. One difference between our study and that of Dekkers  is that Dekkers assumed that accuracies of g PB and g CB were equal whereas we assumed an equal number of phenotypes, which led to a lower prediction accuracy of g CB versus g PB. If we assumed accuracies of g PB and g CB were equal, the index accuracy increased from 0.49 to 0.60 (23% increase) with a CB versus a PB reference population, instead of an increase from 0.49 to 0.55 (13% increase; Figure 1). One could argue that in practice a breeder would be interested in improving both selection lines, in which case the choice is between investing in genotyping n PB individuals in each line, or 2n CB individuals, the latter providing information for both lines. In that case, the molecular breeding values g PB and g CB would be equally accurate, and the benefit of investing the same genotyping effort in CB rather than PB animals would be higher than shown here. However, we also showed that knowledge of the rPC is critical in assessing this benefit.
Another difference is that Dekkers  assumed accuracies of 0.6 for predictions of g PB and g CB , which is higher than what we used here. Higher genomic prediction accuracy gives more gain in accuracy when replacing PB by CB animals in the reference population, as shown in Figure 1. Furthermore, Dekkers  compared the gains from using molecular information only, or adding molecular information to PB phenotypic information on the selection candidates. We assumed that some phenotypic information was known on 10 CB half-sibs, which makes the relative value of adding molecular information based on CB phenotypes smaller. If the phenotypic information on 10 CB half-sibs was discarded, the increase in index accuracy due to introducing genomic selection was 17% (from 0.445 to 0.520) if it was based on a CB reference population) but only 9% (from 0.445 to 0.486) if it was based on a PB reference population. If CB performance is measured, then there is some advantage in genotyping the same individuals, but similarly, if CB performance is measured for the reference population, then this information should be used to predict the breeding value of PB relatives. Genomic information can be used to predict relatedness, which alleviates the need for pedigree recording in the CB animals. Hence, genomic testing and phenotyping of CB individuals might provide more information than predicted by the Daetwyler et al.  formula (which assumes prediction from unrelated individuals), since it facilitates the use of information on close CB relatives.
We used a selection index approach to predict the additional gains from genomic information in breeding programs. The selection index model combines information from phenotypic data on relatives with genomic prediction, and assumes that these sources of information are independent. In practice, this is often not the case, e.g. CB animals can contribute to a reference population, but they can also directly affect predictions of relatives through pedigree relationships. A dependency between these sources of information needs to be accounted for in genetic evaluation procedures, e.g. in ad-hoc “blending” methods , or more appropriately in the so-called “single-step” method , to avoid bias and inflation of the accuracy of the combined breeding value. Such dependence may exist for some individuals, but not for others, so it is difficult to account for it in a deterministic modelling study. Moreover, ignoring this dependence may inflate genomic prediction accuracy, and therefore the benefit of genomic information, but it is unlikely to affect the relative value of CB vs PB records in the reference population.
In practice, breeding objectives often put emphasis on PB performance, in which case the added value of adding CB animals in the reference population becomes very small. The benefit of increasing the percentage of CB individuals in the reference population depends on rPC, as well as on PB_EV. In our examples, we found that if rPC is not very low (~0.8) and PB_EV is greater than 0, which is a realistic scenario, then the benefits of using PB or CB animals in the reference population are similar.
In this study, we assumed an effective population size of 100. With Ne = 500, the added value of genomic selection based on a PB reference population was only 1.7% and increased to 2.9% if the reference population consisted of only CB animals. The accuracy of genomic prediction of g PB based on 6000 PB animals was predicted to be 0.22 for Ne = 500, whereas that of g CB was only 0.16 based on 6000 CB animals. Both predictions assume no direct pedigree relationships between the reference population and selection candidates and probably underestimate the accuracy achieved in practice.
The effective population size affects the estimated number of effective chromosome segments or independent loci (ng), and this can have a large effect on the predicted accuracy of molecular breeding values (r a,g ). However, estimating ng from Ne is not straightforward and approximations used in the literature differ [9,15-17]. We used ng = 2NeL, following Hayes et al. . Other approximations were proposed, e.g. Goddard  suggested ng = 2NeL/log(4NeL), Goddard et al.  used ng = 2NeLk/log(2Ne.L) and Meuwissen et al.  proposed ng = 2NeLk/ln(2Ne), where L is the number of chromosomes and k is the average length per chromosome (i.e. their Lk is the same as our L). The latter approximation leads to a considerably smaller estimate for ng and therefore higher accuracies. It might be useful to use empirical evidence to verify these predictions. However, the approximation of r a,g has an impact on the benefit of genomic selection in breeding programs by affecting the size of the reference population required to achieve a certain benefit, but we showed that it has only a small impact on the optimal composition in terms of the proportion of CB phenotypes in the reference population.
This study did not take the Bulmer effect in consideration, which would reduce the gain due to the phenotypic information on CB half-sibs . Hence, the benefit of genotyping CB animals rather than just using their phenotypes will be higher in reality than presented in this paper. Also, the effect on inbreeding will be favourable when applying genomic selection in crossbreeding programs. These advantages have been pointed out by Dekkers . The main focus of this paper was to look at the effect of different proportions of PB and CB animals in the reference population, when applying genomic selection in CB programs, and these would not be much affected by the Bulmer effect.
Based on deterministic simulations, this paper shows the potential benefit of including CB information in crossbreeding programs, using pig breeding as an example. The main aim of the paper was to evaluate this benefit in relation to some key parameters. The actual value for these parameters, and therefore the value of using CB information will be case dependent. Some recent papers [18,19] discussed the use of genomic selection in pig breeding programs and the opportunities and challenges it brings along.
Genomic selection for CB performance can significantly increase rates of genetic gain in crossbreeding programs. The rate of genetic gain increases more when CB animals are included in the reference population compared to a PB reference population. However, we found that the benefit of replacing PB animals with CB animals in the reference population is small, unless the correlation between PB and CB performance is lower than ~0.8 and PB performance is not considered in the breeding objective.
We thank Brian Kinghorn for useful discussion on the topic. We would like to acknowledge the Science and Technology Foundation of the Netherlands Organization for Scientific Research (NWO-STW, grant number 11116), with the co-financers Institute for Pig Genetics BV (IPG BV) and Product Board Animal Feed (PDV) for the financial support of the project. We would also like to acknowledge TOPIGS BV, The Netherlands, for providing knowledge from practice.
- Wei M, van der Werf JHJ. Genetic correlation and heritabilities for purebred and crossbred performance in poultry egg production traits. J Anim Sci. 1995;73:2220–6.PubMedGoogle Scholar
- Wei M, van der Werf JHJ. Maximizing genetic response in crossbreds using both purebred and crossbred information. Anim Prod. 1994;59:401–13.View ArticleGoogle Scholar
- Bijma P, Arendonk JAM. Maximizing genetic gain for the sire line of a crossbreeding scheme utilizing both purebred and crossbred information. Anim Sci. 1998;66:529–42.View ArticleGoogle Scholar
- Daetwyler HD, Villanueva B, Bijma P, Woolliams JA. Inbreeding in genome-wide selection. J Anim Breed Genet. 2007;124:369–76.View ArticlePubMedGoogle Scholar
- Clark SA, Kinghorn BP, Hickey JM, van der Werf JHJ. The effect of genomic information on optimal contribution selection in livestock breeding programs. Genet Sel Evol. 2013;45:44.View ArticlePubMed CentralPubMedGoogle Scholar
- Dekkers JCM. Marker-assisted selection for commercial crossbred performance. J Anim Sci. 2007;85:2104–14.View ArticlePubMedGoogle Scholar
- Meuwissen THE, Hayes BJ, Goddard ME. Prediction of total genetic value using genome-wide dense marker maps. Genetics. 2001;157:1819–29.PubMed CentralPubMedGoogle Scholar
- Daetwyler HD, Villanueva B, Woolliams JA. Accuracy of predicting the genetic risk of disease using a genome-wide approach. PLoS One. 2008;3:e3395.View ArticlePubMed CentralPubMedGoogle Scholar
- Goddard ME. Genomic selection: prediction of accuracy and maximisation of long term response. Genetica. 2009;136:245–57.View ArticlePubMedGoogle Scholar
- Dekkers JCM. Prediction of response to marker‐assisted and genomic selection using selection index theory. J Anim Breed Genet. 2007;124:331–41.View ArticlePubMedGoogle Scholar
- Engblom L, Lundeheim N, del Schneider MP, Dalin AM, Andersson K. Genetics of crossbred sow longevity. Animal. 2009;3:783–90.View ArticlePubMedGoogle Scholar
- Hayes BJ, Visscher PM, Goddard ME. Increased accuracy of artificial selection by using the realized relationship matrix. Genet Res. 2009;91:47–60.View ArticleGoogle Scholar
- Harris BL, Johnson DL. Genomic predictions for New Zealand dairy bulls and integration with national genetic evaluation. J Dairy Sci. 2010;93:1243–52.View ArticlePubMedGoogle Scholar
- Misztal I, Legarra A, Aguilar I. Computing procedures for genetic evaluation including phenotypic, full pedigree, and genomic information. J Dairy Sci. 2009;92:4648–55.View ArticlePubMedGoogle Scholar
- Goddard ME, Hayes BJ, Meuwissen THE. Using the genomic relationship matrix to predict the accuracy of genomic selection. J Anim Breed Genet. 2011;128:409–21.View ArticlePubMedGoogle Scholar
- Meuwissen THE, Hayes BJ, Goddard ME. Accelerating improvement of livestock with genomic selection. Annu Rev Anim Biosci. 2013;1:221–37.View ArticlePubMedGoogle Scholar
- Van Grevenhof EM, Van Arendonk JAM, Bijma P. Response to genomic selection: The Bulmer effect and the potential of genomic selection when the number of phenotypic records is limiting. Genet Sel Evol. 2012;44:26.View ArticlePubMed CentralPubMedGoogle Scholar
- Van Eenennaam AL, Weigel KA, Young AE, Cleveland MA, Dekkers JCM. Applied animal genomics: results from the field. Annu Rev Anim Biosci. 2014;2:105–39.View ArticlePubMedGoogle Scholar
- Ibáñez-Escriche N, Forni S, Noguera JL, Varona L. Genomic information in pig breeding: Science meets industry needs. Livest Sci. 2014;166:94–100.View ArticleGoogle Scholar
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.