Implications
The practical use of genomic information to select animals, or genomic selection, is currently revolutionizing the organization of dairy cattle breeding schemes. The success of this new technology is because of increased genetic progress on both the bull and cow genetic pathways by reducing costs compared with conventional selection schemes, but also by the potential of using this rich source of information to manage genetic resources. Practical results of the implementation of genomic selection in dairy cattle breeding schemes are of great importance for other livestock species in which genomic selection is now under consideration.
Introduction
Use of molecular information to make selection decisions in breeding programmes was envisaged decades ago (Smith, Reference Smith1967; Soller, Reference Soller1978). Although conceptually simple, the implementation of genetic markers into breeding programmes has been rather limited for technological reasons. The recent availability in dense panels of single nucleotide polymorphism (SNP) markers has offered new opportunities to do so. The basic concept underlying such an approach is to use dense genetic markers as proxies to detect genome regions involved in a trait – the quantitative trait loci (QTL) – and exploit them for selection purposes. SNP genotyping technology has enabled profiling many animals for thousands of marker loci in a single analysis, thus with a low cost per marker (Williams, Reference Williams2005). Capitalizing on this rich source of information permitted estimation of breeding values for young candidates with higher accuracy than before. The principle of genomic evaluation models is to take advantage of both genotypic and phenotypic data available in a training (also called ‘reference’) population to build prediction equations of the genetic merit of individuals (Meuwissen et al., Reference Meuwissen, Hayes and Goddard2001). These equations can then be applied to selection candidates having genotypes but no phenotypes. Diverse approaches have been proposed to estimate genomically enhanced breeding values (GEBV), as reviewed by Hayes et al. (Reference Hayes, Bowman, Chamberlain and Goddard2009).
The use of genomic information to make selection decisions, or genomic selection, was shown to greatly increase the technical and economic efficiency of dairy cattle breeding programmes (Schaeffer, Reference Schaeffer2006; König et al., Reference König, Simianer and Willam2009). Indeed, genetic progress was achieved in conventional progeny testing (PT) schemes via the wide use of the very best progeny-tested bulls, which was enabled by means of artificial insemination (AI). Because selection in dairy cattle is undertaken on traits expressed by females, the PT step was necessary to generate a daughter group whose performance was used to predict the genetic merit of bulls with high accuracy. However, PT implies long generation intervals and huge costs related to bull maintenance and progeny-group constitution. Furthermore, only a limited number of young sires can be progeny tested each year owing to structural constraints. The choice of young males was therefore critical, whereas the accuracy of mid-parental Best Linear Unbiased Prediction (BLUP) breeding values was low at a young age. Because genomic selection alleviates some of these costs and technical constraints, the dairy cattle breeding industry has rapidly integrated genomic information into selection programmes. Many efforts have been devoted (so far) to improving the statistical models used for genomic evaluations (Hayes et al., Reference Hayes, Bowman, Chamberlain and Goddard2009). Research has also been undertaken to answer questions arising about the practical use of genomic selection in breeding programmes. Such questions are also likely to emerge in other populations in which application of genomic selection is being considered.
To date, two main approaches have been used to predict the response to genomic selection. Hence, the first objective of this review was to describe these methods, which rely on either infinitesimal or finite locus models (FLMs), and to discuss limitations in their application. The second aim was to summarize the main conclusions reached on the subject by considering the example of dairy cattle populations. Critical factors influencing the efficiency of genomic selection were reviewed, as were future challenges faced in designing breeding schemes. The present review mainly focused on factors affecting the technical efficiency of programmes that was assessed through three criteria, namely (i) the annual genetic gain (ΔG), (ii) the variability of this response and (iii) the impact of selection procedures on maintenance of genetic diversity accessed via inbreeding rates (ΔF).
Modelling the efficiency of genomic breeding programmes
Modelling genomic selection with the selection index theory
The selection index theory was proposed to model the overall gain in accuracy expected from using genomic information at some selection stages (Lande and Thompson, Reference Lande and Thompson1990; Dekkers, Reference Dekkers2007). To do so, SNP information of genotyped individuals is summarized into the direct genomic value (DGV). In practice, DGV is modelled as an infinitesimal indicator trait, which is highly heritable and genetically correlated with the selected trait. This indicator trait is modelled with a heritability of 1, meaning that genotyping errors are ignored and genotyped animals will receive no information from DGV of relatives. In this approach, the genetic correlation considered between the DGV and the selected trait reflects the desired level of accuracy for genomic predictions (Dekkers, Reference Dekkers2007). Assuming the infinitesimal model and multivariate normal distribution of genotypic and phenotypic values, the selection index theory offers a convenient framework to optimally combine both information sources. Because no molecular data are simulated in this approach, the inbreeding rate generated by a scheme must be assessed with analytical formulas in deterministic models or with pedigrees in stochastic simulations. The main appeal of this approximate approach lies in its computational efficiency and straightforward implementation into existing software using either deterministic prediction models or stochastic simulations. It was extended to multi-trait selection by Dekkers (Reference Dekkers2007), and was also referred to as pseudo-genomic selection by Buch et al. (Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a).
Simulations based on FLMs
Another approach entails simulating individuals and their genotypes at a finite number of markers and QTL. This approach allows modelling the dynamics of genetic diversity at neutral markers and genetic variance at QTL by explicitly considering the discrete nature of genotypes and the finite size of genomes (Dekkers et al., Reference Dekkers, Gibson, Bijma and Van Arendonk2004). In FLM simulations, breeding programmes are generally not started from a founder population with genotypes sampled at random. It is recommended to first simulate an ancestral population that has reached equilibrium with respect to mutation and genetic drift to ensure that a realistic structure of linkage disequilibrium (LD) exists between syntenic loci (Muir, Reference Muir2007). Before simulating the breeding scheme, a base population is constituted by sampling individuals from the ancestral population. The true breeding values (TBV) of individuals are obtained as the sum of effects of QTL alleles that they carry on their chromosomes plus a polygenic term, if simulated QTLs are not assumed to explain all of the genetic variance. Following conventional principles of stochastic simulations, phenotypic observations are reconstructed from TBV by adding a residual term sampled from known statistical distributions to achieve the desired heritability (Dekkers et al., Reference Dekkers, Gibson, Bijma and Van Arendonk2004). GEBVs are predicted by performing a genomic evaluation with simulated phenotypes and genotypes as inputs and by using the variance components assumed in the simulation. Hence, the accuracy of genomic predictions is not a prior in the model, but rather a result of simulations.
Finally, inbreeding rates can be estimated in FLM simulations by using either recorded pedigrees or marker information. Although the pedigree-based estimator of inbreeding ignores the effects of linkage and selection around QTL regions, its use is expected to give accurate estimations of ΔF for long genomes, that is, >10 M (Villanueva et al., Reference Villanueva, Pong-Wong, Fernandez and Toro2005). However, the use of pedigree-based inbreeding estimates does not provide information about reductions in genetic diversity around QTL regions under selection (Sonesson et al., Reference Sonesson, Woolliams and Meuwissen2010). An advantage of FLM simulations is that they provide information at simulated markers to monitor the evolution of autozygosity along the genome. Indeed, the proportion of simulated marker loci that is homozygous by descent with respect to the base population gives an expectation of the ‘realized’ inbreeding coefficient.
Limits of application
Pseudo-genomic selection and FLM simulations rely on different genetic models. Given a certain genetic architecture, the FLM simulation is expected to be the most accurate approach because it models the effects of biological processes and selection procedures on both genetic diversity and genetic variance. Although most breeding programme simulations have focused on additive effects, sophisticated architectures can be simulated with QTL exhibiting dominance and epistatic effects. However, the choice of parameters is expected to influence the level and maintenance of genetic variance and diversity (Hu and Li, Reference Hu and Li2006). Hence, it is safe to test the sensitivity of predictions to different architectures when using this prediction method.
Pseudo-genomic selection is an approximate approach that requires a set of assumptions with various consequences on predictions. First, it assumes the infinitesimal model and multivariate normal distribution of DGV and phenotypic values to combine them optimally into a genomic selection index. This supposes that the genetic architecture of selected traits complies with the hypotheses of the infinitesimal model, that is, traits are controlled by a large number of genes with individually small additive effects (Bulmer, Reference Bulmer1980). Research on model species confirmed the relevance of using the infinitesimal model to describe the genetic mean and variance of a population for most quantitative traits because QTL alleles with large effects are rare and most of genetic variation is generally due to many loci with small effects (Mackay et al., 2009). Thus, this model is valid over a short to medium time period, ignoring long-term changes in genetic variance due to fluctuations of allele frequencies. In addition, when dense marker maps are used, DGVs are expected to follow a multivariate normal distribution as a result of the central limit theorem (Lande and Thompson, Reference Lande and Thompson1990).
Second, in pseudo-genomic selection, the accuracy of GEBVs is a prior in the model and assumed to remain constant over time and to be the same for all individuals. In practice, GEBV accuracy is expected to increase depending on the size of the training population. Fixing GEBV accuracies to currently achieved values consequently depicts a rather pessimistic scenario, which is safe when genomic schemes are compared with conventional schemes. However, accuracy of GEBVs was shown to be sensitive to many intermingled parameters comprising the genetic architecture of traits (Daetwyler et al., Reference Daetwyler, Pong-Wong, Villanueva and Woolliams2010), the LD existing between markers and QTL (Goddard, Reference Goddard2009; Goddard et al., Reference Goddard, Hayes and Meuwissen2011), the statistical model of genomic evaluation (Daetwyler et al., Reference Daetwyler, Pong-Wong, Villanueva and Woolliams2010; Bastiaansen et al., Reference Bastiaansen, Costar, Calus, van Arendonk and Bovenhuis2012), the frequency of updating prediction equations (Muir, Reference Muir2007) and the composition of the training population and its relationship with selection candidates (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011; Pszczola et al., Reference Pszczola, Strabel, Mulder and Calus2012). As a consequence, different designs of genomic schemes can lead to different evolutions of GEBV accuracy (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011). Because pseudo-genomic selection ignores the influence of all of these parameters on GEBV accuracy, this approach may lead to biased comparisons of genomic schemes on the basis of predictions of genetic gain.
Finally, computational requirements to achieve calculations also influence the choice of prediction method. Pseudo-genomic selection makes the optimization of breeding schemes more tractable than in FLM simulations, especially when selection is carried out on a breeding objective that includes several genetically correlated traits. However, Meuwissen (Reference Meuwissen2009) showed that it was possible in FLM simulations to scale down by the same factor the effective size of the population, the genome size and the number of training records without modifying the achieved level of GEBV accuracy. This allows comparison of expected genetic gain from different genomic selection schemes at lower computational expense.
Impact of genomic selection on selection response and its variability
Description of investigated genomic selection schemes
Most studies have compared genomic selection schemes with the conventional PT scheme, which was the norm in dairy cattle populations until recently. Genomic selection was integrated into breeding schemes in two different ways, leading to genomic pre-selection (PS) and genomic juvenile (JS) schemes. The PS scheme entails using GEBVs to pre-select young males for PT. All subsequent steps for the selection of males remain the same as in PT schemes. Compared with PT schemes, PS schemes allow increasing the selection accuracy of young male candidates by using their genotypic information. In JS schemes, the use of genomic information is more aggressive. AI sires are selected by using GEBVs among young genotyped males able to produce semen. Although GEBVs of young sires are less accurate than conventional breeding values estimated for progeny-tested bulls, the loss in selection accuracy is compensated by a huge reduction in generation intervals, as there is no need for progeny testing. Genetic gain is then rapidly cumulated over shortened generations. In both genomic schemes, females can be genotyped to increase both accuracy and intensity for the selection of females.
Large variations in predicted ΔG were found across studies, as described by Pryce and Daetwyler (Reference Pryce and Daetwyler2012). Compared with conventional schemes, gains in ΔG ranged from +9% for PS schemes to more than +100% for JS schemes. However, these figures must be interpreted with caution, considering the specific set of parameters used in each study, for example, the heritability of the trait, the accuracy of GEBV and the number of genotypings allocated to males and females. For instance, for a trait with moderate heritability (0.3) – typical of a production trait – setting up a PS scheme increased ΔG by 9% (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011), 13% (Buch et al., Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a) and 16% (Pryce et al., Reference Pryce, Goddard, Raadsma and Hayes2010) compared with the former PT scheme. JS schemes were shown to generate higher selection response, with increases of 29% (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011), 59% (Pryce et al., Reference Pryce, Goddard, Raadsma and Hayes2010), 65% (Buch et al., Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a) and 86% (Colleau et al., Reference Colleau, Fritz, Guillaume, Baur, Dupassieux, Boscher, Journaux, Eggen and Boichard2009) compared with PT schemes.
Increases in ΔG because of genomic selection were the largest when selection was carried out on traits of low heritability, because genomic data added relatively more information to predict breeding values for these traits. Compared with a traditional PT scheme, Lillehammer et al. (Reference Lillehammer, Meuwissen and Sonesson2011) showed that ΔG was increased by 29%, 40% and 70% in JS schemes for heritability values of 0.30, 0.05 and 0.01, respectively. However, levels of GEBV accuracy remained generally moderate to low for traits of low heritability. Phenotypic information of an indicator trait genetically correlated with the selected trait and recorded on a large scale can be integrated in the genomic evaluation model to improve the accuracy of predictions for traits of low heritability when the genetic correlation between both traits is large, that is, >0.5 (Calus and Veerkamp, Reference Calus and Veerkamp2011; Buch et al., Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a).
GEBV reliabilities larger than 0.5 to 0.6 are already achieved in homogeneous dairy cattle populations with large reference populations (e.g. Holstein populations) for traits with moderate to high heritability (Harris and Johnson, Reference Harris and Johnson2010; Su et al., Reference Su, Guldbrandtsen, Gregersen and Lund2010). By contrast, reliabilities of at least 0.8 are obtained for breeding values estimated for progeny-tested bulls. In breeds with large effective sizes or small reference populations, reliability of GEBVs is generally lower (<0.5) even for production traits.
More emphasis placed on bull dam selection
In schemes integrating genomic selection, an important part of genetic gain is achieved through the selection of bull dams (Schaeffer, Reference Schaeffer2006). Indeed, the reliability of breeding values is strongly increased for genotyped females to a level comparable with that achieved for genotyped males. Furthermore, GEBVs of bull dams are expected to be less biased, because genomic information reduces the weight attributed to own performance, which may be subject to preferential treatment. With a training population of sufficient size, large gains are expected from a more accurate selection of breeding cows, even when only a small proportion is genotyped (Sørensen and Sørensen, Reference Sørensen and Sørensen2009). Therefore, it is relevant to define strategies to optimally allocate genotyping capacities between young males and females in order to maximize genetic gain. Sørensen and Sørensen (Reference Sørensen and Sørensen2009) showed that allocating larger proportions of genotypings to females than males resulted in larger selection responses. Indeed, the proportion of males kept for reproduction is small in dairy cattle populations. If extra genotyping capacity is spent on males, the second best males according to parent average will be genotyped, although they have a lower chance of being selected. On the contrary, if extra genotyping capacity is spent on females, there are greater chances of detecting interesting females among the cow population to generate high genetic merit bull calves.
Influence of reference population constitution on genomic prediction accuracies
The need for large reference populations to increase GEBV accuracy has encouraged collaboration between breeding organizations and countries to exchange genotypic data (Lund et al., Reference Lund, de Roos, de Vries, Druet, Ducrocq, Fritz, Guillaume, Guldbrandtsen, Liu, Reents, Schrooten, Seefried and Su2011; Wiggans et al., Reference Wiggans, VanRaden and Cooper2011). Collaboration has quickly appeared to be one of the cheapest ways to achieve higher GEBV reliabilities, although competitive interests often impeded the process (VanRaden et al., Reference VanRaden, Wiggans, van Tassel, Sonstegaard and Schenkel2009). In such collaborations, all genotypes of selected and unselected candidates have to be shared to avoid biases in the estimation of genomic breeding values, at both the national and international scales (Patry et al., Reference Patry, Jorjani and Ducrocq2011).
Constitution of the reference population also influences the level of GEBV accuracy (Habier et al., Reference Habier, Fernando and Dekkers2007) and its persistence over time (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011). Hence, breeding programmes should be designed to minimize the average relationship within the reference population and maximize relationships between candidates and the reference population (Pszczola et al., Reference Pszczola, Strabel, Mulder and Calus2012). On the basis of FLM simulations of scenarios with reference populations of equivalent sizes, Lillehammer et al. (Reference Lillehammer, Meuwissen and Sonesson2011) showed that PS and JS schemes led to different levels of GEBV accuracy. Indeed, in PS schemes, all sires of male candidates have been progeny tested and are included in the reference population. This implies larger genetic ties between candidates and reference bulls than in JS schemes, in which only grandsires of candidates and older ancestors are included. Reducing the number of progeny-tested bulls in PS schemes led to a decrease in GEBV accuracy because fewer bulls with daughter information could be added each year to the reference population. However, Lillehammer et al. (Reference Lillehammer, Meuwissen and Sonesson2011) showed that the resulting loss in ΔG could be compensated by increasing the progeny group size. Thus, in PS schemes, maintaining maximal testing capacity is required to ensure accurate EBVs for bulls in the training population, and subsequently, accurate GEBVs for candidates, especially for traits of low heritability.
Including cows in the reference population has recently been suggested to further increase the reliability of genomic predictions because the number of AI sires is generally limited in dairy cattle populations (McHugh et al., Reference McHugh, Meuwissen, Cromie and Sonesson2011; Buch et al., Reference Buch, Kargo, Berg, Lassen and Sørensen2012b). Such a strategy is gaining some interest given the increasing popularity of genomic selection in the industry and the decreasing costs of genotyping (Strauss, Reference Strauss2010). Actually, the recent release of cheaper SNP platforms is likely to extend the use of genomic selection to a wider female population. Use of imputation techniques makes it feasible to use sparser SNP chips to estimate GEBVs with only moderate losses in prediction accuracy (Weigel et al., Reference Weigel, de Los Campos, Vazquez, Rosa, Gianola and Van Tassel2010; Dassonneville et al., Reference Dassonneville, Brøndum, Druet, Fritz, Guillaume, Guldbrandtsen, Lund, Ducrocq and Su2011). However, before practical inclusion of female information in reference populations, the issue of preferential treatment, which frequently occurs with elite cows, must be tackled because it can introduce bias in the estimation of SNP effects, potentially decreasing the accuracy of genomic predictions.
Impact on variability of selection response
In PT schemes, the risk of using a bull with poor genetic merit is minimized because of PT. In PS schemes, this risk is further reduced owing to the genomic pre-selection step, which allows distinguishing with higher accuracy the best young males within families before PT (Buch et al., Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a). In JS schemes, the accuracy of breeding values used to select AI sires is lower than that obtained after PT. Considering a fixed number of AI sires selected each year, the variability of selection response obtained by implementing a JS scheme is increased relative to conventional schemes (Buch et al., Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a). Thus, in practice, increasing the number of AI sires in JS schemes is recommended to mitigate the risk of using bulls with poor genetic merit. Another recommendation is to use young sires in ‘teams’ because the reliability of the mean GEBV of a group of bulls increases with group size. Schefers and Weigel (Reference Schefers and Weigel2012) reported that a team of five young bulls, each having a GEBV with a reliability of 0.70, will have an average GEBV with a reliability of 0.94.
New opportunities to monitor and preserve genetic diversity
Impact of genomic selection on inbreeding rates
Exploiting genomic information enables estimation of the Mendelian sampling term of young individuals without any phenotypic information. Therefore, genomic selection is expected to reduce the weight of family information in selection decisions by placing the emphasis on Mendelian sampling information of young candidates (Daetwyler et al., Reference Daetwyler, Villanueva, Bijma and Woolliams2007). The largest reductions in inbreeding rates due to the use of genomic selection were observed for traits of low heritability (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011) and when a large part of variance was explained by markers (de Roos et al., Reference de Roos, Schrooten, Veerkamp and Van Arendonk2011). By screening a large population of candidates, genomic selection also facilitates the identification of the least related animals having high genetic merit with a higher accuracy than before. In particular, McHugh et al. (Reference McHugh, Meuwissen, Cromie and Sonesson2011) showed that genotyping a large number of females had a very beneficial impact on ΔF reduction.
The use of genomic selection to pre-select males for PT (i.e. PS scheme) resulted in a clear reduction of per generation inbreeding rates compared with PT schemes, for only slight modifications of the generation interval (Pryce et al., Reference Pryce, Goddard, Raadsma and Hayes2010; de Roos et al., Reference de Roos, Schrooten, Veerkamp and Van Arendonk2011; Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011; Buch et al., Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a). Setting up JS schemes also generally led to reductions in per generation ΔF compared with PT schemes. However, because the generation interval was reduced as well, annual ΔF may not be decreased compared with conventional schemes (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011; Buch et al., Reference Buch, Sørensen, Berg, Pedersen and Sørensen2012a). Increasing the number of young AI sires used in JS schemes is an option to curb inbreeding rates without markedly reducing genetic gain (Lillehammer et al., Reference Lillehammer, Meuwissen and Sonesson2011).
Differentiating the number of matings per young bulls on the basis of GEBVs is not a sustainable option; it leads to a slight increase in genetic gain at the expense of a drastic increase in ΔF (Sørensen and Sørensen, Reference Sørensen and Sørensen2009). Finally, schemes with mixed use of proven and young genomic bulls were studied by Colleau et al. (Reference Colleau, Fritz, Guillaume, Baur, Dupassieux, Boscher, Journaux, Eggen and Boichard2009) to mimic the situation in which some breeders did not accept the risk of using young AI bulls having a GEBV of moderate reliability. The authors showed that allocating 50% of inseminations in a JS scheme to the 20 best bulls having milking daughters led to a small increase in genetic gain and a large increase in inbreeding rate. Thus, this practice was harmful for genetic diversity unless strict rules were defined for the management of genetic resources.
The conclusions about the impact of genomic selection schemes on genetic diversity critically depend on the time unit chosen to express inbreeding rates. In practice, the inbreeding rate should be kept under 1% per generation in conservation and breeding schemes to avoid undesirable effects of inbreeding on fitness (Food and Agriculture Organization (FAO), 1998). This recommendation indicates that an effective population size of at least 50 individuals is required so that factors such as natural selection, recombination and mutation – which intervene at every generation – can counterbalance effects of inbreeding depression (FAO, 1998). When the selected trait and fitness were not genetically correlated, Meuwissen and Woolliams (Reference Meuwissen and Woolliams1994) estimated that a critical effective size of 30 to 250 individuals was required in a selection scheme so that natural selection could overcome the effects of inbreeding depression on fitness, which is consistent with FAO recommendations. However, when the selected trait and fitness were negatively correlated, even weakly (−0.2), Meuwissen and Woolliams (Reference Meuwissen and Woolliams1994) showed that it was almost impossible to prevent a decline in fitness. In such situations, shortening the generation interval makes the accumulation of inbreeding depression effects bigger. Thus, in our opinion, it is safe to compare breeding schemes by using annual inbreeding rates. In addition, the inbreeding rate can be considered a measure of the risk of a scheme. Therefore, it seems relevant to express it on the same time scale as genetic gain, that is, on an annual basis.
Although the design of breeding schemes is important to limit the rise in inbreeding in a population, there is also a need to develop efficient recording systems that are able to detect emerging genetic defects (Agerholm et al., Reference Agerholm, Bendixen, Andersen and Arnbjerg2001). In particular, when the generation interval is shortened, it is essential to take rapid measures to control the spreading of the defect by restricting the use of bulls carrying it. The example of Holstein BLAD, CVM and Brachyspina genetic defects illustrated how fast a lethal mutation can spread in a population of small effective size.
Including genomic information in the optimal contribution selection (OCS) method
In the near future, JS schemes will likely be preferred to other designs because they generate larger genetic gain at moderate inbreeding rates, if practical recommendations for the use of young AI sires are followed. However, it may become difficult for breeding managers to select and mate animals while monitoring their relationships, as many more AI bulls are used and for a shorter time period. It is therefore important to develop and promote the use of tools to control inbreeding rates. The standard approach used to determine contributions of breeding animals, maximizing genetic gain at a desired rate of inbreeding, is referred to as OCS (Meuwissen, Reference Meuwissen1997). In OCS, the contribution of a parent is the result of a trade-off between its genetic merit and its relationship to other individuals (Fernández et al., Reference Fernández, Meuwissen, Toro and Mäki-Tanila2011). Recently, Sonesson et al. (Reference Sonesson, Woolliams and Meuwissen2010) showed with FLM simulations that OCS was not able to maintain genetic diversity across the whole genome in selected populations when relationship coefficients were estimated from pedigree information. In fact, it led to a strong reduction in diversity around QTL regions by favouring alleles with the largest effects. To circumvent this flaw, these authors proposed using marker-based relationships, because they more accurately reflect the genome sharing between individuals than pedigree-based expectations. Genomic Optimal Contribution Selection (GOCS) enabled a better control of genomic inbreeding across the genome. However, it also strongly reduced genetic gain. The study design may have influenced the results; in particular, the number of simulated QTL was relatively limited, which could have implied a strong effect of selection on QTL regions. Nevertheless, this result questions whether large ΔG can be achieved when an acceptable constraint on ΔF is considered. It also emphasizes the need to determine an acceptable rate of inbreeding in practical breeding programmes by integrating recent knowledge gained from genomic data. In addition, the ability of GOCS to maintain expected and observed heterozygosity in a conservation programme was shown to be sensitive to the estimator used to access genomic relationships (de Cara et al., Reference de Cara, Fernandez, Toro and Villanueva2011). This stresses the importance of identifying accurate marker-based relationship estimators to maximize the efficiency of GOCS in practical applications.
Finally, after having inferred the relationship between chromosome segments of individuals, one can envisage strategies to maintain genetic diversity at the chromosome level, for instance, to prioritize preservation in genome regions exhibiting acute reduction in diversity (Engelsma et al., Reference Engelsma, Veerkamp, Calus and Windig2011) or to maintain breed-specific genetic variation in a hybrid population. It would be interesting to assess whether benefits of these approaches are large relative to GOCS, which uses genome-wide relationship estimates.
Impact of using genomic selection with reproductive biotechnologies
One of the main advantages of genomic selection is to capitalize on SNP information to discriminate within a family the individuals with the largest Mendelian sampling terms. Hence, the greatest benefits of genomic selection are expected when selection is carried out in large full-sib families. Thus, in dairy cattle, there is enormous interest in increasing the number of offspring obtained from elite bull dams by means of reproductive biotechnologies such as multiple ovulation and embryo transfer (MOET) or ovum pick-up followed by in vitro fertilization. The advantages of using genomic selection together with MOET are twofold; it enables the discrimination of the best heifers to be flushed and the early differentiation of the best prospective males obtained from flushings (Sørensen and Sørensen, Reference Sørensen and Sørensen2009; Pedersen et al., Reference Pedersen, Kargo, Berg, Voergaard, Buch and Sørensen2012). However, Pryce et al. (Reference Pryce, Goddard, Raadsma and Hayes2010) showed that an aggressive use of such techniques in a closed nucleus could lead to large increases in inbreeding rates. Exploiting a closed rather than an open MOET scheme was advocated to be more cost-effective and safer with respect to sanitary status, because no heifers have to be purchased from the outside (Pryce et al., Reference Pryce, Goddard, Raadsma and Hayes2010). However, keeping the MOET nucleus open is generally recommended to maintain inbreeding rates at a reasonable level, especially for small-sized nuclei (Korpiaho et al., Reference Korpiaho, Strandén and Mäntysaari2003).
A practical limitation in the use of embryo transfer technologies is the lack of large numbers of recipient cows (Humblot et al., Reference Humblot, Le Bourhis, Fritz, Colleau, Gonzalez, Guyader-Joly, Malafosse, Heyman, Amigues, Tissier and Ponsart2010). Having GEBV for embryos could alleviate this limitation by discriminating the embryos with desirable genotypes before implantation. Colleau et al. (Reference Colleau, Fritz, Ponsart, Le Bourhis, Lacaze, Tissier, Mervant, Amigues, Druet, Malafosse, Boichard and Humblot2008) assessed interest in genotyping embryos for the French marker-assisted selection scheme under a fixed cost constraint. Their results indicated that benefits of embryo genotyping were partially offset by the poor quality of genotypic data from small biopsies of embryos. However, in their study, embryos were genotyped for microsatellite markers. SNP technologies have proven to be more robust to genotyping errors and less demanding in terms of DNA quality and quantity than microsatellite technologies. Promising results were reported by Humblot et al. (Reference Humblot, Le Bourhis, Fritz, Colleau, Gonzalez, Guyader-Joly, Malafosse, Heyman, Amigues, Tissier and Ponsart2010) in terms of genotyping calling rates and pregnancy rates obtained from biopsied embryos under experimental conditions. If such results can be extrapolated to field conditions, embryo genotyping might facilitate exploitation of increased prolificacy of flushed females and reduce the costs related to maintenance of recipient cows in breeding nuclei. This could give rise to the ‘velogenetics schemes’, which were conceptualized 20 years ago by Georges and Massey (Reference Georges and Massey1991).
Finally, the use of sexed semen was shown to be able to increase genetic gain in conventional dairy cattle breeding schemes (Abdel-Azim and Schnell, Reference Abdel-Azim and Schnell2007). However, most benefits were achieved in commercial herds by increasing selection intensity and reducing generation intervals on the cow-to-cow pathway. By simulating a great range of dairy cattle breeding schemes integrating genomic selection, Pedersen et al. (Reference Pedersen, Kargo, Berg, Voergaard, Buch and Sørensen2012) confirmed that benefits of using sorted semen were slight or null relative to those of MOET in terms of genetic gain and preservation of genetic diversity. The most worthwhile scenario was obtained when Y-sorted semen and MOET were used on elite bull dams. No increase in genetic gain was observed, but the negative effects on inbreeding rates usually associated with MOET were partially offset because of increased within-family selection among larger full-sib families (Pedersen et al., Reference Pedersen, Kargo, Berg, Voergaard, Buch and Sørensen2012).
Defining selection objectives and integrating new selection criteria
Compared with conventional schemes, genomic selection has the potential of improving the balance of achieved genetic gain between production, conformation and functional traits. Accordingly, achieved genetic gain should better reflect economic weights considered in the total merit index (Boichard et al., Reference Boichard, Ducrocq, Fritz and Colleau2010). However, a large difference in genetic gain is still expected between highly and lowly heritable traits if the latter receive insufficient weight in the selection objective (König and Swalve, Reference König and Swalve2009). This phenomenon can be strengthened if antagonistic genetic correlations exist between functional and production traits (Ytournel et al., Reference Ytournel, Willam and Simianer2011).
Although more complex objectives (e.g. environmental efficiency and animal welfare) are promoted by society for inclusion in breeding programmes, genomic selection offers an interesting opportunity to add new selection criteria to breeding objectives (Boichard and Brochard, Reference Boichard and Brochard2012). Indeed, selecting new phenotypes seems to be more feasible, especially for traits that cannot be recorded on a large scale because they are difficult or expensive to measure. These new phenotypes are mainly related to health (lameness, disease resistance, metabolic and reproductive disorders), feed efficiency, milk composition or adaptation to different environments (Boichard and Brochard, Reference Boichard and Brochard2012). Experimental herds, MOET herds or test stations could be exploited to record such phenotypes. The quantity and quality of phenotypic data are critical parameters for accurate genomic predictions, and hence for achieving genetic gain in the whole population. Veerkamp et al. (Reference Veerkamp, Berry, Wall, de Haas, McParland, Coffey and Calus2011) showed the feasibility of combining data sets from several experimental herds to increase the accuracy of genomic predictions for a new trait. Even if herds displayed patent differences in recording and management, genomic predictions of reasonable accuracy were obtained. For new traits that can be recorded on a larger scale, Calus et al. (Reference Calus, de Haas, Pszczola and Veerkamp2011) estimated that selection accuracy could be greatly increased by including directly genotyped females in the training population. These studies suggested that breeding companies should contract out herds to expand the recording of new traits, but the issue of recording costs must also be considered.
World-wide use of genomic selection in dairy cattle breeding schemes
To date, genomic selection has only been implemented in a few countries and mainly in connection with breeding programmes of Holstein cattle (Pryce and Daetwyler, Reference Pryce and Daetwyler2012). This breed was a good case study for developing genomic selection because it has been intensively selected for decades, hence strengthening statistical associations between markers and QTL. Furthermore, many large breeding programmes have been set up world wide, facilitating the constitution of large reference populations. Thus, international collaborations to exchange genotypes allowed further improving the accuracy of genomic predictions (Lund et al., Reference Lund, de Roos, de Vries, Druet, Ducrocq, Fritz, Guillaume, Guldbrandtsen, Liu, Reents, Schrooten, Seefried and Su2011; Wiggans et al., Reference Wiggans, VanRaden and Cooper2011).
The first genomic evaluations were officially released for a few Holstein populations in 2009. Different strategies have since been adopted to integrate genomic selection into existing breeding programmes. Focusing on the Holstein breed, Pryce and Daetwyler (Reference Pryce and Daetwyler2012) summarized the strategies adopted in eight countries by April 2011. Advancements have been rapid and a few countries – such as Italy and the United Kingdom – have now released GEBV for large populations of Holstein bulls.
Utilization of genomic selection has accelerated among Holstein breeders (Strauss, Reference Strauss2010). For example, in France, genomically tested bulls without milking daughters represented 39% of market shares in 2011 (P. Le Mezec, personal communication). In most countries, PT continued to be carried out during this transition period (Pryce and Daetwyler, Reference Pryce and Daetwyler2012). France was an exception, with PT officially ceasing in 2009 (Fritz et al., Reference Fritz, Guillaume, Croiseau, Baur, Hozé, Dassonneville, Boscher, Journaux, Boichard and Ducrocq2010; Pryce and Daetwyler, Reference Pryce and Daetwyler2012). However, the number of AI sires used each year was increased, their diffusion was constrained to a few thousand straws per sire and farmers were encouraged to use teams of at least five young bulls (Fritz et al., Reference Fritz, Guillaume, Croiseau, Baur, Hozé, Dassonneville, Boscher, Journaux, Boichard and Ducrocq2010).
In other dairy cattle breeds, several difficulties have impeded the integration of genomic selection into breeding programmes. First, breeding programmes are generally of smaller size than Holstein programmes, which makes it more difficult to gather large reference populations. Several initiatives have, however, been created to exchange genotypes across countries, for example, the InterGenomics consortium for the Brown Swiss breed (Jorjani et al., Reference Jorjani, Zumbach, Dürr and Santus2011) and the collaboration between breeding schemes of the Nordic Red Dairy cattle in Denmark, Sweden, Finland and Norway. Second, some of these breeds may have larger effective population sizes, resulting in weaker associations between markers and QTL. Both of these factors tend to reduce the accuracy of genomic predictions, which are generally lower than in Holstein populations. The recent use of high-density SNP panels (e.g. Illumina Bovine HD (777K)) was a promising option to increase GEBV accuracy in such populations. Having large densities in SNPs reduces the physical distance between markers and QTL, and hence should strengthen the statistical association between them. At such marker density, associations between markers and QTL may be also maintained across breeds, making it possible to build across-breed prediction equations and to capitalize on reference populations of several breeds. However, preliminary analyses of high-density chips with the genomic BLUP evaluation model only resulted in marginal gains within breeds (Su et al., Reference Su, Brondum, Ma, Guldbrandtsen, Aamand and Lund2011) and across breeds (Harris et al., Reference Harris, Creagh, Winkelman and Johnson2011).
Despite these difficulties, genomic selection has been integrated into the breeding programmes of a few large dairy cattle populations other than Holstein. In France, the first official GEBVs were released in 2009 for the Montbéliarde and Normande breeds (Fritz et al., Reference Fritz, Guillaume, Croiseau, Baur, Hozé, Dassonneville, Boscher, Journaux, Boichard and Ducrocq2010). Lower GEBV reliabilities were achieved in these two breeds than in Holstein because of the smaller size of their reference populations, comprising 1500 bulls for the Montbéliarde breed and 1250 bulls for the Normande breed in June 2010 (Fritz et al., Reference Fritz, Guillaume, Croiseau, Baur, Hozé, Dassonneville, Boscher, Journaux, Boichard and Ducrocq2010). The PT of bulls was discontinued in 2009, and rules were defined for the commercialization of genomic sires (Fritz et al., Reference Fritz, Guillaume, Croiseau, Baur, Hozé, Dassonneville, Boscher, Journaux, Boichard and Ducrocq2010). The uptake of genomic selection was lower in these breeds than in the Holstein breed, with 21% of market shares achieved for young Normande genomic sires in 2011 (P. Le Mezec, personal communication).
Genomic selection has also recently been adopted into the breeding programme of Red dairy cattle breeds in Denmark, Sweden and Finland. Officially released in 2011, genomic breeding values displayed moderate reliability, given the size of the reference population, which included 4400 bulls in 2011 (Su et al., Reference Su, Brondum, Ma, Guldbrandtsen, Aamand and Lund2011). Indeed, effects of recurrent genetic admixture have weakened associations existing between SNPs and QTL in this composite population, hampering the accuracy of genomic predictions (Brøndum et al., Reference Brøndum, Rius-Vilarrasa, Strandén, Su, Gulbrandtsen, Fikse and Lund2011). Reliabilities of genomic breeding values are nevertheless comparable with those achieved for the French Normande and Montbéliarde breeds (Fritz et al., Reference Fritz, Guillaume, Croiseau, Baur, Hozé, Dassonneville, Boscher, Journaux, Boichard and Ducrocq2010). Given the moderate GEBV reliabilities, prudent use of genomic selection was chosen. Hence, the PT of bulls did not cease during this transition period, and the semen of the best 20 to 25 young sires was commercialized at the age of 2 years.
Conclusions
Genomic selection has been rapidly adopted by the dairy cattle breeding industry, particularly in the Holstein breed, and its use in other cattle breeds and livestock species is under investigation (Ibanez-Escriche and Gonzalez-Recio, Reference Ibanez-Escriche and Gonzalez-Recio2011). In this transition period, strategies to implement genomic selection have been rather prudent, although the largest gains could be obtained by extensive use of genomic information. Success of this new selection method critically depends on the availability of large databases of phenotypes and genotypes. Hence, breeding organizations will have to work more closely with recording organizations to facilitate the large-scale recording of currently selected traits and to extend the recording of new traits. The link between breeding companies and breeders might also be strengthened to secure access to elite breeding animals.
Genomic selection also appears to be a promising tool to curb inbreeding rates because it allows exploitation of within-family genetic variance from the earliest selection stages. However, to be effective, breeders must consider using a diversified panel of genomically evaluated bulls, both to mitigate the risk of using bulls with poor merit and to limit the erosion of genetic diversity. To this end, the further development and application of tools, such as genomic optimal contribution selection, are critical for sustainable use of genetic resources.
Acknowledgement
This work was funded by FABA Oy (Vantaa, Finland) through TEKES grant no. 433/31/09.
