Published November 25, 2014

Genomic selection for two traits in a maternal pig breeding scheme1 M. Lillehammer,*2 T. H. E. Meuwissen,† and A. K. Sonesson* *Nofima AS, N-1430 Ås, Norway and †Department of Animal and Aquacultural Sciences, Norwegian University of Life Sciences, N-1430 Ås, Norway

ABSTRACT: The objective of this study was to compare different implementations of genomic selection to a conventional maternal pig breeding scheme, when selection was based partly on production traits and partly on maternal traits. A nucleus pig breeding population with size and structure similar to Norwegian Landrace was simulated where equal weight was used for maternal and production traits. To genotype the boars at the boar station and base the final selection of boars on genomic breeding values increased total genetic gain by 13% and reduced the rate of inbreeding by 40%, without significantly affecting the relative contribution of each trait to total genetic gain. To increase the size of the reference population and thereby accuracy of selection, female sibs in the selected litters can also be genotyped to increase genetic gain for maternal traits more than for production traits, thereby resulting in an increased relative contribution of maternal traits to total genetic gain. Genotyping 2,400 females each year increased the rela-

tive contribution of maternal traits to total genetic gain from 16 to 32%. Performing preselection of males by allowing genotyping of 2 males per litter and allowing for selection across and within litters before the boar test increased genetic gain by 5 to 11%, compared with genotyping the boars at the boar station, without significant effects on the relative contribution of each trait to total genetic gain. Genotyping more animals consequently increased genetic gain. Genotyping females to build a larger reference base for maternal traits gave similar genetic gain as genotyping the same amount of additional males but with a lower rate of inbreeding and a greater contribution of maternal traits to total genetic gain. In conclusion, genotyping females should be prioritized before genotyping more males than the tested boars if the breeding goal is to increase maternal traits specifically over production traits or genomic selection is used as a tool to reduce the rate of inbreeding.

Key words: breeding design, genomic selection, multitrait selection, pig © 2013 American Society of Animal Science. All rights reserved. Introduction Pig breeding is usually based on specialized maternal and paternal breeds or lines (Visscher et al., 2000). For maternal breeds, considerable weight in the breeding goal is put on maternal traits, such as litter size, litter weight, and female reproduction. These traits are, however, hard to improve because of low heritability 1The authors thank Dan Olsen and Ina Andersen-Ranberg in Norsvin for valuable input about population structure and variance and breeding goal parameters for Norwegian Landrace. This project, led by Norsvin, was funded by the Norwegian Research Council, project number 186862, and the breeding companies Norsvin, Geno, and AquaGen AS. Calculations were done on the TITAN computer cluster at University of Oslo, Norway 2Corresponding author: [email protected] Received January 11, 2012. Accepted April 3, 2013.

J. Anim. Sci. 2013.91:3079–3087 doi:10.2527/jas2012-5113

and because no information of the traits is available on either sex at the time of selection and no information on maternal traits is available on the male selection candidates until their daughters start producing litters. Production traits, which also have considerable weight in the breeding goal, have greater heritability as well as more information of the traits on male selection candidates available at the time of selection. Selection for production traits is therefore much more effective than selection for maternal traits under a conventional breeding program. In Norwegian Landrace, negative correlations have been found between maternal traits and production traits (Holm et al., 2004), which makes the maternal traits even harder to improve through a conventional breeding program. Lillehammer et al. (2011) showed that use of genomic selection could considerably increase accuracy

3079

3080

Lillehammer et al.

of the breeding values of young boars for maternal traits. However, in that study it was assumed that selection was on maternal traits only. It was therefore not possible to quantify the relative increase in maternal traits to other traits and estimate the impact of genomic selection on total genetic gain. The objective of this study was to compare different implementations of genomic selection to a conventional maternal pig breeding scheme when selection was jointly for production and maternal traits. The basic scenario contained equal weight on the 2 groups of traits, similar to what is used for Norwegian Landrace, where these production and maternal traits together define most of the breeding goal (http://www.norsvin.no/). The schemes were compared on total genetic gain and on the relative contribution of maternal traits to the overall genetic gain. The comparison was based on stochastic simulation of a pig population with size and structure similar to Norwegian Landrace. In Norway, crosses between Norwegian Landrace and Yorkshire are used as dams, and sires of the slaughter pigs are Duroc pigs. Materials and Methods Animal Care and Use Committee approval was not obtained for this study because no animals were used. Historical Population A historical pig population was simulated by random mating of 200 individuals for 2,000 generations according to the Fisher-Wright population model (Fisher, 1930; Wright, 1931) to create mutation-drift balance. The genome consisted of 18 pairs of 100 cM chromosomes. During these generations, polymorphisms and recombinations were sampled at random positions as in Sonesson and Meuwissen (2009), such that each mutation occurred at a unique position and created a biallelic SNP. After 2,000 generations, random selection and mating was used to create 600 females and 25 males, which served as the base generation for the selection schemes. Within each chromosome, 100 SNP among the SNP with a minor allele frequency above 0.05, measured in the last of the 2,000 historical generations, were randomly assigned as QTL. From the remaining SNP, the 500 SNP with the greatest minor allele frequencies from each chromosome were assigned as markers. True and Estimated Breeding Values Two traits were simulated, 1 production trait (PROD) and 1 maternal trait (MAT). Each of these traits represented an index of traits that could be measured on either all candidates before 1 yr of age (PROD) or on females after the first litter (MAT), ignoring traits measured later

in life. The QTL effects were sampled from a multivariate normal distribution, assuming that the QTL effects were normally distributed, with mean 0 and variance  σ2 gMAT  V = 1 / 1800 ×   σgMAT,gPROD

σ



gMAT,gPROD 

σ2

gPROD

 = G / 1800 , 

in which σ2gMAT is the genetic variance of MAT, σ2gPROD is the genetic variance of PROD, and σgMAT,gPROD is the genetic covariance between the 2 traits, adjusted to give a genetic correlation between MAT and PROD of either –0.3 or 0. The total number of QTL is 1,800. True breeding values for MAT (TBVMAT) and PROD (TBVPROD) were calculated as the sum of the QTL effects for each trait for each individual. The records (yi = [yiMAT yiPROD]′) of individual i were simulated as yi = TBVi + ei, where TBV is a vector of TBVMAT and TBVPROD for individual i, and ei is a random deviate sampled from a multivariate normal distribution N(0,R) with  σ2 eMAT  R=  σeMAT,ePROD

σ



eMAT,ePROD  ,  σ2  ePROD

in which σ2eMAT and σ2ePROD are the residual variances of MAT and PROD, respectively, adjusted to give heritability of 0.1 (MAT) or 0.3 (PROD), and σeMAT,ePROD is the covariance between the traits, adjusted to give the same residual correlation between the traits as the genetic correlation, assumed to be either –0.3 or 0. Other scenarios tested the sensitivity of the results to variations in heritability by adjusting σ2ePROD, and σeMAT,ePROD to make the heritability of PROD (h2PROD) 0.2 or 0.4. Daughter yield deviations (DYD) were simulated for selected sires, assuming each sire produced 40 effective half sib daughters with 1 record each. The DYD were simulated using the formula

DYDi = 1/2TBVi + ri, in which ri = [riMAT riPROD]′ is a standard normal random deviate, sampled from

MVN ~ (0,(0.75G+R)/40). Ungenotyped animals were assigned BLUP breeding values for both traits, estimated using the multitrait animal model yi = μ + ui + ei (Henderson, 1984), in which yi was a vector of phenotypic records for individual i or 2 × DYDi if individual i was a sire with 40 offspring, μ was a vector of the overall means for the 2 traits, ui was a vector containing the breeding values of

Genomic selection for 2 traits in pigs

animal i, with variance G, and ei was a vector of environmental effects with variance  σ2 eMAT  R=  σeMAT,ePROD

σ



eMAT,ePROD  ,  σ2  ePROD

for individual records or (3G + 4R)/40 for DYD. For all implementations of genomic selection, the animals in the base generation were assumed to have known marker genotypes and DYD. This should reflect a situation where older boars are genotyped to create a reference population before starting a genomic selection breeding scheme and all females in the base population were genotyped. All new animals with known marker genotype and phenotype or DYD were added to the reference population every selection round. The reference population was used to estimate marker effects for each trait for every round of selection, using the BLUP method of Meuwissen et al. (2001) with the single trait statistical model y =µ + ij

j

9000

∑

k =1

X a +e , ik kj

ij

in which yij was the DYD or phenotypic record for individual i and trait j, eij and μj were defined as above, Xik was the marker genotype, and akj was the random effect of the kth marker on trait j, with variance equal to the total genetic variance for trait j divided by 9,000 (the total number of SNP markers). This difference in variances of eij was accounted for by weighting the records and using the inverse of these variances as weights [i.e., the weights were, after multiplication with the constant σe2, 1, and 40/(3/λj + 4)] for animals with own phenotype and boars with offspring, respectively, in which λj = σe2j/σg2j. The linkage disequilibrium in the simulated data was rather high (r2 = 0.452; Lillehammer et al., 2011). This seems to correspond to markers less than 100 kb apart in real comparable pig populations (Du et al., 2007; Badke et al., 2012). For nongenotyped animals, a conventional total merit index (C-TMI), which was a sum of the conventional BLUP breeding values for the 2 traits, was used as selection criterion. For genotyped animals, a genomic total merit index (G-TMI) was estimated as the sum of the genomic breeding values (GEBV) for the 2 traits and used as selection criterion. In both cases, equal weights on the 2 traits were used. To be able to compare conventional BLUP breeding values and genomic breeding values in schemes where only a fraction of the females were genotyped, both types of breeding values were res-

3081

caled so that the regression of TBV on EBV was equal to 1, using the formula

(G)EBVs = TBV + b ((G)EBV - (G)EBV ) , in which (G)EBV denotes genomic or conventional EBV, subscript s denotes after scaling, a bar denotes the average, and b is the regression of TBV on (G)EBV, that is, b = Cov[TBV,(G)EBV]/Var[(G)EBV] (Lillehammer et al., 2011). Here, b was assumed known but needs to be estimated when using real data. Population under Selection For all selection schemes, each year consisted of 4 time steps. Each time step, the 300 dams with the greatest breeding values, among those eligible for selection, produced litters further referred to as “selected litters” through random hierarchical mating with the 25 greatest indexing males. It was assumed that all dams had phenotypic records for MAT and PROD before this selection by simulating MAT and PROD records for all dams and then offspring for the greatest ranking dams only. This was done to reflect a situation where all females get a litter before some litters are selected. All male and female selection candidates came from the selected litters. From each selected litter, 2 females were picked randomly and used to replace culled females in the population, assuming that no individual information about the piglets within a litter was available. The selected females got PROD records when they were 3 time steps old (9 mo) and got MAT records when they were 1 yr (4 time steps) old. At 1 yr of age, the greatest indexing females also got their first litter. They remained in the population until the age of 2 yr old and were eligible for selection every second time step within that period, so dams of selected litters were chosen among 3 age classes of females (i.e., females that were 1.0, 1.5, or 2.0 yr old). Twenty-five percent of the females were randomly culled after their first litter and 25% after their second litter, to reflect culling for other reasons than age or low total merit index (TMI), which is likely to happen in Norway because the farmers themselves make the decisions on which sows to keep. Male selection candidates were also picked randomly, 1 from each selected litter. These males were tested and received a PROD record when they were 9 mo (3 time steps) old. Twenty-five sires to produce the new selected litters were selected among tested boars that were 9 or 12 mo old.

3082

Lillehammer et al.

Compared Breeding Schemes Within the conventional breeding scheme (CONV), to which all other breeding schemes were compared, all selection decisions were made based on C-TMI values and no animals were genotyped. The BLUP breeding values for all selection candidates were reestimated every time step to take advantage of new relatives obtaining phenotypes. When genomic selection was implemented without changing the breeding structure, the schemes were called GS_x, in which x denotes the number of females genotyped per year and varied among 0, 1,200, and 2,400. In GS_x schemes, the 1,200 males (1 per litter) that were tested at the boar station and a number of the replacement females were genotyped each year. Genotyped animals, males or females, were assigned G-TMI after genotyping before selection of sires and dams to produce selected litters was performed. All genotyped animals contributed to increase the reference population that was used to estimate SNP effects on PROD whereas only genotyped females or males with daughters in production contributed to the reference population for MAT. The total number of animals genotyped each year was 1,200 (males) + x (females). Females to be genotyped were selected randomly among those selected as replacement females, to ensure that all genotyped females had phenotypes. Genomic selection could also be used to select between full-sib selection candidates to use the within-litter genetic variation for selection. This was tested in withinlitter (WL) schemes, denoted WL_x, in which x is again the number of females genotyped per year, which was 0, 1,200, or 2,400. In these schemes, 2 male candidates were picked randomly from each selected litter, instead of 1 as in the (GS_x) schemes. These boars got genomic breeding values for both traits, MAT and PROD, shortly after birth. Based on the G-TMI, one-half of the males were preselected. The preselected males entered the boar station and obtained a PROD record at 9 mo of age. Final selection of males to produce the next generation of selected litters were made among 9-mo-old boars and 12-mo-old boars based on G-TMI values. All these boars had own information and full-sib information on PROD, and the oldest ones also had full-sib information on MAT. Again it was assumed that MAT records were available before litters were selected, to mimic a situation where more litters are born than the selected litters without having to simulate offspring from litters where no candidates are selected. There was no restriction against co-selection of full-sibs in the preselection step or in the final selection step. Hence, the WL strategy gave the opportunity to co-select full sibs from high performing litters as well as to use the within-litter genetic variance for selection (Lillehammer et al., 2011).

To test the sensitivity of the results to the assumptions made, some alternative scenarios were tested. One alternative scenario included PROD records only for males. In reality, several production traits are only measured on tested boars or with a greater quality in tested boars than in females. A scenario in which the economic weight for MAT was 4 times that of PROD was used to test the sensitivity of the results to changes in economic weights. All selection schemes were run for 10 yr. The first years of selection were affected by that the base population was used as selection candidates until the scheme had produced own selection candidates and by the shift from discrete generations in the historical population to overlapping generations in the population under selection. Therefore, the first 5 yr of selection were omitted to make sure that a steady state was reached before results were reported. Average total genetic gain (ΔG), measured in σg units was calculated as

( g10MAT − g5MAT ) + ( g10PROD − g5PROD ) 5 σ2gMAT + σ2gPROD + 2σ gMAT , gPRODD in which g10 and g5 are the genetic levels in yr 10 and 5, respectively. Average percentage contribution of the maternal trait to total genetic gain (MAT%) per round of selection from yr 5 to 10 was calculated as (100 × ΔGMAT)/ (ΔGMAT + ΔGPROD). Average rate of inbreeding (ΔF) per year from yr 5 to 10 and accuracy of male selection (Acc), computed empirically as correlations between estimated and true breeding values after 10 yr of selection were also reported. All results are the averages of 50 replicates. A cost–benefit analysis for the different genomic selection breeding schemes was conducted, based on the results from the basic scenario where h2PROD was 0.3 and the genetic correlation between MAT and PROD was –0.3. It was assumed that the cost of genotyping 1,200 animals was about 2 million Norwegian Kroner (NOK), including tissue sampling and lab work. Other additional costs of applying genomic selection were ignored. Benefit per slaughter pig produced was assumed to be 140 NOK per genetic SD, for both MAT and PROD, based on Norsvin’s economic weights for the different traits in the breeding goal (D. Olsen, personal communication), ignoring differences in heritability and genetic correlations between traits within the MAT and PROD categories. The benefit of a given scheme was estimated as the average additional genetic gain, compared with CONV, from yr 5 to 10 of the simulation. Break-even point indicates the number of slaughter pigs that must be produced to pay back the investment of 1 yr of genotyping.

3083

Genomic selection for 2 traits in pigs

Table 1. Total genetic gain (ΔG), measured in genetic SD, relative contribution of maternal trait to total genetic gain (MAT%), and rate of inbreeding (ΔF), measured as an average for the last 5 yr of selection in the basic scenario, where the production trait had a heritability of 0.3 and the genetic and environmental correlation between production and maternal trait was –0.3 Scheme1

ΔG (SE)

MAT% (SE)

ΔF (SE)

CONV GS_0 GS_1200 GS_2400 WL_0 WL_1200 WL_2400

0.618 (0.007) 0.698 (0.010) 0.788 (0.009) 0.865 (0.009) 0.760 (0.009) 0.878 (0.011) 0.956 (0.007)

18 (1) 16 (1) 26 (1) 32 (1) 16 (1) 27 (1) 33 (1)

0.0088 (0.0004) 0.0053 (0.0002) 0.0048 (0.0002) 0.0040 (0.0001) 0.0062 (0.0003) 0.0057 (0.0003) 0.0045 (0.0002)

1CONV

= conventional breeding scheme; GS_x denotes a genomic selection scheme where x females and the boars at the boar station (1,200) were genotyped each year, and WL_x denotes a genomic selection scheme where x females and 2 boars per litter (2,400) were genotyped each year.

Results In the basic scenario, all genomic selection schemes increased ΔG, compared with CONV (Table 1). The cheapest implementation of genomic selection, where only males at the boar station were genotyped (GS_0), gave an increase in genetic gain of 13%, compared with CONV. For WL_2400, where 2 males from each litter as well as 2,400 females were genotyped every round of selection, ΔG was 55% greater than in CONV. Whether males or females were prioritized for genotyping had little impact on ΔG, resulting in similar ΔG for schemes with the same number of animals genotyped. Using CONV, MAT% was 18%, which is low, compared with the relative economic weight of 50%, because of the low heritability of MAT and the negative correlation between MAT and PROD. Increasing the number of genotyped females in the GS_x schemes increased MAT% up to 32 to 33% when 2,400 females were genotyped per year. Genotyping males did not affect MAT% significantly, causing schemes with the same number of genotyped females to have similar MAT%. Rate of inbreeding was reduced by 30 to 55% when genomic selection was applied, compared with CONV. Among the genomic selection schemes, genotyping more females consequently caused a reduction in rate of inbreeding. Genotyping more males (i.e., moving from GS to WL keeping the number of females constant) caused an increase in rate of inbreeding. In the scenario where MAT and PROD were assumed uncorrelated, MAT% increased for all schemes, from 28% in CONV to 40% in GS_2400 (Table 2). Uncorrelated traits gave similar or greater ΔG than negatively correlated traits for all schemes. The difference

Table 2. Total genetic gain (ΔG), measured in genetic SD, relative contribution of maternal traits (MAT%), and rate of inbreeding (ΔF), measured as an average for the last 5 yr of selection in the scenario where the production and maternal traits are uncorrelated1 Scheme2

ΔG (SE)

MAT% (SE)

ΔF (SE)

CONV GS_0 GS_1200 GS_2400 WL_0 WL_1200 WL_2400

0.661 (0.008) 0.723 (0.008) 0.812 (0.006) 0.869 (0.009) 0.806 (0.007) 0.900 (0.004) 0.956 (0.004)

28 (1) 29 (1) 36 (1) 40 (1) 29 (1) 36 (1) 39 (1)

0.0092 (0.0003) 0.0050 (0.0002) 0.0047 (0.0002) 0.0043 (0.0002) 0.0061 (0.0002) 0.0057 (0.0002) 0.0047 (0.0002)

1Heritability of the maternal trait was 0.1 and for the production trait was 0.3. 2CONV = conventional breeding scheme; GS_x denotes a genomic selection scheme where x females and the boars at the boar station (1,200) were genotyped each year, and WL_x denotes a genomic selection scheme where x females and 2 boars per litter (2,400) were genotyped each year.

in ΔG between schemes was smaller with uncorrelated traits than with negatively correlated traits, but the schemes ranked the same. Schemes with the same total number of animals genotyped gave similar ΔG whereas schemes with the same number of genotyped females gave similar MAT%, also when the traits were uncorrelated. Genotyping more females decreased ΔF whereas genotyping more males increased ΔF, similarly to what was found for negatively correlated traits. With increasing h2PROD, ΔG increased and MAT% decreased within all breeding schemes (Table 3). The ranking of the schemes, with respect to ΔG, MAT%, or ΔF were robust to changes in h2PROD. However, because CONV was more sensitive to changes in h2PROD than the genomic selection schemes, the relative benefit of genomic selection compared with CONV was reduced when h2PROD increased. In CONV, Acc spanned from 0.39 to 0.47 (Table 4). Selection accuracy increased when the traits were uncorrelated and with increasing h2PROD. Even the scheme where only boars at the test station was genotyped (GS_0) gave greater Acc than CONV, with a span from 0.49 to 0.57. When females were genotyped, Acc increased further, up to 0.66 and 0.69 in the scenarios where 1,200 and 2,400 females were genotyped each year. Genotyping more males would not affect selection accuracy, as long as these males do not obtain records, so they do not contribute to the reference population (Lillehammer et al., 2011). The results for the scenario where MAT had 80% of the weight in the breeding goal are shown in Table 5 for some key schemes. In this scenario, MAT% using CONV exceeded 100% because of a negative genetic gain for PROD. Thereby, genetic gain for MAT was greater than the sum of the genetic gain for the 2 traits (see formula

3084

Lillehammer et al.

Table 3. Total genetic gain (ΔG) and relative contribution of maternal traits (MAT%) and rate of inbreeding (ΔF), measured as an average for the last 5 yr of selection in scenarios where maternal and production traits are negatively correlated and the heritability of the production trait (h2PROD) is varied h2PROD = 0.2

h2PROD = 0.4

Scheme1

ΔG (SE)

MAT% (SE)

ΔF (SE)

ΔG (SE)

MAT% (SE)

ΔF (SE)

CONV GS_0 GS_1200 GS_2400 WL_0 WL_1200 WL_2400

0.585 (0.009) 0.667 (0.008) 0.764 (0.008) 0.837 (0.008) 0.739 (0.008) 0.846 (0.008) 0.921 (0.008)

25 (1) 21 (1) 31 (1) 36 (1) 19 (1) 31 (1) 34 (1)

0.0101 (0.0005) 0.0058 (0.0002) 0.0049 (0.0002) 0.0045 (0.0002) 0.0072 (0.0004) 0.0057 (0.0002) 0.0049 (0.0002)

0.638 (0.008) 0.710 (0.008) 0.796 (0.009) 0.861 (0.007) 0.766 (0.008) 0.890 (0.009) 0.953 (0.011)

13 (1) 10 (1) 23 (1) 29 (1) 11 (1) 23 (1) 32 (1)

0.0082 (0.0004) 0.0049 (0.0001) 0.0047 (0.0001) 0.0041 (0.0001) 0.0056 (0.0002) 0.0051 (0.0002) 0.0047 (0.0002)

1CONV = conventional breeding scheme; GS_x denotes a genomic selection scheme where x females and the boars at the boar station (1,200) were genotyped each round of selection, and WL_x denotes a genomic selection scheme where x females and 2 boars per litter (2,400) were genotyped every year.

in Materials and Methods section). The tested schemes ranked the same as in the basic scenario, with WL_1200 and GS_2400 giving similar ΔG, but WL_1200 still gave lower MAT% and greater ΔF than GS_2400. In this scenario, however, the lower MAT% obtained by WL_1200 could be seen as an advantage of that scheme because that results in WL_1200 to be more in concordance with the breeding goal than GS_2400. Still, the difference in ΔF between the schemes would favor GS_2400. The results from the scenario where only males received PROD records are shown in Table 6. In this scenario also, WL_1200 gave a slightly greater ΔG than GS_2400 but a greater ΔF. Hence, the overall ranking of the schemes would be the same as before. The effects of the number of genotyped animals and which animals (males or females) that are genotyped were hence shown to be robust against different genetic correlations between traits, heritabilities, economic weights, and how much phenotypic information can be obtained from the females. The results from the cost–benefit analysis is shown in Table 7, and the number of slaughter pigs necessary to produce to pay back the investment varied from 140,000 for WL_1200 to 169,000 for WL_0. The GS schemes Table 4. Accuracy of the final boar selection after 10 yr of selection when correlation between the traits (Corr) and heritability of the production trait (h2PROD) were varied1 Corr (= 0

Corr = –0.3

Scheme2

h2PROD = 0.3

CONV GS_0 GS_1200 GS_2400

0.47 (0.007) 0.57 (0.007) 0.66 (0.004) 0.69 (0.005)

h2PROD = 0.2 h2PROD = 0.3 h2PROD = 0.4 0.39 (0.009) 0.42 (0.008) 0.44 (0.009) 0.49 (0.007) 0.52 (0.007) 0.52 (0.007) 0.62 (0.005) 0.63 (0.005) 0.63 (0.005) 0.66 (0.005) 0.67 (0.005) 0.68 (0.005)

1Heritability of maternal trait (MAT) was constantly 0.1. Standard errors are given in parentheses. 2CONV = conventional breeding scheme; GS_x denotes a genomic selection scheme where x females and the boars at the boar station (1,200) were genotyped every year.

were less variable than the WL schemes and varied in break-even point from 142,000 to 149,000. These results show that all the tested schemes would be profitable in a commercial scale breeding program. Discussion This study compared different implementations of genomic selection to describe the effect of genotyping more animals, males or females, on genetic gain and ΔF. In addition to the total genetic gain, the effect of genotyping strategies on the relative contribution of each trait to total genetic gain was evaluated. The schemes compared represented 2 main strategies: 1 where the males at the boar station were genotyped and 1 where 2 males from each litter were genotyped before preselecting males to be tested at the boar station. Within both strategies, the effect of genotyping females was investigated. The more animals that were genotyped, the greater ΔG was obtained. The reason why ΔG increased was difTable 5. Total genetic gain (ΔG), relative contribution of maternal traits to total genetic gain (MAT%) and rate of inbreeding (ΔF) from yr 5 to 10 of selection, obtained by the main schemes when 80% of the weight in the breeding goal was on maternal traits1 Scheme2 CONV GS_0 GS_2400 WL_1200

ΔG (SE) 0.362 (0.010) 0.530 (0.008) 0.645 (0.008) 0.670 (0.009)

MAT% (SE) 117 (4) 75 (2) 96 (1) 84 (1)

ΔF (SE) 0.0118 (0.0006) 0.0063 (0.0002) 0.0047 (0.0002) 0.0062 (0.0003)

1Heritabilities were assumed to be 0.3 and 0.1 for the production trait (PROD) and maternal trait (MAT), respectively, and genetic correlation between MAT and PROD was –0.3. 2CONV = conventional breeding scheme; GS_x denotes a genomic selection scheme where x females and the boars at the boar station (1,200) were genotyped each year, and WL_x denotes a genomic selection scheme where x females and 2 boars per litter (2,400) were genotyped each year.

3085

Genomic selection for 2 traits in pigs

Table 6. Total genetic gain (ΔG), relative contribution of maternal traits to total genetic gain (MAT%), and rate of inbreeding (ΔF) from yr 5 to 10 of selection when females did not obtain production trait (PROD) records1 Scheme2

ΔG (SE)

MAT% (SE)

ΔF (SE)

CONV GS_0 GS_2400 WL_1200

0.271 (0.007) 0.655 (0.008) 0.795 (0.009) 0.810 (0.010)

27 (1) 19 (1) 37 (1) 33 (1)

0.0105 (0.0004) 0.0067 (0.0002) 0.0046 (0.0002) 0.0059 (0.0002)

1Heritabilities were assumed to be 0.3 and 0.1 for PROD and maternal trait (MAT), respectively, and genetic correlation between MAT and PROD was –0.3. 2CONV = conventional breeding scheme; GS_x denotes a genomic selection scheme where x females and the boars at the boar station (1,200) were genotyped each year, and WL_x denotes a genomic selection scheme where x females and 2 boars per litter (2,400) were genotyped every year.

ferent whether more males or females were genotyped, and the choice of design thereby had consequences on MAT% and ΔF in addition to ΔG. Genomic preselection of males increased ΔG for 2 reasons: use of both between-litter and within-litter genetic variance for selection and greater selection intensity through the possibility to select 2 males from the greatest indexed litters. Greater selection intensity and use of all genetic variance together resulted in 8 to 12% increase in ΔG without changing MAT%, irrespective of h2PROD and the correlation between MAT and PROD. Although within-litter selection does not affect ΔF (Daetwyler et al., 2007), coselection of full sibs increases the ΔF. The increased ΔF in WL schemes, compared with GS schemes with the same number of genotyped females, indicates that coselection of full sibs explained parts of the increase in ΔG obtained by genotyping 2 males per litter. Genotyping females increased the reference population relatively more for MAT than for PROD contributing to increased ΔG, reduced ΔF, and increased MAT%. With increasing number of females genotyped, the difference in ΔF between WL schemes and GS schemes decreased, indicating that more of the increase in ΔG came from increased accuracy rather than co-selection of full sib males. Allowing co-selection of full-sibs thereby seems possible without increasing rate of inbreeding significantly, as long as the selection accuracy is high, that is, that the reference population for each trait is continuously updated with genotyped animals with phenotypes. It would have been possible to select more piglets from the greatest indexing litters also when using the GS strategy. This would, however, have reduced the number of litters from which we could test boar candidates, assuming the test capacity on the station to be constant. In the GS strategy, the information on each piglet was highly limited before the boar test, and it seemed thereby appropriate to test piglets from as many litters

Table 7. Cost–benefit analysis estimated for each of the genomic selection schemes under the assumptions that maternal trait (MAT) had heritability 0.1, production trait (PROD) had heritability 0.3, and the genetic and phenotypic correlations between the traits were –0.3 Scheme1

Cost (NOK)2

GS_0 GS_1200 GS_2400 WL_0 WL_1200 WL_2400

2 million 4 million 6 million 4 million 6 million 8 million

Benefit (NOK/pig)3 Break-even point4 13 28 41 24 43 56

149,000 142,000 146,000 169,000 140,000 143,000

1GS_x denotes a genomic selection scheme where x females and the boars at the boar station (1,200) were genotyped and WL_x denotes a genomic selection scheme where x females and 2 boars per litter (2,400) were genotyped. 2Costs of genotyping for this scheme per year. NOK = Norwegian Kroner. 3Extra gain per slaughter pig produced for this scheme. NOK = Norwegian Kroner. 4Number of slaughter pigs required to pay back the genotyping costs of 1 yr.

as possible. Another possibility not considered was to genotype more females and perform genomic preselection of females. It was shown that genomic preselection of males gave limited increase in genetic gain and increased the rate of inbreeding, compared with increasing the reference population. Genotyping female selection candidates, instead of selected females, would be even less effective because selection intensity was assumed lower for females than for males. However, if considering genotyping more animals than the numbers tested in this study, genomic preselection of females as well as for males could be considered. The relative ranking of the different schemes was robust to changes in heritability, correlation between traits, and economic weights. The quantitative changes differed, however, between scenarios. Generally, genomic selection was more beneficial when the conventional breeding program gave low genetic gain (i.e., when the traits under selection were negatively correlated and when h2PROD was low). Whether more male selection candidates or females with phenotypes were prioritized for genotyping influenced mainly the relative contribution of the different traits to total genetic gain. This can be used strategically to shift the genetic gain towards a desired direction to achieve long term sustainability of the breeding program, without compromising the total genetic gain, which is economically important for the farmers and the breeding organization. In the studied case of Norwegian landrace, GS_2400 and WL_1200 had similar genetic gain and would hence give similar profit. However, GS_2400 would probably be more sustainable, first of all because of the lower rate of inbreeding but also because of a more balanced relative genetic gain of the different traits. The

3086

Lillehammer et al.

use of genomic selection to increase MAT% was more efficient when CONV gave low genetic gain for MAT due to stronger competition from PROD. The impact of genomic selection on MAT% therefore decreased when the traits were uncorrelated and with decreasing h2PROD. In cases where economic weights are uncertain or change over time, a breeding strategy that assures genetic gain for all traits under selection is most sustainable (Gibson, 1989), especially when farmer preferences result in an underweighting of robustness traits due to a short-term perspective (Amer, 2012). For maternal breeds, it is obvious that genetic gain for maternal traits is needed to stay competitive and that additional gain for traits that are difficult to improve by conventional methods can be a competitive advantage. For dairy cattle, it has been shown that genomic selection could move the selection pressure away from functional traits towards production traits because of the gap between selection accuracy for functional and production traits, respectively, being larger for genomic breeding values than for conventional progeny test breeding values (König and Swalve, 2009). In pig breeding, breeding values for traits measured directly on the selection candidates are more accurate than breeding values for traits measured on relatives due to phenotypic testing rather than progeny testing. In this case, genomic selection would increase the relative contribution to total genetic gain of traits with low heritability and no phenotypic information from the boar selection candidates, such as maternal traits. These results are similar to those of other studies, which have shown that genomic selection would favor lowly heritable traits (e.g., Togashi et al., 2011; Buch et al., 2012). To achieve a significant increase in relative contribution of maternal traits to total genetic gain, we found that it is essential to genotype females with maternal records. Maternal and production traits sum to about 80% of the breeding goal for Norwegian Landrace, so some traits in the breeding goal do not fit into any of these categories. The impacts of genomic selection on these traits are dependent on the available phenotypic information. Traits that can be measured directly on selection candidates are comparable with PROD and are expected to increase in genetic gain similar to what was found in this study. Traits that are measured on sibs or other close relatives can benefit more from genomic selection, similar to what was found for MAT, due to a shift in genetic gain towards such traits. This is, however, dependent on having a sufficiently large reference population for these traits, which are updated regularly, using close relatives to the selection candidates. The ranking of the schemes, with respect to return on economic investment, is dependent on the size of the production of the breed under selection. Generally, the larger production, the larger investment, that is, number of ani-

mals genotyped, can be justified. In addition to the values presented in Table 7, it should be noted that genetic gain will accumulate over time whereas the costs remain constant, thereby increasing the benefit over cost ratio over time. Also, this analysis ignored other traits in the breeding goal that could also benefit from genomic selection, without adding more costs if they are measured on animals already genotyped. All in all, the numbers show that genomic selection would easily pay back the investment for a breeding program of commercial size. The differences in break-even point between the schemes were small, showing that the schemes will all pay back the investment approximately in the same time. After the investment is paid back, schemes with the greatest additional value of each pig will generate the largest output and hence make the more expensive schemes, with high input and high output, most profitable. In practice, breeding goals contain more than 2 traits and a simplified situation where all traits are captured into the MAT and PROD indexes would not capture all traits in the breeding goal. For maternal traits, most traits would be possible to capture in a MAT index although some traits are measured later in life than assumed here (e.g., litter size in later parities). For production traits, some traits including growth could be measured on all candidates as was assumed for the entire PROD index in this study. Other traits, however, are only measured on tested boars or at least have a greater reliability of measurement on the boar test than on field data. The sensitivity of the results to different amounts of phenotypic information for PROD was tested when only males got PROD records. In this alternative scenario MAT% increased both for CONV and for the genomic selection schemes. The ranking of the schemes, with respect to genetic gain and rate of inbreeding, however, was unchanged. As in the other scenarios, WL was shown to give the greatest genetic gain for PROD whereas GS gave the greatest genetic gain for MAT, when comparing schemes with the same total number of genotyped animals. The choice of which and how many animals to genotype was hence unaffected by fewer records for PROD and should be decided based on the breeding goal and secure that all important traits obtain regularly updated reference populations. To make the decisions on which and how many animals should be genotyped, it is important not only to have a strategy for genotyping selection candidates but also to genotype animals that contribute to the estimation and updating of SNP effects. The results showed that an increase in ΔG of 13% can be achieved by genotyping only the males at the boar test station. Simultaneously, ΔF was reduced by 40%. It is thereby possible for the breeding companies to take advantage of genomic selection without going through a complete restructur-

Genomic selection for 2 traits in pigs

ing of the breeding scheme and by genotyping only 1,200 animals every year. A further increase in ΔG, up to 55%, and a further reduction in ΔF could be obtained by genotyping more animals. Genotyping females should be prioritized before genotyping more males from each litter, as that would give similar or greater genetic gain and reduced rate of inbreeding. LITERATURE CITED Amer, P. J. 2012. Turning science on robust cattle into improved genetic selection decisions. Animal 6:551–556. Badke, Y. M., R. O. Bates, C. W. Ernst, C. Shwab, and J. P. Steibel. 2012. Estimation of linkage disequilibrium in four US pig breeds. BMC Genomics 13:24. Buch, L. H., M. K. Sørensen, P. Berg, L. D. Pedersen, and A. C. Sørensen. 2012. Genomic selection strategies in dairy cattle: Strong positive interaction between use of genotypic information and intensive use of young bulls on genetic gain. J. Anim. Breed. Genet. 129:138–151. Daetwyler, H. D., B. Villanueva, P. Bijma, and J. A. Woolliams. 2007. Inbreeding in genome-wide selection. J. Anim. Breed. Genet. 124:369–376. Du, F.-X., A. C. Clutter, and M. M. Lohuis. 2007. Characterizing linkage disequilibrium in pig populations. Int. J. Biol. Sci. 3:166–178.

3087

Fisher, R. A. 1930. The genetical theory of natural selection. Clarendon Press, Oxford, UK. Gibson, J. P. 1989. Selection strategies and artificial evolution. Theor. Appl. Genet. 78:87–92. Henderson, C. 1984. Applications of linear models in animal breeding. Guelph Univ. Press, Guelph, Canada. Holm, B., M. Bakken, G. Klemetsdal, and O. Vangen. 2004. Genetic correlations between reproduction and production traits in swine. J. Anim. Sci. 82:3458–3464. König, S., and H. H. Swalve. 2009. Application of selection index calculations to determine selection strategies in genomic breeding programs. J. Dairy Sci. 92:5292–5303. Lillehammer, M., T. H. E. Meuwissen, and A. K. Sonesson. 2011. Genomic selection for maternal traits in pigs. J. Anim. Sci. 89:3908–3916. Meuwissen, T. H. E., B. J. Hayes, and M. E. Goddard. 2001. Prediction of total genetic value using genome-wide dense marker maps. Genetics 157:1819–1829. Sonesson, A. K., and T. H. E. Meuwissen. 2009. Testing strategies for genomic selection in aquaculture breeding programs. Genet. Sel. Evol. 41:37. Togashi, K., C. Y. Lin, and T. Yamazaki. 2011. The efficiency of genomic-wide selection for genetic improvement of net merit. J. Anim. Sci. 89:2972–2980. Visscher, P., R. Pong-Wong, C. Whittemore, and C. Haley. 2000. Impact of biotechnology on (cross)breeding programmes in pigs. Livest. Prod. Sci. 65:57–70. Wright, S. 1931. Evolution in Mendelian populations. Genetics 16:97–159.