Get Complete Project Material File(s) Now! »
Using several breeds
In dairy cattle
Using a multi-breed reference population has a small effect on GEBV accuracy predictions whatever the dairy trait considered (Hayes et al, 2009a; Pryce et al, 2011). In Jersey cows, accuracy for fat percent is 1% higher using a multi-breed than purebred reference population and 13% higher compare to non genomic accuracy (Hayes et al, 2009a). A three-breeds reference population induces a small increase in GEBV accuracy compare to using two breeds. Accuracy obtained in milk yield with the Holstein-Jersey-Fleckvieh reference population is 5% higher than the one calculated with Holstein-Fleckvieh reference population. The accuracy gain obtained for small breeds is not really noticed in large breed. In order to predict genomic gain in Holstein breed using only Holstein or multi-breed (Jersey, Fleckvieh, Holstein) reference populations gives the same accuracies of 0.41 in protein and fat yields (Pryce et al, 2011).
In beef cattle
Purebred training population (Angus) used with multi-breed validation population (Carcass Merit Project including Angus) gives good predictions for all breeds considered. Correlations between predicted and true genotypic values are 0.946, 0.969, and 0.972 for Hereford Brahman and Angus respectively. Multi-breed reference population increase correlations of 1% for simulated phenotypes compare to Angus purebred reference population. Nevertheless correlation between predicted genetic merit and phenotype is better using multi-breed reference population (Kizilkaya et al, 2009).
Training in purebred population to predict genetic merit of multi-breed animals is possible but induce small increase in GEBV accuracies especially for large breeds. It is useful in livestock because of crossbreed or multi-breed lines but not in dairy cattle because accuracies are improved only in small breeds. GEBV are more accurate with a multi-breed than with purebred reference population in case of population without individuals genotyped and phenotyped (Pryce et al, 2011). Results using several breeds in a unique reference population were summed up in Annexe A.
Several models to use multi-breed reference population
Several statistical models tend to predict GEBV of animals like Bayesian models (Bayes A, B, etc) or GBLUP. At this time no any model is assume to be better than others. In genomic evaluation with multi-breed (Holstein-Jersey-Fleckvieh) reference population, GEBV accuracies in dairy traits calculated with GBLUP model (0.4) are really close to the ones calculated with BayesA (0.37) (Pryce et al, 2011). Some studies have shown the superiority of Bayesian model in case of multi-breeds reference population. Bayes A and Bayes_SSVS are more accurate than GBLUP for GEBV of Jersey animals using Holstein-Jersey reference population (Pryce et al, 2011). It seems to be better (16% higher than BLUP on weight at ultrasound scanning) to predict the breed mostly present (Merinos sheep) in the reference population (Daetwyler et al, 2010). On cows fat percent GEBV are 12% higher with Bayesian (0.82) models than with GBLUP (0.73). But BLUP is better for large breed like Holstein cows (Hayes et al, 2009a). GBLUP model with a specific breed effect leads to higher DGV reliabilities than Bayesian models (Makgahlela et al, 2010). A model with SNP effects different but correlated between breeds is 4.1% more accurate than model with same SNP effects for genetically distant breeds (Varona et al, 2010).
A reference population combining populations could improve the reliability of predictions except if marker effects are very different between populations. Markers need to be very close to QTL, the only ones in persistent LD with QTL across breeds or populations. So many projects focus on high density chip like the 800 k chip in ANR GEMBAL project for French beef cattle.
Use of female genotypes
Since 2011, decrease of genotyping cost allows examining the interest of female genotypes like ANR AMASGEN project (with a low density chip) which is expected to increase genetic gain and decrease inbreeding rates (Moser et al, 2010). In average combined male and female genotypes increases GEBV accuracies of 25% compare to progeny testing (Schaeffer, 2006). DGV reliabilities of cows using only female genotypes are 39% higher than EBV on energy balance (Verbyla et al, 2010) whereas genetic gain is more than twice higher using female genotypes in genomic selection than in progeny testing (Mc Hugh et al, 2011). In a simulated dairy cattle breeding scheme (Figure 2), using female genotypes increases the annual genetic gain compare to genomic selection with only male genotypes of 44% (Mc Hugh et al, 2011) or 21 % (Buch et al, 2010) depending on simulations. Using female genotypes to preselect them at birth leads to genetic progress 2.4% low than genomic selection which preselect males at birth (Schrooten et al, 2005).
Moreover increase number of cow genotypes used in reference population (from 1 000 to 2 000 ) improves DGV reliabilities of 40% (0.3 compare to 0.18) on a simulated trait (h²=0.3) (Calus et al, 2011). Indeed use 10% of female genotypes compare to 2% leads to an increase in accuracy of 4% for a simulated trait (h²=0.25). Furthermore annual genetic gain is maximum for 25% of male genotypes in reference population (Figure 3). In this case, increasing proportion of female genotypes in reference population (from 75% to 100%) does not improve the annual genetic gain (Sørensen and S ørensen, 2010). Using top ranking female genotypes does not improve GEBV accuracies compare to female genotypes chosen at random (Jiménez-Monteroet al, 2010). Using female genotypes in the reference population is a good way to accelerate genetic progress and to improve accuracy of GEBV of both sexes. It is interesting because genotyping females would lead to better mating and high merit calves that would compete with current sires (Sørensen and Sørensen, 2010). Nevertheless even if female genotypes are not used, applying selection intensity (multiple ovulation and embryo transfer) in dams of cows leads to an increase in genetic gain (Pryce et al, 2010). Results are summed up in Annexe B.
Phenotypes mostly used for bulls in genomic selection are daughter yield deviations (DYD). DYD are average performance of bull daughters corrected for environmental effects and for their dam genetic merit. For females, yield deviations are used as phenotypes, and defined as YD y p where is yield of cow, β is fixed effects and is permanent environment effect (Szyda et al, 2008). If DYD are not available like for stranger bulls, deregressed proofs (DRP) are considered as phenotypes (Robert-Graniéet al, 2011). The DRP are average performance of daughters calculated from genetic merit of bulls using the inverse of a genetic evaluation method. In Nordic dairy cattle using DRP with single step genomic evaluation improve DGV reliabilities of 6.7%, 12.9% and 10% for milk, protein and fat yields respectively compare to using DYD (Mäntysaari et al, 2011).
Estimated breeding values (EBV) which already contain genetic information can be used as phenotypes. With GBLUP, GEBV reliabilities are slightly higher (+ 0.7%) using EBV than using DYD. However this difference depends on heritability (Guo et al, 2010). EBV easier to use because directly obtained from genetic evaluation contains smaller random error than DYD. Nevertheless EBV includes information of all relatives leading to double counting because predictions of GEBV also include information from relatives. DRP could be a good alternative because it does not take into account relatives and are easier to obtained compare to DYD (Guo et al, 2010).
A “single step approach” is for the moment the only method which analyses directly performance records for genomic predictions (Legarra et al, 2011). Indeed this approach combines phenotypic, genomic and pedigree information into a single simultaneous analysis. Using a one step approach leads to similar accuracy and smaller values (-73%) for prediction error variance and mean square error than using DYD in classical genomic evaluation method (Vitezica et al, 2010). Nevertheless based on coefficients of determination, one-step approach is 2.5% less accurate than multiple-step one using DYD (Aguilar et al, 2010).
Material and methods
The aim of this study is to characterize the feasibility of genomic selection in the French dairy goat population. The interest of genomic selection consists in the decrease of generation interval in a higher accuracy of breeding values. For this study the feasibility of genomic selection was estimated by comparing GEBV’s accuracies to EBV ones. Two populations of genotyped animals (popIA and popQTL) were considered as the two sets of a reference population.
Reference population in French Alpine and Saanen goats
As described above, population structure has an important effect on accuracy of GEBV. Therefore the two populations of genotyped animals were first analyzed in terms of pedigree and genetic structure.
A Description of popQTL and popIA population structures
PopQTL and popIA were the two populations genotyped with the Illumina 50K goat BeadChip. The popQTL was a population initially dedicated to QTL detection. It consisted in 20 AI bucks (9 Saanen and 11 Alpine) born between 1999 and 2004 and their 2 246 daughters with an average of 112 daughters genotyped per male (from 74 to 127). Bucks of popQTL population had high total merit index (ICC, including type and production traits) from 0.97 to 5.76. The popIA population consisted in 1 012 AI bucks (429 Saanen and 582 Alpine) including bucks from popQTL, born between 1985 and 2011 with an ICC merit from – 6.48 to 7.78. Among these 1 012 bucks, only the 850 youngest bucks born after 1998 will be genotyped. Negatives ICC values were obtained for former bucks because indexes were expressed as deviation from a rolling base updated each year with current bucks (with a better genetic level). At this time only 85 genotypes were available including 39 Saanen bucks and 46 Alpine ones.
The description of the population consisted in analysis of pedigree and genetic structure. The pedigree structure of the two populations was designed with Pedigree viewer software (Kinghorn and Kinghorn, 2007) considering genotyped animals and 4 anterior generations (all parents known). Analysis of genetic structure was done with Pedig software (Boichard, 2006) considering 26 generations. The quality of pedigree information, given by the number of equivalent generation known, was evaluated using option “Ngen” of Pedig known, ngp the number of grandparents known, nggp the number of great grandparents known and nn the number of ancestor known at generation n (Danchin-Burge, 2011). The average number of ancestors by individual was also computed. Average inbreeding and parental coefficients were calculated independently for all animals: 1) from popQTL AI bucks pedigree, 2) from their daughter pedigree and 3) from popIA bucks pedigree considering the whole pedigree including 26 generations at maximum. The popQTL pedigree consisted in 34 771 animals, 18% had at list one parent unknown and 14% with both parents unknown. 17 912 animals made up the popIA pedigree, 21% had at least one parent unknown and 18% with both parents unknown. Inbreeding coefficient, calculated with “meuw” option of Pedig software, is the probability at a given locus that an individual received similar alleles from his both parents. It could be expressed as Fj 1 FA where n is the number of individuals on parental line and FA is inbreeding coefficient of common ancestor (Verrier et al, 2001). Average parental coefficients had been calculated with “par” option of Pedig software taking 50 samples of 30 individuals at random. Parental coefficient between I and J is the probability that two alleles of a gene are similar in I and J genome (Verrier et al, 2001). Effective number of founders was calculated for AI bucks of popQTL and popIA with “probability of gene origin” option of Pedig software. This effective nu mber of founders is the inverse of the probability that two genes in the population come from the same ancestor: this is the number of founders that given the same genetic variability if they all contribute at the same level (De Rochambeau et al, 2003).
Phenotypes mostly used in genomic evaluation are DYD for males and YD for females. They are derived from classical genetic evaluation. Figure 4 is a scheme of methods used to obtain DYD and YD in this study. Genetic evaluations did not include only animals of popIA and popQTL populations. DYD and YD were more accurate than expected because animals have already progeny and relatives in the all population. Therefore two classical genetic evaluations were made to obtain two DYD or YD per animal.
YD and DYD
The two evaluations were: (i) an evaluation with all official females (2 885 244 individuals), DYDallfem were well known because all bucks daughters were included and, (ii) an evaluation on the 2 246 females from popQTL only, DYDfemg were not as accurate as in the first case. This second case was similar to a situation where only data from females were available to predict young males (not progeny tested) breeding values. In this case including females genotyped was expected to improve GEBV accuracies. For this case genetic evaluation model was simplified without heterogeneity of variance and permanent environment because of few data. Level of fixed effects like year, region and parity were pooling to allow estimation of effects (Annexe C). To underline the interest of female genotypes, a genomic evaluation using phenotypes obtained in the second case (DYDfemg and YD) and male and female genotypes was compare to another one using only male phenotypes (DYDfemg) and genotypes. YD of genotyped females and their deviations were calculated with preadj option of GENEKIT
– Table 1 shows statistical results on YD and DYD given for animals genotyped and calculated with evaluation on all females. In this case YD of females genotyped were on average of 197.2 256 for milk yield and 0.74 5.8 for fat percent (FP). Average DYD was 76.7 80 for milk yield and 0.52 2.8 for FP. DYD of all males from popIA and popQTL were given for the 860 bucks with daughters in production in 2012. The average number of observations (lactations of daughters) for DYD was 801 2 388 (from 1 to 10 120) per buck.
– Table 2 presents statistical results on YD and DYD obtained with evaluation on the 2 246 females genotyped and given for the 2 266 animals from popQTL. Average YD was – 2.89 362 on milk yield and – 3.04 5.8 on FP. DYD was on average – 0.56 102 on milk yield and 0.15 2.5 on FP. Table 2 also shows correlations between DYD (and YD) obtained with evaluation on all females and those obtained with evaluation on females genotyped. These correlations ranged from 0.75 to 0.94, because model of non official genetic evaluations were different and variance heterogeneity was not taking into account. YD and DYD values in official method were not comparable with the one obtained for evaluation on females genotyped. Nevertheless minimum and maximum of these YD and DYD ranged in the same order than those obtained on all official females.
EBVs for Alpine and Saanen females were calculated together for dairy traits. It was known that there were slight differences between their performances (Bélichonet al, 1999). In this study differences between Saanen YD and Alpine one in the whole population (2.8 millions) were noticed. YD for milk yield were in average of – 32.39 372 in Alpine and 30.35 380 in Saanen (results not shown). Phenotypes of Saanen goats were in average superior than the ones in Alpine for milk protein and fat yield but it was the contrary for fat and protein percent. These differences induced taking into account a fixed breed effect in genomic evaluation model. Standard errors of YD in Alpine and Saanen were similar. So heterogeneity of variance had not to be implemented in the genomic evaluation model. SCS: Somatic Cell Score, FU: Fore Udder, UP: Udder Profile, UFP: Udder Floor Position, TO: Teat Orientation, RUA: Rear Udder Attachment
Table 3- Table 3 presents genetic parameters for several traits, only type traits and SCS are different between the two breeds. Higher heritability is 0.4 for FP and protein percent (Spangler et al, 2008), 0.3 for milk yield, FY and PY, 0.19 for SCS in Alpine and 0.21 in Saanen and from 0.26 to 0.36 for type traits. YD of Saanen and Alpine for those functional traits were calculated separately.
Table 4 shows statistical results on YD and DYD given for animals genotyped for SCS and type traits obtained with evaluation on all females. On all females genotyped only 2 233 had been registered for morphology. YD obtained for type traits were lower than those obtained for milk traits. In average all YD on type traits were around – 0.1 except for fore udder where average YD is close to 0.01. DYD were in average from – 0.007 to 0.027, the highest DYD was obtained for rear udder attachment and the smallest one for fore udder. For somatic cell score only 2 229 females had been recorded and 789 males had daughter with SCS records. Average YD in SCS were 0.117 2 and average DYD were 0.03 0.52 which was in the same order of magnitude than the ones for type traits.
Weight of phenotypes
Weight of female phenotypes taken here was weight of lactation in the official evaluation, 1 for fist lactation and 0.8 for second and third lactation in goats. As in dairy cattle weight of male performances considered here was the effective daughter contribution (EDC). It took into account almost all the above-mentioned factors and was calculated as of lactations of daughter k corrected for contemporary group and njkl is contemporary group size in which daughter k made its lth lactation, and 4 h² where h² is the heritability of h² the trait and r the repeatability of a record (Fikse and Banos, 2001). In this study EDC were calculated using crEDC software (Sullivan, 2010) and separately for the two breeds for SCS and type traits (Table 3). Figure 6 presents the distribution of EDC values for SCS identical to EDC distribution for other traits, almost following Gaussian distribution. Figure 7 is a representation of EDC values linearly proportional to number of daughters in official genetic evaluation. The bucks with less than 100 daughters were the young ones born after 2007. Table 5 shows high correlations (from 0.99 to 0.98) between EDC and number of daughters in official evaluation.
Table 6 presents average EDC calculated with evaluation on all females and on only females genotyped. EDC in official case (413 on FP) were larger than the one obtained only with female genotypes (89 on FP) and were ten times weight of female phenotypes.
Table of contents :
I.1. Structure of the reference population
I.1.A. Extent of LD
I.1.B. Levels of relationship within reference population
I.2. Multi breeds or multi populations reference population
I.2.A. Using several populations
I.2.B. Using several breeds
I.2.B.a. In dairy cattle
I.2.B.b. In beef cattle
I.2.C. Several models to use multi-breed reference population
I.1. Use of female genotypes
I.2. Phenotypes used
II. Material and methods
II.1. Reference population in French Alpine and Saanen goats
II.1.A Description of popQTL and popIA population structures
II.1.B Phenotypes available
II.1.B.a. YD and DYD
II.1.B.b. Weight of phenotypes
II.1.C Genotypes of the reference population
II.1.C.a. Genotypes available
II.1.C.a. Calculation of LD extent and persistence of LD between breeds
II.1.C.b. Simulation of bucks genotyped
II.2. Prediction of GEBV
II.2.A. Prediction of GEBV accuracy
II.2.B. Estimation of GEBV
II.2.B.a. Calculation of accuracy
II.2.B.b. Cases studied
III. Results and discussion
III.1. Structures of the two populations genotyped
III.2. LD extent in the population genotyped
III.2.A. LD within the population
III.2.B. Persistence of LD between Alpine and Saanen goats
III.3. Simulation of bucks genotypes
III.4. Prediction of GEBV accuracy level
III.4.A. In functions of the heritability and population size
III.5.A. Function of LD extent
III.5. Results on genomic evaluation results could indicate that these males are homozygote for a QTL concerning SCS. But further analyses are needed.
III.5.A. Effect of population size on GEBV accuracy
III.5.B. Interest of female genotypes
III.5.C. Interest of genomic selection