Search
Author
Title
Vol.
Issue
Year
1st Page

Journal of Animal Science - Animal Genetics

Genomic selection in the German Landrace population of the Bavarian herdbook1

 

This article in JAS

  1. Vol. 94 No. 11, p. 4549-4557
     
    Received: Apr 27, 2016
    Accepted: Sept 02, 2016
    Published: October 27, 2016


    2 Corresponding author(s): mgertz@tierzucht.uni-kiel.de
 View
 Download
 Share

doi:10.2527/jas.2016-0581
  1. M. Gertz 2*,
  2. C. Edel,
  3. I. Ruß,
  4. J. Dodenhoff,
  5. K.-U. Götz and
  6. G. Thaller*
  1. * Institute of Animal Breeding and Husbandry, Kiel University, 24098 Kiel, Germany
     Institute of Animal Breeding, Bavarian State Research Centre for Agriculture, 85586 Poing, Germany
     Tierzuchtforschung e.V., 85586 München, Germany

Abstract

The aim of our study was to compare different validation methods with respect to their impact on validation results and to evaluate the feasibility of genomic selection in the German Landrace population of the Bavarian herdbook. For this purpose, a sample of 337 boars and 1,676 sows was genotyped with the Illumina PorcineSNP60 BeadChip. Conventional BLUP breeding values for fertility, growth, carcass, and quality traits were deregressed and used as phenotypes in genomic BLUP. The resulting genomic breeding values were also blended with information from the full conventional breeding value estimation to include information from nongenotyped parents. Subsequent validation used forward prediction, realized reliabilities, and theoretical reliabilities. The results indicate that the validation methods showed a relatively large effect on in the displayed reliability levels in our study: forward prediction reliabilities were found to be much lower than the conventional parent-average reliabilities whereas corresponding realized and theoretical reliabilities were found substantially greater. Theoretical reliabilities appear to be the most consistent validation approach tested in our study, because they avoid the use of proxy variables. Generally, our results suggest a substantial potential for a genomic selection implementation for the Bavarian herdbook by using both sows and boars. Theoretical genomic reliabilities of direct genomic values of selection candidates were, on average, 31 to 36% greater than the conventional parent average reliabilities. However, the inclusion of residual information from conventional breeding values had only a marginal effect on reliabilities.



INTRODUCTION

An important objective of the genomic selection (GS) research in pigs is to solve specific challenges associated with practical implementations of GS routines, for example, the need for new genotyping strategies, as the sample size for genomic evaluations in pig is usually very limited (Simianer, 2009). To assess if benefits from GS will actually justify an implementation, different breeding strategies should be examined to estimate the increase in reliabilities for genomic breeding values compared with conventional parent averages. For this purpose, results from breeding value validation (genomic reliabilities) are crucial, but it is well known that validation results can noticeably differ from one validation method to another (VanRaden et al., 2009; Su et al., 2010; Thomasen et al., 2011). Nevertheless, such effects have not yet been examined and discussed for a typical pig breeding scenario. The aim of our study was to compare different validation methods regarding their impact on validation results and to evaluate possible issues of genomic evaluations within the Bavarian herdbook population in terms of genomic reliabilities, the animal sample structure, and implications arising from both.


MATERIALS AND METHODS

Data

Tissue samples of 2,031 pigs from the German Landrace purebred population of the Bavarian herdbook breeding program were collected. The sample consisted of boars from birth years 1995 through 2011 and sows from birth years 2005 through 2011. In this sample, 18% of the genotyped sows originated from the nucleus herd and 82% originated from downstream multiplier farms. The resulting distribution of animal birth dates in the sample is presented in Fig. 1. Samples were genotyped using the Illumina PorcineSNP60 BeadChip (version 1 and version 2; Illumina, San Diego, CA) and genotyping quality was subsequently checked. All statistical analyses were performed using the statistical computing language R (R Development Core Team, 2015). Single nucleotide polymorphism markers were excluded from further analysis when they deviated from Hardy–Weinberg equilibrium with P < 10−5, showed a call rate below 0.95, or had a minor allele frequency lower than 0.01. After editing, 49,856 markers remained for analysis and were used for the consequent evaluation of pedigree consistency by checking the coherence of marker-based identity-by-descent coefficients (Wang, 2007) between the parental genotypes and the offspring genotypes and, second, the coherence between the maternal and paternal grandsire genotypes and the offspring genotypes. If threshold values were exceeded, conflicting parents or grandparents were set to missing in the pedigree and, therefore, excluded from further evaluations. After final editing, 337 boars and 1,676 sows remained for analysis (2,013 in total).

Figure 1.
Figure 1.

Number of boars and sows in the sample (y-axis) sorted by year of birth (x-axis).

 

The traits taken into account in this evaluation were number of piglets born alive (LBP), proportion of piglets weaned (PW), average daily gain (ADG), feed conversion ratio (FCR), lean meat content (LP), belly meat content (BMC), eye muscle area (EMA), fat-to-lean ratio (FLR), carcass length (CL), pH after 45 min (PH1), drip loss(DL), and intramuscular fat content (IMF). The respective heritabilities, provided from the current Bavarian breeding scheme, are listed in Table 1.


View Full Table | Close Full ViewTable 1.

Traits in analysis with the respective abbreviations, heritabilities (h2),1 average reliabilities of calibration animals,2 and average reliabilities of validation animals3

 
Trait group Trait Abbr.4 h2 r2C r2V
Fertility Number of piglets born alive LBP 0.24 0.56 0.41
Proportion of piglets weaned PW 0.18 0.41 0.31
Growth Average daily gain ADG 0.31 0.37 0.24
Feed conversion ratio FCR 0.48 0.40 0.32
Carcass Lean meat content LP 0.66 0.53 0.43
Belly meat content BMC 0.59 0.50 0.40
Eye muscle area EMA 0.72 0.56 0.47
Fat-to-lean ratio FLR 0.69 0.54 0.45
Carcass length CL 0.72 0.50 0.39
Quality pH after 45 min PH1 0.19 0.31 0.18
Drip loss DL 0.26 0.25 0.22
Intramuscular fat content IMF 0.70 0.54 0.46
1 h2 = heritability.
2r2C = average deregressed breeding value reliabilities of the calibration animal subset.
3r2V = average deregressed breeding value reliabilities of the validation animal subset.
4Abbr. = abbreviated trait names.

Prediction

All conventional breeding values were obtained from single-trait models by BLUP (Henderson, 1984). Results of Ostersen et al. (2011) indicated that the use of deregressed breeding values (dEBV) can increase the reliabilities of genomic breeding values in a 2-step approach. Hence, EBV were deregressed after the approach of Garrick et al. (2009) by removing the parent average (PA) and subsequent scaling of the remaining EBV. The direct genomic breeding values (DGV) were then calculated using genomic BLUP (gBLUP) where the genomic relationship matrix G was calculated in accordance to “Method 1” of VanRaden (2008) asin which M is the genotype matrix with dimensions number of individuals (n) by the number of loci (m) and M is calculated as M = Z − 2(pk − 0.5), with the marker matrix Z containing the elements −1 and 1 for homozygous and 0 for heterozygous genotypes, respectively, and pk is the base allele frequency of locus k calculated according to the approach of Gengler et al. (2007), using all animals remaining after quality check for estimation. Subsequently, G was combined to some proportion with the numerator relationship matrix A with

The proportions of G and A are hereby based on the experience of the authors, aiming for an improved numerical stability. This step was considered for the evaluation because similar approaches are frequently applied in commercial breeding and derived genomic reliabilities should reflect a practical application best possible. The model used to predict the DGV was thenin which y is a vector with dEBV of calibration animals, μ is the fixed mean, g is the random effects vector of DGV of all animals as elements, and e is the random residual term. Prediction of was done using the gBLUP formula,in which is the vector of estimated breeding values, σa2 is the additive genetic variance, V−1 is the inverse of the variance–covariance matrix of y, and is the generalized least squares estimate of μ as

The variance–covariance matrix was calculated asin which σe2 is the residual error variance and W is a diagonal matrix of reciprocal weights from the deregression procedure for the respective elements in y. These weights account for heterogeneous residual variances, which are a result of deregressing EBV with nonuniform reliabilities (Garrick et al., 2009). The so-calculated DGV were subsequently blended with the conventional breeding value to obtain the genomically enhanced breeding values (GEBV). If the true breeding value is assumed to be distributed as multivariate normal together with the estimated breeding value, then the covariance can be written aswith the true breeding value a, the estimated breeding value , the additive genetic variance σa2, and the reliabilities of the estimated breeding value r2 (VanRaden et al., 2009). In the case of multiple sources of information (DGV, PA, and EBV based on only a subset of animals), the resulting multivariate distribution has the following structure (Edel et al., 2010):

In accordance with VanRaden et al. (2009) and Edel et al. (2010), the variance–covariance matrix V can be written aswith reliabilities of DGV equal to rDGV2, reliabilities equal to EBV to rEBV2, and reliabilities of PA equal to rPA2. The EBV in this distribution, the so-called subset EBV, is a conventional BLUP EBV obtained by including only the pedigree of genotyped animals for its estimation (Edel et al., 2011). The resulting reliabilities of this subset EBV can then be interpreted as the covariance between the DGV and the PA. The weights for index combination of information sources were calculated byin which b is the vector of weighting coefficients and V−1 is the inverse of V. The GEBV was then calculated by weighted combination as

Validation

The focus of our GS implementation study was on comparing 3 different validation approaches. Before an actual GS implementation, calculations for assessing the monetary aspects are usually applied and include several variables (e.g., Tribout et al., 2013). The expected benefit thereby largely depends on the increase of genomic reliabilities compared with the reliabilities of conventional PA. Therefore, it appeared relevant to examine differences of the validation-method in the displayed genomic reliabilities, because such differences may end up in opposing economic assessments and, of course, would alter an ongoing routine evaluation. Therefore, we first calculated the model reliabilities from inversion of the mixed-model gBLUP equations (Mrode, 2005), hereafter referred to as “theoretical reliability.” Theoretical reliabilities of the GEBV were computed in accordance with Edel et al. (2010) as

Second, we applied the forward prediction approach of Mäntysaari et al. (2011), which is the official Interbull evaluation method and considers the timing of events in a real breeding scheme by regressing EBV with high reliabilities on the DGV at time of birth. And third, we applied the method of VanRaden et al. (2009), which combines the “best of both worlds” to some extent by transforming the observed reliability from forward prediction to the scale of model-based reliability.

The forward prediction validation follows the principle of Mäntysaari et al. (2011). In contrast to standard cross-validation procedures, forward prediction accounts for information available at the time of prediction and at the time of validation in a real breeding scheme. This allowance for the dimension of time was the main reason why we decided to apply forward prediction. Forward prediction requires 2 data sets, 1 full data set and 1 reduced data set. The full data set includes all available phenotypic information for breeding value estimation, whereas in the reduced data set, phenotypes are removed in such a way that a defined group of validation animals has only their parent’s information available for breeding value estimation. We defined all sows born since the year 2011 as the validation set. The so-defined calibration groups consisted of 1,393 sows and 337 boars for the fertility traits (LBP and PW) and 1,037 sows and 337 boars for the remaining traits. The resulting average reliabilities of calibration and validation groups of the respective traits are presented in Table 1. Subsequently, validation was done by applying the following regression model:in which y is a vector with dEBV based on the full data set, x is a vector with DGV or GEBV based on the restricted data set, b0 is the intercept, b1 is the slope of regression, and e is the vector of the random residual errors, whereas the corresponding weights from the deregression procedure were used to scale the residuals within which σe2 is the constant residual variance and w the vector of individual weights. The coefficient of determination of this regression is then the estimated reliability of the DGV or GEBV and the slope of regression could be interpreted as an estimate for the consistency of the DGV’s or GEBV’s distribution with the distribution of the response variable. The resulting forward prediction reliability is hereafter referred to as “validation reliability.”

We also applied the validation approach of VanRaden et al. (2009), which should allow for a better comparison of reliabilities from a forward prediction with theoretical reliabilities. The assumption made is that the theoretical PA and the validation PA should have the same reliabilities and the same scale. This approach consists of 2 steps. First, the validation reliabilities are divided by the average reliability of the response variable. In theory, this step accounts for the error induced by not using the true breeding value in forward prediction regression. Second, the difference between the theoretical PA reliabilities and the separately calculated validation PA reliabilities (results not shown) is added to the adjusted reliabilities from step 1. In theory, this fits the scale of validation reliabilities to the base of theoretical reliabilities. This reliability is hereafter referred to as “realized reliability.”

Population and Kinship Structure

It is well known that besides the size of the animal sample and its average reliability, further factors affect the outcome of genomic evaluations. Several findings from the literature (e.g., Su et al., 2010; Thomasen et al., 2011; Pszczola et al., 2012; Wientjes et al., 2013; Wu, 2014) indicate the importance of the kinship and population structure between the calibration and the validation sample as well as within the calibration sample. Therefore, we analyzed our sample with respect to its genomic kinship structure and population stratification. Average kinship was calculated within and between the respective subsets of validation and calibration animals by equating the respective diagonal and off-diagonal elements from the genomic relationship matrix. Principle components of the genomic relationship matrix were calculated by using the standard R function prcomp.


RESULTS

For all traits, theoretical reliabilities of both the DGV and the GEBV were larger than PA reliabilities (Table 2). They ranged from 0.42 to 0.54 for DGV and from 0.44 to 0.56 for GEBV whereas PA reliabilities ranged from 0.31 to 0.42. The use of genomic breeding value estimation, therefore, resulted in an average reliability increase of 31% for DGV and of 36% for GEBV when compared with traditional, pedigree-based reliabilities of selection candidates. Inclusion of polygenic information, as done by the described blending methodology, only slightly increased the reliabilities for all traits.


View Full Table | Close Full ViewTable 2.

Validation reliabilities (VR),1 realized reliabilities (RR),2 and theoretical reliabilities (TR)3 of the parent averages (PA),4 the direct genomic breeding values (DGV),5 and the genomically enhanced breeding values (GEBV)6

 
PA
DGV
GEBV
Trait TR VR RR TR VR RR TR
Number of piglets born alive 0.38 0.18 0.52 0.54 0.16 0.48 0.56
Proportion of piglets weaned 0.31 0.07 0.37 0.47 0.07 0.39 0.48
Average daily gain 0.36 0.07 0.60 0.46 0.06 0.53 0.48
Feed conversion ratio 0.39 0.09 0.49 0.49 0.08 0.45 0.51
Lean meat content 0.41 0.05 0.38 0.53 0.05 0.38 0.56
Belly meat content 0.41 0.03 0.38 0.52 0.03 0.37 0.54
Eye muscle area 0.42 0.10 0.35 0.54 0.12 0.40 0.56
Fat-to-lean ratio 0.42 0.06 0.38 0.54 0.06 0.38 0.56
Carcass length 0.40 0.11 0.35 0.52 0.13 0.39 0.54
pH after 45 min 0.33 0.09 0.55 0.43 0.07 0.46 0.45
Drip loss 0.33 0.18 0.22 0.42 0.16 0.11 0.44
Intramuscular fat content 0.42 0.17 0.45 0.54 0.18 0.48 0.56
1Calculated using the forward-prediction method Mäntysaari et al. (2011).
2Calculated using the VanRaden et al. (2009) approach.
3Calculated using the prediction error variance approach (Mrode, 2005; Edel et al., 2010).
4The average reliability of the parental breeding value average. The PA hereby refers to the PA of the selection candidates.
5The average reliability of the DGV.
6The average reliability of the GEBV.

In direct comparison with the theoretical reliabilities for DGV and GEBV, resulting validation reliabilities were considerably lower, with 0.03 to 0.18 for DGV and 0.03 to 0.18 for GEBV. Average reliability of the response variable, used for calculation of the latter results, was rather low and ranged from 0.42 to 0.62 before deregression (EBV reliabilities; Table 1). Blending DGV with PA increased the reliabilities only for the traits EMA, CL, and IMF whereas the other traits had either equal or lower reliabilities compared with DGV reliabilities.

The realized reliabilities were, on average, slightly lower than theoretical reliabilities but greater than validation reliabilities. They ranged from 0.22 to 0.60 for the DGV and from 0.11 to 0.53 for the GEBV. Realized reliabilities matched well to theoretical reliabilities for the traits LBP, PW, and FCR but were lower for all the carcass traits. For the traits ADG and PH1, realized reliabilities were substantially greater than theoretical reliabilities. When comparing the realized DGV and GEBV reliabilities, blending was beneficial only for the traits CL, EMA, and IMF. The other traits had either an equal or lower realized reliabilities.

The age and sex distribution within the sample is presented in Fig. 1. Analysis of the kinship structure derived from the genomic relationship matrix showed calibration animals to be, on average, 0.08 related and the calibration and validation animals to be, on average, 0.05 related. Visual inspection of the first 2 principle components indicates that the top 10 insemination boars (sires) match with the sample subclusters (Fig. 2). Further examination suggests that the farm of origin, sex, and birth date were not very well distributed within the Cartesian plane (results not shown). The first 3 principle components of the kinship matrix explained approximately 30% of the total variance (13, 10, and 7%), and all subsequent principle components decreased relatively gradually from 4 to about 0% of explained variance (Fig. 3).

Figure 2.
Figure 2.

Scatterplot of the first 2 principle components (PC) of the genomic relationship matrix. Each dot represents a single animal (n = 2,013 in total) in the sample and each color corresponds to 1 of the top 10 insemination boars. All other boars are summarized within the magenta color. The percent values on the axis labels account for the amount of variance explained by the respective principle component of the genomic relationship matrix.

 
Figure 3.
Figure 3.

Cumulative proportion of explained variance of the first 500 principle components of the genomic relationship matrix in decreasing order.

 

DISCUSSION

The theoretical DGV and GEBV reliabilities of selection candidates in our study were found to be, on average, greater than the conventional PA reliabilities, which suggests considerable potential for the implementation of GS in the Bavarian herdbook. Our results seem hereby in coherence with results of Lillehammer et al. (2013) and Tusell et al. (2013), who indicated the potential of using boars and sows together in genomic evaluation by increasing the genetic progress and genomic reliabilities.

Nevertheless, our results indicate that the validation methods applied showed a relatively large effect on reliability levels: forward prediction reliabilities were found to be much lower than the conventional PA reliabilities whereas corresponding realized and theoretical reliabilities were substantially greater than the PA reliabilities. The very low reliabilities from forward prediction are hereby somewhat unexpected. A possible explanation for the observed result is that this method uses the regression of some proxy for the true breeding value on the genomic breeding value, which requires a high reliability of the proxy variable to be meaningful by itself. Regarding the validation sample in this study, it consisted of only a comparatively small number of sows (Fig. 1) with relatively low average reliability of the response variable (Table 1). Nevertheless, this sample seems realistic for a putative implementation of a GS system in pigs but, however, may not provide enough reliability of conventional breeding values to calculate a proxy variable that matches the requirements (Mäntysaari et al., 2011) of forward prediction. Further analysis indicated that even under more optimistic assumptions regarding the reliabilities of the validation group (by excluding sows with a dEBV reliability less than 0.30), the distinctive difference between the methods remains nearly the same (results not shown). Interestingly, the outcome of simulations we conducted in preparation of this study, considering the Bavarian herdbook population parameters, indicated, that true genomic reliabilities could be substantially underestimated by forward prediction (Gertz, 2015). Because this method was initially developed for genomic evaluation in cattle, where usually a reliable proxy and also many more validation animals are available, forward prediction seems to be neither a suitable method for validation nor appropriate for the assessment of economic aspects of GS in pigs.

In contrast to validation reliabilities, realized reliabilities are adjusted for the reliability of the proxy for the true breeding value as well as for differences in the scale of reliabilities between the full and reduced model (VanRaden et al., 2009) and should, therefore, show a better agreement with theoretical reliabilities. However, our results suggest that this was true for only a few traits in this study. For most traits, realized reliabilities did not match very well with theoretical reliabilities. Additionally, we found the traits ADG and PH1 behaving strangely in a sense that realized reliabilities were substantially greater than theoretical reliabilities. Such an outcome seemed not plausible and might indicate a conceptual weakness of method of VanRaden et al. (2009). A theoretical example may clarify this: if the average reliability of the proxy response variable (denominator) is equal to or even lower than the reliabilities obtained from forward prediction (enumerator), realized reliability (after division) would be close to or even above 1. Such a result obviously cannot be true and is “misadjusted.” Therefore, we conclude that realized reliabilities can give a more realistic impression of the true genomic reliabilities than validation reliabilities, but they also include the risk of misinterpretation if the validation sample is of low average reliability. As our data suggest, such risk seems high in genomic evaluations, consisting of a lowly reliable validation group (Table 1; e.g., ADG and PH1). Therefore, realized reliabilities should not be the main criterion in a routine pig breeding system but may be useful as a statistical examination tool.

Theoretical model reliabilities are broadly used in conventional animal breeding and have a strong foundation in theory by using the matrix of prediction error variances (Mrode, 2005). This method has the advantage that population structure and prior selection is properly addressed by the relationship matrix and that no proxy response variable or additional correction is needed for the calculation of genomic reliabilities. Despite their theoretical foundation, model-based reliabilities are not error free, for example, if genomic relationship matrix is not properly scaled or if a nonrepresentative and/or a small sample from a selected population is used for genomic estimation (VanRaden et al., 2009; Gorjanc et al., 2015). Nevertheless, it appears that in this study, theoretical reliabilities represent the most consistent approach for the estimation of genomic reliabilities as no proxy variable is needed. Computational issues regarding the inversion of the variance–covariance matrix are not expected because sample sizes in pig breeding should not become as large as in cattle, at least not in the near future.

All of the here-regarded validation methods give slightly different answers and, at the same time, do have shortcomings, which may result in serious misinterpretations. The combined use of different validation approaches seems to be a valuable approach for reducing the risk of misjudgment in genomic evaluations, but nonetheless, dealing with the relatively low reliabilities of sows, especially by using them in the validation group, remains a challenging task for pig GS.

The analysis of population structure via principal components revealed some degree of stratification in our sample, pointing toward 2 more general issues: first, a chronological stratification, as it was necessary to draw an animal sample consisting of individuals from a relatively long time span and unequal periods in our case, and second, a spatial stratification, because the farms involved in this study were geographically separated and served by different AI stations and, therefore, slightly different managements. As a consequence of such structure, the number of shared haplotypes (Wientjes et al., 2013) or the “matching” of genotypes (Su et al., 2010) in genomic evaluations and, therefore, also genomic reliabilities will likely to be reduced. The analysis of the kinship structure further indicates a related but different issue: when comparing the resulting kinships in the calibration and validation sample with the results provided by Pszczola et al. (2012), the kinship structure in our sample seems close to be the opposite of the “ideal” for GS. Pszczola et al. (2012) proposed a low kinship within calibration but a high kinship between calibration and validation as an ideal scenario for GS. The relatively high kinship within the calibration sample may, therefore, be able to explain the relatively low GEBV reliabilities observed in this study. After blending the DGV with conventional PA, the reliabilities of the GEBV only marginally increased. Generally, blending is a technique that makes use of the PA information of ungenotyped animals, but in our case, we genotyped most of the female ancestors of the validation animals so that sires and dams of validation animals are included with information in the subset EBV and resulted in relatively high kinship. As a consequence, for a GS system where all base animals are genotyped, blending of DGV seems unnecessary.

The here-presented findings regarding the sample structure are suboptimal in view of an efficient genomic evaluation but illustrate the common challenges of GS in pigs, which seem not limited to only the Bavarian herdbook. Therefore, our results also indicate that further research on the management and optimization of kinship structure and population homogeneity may become valuable for pig GS in the future. A “strategic genotyping” approach seems hereby an attractive perspective. Interesting results in this regard were provided by Rincent et al. (2012), indicating that the reliabilities of genomic breeding values can be increased by using careful preselection of the calibration sample instead of using an increased calibration sample. However, this requires strategic sampling of DNA before the genotyping project starts. In our case, DNA was available for only historical AI boars and availability of DNA for phenotyped individuals was more of a limiting factor than genotyping costs. Another appealing approach would be the combination of different populations in 1 large animal sample. Recent endeavors of several German pig breeding organizations had aimed at such an increase by cooperation, but it turned out that the genetic connections between the different subpopulations were unfavorable in terms of stratification and relationships. Nevertheless, new results from Fangmann et al. (2015) and Xiang et al. (2016) may provide perspectives for such cooperation in the future by the use of multi-subpopulation reference sets. In addition, several studies suggest an increased efficiency of genomic evaluations in pigs by using the single step approach (e.g., Guo et al., 2015; Xiang et al., 2016). The main benefit from using single step approach would be the efficient inclusion of all nongenotyped animals of the pedigree (Fernando et al., 2014) and, therefore, a possibly improved genetic connection but also the reduced accumulation of errors from deregression and 2-step estimation.

Conclusion

We demonstrated that the 3 different validation methods lead to quite different results, which, to some extent, can be explained by the properties of a typical pig sample. Realized reliabilities and forward prediction are often used in cattle genomic evaluations but do not appear to be suitable validation methods in this study, because the low average reliabilities of the validation group make it difficult to obtain consistent results. Theoretical reliabilities seem to be the more consistent validation approach, mainly because this avoids the use of some proxy variable for the true breeding value. But nevertheless, all applied methods exhibit weaknesses, so that we suggest a combined use of validation methods to avoid potential misinterpretations. Our findings further indicate that the relationship structure of the animal sample was not ideal but worked out quite well. Therefore, it seems worth to examine the effects of sample composition and management more closely in the future. Nevertheless, overall results are encouraging and indicate that the application of GS for the Bavarian herdbook can result in a substantial increase of selection candidate reliabilities and that the inclusion of sows is a valid approach to increase the animal sample in pig GS, provided that there is systematic progeny testing for sows.

 

References

Footnotes


Comments
Be the first to comment.



Please log in to post a comment.
*Society members, certified professionals, and authors are permitted to comment.