selection on resilience improves disease resistance and tolerance to ...

Published online August 3, 2017

Selection on resilience improves disease resistance and tolerance to infections1 H. A. Mulder2 and H. Rashidi Wageningen University & Research Animal Breeding and Genomics, P.O. Box 338, 6700 AH, Wageningen, the Netherlands

Abstract: Response to infection in animals has 2 main mechanisms: resistance (ability to control pathogen burden) and tolerance (ability to maintain performance given the pathogen burden). Selection on disease resistance and tolerance to infections seems a promising avenue to increase productivity of animals in the presence of disease infections, but it is hampered by a lack of records of pathogen burden of infected animals. Selection on resilience (ability to maintain performance regardless of pathogen burden) may, therefore, be an alternative pragmatic approach, because it does not need records of pathogen burden. Therefore, the aim of this study was to assess response to selection in resistance and tolerance when selecting on resilience compared with direct selection on resistance and tolerance. Monte Carlo simulation was used combined with selection index theory to predict responses to selection. Using EBV for resilience in the absence of records for pathogen burden resulted in favorable responses in resistance and tolerance to infections, with higher responses in tolerance than in resistance. If resistance

and tolerance were unfavorably correlated, lower selection responses were obtained, especially in resistance. When the genetic correlation was very unfavorable, the selection response in tolerance became negative. Results showed that lower selection responses in resistance and tolerance were obtained when the frequency of disease outbreaks was 10% rather than 50% of the contemporary groups. The efficiency of selection on EBV for resilience compared with selection on EBV for resistance and tolerance was, however, not affected by the frequency of disease outbreaks. When records on pathogen burden were available, selection responses in resistance, tolerance, and the total breeding goal were 3 to 28%, 66 to 398%, and 2 to 11% higher, respectively, than when using the EBV for resilience, showing a clear benefit of recording pathogen burden. This study shows that selection on resilience is a pragmatic way of increasing disease resistance and tolerance to infections in the absence of records on pathogen burden, but recording pathogen burden would yield higher selection responses in resistance and tolerance.

Key words: breeding program, disease resilience, disease resistance, response to selection, selection index, tolerance to infections © 2017 American Society of Animal Science. All rights reserved. Introduction Infectious diseases in farm animals impose economic losses for the farmer, reduce animal welfare, and 1This research was part of the NematodeSystemHealth project,

financed by Marie Curie Initial Training Networks (FP7-People2010-ITN) and co-financed by TOPIGS Norsvin BV, the Netherlands, and the Dutch Ministry of Economic Affairs, Agriculture, and Innovation (public–private partnership “Breed4Food” code KB12-006.03-004-ASG-LR and KB-12-006.03-005-ASG-LR). 2Corresponding author: [email protected] Received February 16, 2017. Accepted June 7, 2017.

J. Anim. Sci. 2017.95:3346–3358 doi:10.2527/jas2017.1479

create public concerns. Therefore, selective breeding to increase disease resistance has been part of breeding programs for decades (Albers et al., 1987; Bishop and Morris, 2007; Morris et al., 2010; Bishop, 2012b). Animals respond to infections with 2 mechanisms: resistance and tolerance. Resistance is the ability of animals to restrict the invading pathogen’s life cycle. An animal with greater resistance will have a lower pathogen burden. Pathogen burden is the amount of pathogen in the animal’s body. Tolerance is the animal’s ability to minimize the detrimental impact of infection on performance. A tolerant animal maintains performance despite the pathogen burden. Tolerance, therefore, is measured as the regression of animal

3346

Genetic improvement in response to disease

performance on pathogen burden (Simms, 2000). If the pathogen burden is known, breeding values for resistance and tolerance can be estimated (Kause, 2011). In practice, however, pathogen burden is not recorded at the individual level, because it is laborious and costly and may vary over the time course of an infection (Doeschl-Wilson et al., 2012). Breeders, therefore, use resilience instead of tolerance or resistance. Resilience is the animal’s ability to maintain performance when exposed to infectious pathogens and does not need records of pathogen burden (Albers et al., 1987; Bisset and Morris, 1996). A resilient animal might, therefore, be resistant, tolerant, or both (Doeschl-Wilson et al., 2012). It is unknown, however, to what extent resistance and tolerance can be improved in breeding programs when selecting on EBV for resilience. The aim of this study was to assess response to selection in resistance and tolerance when using EBV for resilience when pathogen burden is not recorded compared with using EBV for resistance and tolerance when pathogen burden is recorded. Genetic parameters, breeding goals, and the proportion of infected animals were varied. Material and methods This study does not use any animal material, and therefore, no approval of animal care and use committee was required. Outline of Breeding Scheme A pig breeding scheme in a dam line that focused on increasing dam performance traits such as litter size, piglet birth weight, and survival was considered. Due to disease outbreaks such as porcine reproductive and respiratory syndrome (PRRS), the breeding goal may be extended with resistance and tolerance to infections to reduce the loss in performance due to infections. However, EBV for resistance and tolerance can be estimated only when a measure of pathogen burden is recorded. In the absence of a measure of pathogen burden, the EBV for resilience may be used. Because sow traits can be measured only in females, boars were selected based on half-sib information. Sows in the breeding nucleus were selected based on their own performance and half-sib information and were always kept under healthy conditions to mimic nucleus circumstances. Half-sib sows were, however, in diseased and healthy conditions, for example, in multiplier herds. To predict responses to selection in resistance and tolerance, Monte Carlo simulation was used to simulate true breeding values and phenotypes and breeding values were estimated with ASReml (Gilmour et al., 2009). Monte Carlo simulation was used because deterministic

3347

prediction equations are not available for this complex case of resistance and tolerance to infections. Monte Carlo Simulation A half-sib family structure was simulated, although in practice, small numbers of full-sibs may exist as well. In the base generation, 100 sires and 10,000 dams were simulated that were all unrelated. These 100 sires were mated each with 100 dams and each produced 1 female offspring, which resulted in 100 half-sib families. Breeding values were sampled for the base generation and the generation of offspring, whereas only the phenotypes of the offspring were used in breeding value estimation. A performance trait, that is, litter size, was simulated that was affected by pathogen burden when the animal was diseased and the trait pathogen burden, which was the inverse of disease resistance, that is, if pathogen burden was lower (higher) the animal was more (less) resistant. When animals were not infected (intercept), the performance Pint was a function of the mean performance in the absence of disease outbreaks (μint); a breeding value when the animal was not infected (Aint), that is, performance when pathogen burden was 0; and an environmental effect (Eint) following the classical quantitative genetic model (Falconer and Mackay, 1996): Pint = μint + Aint + Eint. 

[1]

When animals were infected, the performance of the animal depended not only on the components in Eq. [1] but also on the response of the animal (Psl) to pathogen burden (PPB), that is, the reaction norm on infection (slope), which indicated tolerance to infection (Kause, 2011): Pinf = Pint + Psl × PPB, 

[2]

in which the phenotypes for tolerance (Psl) and pathogen burden (PPB) were defined as Psl = μsl + Asl + Esl and 

[3]

PPB = μPB + APB + EPB, 

[4]

in which μsl is the average decrease in performance of infected animals compared with not-infected animals, because it is expected that, on average, animals decrease in performance during an infection; μPB is the average pathogen burden; Asl is the breeding value for the slope, that is, tolerance; Esl is the environmental effect for the slope; APB is the breeding value for pathogen burden;

3348

Mulder and Rashidi

and EPB is the environmental effect for pathogen burden. The 3 breeding values were assumed to follow a multivariate normal distribution (MVN)  a int  a  ~  sl  a PB  s Aint , Asl 0  s2Aint    MVN  0 , A ⊗  s2Asl   0    symmetric

sl

sl

int

PB

sl

s Eint , EPB  



s Esl , EPB  



s EPB  

in which eint, esl, and ePB are vectors with the environmental values Eint, Esl, and EPB for all animals; 2 2 2 I is the identity matrix; and sE , sE , sE , s Eint , Esl , s Eint , EPB , and s Esl , EPB are the environmental variances and covariances among the 3 environmental effects. Environmental covariances were assumed to be 0 to limit the number of possible scenarios to test and focus on the effect of additive genetic correlations. The phenotypic variances for Pint, Psl, and PPB were always 1. The offspring were randomly allocated to 100 contemporary groups (CG) of equal size. In the default situation, half of the CG were not infected and phenotypes for performance of offspring in these CG were simulated according to Eq. [1]. Phenotypes for performance of offspring in infected CG were simulated according to equations 2, 3, and 4; all infected CG had the same average pathogen burden. Parameters were chosen in such a way that complete resistance, that is, PPB = 0, did not occur. Furthermore, the default heritability for pathogen burden was set to 0.3 to mimic heritabilities found for viremia due to PRRS (Boddicker et al., 2012) and the default heritability for tolerance was 0.05 to mimic moderate genotype × environment interaction between performance of infected animals at average PPB and uninfected animals; that is, the genetic correlation is 0.52. Table 1 shows all the parameters used with their default and varied values; 100 replicates were simulated per situation. int

sl

PB

Parameter values Alternative – – – –

0.05

0, 0.1, 0.2, 0.3, 0.4, and 0.5 0, 0.1, 0.2, 0.3, 0.4, and 0.5 –

0.3

s2APB

PB

2

Basic 11 −1 4 0.3

s2Asl

PB

s Eint , Esl 0  s2Eint  eint   2  e  ~ MVN  0  , I ⊗  s Esl     sl  0  e PB    symmetric

Parameter1 μint μsl μPB

s2Aint

s Aint , APB   , s Asl , APB    s2APB  

in which aint, asl, and aPB are vectors with the breeding values Aint, Asl, and APB for all animals; A is the 2 2 2 numerator relationship matrix; and s Aint , s A , s A , s A , A , s A , A , and s A , A are the additive genetic variances and covariances among the 3 breeding values. The 3 environmental effects were assumed to follow a multivariate normal distribution int

Table 1. Parameter values used in the basic situation and in alternative situations

s2Pint s2Asl , and s2APB ,

1

rAsl , APB

0

v Aint

1

−0.75, −0.50, −0.25, 0.25, 0.50, and 0.75 0

vA l

1

0

v APB

−1

0

s

Number of sires 100 Number of dams 10,000 Number of half-sib progeny 100 Selected proportions sires 0.05 Selected proportions dams 0.2 Number of contemporary groups 100 Number of infected contemporary groups 50

– – – – – 40, 30, 20, and 10

1μ

int = mean performance in the absence of disease outbreaks; μsl = average decrease in performance of infected animals compared with notinfected animals; μPB = average pathogen burden; s2A = additive genetic 2 int variance in performance of not-infected animals; s A = additive genetic sl 2 variance in slope; s APB = additive genetic variance in pathogen burden; rAsl , APB = genetic correlation between slope and pathogen burden; v Aint = economic value of performance of not infected animals; v A l  = economic value of slope; v A = economic value of pathogen burden. s

PB

Breeding Value Estimation Scenarios Two scenarios were considered: 1) both performance and pathogen burden were recorded on each animal and 2) only performance was recorded. In the first scenario, pathogen burden was known for each diseased animal. Therefore, a bivariate model for performance and pathogen burden was used. Data from infected and uninfected CG was used in one analysis. Animals from uninfected CG had missing values for PPB. The model for performance (PPerf; Pinf for infected animals and Pint for uninfected animals) was an animal random regression model in ASReml version 3 (Gilmour et al., 2009) to estimate breeding values for Aint and Asl; the model for pathogen burden (PPB) was a simple animal model: 

PPerf = μ + b × PB + Aint + Asl × PPB + ePerf and [5]

3349

Genetic improvement in response to disease

PPB = μPB + APB + ePB, 

[6]

in which μ is the overall mean, b is the fixed reaction norm on pathogen burden (PB), and ePerf is the residual. Breeding values were assumed trivariate normally distributed as shown before for Eq. [1-4]. The residuals ePerf and ePB for infected animals were assumed bivariate normally distributed 0  se2Perf sePerf ,ePB   e Perf   I ~ MVN , ⊗ in  ,   e  2 0  symmetric s  Pb    e PB     2 2 which sePerf , sePB , and sePerf ,ePB are the residual varianc-

es of ePerf and ePB and the covariance between them, respectively. The environmental covariance was estimated, although the true covariance was 0. Modeling a heterogeneous residual variance for ePerf between infected and healthy CG was not feasible, but simulations showed that for the univariate model in Eq. [5], heterogeneity or homogeneity of residual variance hardly affected the accuracies of breeding values. In the second scenario, pathogen burden was considered to be not recorded, and therefore, in Eq. [5], pathogen burden was replaced by the average performance of the CG ( CG ) as an indirect measure of infection. Theoretically, a mixed model estimate of the CG effect is better, because the CG average may be inflated when some CG have higher average breeding values than others, for example, due to selection or when CG size is small. A mixed model is better able to disentangle the true environmental effects from genetic effects. In our simulation, the CG size was 100 animals and there was no selection. Therefore, differences in average breeding values between CG were very small. Using the simple average performance in each CG rather than a mixed model estimate was, therefore, appropriate. The model is 

PPerf = μ + bres × CG + Aint,2 + Ares × CG + ePerf, [7]

in which bres is the fixed reaction norm on CG ; Aint,2 is the breeding value for intercept, which is different than in Eq. [5]; and Ares is the breeding value for resilience to infection that is the slope of the reaction norm on the average of the CG. The breeding values Aint,2 and Ares were assumed bivariate normally distributed

0  s2Aint,2 s Aint,2 , Ares   a int   , in a  ~ MVN   0  , A ⊗  2  symmetric s      res  A res    2 which s2Aint,2 , s Ares , and s Aint,2 , Ares are the additive genetic

variances of Aint,2 and Ares and the covariance between them, respectively. Furthermore, the residual variance was assumed to be heterogeneous between infected and healthy CG.

Evaluation of Scenarios Per replicate, correlations between EBV and TBV as well as the correlations among EBV for boars based on their offspring and for sows based on their own performance under healthy conditions and the sows’ halfsibs in healthy and infected conditions were estimated. The average correlations and their SD were calculated across the 100 replicates. These average correlations were the inputs for response to selection calculations. Response to Selection The breeding goal (H) was to increase performance in CG with and without infection, and therefore, the aim was to increase performance in the absence of disease outbreaks (Aint), to increase tolerance to infections (Asl) , and to decrease pathogen burden (APB), that is, to increase resistance: H = v Aint Aint + v Asl Asl + v APB APB = v′a ,

[8]

in which v A is the economic value to increase performance in the absence of infection, v A is the economic value for tolerance to infections, v A is the economic value for pathogen burden or resistance, v is the vector with economic values, and a is the vector with breeding values. Three breeding goals were used: 1) increase only tolerance; 2) increase only resistance, that is, reduce pathogen burden; and 3) increase performance in the absence of disease outbreaks, increase tolerance, and increase resistance. In the last breeding goal, all traits had an equal absolute economic value (1.0, 1.0, and −1.0 for Aint, Asl, and APB, respectively), because a formal economic analysis was beyond the scope of this study. Selection was based on an index using EBV. In Aint ,  Asl , and  APB were used in index I1 scenario 1,     Aint,2 and  A A A Ares (a1′ = [ int sl PB ]). In scenario 2,    A A were used in index I2 (a2′ = [ int,2 res ]): int

sl

AB

I1 = b11 A + b12 A + b13 A = b1′a1 and[9] int

PB

sl

I2 = b21 A + b22 A = b2′a2. int,2

res

[10]

The optimal selection index weights in vectors b1 and b2 were calculated using selection index theory (Hazel, 1943): b = P−1Gv.

[11]

The P matrices P1 and P2 contain the variances and covariances between EBV in the selection indices I1 and I2. In this case, it was assumed that these EBV

3350

Mulder and Rashidi

Table 2. Correlations between EBV and true breeding values (TBV) for sires (SD between parentheses) based on 100 half-sib offspring when pathogen burden heritability varied for scenarios in which the pathogen burden (PB) was known or unknown. When PB was known, correlations between EBV and TBV for PB ( rA A ), intercept ( rA A ), and slope ( rA A ); TBV for PB and EBV for intercept ( rA A ); and TBV for PB and EBV for slope ( rA A ) were calculated. When PB was unknown, correlations between EBV and TBV for intercept ( rA A ), TBV for PB and EBV for intercept ( rA A ), TBV for slope and EBV for resilience ( rA A ), and TBV for PB and EBV for resilience ( rA A ) were calculated. Default values used in the simulation can be found in Table 1 PB PB

int PB

sl sl

int int

sl PB

int ,2 int

int ,2 PB

res sl

res PB

Pathogen burden heritability 0.0 0.1 0.2 0.3 0.4 0.5

Pathogen burden known

rA A

rA A

PB PB

rA A

int int

-0.04 (0.12) 0.75 (0.04) 0.84 (0.03) 0.89 (0.02) 0.92 (0.02) 0.94 (0.01)

rA A

int PB

sl sl

0.90 (0.02) 0.90 (0.02) 0.90 (0.02) 0.90 (0.02) 0.90 (0.02) 0.90 (0.02)

Pathogen burden unknown

0.90 (0.02) 0.90 (0.02) 0.90 (0.02) 0.90 (0.02) 0.90 (0.02) 0.90 (0.02)

-0.02 (0.10) -0.02 (0.09) 0.02 (0.09) 0.00 (0.10) 0.00 (0.10) 0.00 (0.10)

were scaled toward a variance of 1. The covariances are then equal to the correlations between EBV: é 1 r Aint ,  Asl ê ê 1 P1 = ê ê êsymmetric ë

ù r Aint ,  APB ú ú r Asl ,  APB ú ú 1 úû

é ù 1 r Aint,2 ,  Ares ú P2 = êê ú, 1 ûú ëêsymmetric

and

[12]

[13]

rA ,  rA ,  in which r are the correlations beA , A ,  A A , and     is the correlation tween Aint , Asl , and APB and r A , A Aint,2 and  Ares . The G matrices G1 and G2 conbetween  Aint ,  Asl , and  APB tain the covariances between the EBV    A A or int,2 and res with the TBV Aint, Asl, and APB in the breeding goal. Because the EBV were standardized with a variance of 1, matrices G1 and G2 were calculated as int

sl

int

sl

PB

PB

int ,2

s  r A A  int int Aint sA G 1 =  r  Asl Aint int  r s  APB Aint Aint

r s Asl A A int sl

r s Asl A A sl sl

r s Asl A A PB sl

res

r s  Aint APB APB    [14] r s Asl APB APB   r s A APB APB PB 

and  r s Aint A A G 2 =  int,2 int r s   Ares Aint Aint

r s Asl A A int,2 sl

r s Asl A A res sl

r s  Aint,2 APB APB   . [15] r s Ares APB APB  

Note that in the case of I1 and without scaling the EBV toward a variance of 1, the index weights in vector b1 were equal to the economic values (Schneeberger et al., 1992). The selection responses for trait j, that is, Aint, Asl, and APB, were calculated as ΔGj = (Rs,j + Rd,j)/(Ls + Ld),

[16]

rA A

sl PB

0.00 (0.10) -0.01 (0.10) 0.00 (0.10) 0.01 (0.10) -0.01 (0.10) 0.00 (0.10)

rA A

int ,2 int

0.73 (0.05) 0.72 (0.05) 0.70 (0.05) 0.69 (0.06) 0.67 (0.06) 0.64 (0.06)

rA A

int ,2 PB

0.01 (0.10) -0.19 (0.10) -0.28 (0.09) -0.34 (0.08) -0.38 (0.08) -0.42 (0.09)

rA A

res sl

0.86 (0.03) 0.82 (0.03) 0.79 (0.04) 0.76 (0.04) 0.73 (0.05) 0.71 (0.05)

rA A

res PB

0.00 (0.10) 0.28 (0.09) 0.39 (0.10) 0.47 (0.07) 0.51 (0.08) 0.56 (0.08)

in which Rs,j and Rd,j are the genetic selection differentials for sires and dams, respectively, and Ls and Ld are the relative generation intervals of sires and dams, respectively. The aim was to simulate a pig breeding program for a dam line based on a sib testing scheme, although for boars, the EBV in the Monte Carlo simulation were based on offspring, that is, a progeny testing scheme. However, in this simplified case, in the absence of the Bulmer effect (Bulmer, 1971), a sib testing scheme and a progeny scheme would yield equal selection responses when Ls is set to 2 and Ld is set to 1, because the accuracy based on half-sibs is exactly half of the accuracy with half-sib offspring. The genetic selection differentials were calculated as Rj = ib′gj/σI,

[17]

in which i is the selection intensity, gj is column j of G1 or G2, and σI = (b′Pb)1/2 is the SD of the index. Selection intensities were calculated assuming an infinite population of selection candidates without correction for correlated index values among relatives (Hill, 1976; Meuwissen, 1991). The selected proportions in boars and sows were assumed to be 5 and 20%, respectively. Selection responses are presented in genetic SD to facilitate comparison across traits. Results Correlations between Estimated and True Breeding Values Correlations between EBV and TBV of boars are shown in Table 2. The correlation between EBV and TBV for the pathogen burden was 0 when the heritability of the pathogen burden was 0 and increased by increasing the heritability of the pathogen burden.

Genetic improvement in response to disease

3351

Figure 1. Genetic gain in the slope (sl) and pathogen burden (PB) in genetic SD for each trait after 1 generation of index selection in sib testing schemes whether PB phenotype was either known or unknown and when the breeding goal was to increase tolerance. Economic values for the intercept (int), slope, and pathogen burden were v Aint = 0, v Asl = 1, and v APB = 0, respectively. Different values of pathogen burden heritability (0, 0.1, 0.2, 0.3, 0.4, and 0.5; panel A) and tolerance heritability (0, 0.05, 0.1, 0.2, 0.3, 0.4, and 0.5; panel B) were simulated. Default values used in the simulation can be found in Table 1.

When the pathogen burden was recorded, correlations Aint and Aint) between EBV and TBV for the intercept (  Asl and Asl) were about 0.9 and did and for the slope (  not change when the heritability of the pathogen burden was increased. However, when the pathogen burden was not known and breeding values were estimated, Aint,2 correlations between EBV and TBV for intercept (   A and Aint) and for resilience and the slope ( res and Asl) decreased with increasing heritability of the pathogen burden. In addition, the correlation between the EBV for resilience and the TBV for the pathogen burden Ares and APB, respectively) increased with increasing ( heritability of the pathogen burden. The key message is that the EBV for resilience is correlated both with the TBV for resistance and the TBV for tolerance. The Effect of Genetic Variance in Resistance and Tolerance on Selection Responses Figures 1, 2, and 3 show selection responses in the slope (tolerance) and in the pathogen burden (resistance) as a function of the heritability of the pathogen burden and slope for different breeding goals. Table 3 shows the effect of measuring the pathogen burden on the increase in response in the breeding goal. If the breeding goal was to select only on tolerance (Fig. 1) or only on the pathogen burden (Fig. 2), there was no response in the pathogen burden or slope if the pathogen burden was recorded and used in breeding value

estimation. This is expected, because the genetic correlation between the pathogen burden and slope was 0. However, when selection was on the EBV for resilience because the pathogen burden was not recorded, both the slope and pathogen burden responded (Fig. 1 and 2), because in both cases, the same animals were selected based on the EBV for resilience. When selection was on resilience and the heritability of the pathogen burden increased, the response in the pathogen burden increased (in absolute terms) at the cost of a lower response in the slope. When selection was on resilience and the heritability of the slope increased, the response in the slope increased at the cost of a lower response in the pathogen burden. In general, when selecting on resilience, the response in the slope was higher, that is, between 0.7 and 1.0 genetic SD, than for the pathogen burden, that is, between 0.2 and 0.7 genetic SD. In other words, selection on resilience places greater selection pressure on the slope than on the pathogen burden, given the set of genetic parameters used. As a consequence, measuring the pathogen burden and using EBV for resistance and tolerance compared with selection on EBV for resilience increased response in the pathogen burden much more (maximum 398%) than in the slope (maximum 27%; Table 3). If the breeding goal contained all 3 traits (Fig. 3), the selection response in tolerance (slope) was higher and the selection response in resistance (pathogen burden) was lower when using the EBV for resilience than when using EBV for the

3352

Mulder and Rashidi

Figure 2. Genetic gain in the slope (sl) and pathogen burden (PB) in genetic SD for each trait after 1 generation of index selection in sib testing schemes whether PB phenotype was either known or unknown and when the breeding goal was to increase resistance (decrease PB). Economic values for the intercept (int), slope, and pathogen burden were v Aint = 0, v Asl = 0, and v APB = −1, respectively. Different values of pathogen burden heritability (0, 0.1, 0.2, 0.3, 0.4, and 0.5; panel A) and tolerance heritability (0, 0.05, 0.1, 0.2, 0.3, 0.4, and 0.5; panel B) were simulated. Default values used in the simulation can be found in Table 1.

slope and pathogen burden. When the pathogen burden was measured and breeding values for resistance and tolerance were estimated, the response in the breeding goal increased by 1.8 to 11% compared with using the EBV for resilience when the pathogen burden was not measured. In conclusion, given the set of genetic parameters used, when the EBV for resilience is used in index selection when the pathogen burden is not recorded, high selection response in tolerance and moderate selection responses in resistance can be achieved. Measuring the pathogen burden increases selection responses in resistance more than in tolerance. The Effect of Genetic Correlations on Selection Responses Figure 4 shows the effect of the genetic correlation between the slope and pathogen burden. A negative genetic correlation is favorable and a positive correlation is unfavorable. When the genetic correlation was negative, that is, −0.75, selection responses were very similar whether the pathogen burden was known or unknown. When the genetic correlation increased, selection responses decreased or became unfavorable, especially the response in the pathogen burden when selecting on resilience. When using the EBV for resilience, selection responses in the slope were higher than in the pathogen burden, as seen before. With in-

creasing genetic correlation, the difference in selection responses with or without pathogen burden increased for all breeding goals. A peculiar result occurred when the breeding goal was solely to decrease the pathogen burden (Fig. 4C) and selection was on the EBV for resilience. In this case, the selection response in the slope suddenly became negative when the genetic correlation was 0.75. The direction of selection on the EBV for resilience suddenly changed from selecting the animals with the highest EBV to selecting the animals with the lowest EBV. When the genetic correlation between the pathogen burden and the slope was very high and positive, the animals with genetically the lowest pathogen burden, that is, the breeding goal, tended to have lower slopes as well. In other words, if animals would have been selected on the highest EBV for resilience, the genetic response in the pathogen burden would have become positive and undesirable given the breeding goal. The sudden change in breeding goal was not a numerical artifact but a sign of lack of genetic variation to simultaneously improving both resistance and tolerance when the genetic correlation between both of them is high and unfavorable. In conclusion, the genetic correlation between the slope and pathogen burden has a high impact on the selection responses in resistance and tolerance and selection on resilience may lead to an unfavorable response in resistance or tolerance depending on the breeding goal.

3353

Genetic improvement in response to disease

Figure 3. Genetic gain in the slope (sl) and pathogen burden (PB) in genetic SD for each trait after 1 generation of index selection in sib testing schemes whether PB phenotype was either known or unknown and when the breeding goal was to increase performance in the absence of disease outbreaks, to increase tolerance and to increase resistance (decrease PB). Economic values for the intercept (int), slope, and pathogen burden were v Aint = 1, v Asl = 1, and v APB = −1, respectively. Different values of pathogen burden heritability (0, 0.1, 0.2, 0.3, 0.4, and 0.5; panel A) and tolerance heritability (0, 0.05, 0.1, 0.2, 0.3, 0.4, and 0.5; panel B) were simulated. Default values used in the simulation can be found in Table 1.

The Effect of the Proportion of Animals Infected on Selection Responses

Table 3. The increase (%) in selection response in the breeding goal when measuring the pathogen burden and having EBV for resistance and tolerance compared with breeding on an EBV for resilience when varying the pathogen burden heritability or the tolerance heritability for 3 different breeding goals (increase all traits performance, tolerance, and resistance; increase tolerance; or increase resistance) Changed parameter Pathogen burden heritability

Tolerance heritability

1Default

Value1 0.0 0.1 0.2 0.3 0.4 0.5 0.0 0.05 0.1 0.2 0.3 0.4 0.5

All traits 1.8 2.2 4.0 8.2 10.3 11.0 2.3 6.5 9.3 8.2 7.9 7.5 7.4

Breeding goal Tolerance 4.2 10.0 14.3 18.1 23.6 27.6 0.0 18.1 8.1 5.8 4.1 3.4 2.9

Resistance 0.0 146.3 110.5 92.1 77.3 66.8 0.0 88.2 160.6 220.7 299.7 348.1 398.1

values used in the simulation can be found in Table 1.

In the previous section, it was assumed that half of the CG were infected and the other half were not, but fortunately, disease outbreaks are often less frequent than 50%. Therefore, the percentage of CG infected was varied between 10 to 50% (Fig. 5). As expected, selection responses in the slope and pathogen burden were higher when the proportion of infected CG increased. If 10% of the CG was infected, the responses were about 57 to 73% of the responses when 50% of the CG were infected. The figures show that the proportion of infected CG hardly affected the efficiency of selection on EBV for resilience compared with selection on EBV for the pathogen burden and slope; it affected only the absolute selection responses. Discussion Model and Results The aim of the current study was to assess response to selection in resistance and tolerance when selection was on resilience (unknown pathogen burden) compared with when selection was on resistance and tolerance (known pathogen burden). Using the EBV for resilience resulted in favorable selection responses in resistance and tolerance, but it was not as effective

3354

Mulder and Rashidi

Figure 4. Genetic gain in the slope (sl) and pathogen burden (PB) in genetic SD for each trait after 1 generation of index selection in sib testing schemes whether PB phenotype was either known or unknown as a function of the genetic correlation between the slope and pathogen burden ( rA , A ) for sl PB 3 breeding goals (panels A, B, and C). Economic values for the intercept (int), slope (sl), and pathogen burden (PB) were v Aint = 1, v Asl = 1, and v APB = v v v v v v −1, respectively (panel A); Aint = 0, Asl = 1, and APB = 0, respectively (panel B); and Aint = 0, Asl = 0, and APB = −1, respectively (panel C). Default values used in the simulation can be found in Table 1.

as selection on EBV for resistance and tolerance. Selection responses in resistance were smaller than in tolerance. The selection responses in resistance and tolerance depended on the genetic variances in these traits as well as the genetic correlation between the 2. However, the comparison of index selection using the EBV for resilience with index selection on EBV for resistance and tolerance was not much affected by the genetic variances in the pathogen burden or slope. This indicates that selection on the EBV for resilience was quite robust and yielded favorable responses at least in tolerance and mostly also in resistance. To our knowledge, this is the first study investigating the efficiency of selection on resilience on the underlying genetics of resistance and tolerance. The results indicate that selection on an EBV for resilience can be considered to be index selection for resistance and tolerance. In this study, a combination of Monte Carlo simulation and selection index theory was used to predict responses to selection. This was a fast and accurate way of predicting selection responses, because deterministic predictions for elements in the P and G matrices were not needed, which are expected to be complex because of involvement of a product of tolerance and pathogen burden. One generation of selection was simulated and gametic phase disequilib-

rium was not accounted for (Bulmer, 1971). Ignoring gametic phase disequilibrium or the so-called Bulmer effect, however, is expected to not affect the comparison of using EBV for resilience with EBV for tolerance and resistance, because the breeding scheme was considered constant in this study. Accounting for the Bulmer effect is especially important when comparing different breeding schemes such as sib testing, progeny testing, or genomic selection (Mulder and Bijma, 2005; Van Grevenhof et al., 2012). In this study, the focus was on resistance and tolerance to one disease that affected the CG mean. In real life, many diseases or other environmental disturbances could affect the CG mean. Therefore, breeding on an EBV for resilience will target general disease resistance and tolerance, and therefore, it is expected that selection responses in specific resistance and tolerance to a certain disease will be lower than those found in this study, because general disease resistance and tolerance is expected to explain only part of the total genetic variation in resistance and tolerance to a specific disease. Furthermore, the effectiveness of selection on an EBV for resilience depends on the incidences of different diseases and other disturbances, the genetic architecture of general and specific disease resistances (see also discussion on “Selection on Resilience to

Genetic improvement in response to disease

3355

Figure 5. Genetic gain in the slope (sl) and pathogen burden (PB) in genetic SD for each trait after 1 generation of index selection in sib testing schemes whether PB phenotype was either known or unknown as a function of the percentage of infected contemporary groups. Economic values for the intercept (int), slope (sl), and pathogen burden (PB) were v Aint = 1, v Asl = 1, and v APB = −1, respectively (panel A); v Aint = 0, v Asl = 1, and v APB = 0, respectively (panel B); and v Aint = 0, v Asl = 0, and v APB = −1, respectively (panel C). Default values used in the simulation can be found in Table 1.

Infections”), presence of complete resistance, and whether the true tolerance reaction norm is linear or curvilinear. This study makes a first attempt to study the effectiveness of selection on an EBV for resilience, and results show that the effectiveness depends on the genetic parameters of resistance and tolerance. In this study, selection responses after 1 generation of selection were considered. Responses in resistance would affect the amount of pathogen burden shed to group mates. As a consequence, selection for improved resistance will also improve the performance of lessresistant animals, because the probability of getting infected decreases. Accounting for these indirect benefits of selection on resistance would require a genetic– epidemiological approach such as in Bishop and Stear (1997) for nematode resistance in sheep. Therefore, the benefits of selection on resistance are expected to be larger than those found here in this study due to ignoring the epidemiological consequences of selection. Selection on tolerance does not stop the spread of the infection and is, therefore, not suitable to control highly infectious pathogens and zoonotic diseases (Bishop, 2012a). On the other hand, selection on resistance may be less effective, because it would put selection pressure on the pathogen, which may overcome the resistance mechanism. Selection on tolerance would not put selection pressure on the pathogen. Therefore, re-

sistance and tolerance may have different effects on the epidemiology of infectious diseases and host–parasite coevolution (Råberg et al., 2007, 2009). Based on results in this study, selection on resilience targets both tolerance and resistance but targets tolerance more than resistance. To study the long-term effects of selection on resistance, tolerance, or resilience, a genetic–epidemiological approach is required. The Value of Recording the Pathogen Burden Until recently, the genetics of tolerance to infections has not attracted much attention in animal breeding, whereas breeding for resistance has been on the research agenda in animal breeding for several years (Doeschl-Wilson et al., 2012). The main hurdle with breeding for tolerance is the need for a measure of the pathogen burden. Our study shows that measuring the pathogen burden increases the response in resistance by 66 to 398% (Table 3), whereas the response in tolerance increases by 3 to 28%. The increase in selection response in the breeding goal, though, is limited: 2 to 11%. These results show that measuring the pathogen burden can increase genetic gain in breeding programs compared with selection on EBV for resilience. Although measuring the pathogen burden has clear advantages for genetic improvement, it is often difficult

3356

Mulder and Rashidi

to obtain measures of pathogen burden in commercial animals, because animals are infected at different time points, registration of diseases is limited, measuring the pathogen burden is costly and laborious, and the pathogen burden changes during the course of infection (Doeschl-Wilson et al., 2012). In some cases, it is feasible to obtain indicators for pathogen burden such as fecal egg count for nematode infections in sheep (Albers et al., 1987; Stear et al., 1995; Bishop et al., 1996) or somatic cell count in milk as an indication for the severity of mastitis infection (Detilleux et al., 2012). Even though the pathogen burden may be difficult to obtain, recording which animals are infected and which are not infected would be already of great value, especially for endemic diseases, such as mastitis. Presence or absence of infection would allow estimating breeding values for resistance as well as for tolerance which is the change in performance when animals are infected. With challenge experiments, for instance, the PRRS host consortium trials at Kansas State University (Lunney et al., 2011; Rowland et al., 2012), it is feasible to obtain measures of viremia at different points after infection (Boddicker et al., 2012; Islam et al., 2013). In other words, these measures of viremia can be used as the pathogen burden to estimate genetic variation in tolerance. Even though many data have been recorded on these infected pigs, it has proven be difficult to find genetic variation in tolerance (Lough et al., 2017). One reason might be that viremia in blood is not a good measure of the pathogen burden. Another important issue is unbiased estimation of the intercept of the reaction norm when animals are not infected (Kause, 2011; Doeschl-Wilson et al., 2012). Therefore, the experiment should ideally contain both relatives that are infected and relatives that are not infected. Although challenge experiments are very useful for research on genetics of disease resistance and tolerance to infections, the value for commercial breeding programs may be limited because the challenge environment may still be very different from the commercial environments. Selection on Resilience to Infections Although selection for increased tolerance seems to be still challenging because of a lack of records of pathogen burden, our study shows that using EBV for resilience is an effective way to increase tolerance and resistance by selection, provided that both are not strongly unfavorably correlated. This is in contrast to Albers et al. (1987), who concluded that the heritability of resilience is too small to obtain direct selection responses to mass selection. Indeed, mass selection will yield small selection responses in resilience (Kolmodin and Bijma, 2004; Sae-Lim et al., 2015), but using in-

formation of half-sibs that are infected, for example, in multiplier herds with PRRS outbreaks and half-sibs that are not infected can greatly increase the selection responses in resilience, as observed in this study. An important drawback of using the EBV for resilience is that resilience is a “black box”: the emphasis on tolerance and resistance depends on the parameters. Furthermore, obtaining correct measures of the CG mean may be statistically challenging and may lead to biased estimates of the genetic variance in resilience, especially to disentangle the genetic trend from the CG means (Knap and Su, 2008). If CG are large, which is generally the case in pig breeding, bias is expected to be small or absent. Mixed model estimates of CG means could be used (Rashidi et al., 2014; Silva et al., 2014). In previous studies, we showed that such CG effects could be used as well to detect disease outbreaks (Mathur et al., 2014; Rashidi et al., 2014). Another drawback is that the EBV for resilience will mainly pick up resistance and tolerance to epidemic diseases. For endemic diseases, the approach is less useful because there are animals continuously infected and, therefore, the CG mean is not a good indicator for presence of infections. In those cases, presence or absence of infection at the animal level could be used as a covariate in the random regression model. An important advantage of using the EBV for resilience is that it targets general resilience (Guy et al., 2012). Multiple diseases may decrease the CG means. Therefore, selection on the EBV for resilience will target general disease tolerance and resistance rather than specific disease resistance or tolerance. In addition to diseases, there may be other environmental factors that decrease performance, such as heat stress (Bloemhof et al., 2008) or seasonality (Sevillano et al., 2016). It is likely that general mechanisms related to dealing with stress situations are involved. In laboratory species, heat shock proteins are found to control effects of stress (Queitsch et al., 2002; Sangster et al., 2008). Genomewide associations can help unraveling the genetic background of resilience (Silva et al., 2014; Sell-Kubiak et al., 2015), and genomic selection can help to further increase response to selection in resilience (Silva et al., 2014; Mulder, 2016). In this study, it was shown that using EBV for resilience in the absence of pathogen burden records led to favorable responses in resistance and tolerance to infections. The selection responses in resistance and tolerance depended on the heritabilities of resistance and tolerance and the genetic correlation between resistance and tolerance. If resistance and tolerance were unfavorably correlated, responses decreased, especially in resistance. Although using EBV for resilience resulted mostly in favorable responses in resistance and tolerance, more genetic gain could be achieved when the pathogen burden is recorded.

Genetic improvement in response to disease

Literature Cited Albers, G. A. A., G. D. Gray, L. R. Piper, J. S. F. Barker, L. F. Lejambre, and I. A. Barger. 1987. The genetics of resistance and resilience to Haemonchus contortus infection in young Merino sheep. Int. J. Parasitol. 17:1355–1363. doi:10.1016/0020-7519(87)90103-2 Bishop, S. C. 2012a. A consideration of resistance and tolerance for ruminant nematode infections. Front. Genet. 3:168. doi:10.3389/ fgene.2012.00168 Bishop, S. C. 2012b. Possibilities to breed for resistance to nematode parasite infections in small ruminants in tropical production systems. Animal 6:741–747. doi:10.1017/S1751731111000681 Bishop, S. C., K. Bairden, Q. A. McKellar, M. Park, and M. J. Stear. 1996. Genetic parameters for faecal egg count following mixed, natural, predominantly Ostertagia circumcincta infection and relationships with live weight in young lambs. Anim. Sci. 63:423–428. doi:10.1017/S1357729800015319 Bishop, S. C., and C. A. Morris. 2007. Genetics of disease resistance in sheep and goats. Small Rumin. Res. 70:48–59. doi:10.1016/j. smallrumres.2007.01.006 Bishop, S. C., and M. J. Stear. 1997. Modelling responses to selection for resistance to gastro-intestinal parasites in sheep. Anim. Sci. 64:469–478. doi:10.1017/S1357729800016088 Bisset, S. A., and C. A. Morris. 1996. Feasibility and implications of breeding sheep for resilience to nematode challenge. Int. J. Parasitol. 26:857–868. doi:10.1016/S0020-7519(96)80056-7 Bloemhof, S., E. H. van der Waaij, J. W. M. Merks, and E. F. Knol. 2008. Sow line differences in heat stress tolerance expressed in reproductive performance traits. J. Anim. Sci. 86:3330–3337. doi:10.2527/jas.2008-0862 Boddicker, N., E. H. Waide, R. R. R. Rowland, J. K. Lunney, D. J. Garrick, J. M. Reecy, and J. C. M. Dekkers. 2012. Evidence for a major QTL associated with host response to porcine reproductive and respiratory syndrome virus challenge. J. Anim. Sci. 90:1733–1746. doi:10.2527/jas.2011-4464 Bulmer, M. G. 1971. The effect of selection on genetic variability. Am. Nat. 105:201–211. doi:10.1086/282718 Detilleux, J., L. Theron, J. M. Beduin, and C. Hanzen. 2012. A structural equation model to evaluate direct and indirect factors associated with a latent measure of mastitis in Belgian dairy herds. Prev. Vet. Med. 107:170–179. doi:10.1016/j.prevetmed.2012.06.005 Doeschl-Wilson, A. B., B. Villanueva, and I. Kyriazakis. 2012. The first step towards genetic selection for host tolerance to infectious pathogens: Obtaining the tolerance phenotype through group estimates. Front. Genet. 3:265. doi:10.3389/fgene.2012.00265 Falconer, D. S., and T. F. C. Mackay. 1996. Introduction to quantitative genetics. 4th ed. Pearson Education Limited, Essex, UK. Gilmour, A. R., B. J. Gogel, B. R. Cullis, and R. Thompson. 2009. ASReml user guide release 3.0. VSN Int., Hemel Hempstead, UK. Guy, S. Z. Y., P. C. Thompson, and S. Hermesch. 2012. Selection of pigs for improved coping with health and environmental challenges: Breeding for resistance or tolerance? Front. Genet. 3:281. doi:10.3389/fgene.2012.00281 Hazel, L. N. 1943. The genetic basis for constructing selection indexes. Genetics 28:476–490. Hill, W. G. 1976. Order statistics of correlated variables and implications in genetic selection programmes. Biometrics 32:889–902. doi:10.2307/2529272

3357

Islam, Z. U., S. C. Bishop, N. J. Savill, R. R. Rowland, J. K. Lunney, B. Trible, and A. B. Doeschl-Wilson. 2013. Quantitative analysis of porcine reproductive and respiratory syndrome (PRRS) viremia profiles from experimental infection: A statistical modelling approach. PLoS One 8:e83567. doi:10.1371/journal.pone.0083567 Kause, A. 2011. Genetic analysis of tolerance to infections using random regressions: A simulation study. Genet. Res. 93:291–302. doi:10.1017/S0016672311000176 Knap, P. W., and G. Su. 2008. Genotype by environment interaction for litter size in pigs as quantified by reaction norms analysis. Animal 2:1742–1747. doi:10.1017/S1751731108003145 Kolmodin, R., and P. Bijma. 2004. Response to mass selection when genotype by environment interaction is modelled as a linear reaction norm. Genet. Sel. Evol. 36:435–454. doi:10.1186/12979686-36-4-435 Lough, G., H. Rashidi, I. Kyriazakis, J. C. M. Dekkers, A. Hess, M. Hess, N. Deeb, A. Kause, J. K. Lunney, R. R. R. Rowland, H. A. Mulder, and A. Doeschl-Wilson. 2017. Use of multi-trait and random regression models to identify genetic variation in tolerance to porcine reproductive and respiratory syndrome virus. Genet. Sel. Evol. 49:37. doi:10.1186/s12711-017-0312-7 Lunney, J. K., J. P. Steibel, J. M. Reecy, E. Fritz, M. F. Rothschild, M. Kerrigan, B. Trible, and R. R. Rowland. 2011. Probing genetic control of swine responses to PRRSV infection: Current progress of the PRRS host genetics consortium. BMC Proc. 5(Suppl. 4):S30. doi:10.1186/1753-6561-5-S4-S30 Mathur, P. K., J. M. Herrero-Medrano, P. Alexandri, E. F. Knol, J. Ten Napel, H. Rashidi, and H. A. Mulder. 2014. Estimating challenge load due to disease outbreaks and other challenges using reproduction records of sows. J. Anim. Sci. 92:5374–5381. doi:10.2527/jas.2014-8059 Meuwissen, T. H. E. 1991. Reduction of selection differentials in finite populations with a nested full-half sib family structure. Biometrics 47:195–203. doi:10.2307/2532506 Morris, C. A., S. A. Bisset, A. Vlassoff, M. Wheeler, C. J. West, B. P. Devantier, and A. D. Mackay. 2010. Selecting for resilience in Romney sheep under nematode parasite challenge, 1994-2007. N. Z. J. Agric. Res. 53(3):245–261. doi:10.1080/00288233.201 0.500714 Mulder, H. A. 2016. Genomic selection improves response to selection in resilience by exploiting genotype by environment interactions. Front. Genet. 7:178. doi:10.3389/fgene.2016.00178 Mulder, H. A., and P. Bijma. 2005. Effects of genotype × environment interaction on genetic gain in breeding programs. J. Anim. Sci. 83:49–61. doi:10.2527/2005.83149x Queitsch, C., T. A. Sangster, and S. Lindquist. 2002. Hsp90 as a capacitor of phenotypic variation. Nature 417:618–624. doi:10.1038/ nature749 Råberg, L., A. L. Graham, and A. F. Read. 2009. Decomposing health: Tolerance and resistance to parasites in animals. Philos. Trans. R. Soc. Lond. B Biol Sci. 364:37–49. doi:10.1098/rstb.2008.0184 Råberg, L., D. Sim, and A. F. Read. 2007. Disentangling genetic variation for resistance and tolerance to infectious diseases in animals. Science 318:812–814. doi:10.1126/science.1148526 Rashidi, H., H. A. Mulder, P. Mathur, J. A. M. Van Arendonk, and E. F. Knol. 2014. Variation among sows in response to porcine reproductive and respiratory syndrome. J. Anim. Sci. 92:95–105. doi:10.2527/jas.2013-6889 Rowland, R. R., J. K. Lunney, and J. C. M. Dekkers. 2012. Control of porcine reproductive and respiratory syndrome (PRRS) through genetic improvements in disease resistance and tolerance. Front. Genet. 3:260. doi:10.3389/fgene.2012.00260

3358

Mulder and Rashidi

Sae-Lim, P., H. Mulder, B. Gjerde, H. Koskinen, M. Lillehammer, and A. Kause. 2015. Genetics of growth reaction norms in farmed rainbow trout. PLoS One 10:e0135133. doi:10.1371/ journal.pone.0135133 Sangster, T. A., N. Salathia, S. Undurraga, R. Milo, K. Schelienberg, S. Lindquist, and C. Queitsch. 2008. HSP90 affects the expression of genetic variation and developmental stability in quantitative traits. Proc. Natl. Acad. Sci. USA 105:2963–2968. doi:10.1073/ pnas.0712200105 Schneeberger, M., S. A. Barwick, G. H. Crow, and K. Hammond. 1992. Economic indices using breeding values predicted by BLUP. J. Anim. Breed. Genet. 109:180–187. doi:10.1111/j.1439-0388.1992.tb00395.x Sell-Kubiak, E., N. Duijvesteijn, M. S. Lopes, L. L. G. Janss, E. F. Knol, P. Bijma, and H. A. Mulder. 2015. Genome-wide association study reveals novel loci for litter size and its variability in a Large White pig population. BMC Genomics 16:1049. doi:10.1186/s12864-015-2273-y Sevillano, C. A., H. A. Mulder, H. Rashidi, P. K. Mathur, and E. F. Knol. 2016. Genetic variation for farrowing rate in pigs in response to change in photoperiod and ambient temperature. J. Anim. Sci. 94:3185–3197. doi:10.2527/jas.2015-9915

Silva, F. F., H. A. Mulder, E. F. Knol, M. S. Lopes, S. E. F. Guimarães, P. S. Lopes, P. K. Mathur, J. M. S. Viana, and J. W. M. Bastiaansen. 2014. Sire evaluation for total number born in pigs using a genomic reaction norms approach. J. Anim. Sci. 92:3825–3834. doi:10.2527/jas.2013-6486 Simms, E. L. 2000. Defining tolerance as a norm of reaction. Evol. Ecol. 14:563–570. doi:10.1023/A:1010956716539 Stear, M. J., S. C. Bishop, M. Doligalska, J. L. Duncan, P. H. Holmes, J. Irvine, L. McCririe, Q. A. McKellar, E. Sinski, and M. Murray. 1995. Regulation of egg production, worm burden, worm length and worm fecundity by host responses in sheep infected with Ostertagia circumcincta. Parasite Immunol. 17:643–652. doi:10.1111/j.1365-3024.1995.tb01010.x Van Grevenhof, E. M., J. A. M. Van Arendonk, and P. Bijma. 2012. Response to genomic selection: The Bulmer effect and the potential of genomic selection when the number of phenotypic records is limiting. Genet. Sel. Evol. 44:26. doi:10.1186/12979686-44-26