dynamic A deterministic model to optimize a multiple-trait selection scheme
A. D. Costard, Z. G. Vitezica, C. R. Moreno and J.-M. Elsen
J ANIM SCI 2009, 87:885-894.
doi: 10.2527/jas.2008-0898 originally published online November 21, 2008
The online version of this article, along with updated information and services, is located on
the World Wide Web at:
http://jas.fass.org/content/87/3/885
www.asas.org
Downloaded from jas.fass.org by guest on October 11, 2011
A dynamic deterministic model to optimize
a multiple-trait selection scheme1
A. D. Costard,*2 Z. G. Vitezica, C. R. Moreno,* and J.-M. Elsen*
*lInstitut National de la Recherche Agronomique, SAGA station damlioration gntique animale,
31326 Castanet-Tolosan, France; Universit de Toulouse; Institut National Polytechnique de Toulouse,
Ecole Nationale Vtrinaire de Toulouse; Unit Mixte de Recherche, 1289 Tandem, Tissus Animaux, Nutrition,
Digestion, Ecosystme et Mtabolisme; Ecole Nationale Suprieure Agronomique de Toulouse,
F-31326 Castanet-Tolosan Cedex, France; and lInstitut National de la Recherche Agronomique,
Unit Mixte de Recherche, 1289 Tandem, Tissus Animaux, Nutrition, Digestion, Ecosystme et Mtabolisme;
Chemin de Borde-Rouge, Auzeville, F-31326 Castanet-Tolosan, France
ABSTRACT: A mathematical approach was developed to model and optimize simultaneous selection on
2 traits, a quantitative trait with underlying polygenic
variation and a monogenic trait (e.g., resistance to a
disease). A deterministic model allows global optimization of the selection scheme to maximize the frequency
of the desired genotype for the monogenic trait, while
minimizing the loss of genetic progress on the polygenic
trait. An additive QTL or gene was considered. Breeding programs with overlapping generations, different se-
lection strategies for males and females, and assortative
mating were modeled. A genetic algorithm was used
to solve this optimization problem. This modeling approach may easily be adapted to a variety of underlying
genetic models and selection schemes. This model was
applied to an example where selection on the Prp gene
for scrapie resistance was introduced as an additional
selection criterion in an already existing dairy sheep
selection scheme.
Key words: genetic algorithm, global optimization, marker-assisted selection, selection scheme
2009 American Society of Animal Science. All rights reserved.
INTRODUCTION
J. Anim. Sci. 2009. 87:885894
doi:10.2527/jas.2008-0898
Larzul et al., 1997b; Pong-Wong and Woolliams, 1998)
have explored the interactions between selection methods and time horizon. Dekkers and van Arendonk (1998)
optimized selection on an identified QTL in a simple selection scheme (phenotypic selection, random mating)
using an optimal control approach (Lewis, 1986). The
method was extended by Chakraborty et al. (2002). The
benefit of optimal selection on a single identified QTL
and the differences between optimal and standard QTL
selection were evaluated by Dekkers and Chakraborty
(2001). Manfredi et al. (1998) optimized selection and
mating on a QTL for a sex-limited trait using sequential quadratic programming. In these studies, selection
was for a single quantitative trait affected by 1 identified QTL of meaningful size as well as polygenes. The
objective of this study is to optimize simultaneous selection on 2 traits, a monogenic trait controlled by an
identified QTL or gene and a second independent polygenic trait that is unaffected by the major locus with
a genetic algorithm (Holland, 1973; Goldberg, 1989;
Mongeau, 2003). A deterministic and dynamic model is
presented for breeding programs including overlapping
generations, different selection strategies for males and
females, and nonrandom mating.
Gene or QTL information could improve selection
efficiency and allow the implementation of markerassisted selection. Marker-assisted selection has been
proposed to improve classical selection schemes for
complex traits with low heritability or that are difficult
or expensive to measure (Lande and Thompson, 1990;
Meuwissen and van Arendonk, 1992; Dekkers and van
Arendonk, 1998). Strategies to use QTL or gene information in marker-assisted selection are mainly based
on indices combining the EBV for QTL and the EBV
for the residual polygenic effect. Marker-assisted selection increases the QTL frequency in the short-term,
but can result in less response in the long-term (Lande
and Thompson, 1990). Several studies (Gibson, 1994;
1
This work was supported by the European Union project QLRT2000-01733 Scrapiefreesheep and by the EU Network of Excellence, Neuroprion.
2
Corresponding author: [e-mail]
Received January 23, 2008.
Accepted November 18, 2008.
885
Downloaded from jas.fass.org by guest on October 11, 2011
886
Costard et al.
Figure 1.Selectionschemedescriptionwithindividualclasses(rectangles),selectionsteps(),andselectionproportionWsac (sex s = 1, 2; age
a = 1,...,as; category c = 1,2).
MATERIALS AND METHODS
Animal Care and Use Committee approval was not
obtained for this study because no animals were used.
Genetic Model
A population of infinite size with overlapping generations and different classes of animals defined by sex s (s
= 1 for males and s = 2 for females), age a, and EBV
categories c (elite, not elite) was considered. Animals
of identical sex and age may be classified in different
subsets according to their EBV mean and their use
as reproducers (e.g., the number of progeny per year).
The QTL or gene has 2 alleles (B and b). Genotypes
BB, Bb/bB, and bb are denoted as g1, g2, and g3,
respectively. The polygenic effects are assumed to follow the infinitesimal genetic model with variance 2.
The identified QTL or gene is unlinked with genes that
influence the quantitative trait. For generation t, each
class of animals has a QTL or gene effect with a genotype frequency of sactg, and a polygenic effect denoted
by the within class average EBV sactg. Selection can
involve the EBV of the quantitative trait, the identified
QTL or gene, or both of them. Random mating and
assortative mating according to genotype or to EBV
categories can be used. The deterministic and dynamic
model includes different equations that correspond to a
general scheme presented in Figure 1. The phenomena
underlying the evolution of the population are aging,
selection, and matings. The equations describing the
scheme are as follows:
Without selection, after year t = 1, the genotype frequency and the average EBV for an animal of sex s,
category c, and genotype g do not change between t
and t1.
fsactg = fs (a -1)c(t -1)g
msactg = ms (a -1)c(t -1)g .
Selection. Different selection steps after year t = 1
can be considered for males and females (Figure 1). For
the last selection step, males and females are distributed in 2 categories, elite and not elite. The selection
rate Wsac can vary at each step for males and females.
The expression for Wsac is
Wsac =
g
wsactg fs(a 1)c(t 1)g ,
where wsactg is the proportion of individuals with sex s,
age a, and category c at generation t with genotype g
that are selected to become breeding animals. At each
step, candidates for selection can be evaluated through
their phenotype, genotype, or both.
Assuming that EBV are normally distributed, wsactg
is equal to
wsactg = f(b2g - ms(a -1)c(t -1)g ) - f(b1g - ms(a -1)c(t -1)g ),
Downloaded from jas.fass.org by guest on October 11, 2011
887
Optimizing multiple-trait selection
r -1
fs(a -1)c(t -1)g[k ] < Wsac
and
k =1
r
fs(a -1)c(t -1)g[k ] Wsac
k =1
because there cannot be more individuals in those
classes than in the whole population.
The selection proportions are
wsactg[k ] = 1 if k r 1;
wsactg[k ] = 0 if k > r ; and
r -1
wsactg[r ] =
Figure 2. Selection threshold for genotype g1 (), g2 ( ),
and g3 () under combined genotypic and phenotypic selection (a)
and phenotypic selection (b).
where f is the normal cumulative distribution function
for variance 2, and 1g and 2g are lower and upper
thresholds for genotype g (1g may be and 2g may
be +) defining the interval selected.
When individuals are selected according to EBV and
genotype information, the selection thresholds may
vary according to the genotype (Figure 2a), and the
state variables sactg and sactg are given by
fsactg =
wsactg
Wsac
[1]
fs(a -1)c(t -1)g ;
msactg = ms (a -1)c(t -1)g
+
[j(b1g - ms (a -1)c(t -1)g ) - j(b2g - ms (a -1)c (t -1)g )]
wsactg
rs,
[2]
where j(x ) is the normal density function, is the accuracy or the correlation between true and polygenic
EBV, and the genetic SD. Phenotypic selection is a
special case of these equations, where 1 and 2 are the
same for all genotypes (i.e., 11 = 12 = 13 = 1 and 21
= 22 = 23 = 2; Figure 2b).
In the case of selection for a monogenic trait, the
genotypes are ranked because some genotypes are better than others from a biological point of view (e.g., favorable or unfavorable genotypes for a disease). Let g[k]
be the kth ranked genotype. These constraints must be
verified:
Wsac - fs (a -1)c(t -1)g[k ]
k =1
otherwise.
fs(a -1)c(t -1)g[r ]
Selection within the rth genotype was random. The
genotype frequency sactg and the EBV mean sactg of
the population in generation t are obtained from Eq.
[1] and [2].
Mating. Mating between different classes (selected,
elite, and not elite) of females and males can be considered (Figure 1). The probability for a progeny to be of
genotype g, given that its parents carry the genotypes h
(sire) and k (dam), is ghk. The probability that a parent
of sex s has genotype g is Fs(t1)g and is defined as
Fs(t -1)g =
dsac fsac(t -1)g ,
a ,c
where sac is the proportion of individuals of age a and
class c among the parents with sex s. This quantity
depends on demographic constraints (e.g., reproductive
age) and breeder decisions (e.g., relative use of progeny
tested males). From the parameters ghk and Fs(t1)g, different matings (random or assortative) can be considered.
For random mating, none of the characteristics of the
parents (e.g., EBV or genotype) are considered in matings. The genotype frequency and the EBV mean for
the newborn animals of age 1 and category 1 are
fs11tg =
ms11tg =
1
2 fs11tg
tghk
h ,k
tghk F1h(t -1)F2k(t -1);
h ,k
d1a c f1a c (t -1)h d2a c f2a c
11
a1 ,c1 a2 ,c2
(m1a c (t -1)h + m2a c
11
Downloaded from jas.fass.org by guest on October 11, 2011
11
F1h (t -1)
2 2 (t -1)k
),
22
2 2 (t -1)k
F2k (t -1)
888
Costard et al.
where s = 1 for sire, s = 2 for dam, a1 and a2, and c1
and c2 are the age and category for the sire and dam.
For genotypic assortative mating, the mating may be
organized according to the genotypes of the parents.
The optimization of these matings may be achieved by
choosing a value for hkt, which is the proportion of
newborn animals in generation t from a sire with genotype h and a dam with genotype k. This proportion is
constrained by
ahkt = F1h(t -1);
k
ahkt = F2k(t -1).
h
In this situation, the genotype frequency and the
EBV mean are given by
fs11tg =
ms11tg =
tghk ahkt ;
h ,k
1
2 fs11tg
tghk ahkt
h ,k
d1a c f1a c (t -1)h d2a c f2a c
11
a1 ,c1 a2 ,c2
(m1a c (t -1)h + m2a c
2 2 (t -1)k
11
11
22
F1h (t -1)
2 2 (t -1)k
F2k (t -1)
).
Finally, other assortative matings can be considered. Most often the structure of the parent population
includes different groups involved in mating such as
males on progeny testing, elite males or females, selected non elite males or females. For instance, males in
progeny testing may be mated exclusively to nonelite
females. To include assortative matings in the model,
reproducers are classified according to mating groups
m. Fmsc(t1)g is the probability that a parent of group m,
and sex s has genotype g. It is defined as
Fmsg (t -1) =
dmsac fsac(t -1)g ,
a ,c
where msac is the proportion of animals of age a, class c,
and sex s within the mth group. When mating groups
are considered, the equations of genotype frequency
and EBV mean are modified accordingly. For instance
in genotypic random mating:
fs11tg =
ms11tg =
tghk Fm 1h(t -1)Fm 2k(t -1);
h ,k
m1 m2
2 fs11tg
dm 1a c f1a c (t -1)h dm
1
m1 m2 a1 ,c1 a2 ,c2
(m1a c (t -1)h + m2a c
11
2
1
tghk
h ,k
1
22
).
(t -1)k
11
11
Fm 1h (t -1)
1
f
2 2a2c2 2a2c2 (t -1)k
Fm
2 2h (t -1)
Optimization
The deterministic model described above allows global optimization of the selection scheme to maximize
the frequency of the desired genotype for a monogenic
trait, while minimizing the loss of genetic progress on
a polygenic trait. These objectives may be achieved by
taking into account the whole population (male and female) or a part of the population (female or male) for a
given age. The decrease in genetic progress for the polygenic trait may be considered as a penalty term, which
may be defined as a strict or as a progressive penalty.
Under a strict penalty, the objective function is put to
0 when the genetic progress is below a threshold. Under
a progressive penalty, this term is proportional to the
decrease in genetic progress.
The frequencies of the favorable genotype and the average EBV in the population may be considered during
all periods of selection (e.g., in the use of discounting)
or only in the last year of selection.
The decision variables wsactg and hkt, describing the
evolution of the population, have to be optimized. In
our approach, we did not calculate the derivatives of
the objective function, differing from the approaches
used by Dekkers and van Arendonk (1998) and Manfredi et al. (1998). A genetic algorithm (GA) was used
to find the optimum solutions (Goldberg, 1989). A GA
is an optimization technique based on an evolution of
a population of artificial individuals. Each individual
is a point in the research space and is characterized by
its chromosome Ck, which determines the individual
fitness h(Ck); k = 1,..., n; n is a population size. The
evolution process consists of successive generations.
At each generation, individuals with high fitness are
selected; the chromosomes of selected individuals
are recombined and subjected to small mutations. A
GA implemented as in Carroll (2001) was used in our
study.
APPLICATION
This model was applied to an example, where selection for scrapie resistance, which is controlled by a
single gene (the PrP gene), was introduced as an additional selection criterion in an already existing dairy
sheep selection scheme. It is expected that genetic progress in the polygenic trait (e.g., milk production) will
be reduced under this selection objective.
Scrapie is a spongiform encephalopathy in sheep similar to bovine spongiform encephalopathy in bovine and
Creutzfeldt-Jakob disease in humans. In sheep, genetic
variability of the resistance to scrapie has been demonstrated and is mainly determined by the PrP polymorphisms (Elsen et al., 1998; Fabre et al., 2006). Thus,
selection on the PrP locus was proposed to improve
scrapie resistance in sheep populations. Equations corresponding to this example are given in the Appendix.
The population (Figure 3) was composed of animals,
males and females, aged from 0 to 6 yr. Females were
Downloaded from jas.fass.org by guest on October 11, 2011
889
Optimizing multiple-trait selection
Figure 3. Population description with 7 categories (rectangles), 7 mating paths (thick lines), selection steps (dotted lines), and 5 selection
proportions Wsac (sex s=1, 2; age a=1, ..., as; category c=1,2).
selected on their own performance (with a selection
rate W221) after their first lactation between 2 and 3
yr of age (Figure 3). The selected females were distributed in 2 categories, elite (W232) and nonelite (1
W231). Among the young males born from matings
between elite reproducers, a fraction W111 was mated to
nonelite females and their EBV for the polygenic trait
were estimated after a progeny test. A fraction W131 of
these males was selected 2 yr later. Among the selected
males, a proportion W142 was mated to the elite females
to produce young male candidates for selection in the
next generation. The replacement of females resulted
from mating between nonelite females and elite, nonelite males or males in progeny testing; and between
elite females and elite males. Two alleles (favorable vs.
unfavorable) were considered for the Prp gene; thus, 3
genotypes exist (g = 1 favorable genotype, g = 2 intermediary genotype, and g = 3 unfavorable genotype).
The PrP genotypes were known for females and males.
Only males were selected on their PrP genotype and
on their EBV value before and after progeny testing.
The proportion of males in progeny test among reproductive males was equal to 0.2. The proportion a of
females with age a (4, 5, and 6 yr) among reproductive
females was 0.4, 0.3, and 0.3, respectively. This selection scheme was applied during 15 yr (t = 15) with the
following parameter values: W221 = 0.8; W232 = 0.1;
W111 = 0.225; W131 = 0.2; W142 = 0.3; = 36. The accuracy for females (f) was 0.73. The accuracy for males
before and after progeny test was 0.56 (m1) and 0.77
(m2), respectively. The accuracy for elite males (m3)
was 0.89. The GA parameters were as follows: population size = 100; number of generations = 1,000; creep
mutation probability = 0.02; crossover probability =
0.5; jump mutation probability = 0.01.
The aim was to maximize scrapie resistance, minimizing the loss of genetic progress in milk production. The
objective function chosen was the frequency of homozygous favorable females (s = 2, g = 1) with a penalty on
the genetic progress, constraining it not to be less than
a percentage of the progress obtained without scrapie
selection (GT). The objective function can be written
as follows:
Fobj
6 21
pDGT
f2acT 1
a =1 c=1 a
- 100
=
6
1
a
a =1
3
-
f
211Tg
m211Tg
g =1
pDGT
.
The 1st term is the mean of homozygous favorable
females of the elite (c = 1) and non elite (c = 2) group
1
weighted by the coefficient
(a is the age), which
a
gives more importance to younger females. All of the
females were considered in the objective function because all of them will participate in the renewal. The
second term is proportional to the difference in genetic
progress obtained in this selection scheme and without
selection on scrapie resistance.
First, the model was used with different initial frequencies of the favorable genotype without any constraint on genetic progress ( = 100%). Three initial
frequencies of the favorable genotype were considered
(0.03, 0.3, and 0.69). Then, situations with different
constraints on genetic progress were considered while
fixing the initial frequencies. Finally, simplifications of
the strategy were explored to make the optimized selection procedures more practicable for farms.
Effect of Different Initial Frequencies
Without Constraint on Genetic Progress
Whatever the initial frequency, the unfavorable genotype disappeared after 15 yr of selection (Table 1). The
Downloaded from jas.fass.org by guest on October 11, 2011
890
Costard et al.
Table 1. Final genotype frequencies for young females, objective function, and loss of genetic progress for 3 initial
frequencies 1, (genotype favorable genotype; genotype 2, intermediary genotype; genotype 3, unfavorable genotype)
without any constraint on loss of genetic progress
Initial frequency
Genotype 1
0.03
0.30
0.699
Final frequency for young females
Genotype 2
Genotype 3
Genotype 1
Genotype 2
Genotype 3
Objective function
Loss of genetic
progress, %
0.28
0.50
0.28
0.69
0.20
0.03
0.77
0.91
0.97
0.23
0.09
0.03
0.00
0.00
0.00
0.68
0.88
0.96
18.10
6.40
2.00
loss of genetic progress was more important (18.1%)
with a low initial frequency of the favorable genotype
than with a high initial frequency (2.0%). This loss is
much greater in the first few years, and impacts in the
first few years are very important if counting discounted returns. The optimal selection proportions and mating rules evolve during the 15 yr process. The favorable
genotype frequency increased, keeping a high selection
proportion of the homozygous-resistant males before
and after their progeny test. When the initial frequency
of the favorable genotype (0.03 or 0.3) was low, the selection proportions of males carrying this genotype were
high (0.9 and 0.7) during the first years to rapidly fix
this genotype in the sires. These selection proportions
decreased progressively when all sires became resistant.
When the initial frequency of the favorable genotype
(0.69) was high, the selection proportions for the favorable genotype were less, in particular before progeny
testing (0.32 instead of 0.9 and 0.7). Moreover, the selection proportions after the progeny test were constant
from the beginning, whereas they were constant only
after yr 9 and 4 for the case of low frequencies (0.03 and
0.3). Whatever the initial frequencies of the favorable
genotype, none of the susceptible males were selected,
eliminating all of them from the beginning. A fully favorable population of sires was reached more rapidly
with a high initial frequency of the favorable genotype
(yr 3) than with a low frequency (yr 8).
The reduction of the unfavorable genotype frequencies in the dam population occurred very slowly because (i) dams were selected just once, (ii) the selection
threshold was the same across genotype classes, and
(iii) the selection pressure was low. With a high initial
frequency of the favorable genotype, the objective was
reached and the dam population was fully resistant.
However, with a low frequency, only sires were resistant
at the end of yr 15, with the unfavorable genotype still
segregating in the dam population.
Effects of Different Constraints on the Loss
of Genetic Progress
In contrast to the situation described above in which
no constraint was put on genetic progress, a 10% maximum loss of genetic progress at the end of the process
was allowed ( = 90%) for the situation with the decreased initial frequency (0.03). The case with the ini-
tial low frequency of the favorable genotype was studied
because the loss of genetic progress with this frequency
was the most important (18.1%). In this situation, the
optimal selection proportions of males with a favorable
genotype were less and fluctuated between high and low
values (Figure 4). These fluctuations can be explained
by the fact that the genetic progress must be kept over
a threshold. Each year, the selection intensity varies to
compensate for the loss in genetic progress. Contrary
to the previous case, the selection proportions of favorable and heterozygous genotypes were high and similar
(0.7). The selection proportions for homozygous unfavorable sires were null. After the elimination of unfavorable genotype carriers, the selection proportions of the
favorable genotype (20 to 30%) were greater than for
the heterozygous genotype (5 to 20%). The heterozygous genotype was conserved for a longer time than
without any constraint on the genetic progress. The objective of achieving a fully resistant population of sires
was reached less rapidly with 10% of maximum loss of
genetic progress (yr 11) than without any constraint on
genetic progress (yr 8). In the dam population, proportions of favorable homozygous and heterozygous genotypes were similar. Dams with an unfavorable genotype
were almost entirely eliminated, being kept at less than
5%. Thus, the optimal strategy of selection depends on
the constraints put on genetic progress. If the priority
is to reach a population fully resistant to the disease,
the loss of genetic progress may not be essential; the
selection on genotype information is drastic at the beginning of the process and the animals are quickly fully
resistant. If a more important penalty is put on the loss
of genetic progress, the selection changes, with cycles
of high and low selection proportions for the favorable
genotype. Eventually, the unfavorable genotype disappears and the population is divided between favorable
homozygous and heterozygous genotypes.
In the previous case (Figure 4a), the selection proportions decreased gradually as the population became
progressively resistant. With a constraint on genetic
progress, the fluctuation of selection proportions was
larger (Figure 4b). The selection proportions of the favorable genotype were high during the first years, inducing an increase in resistance during this period; they
were less than with previous cases to maintain acceptable genetic progress. This optimal strategy would not
be practicable for breeders because of the annual varia-
Downloaded from jas.fass.org by guest on October 11, 2011
Optimizing multiple-trait selection
tions of the selection proportions. However, a detailed
analysis of selection proportions and assortative mating
during the 15 selection years suggests some simplifications.
Simplifications of the Optimal Strategy
Two cases were analyzed: (1) a low initial frequency
of the favorable genotype and a maximum 10% loss of
genetic progress accepted after 15 selection years and
(2) a low initial frequency of the favorable genotype
without any constraint on the loss of genetic progress.
The results are given in Table 2. The first case shows
891
the evolution of the objective function without any penalty on loss of genetic progress. The second case shows
the evolution of the objective function with a maximum
loss of 10% in the genetic progress.
In the optimal strategy, the selection proportions of
the males before progeny testing (W111, Figure 3) varied between years and thus needed to be optimized. On
the contrary, some selection proportions were constant
after a few years of selection, suggesting some simplifications.
1. Within genotype selection proportions of males
after progeny testing: The stabilization of these
Figure 4. Selection rate evolution for the favorable genotype (g = 1) without (a) and with (b) constraint in loss of genetic progress before
progeny test (), after progeny test (.), and for elite males ().
Downloaded from jas.fass.org by guest on October 11, 2011
892
Costard et al.
Table 2. Value of the objective function with some simplifications in 2 cases without
constraint on loss of genetic progress (L-0) and with 10% of loss of genetic progress
(L-10)
Item
L-0
L-10
Fully optimized scheme
Constant selection rate for elite males
Constant selection rate after progeny testing
Random mating
Constant selection rate for elite males + random mating
0.68
0.67
0.56
0.68
0.67
0.53
0.53
0.48
0.52
0.52
selection proportions is late, after the 11th selection year. However, the fixation of these selection proportions was tested. The values of the
objective function were less than in the optimal
strategy. Thus, these selection proportions have
to be optimized.
2. Within genotype selection proportions for elite
males: The results of the previous optimization
suggest a suboptimal strategy where these selection proportions are constant. The objective
function value obtained with this simplification
was similar to the optimal one (Table 2). Consequently, it appears that it is not critical for
within genotype selection proportions for elite
males to be optimized every year.
Another question was to check if assortative mating
based on the genotypes was necessary or not. The slight
difference in the objective function value between the
fully optimized scheme and the simplified one shows
that mating optimization was not necessary.
Globally, 2 simplifications were thus possible: fixed
selection proportions for elite males and random mating. The case combining simplifications gave a value
of the objective function similar to the optimal one.
Moreover, the penalty weight of the loss on genetic
progress was identical with or without simplifications.
Implementing other simplifications could give a reduced
value of the objective function than the one obtained
in the fully optimized scheme, but these were not considered here.
In our example, we proposed that the objective was
to maximize the genotypic frequency of the favorable
female genotypes with a penalty on genetic progress.
Avoiding loss of genetic progress increases the time
needed to reach fixation of the favorable genotype. For
example, sires are all resistant at the end of yr 8 or
after yr 11 without any constraint or with a maximum
loss of 10% of genetic progress, respectively. Other objective functions are possible and must be adapted to
the aim of the breeder. For example, the frequency of
males may be considered instead of the frequencies of
females. Males have a more important impact on the
selection scheme. In this example, the objective was to
keep the homozygous genotype; however, in the case of
breeds with a risk of extinction, it could be preferable
to keep heterozygote genotypes. Another option could
be to consider the whole process, rather than only the
final state of the population, possibly using a discounting factor if sooner is better.
DISCUSSION
In this paper, a method was developed to optimize selection on an identified QTL or gene and a quantitative
trait over multiple generations. The breeding program
modeled assumed assortative mating and overlapping
generations. This modeling may easily be adapted to a
variety of situations concerning the underlying genetic
model and the selection scheme. In this study, 2 alleles
were considered, but the methodology can consider a
multi-allelic case. However, the number of alleles (and
genotypes) should be limited. The number of optimized
parameters increases in an exponential manner with
the number of alleles considered. For instance, when
3 genotypes (2 alleles) are considered, 10 parameters
have to be optimized, and if 10 genotypes (4 alleles)
are considered, the optimization problem would involve
108 parameters. Only the case with an additive QTL
or gene was considered here, but dominance could be
included. The method can also be extended to multiple
QTL or genes with any type of linkage between them
(epistasis).
Different methodologies to solve multiple-generation
selection problems in animal breeding are possible. In
this study, a GA was used to solve the optimization
problem. Genetic algorithms are more flexible when
additional constraints have to be added than optimal
control theory, the method used by Dekkers and van
Arendonk (1998), because to optimize selection the derivative of the selection criterion has to be calculated
with optimal control theory. Use of a GA bypasses this
difficulty.
Additive variance was assumed to be constant over
generations, and genotyping errors and relationship errors were assumed to be nonexistent. The reduction of
the additive variance by selection could be considered
in the model.
The strategy developed here consisted of selecting
animals that were resistant to a particular disease to
decrease the number of infected animals. The aim was
to reduce the incidence of the disease, while keeping the
genetic progress on the traditional production traits.
Downloaded from jas.fass.org by guest on October 11, 2011
893
Optimizing multiple-trait selection
In this model, inbreeding was not considered. A possibility for taking this into account would be to add the
notion of family for males. Males of one family participate for the renewal of the males for their family and
males of different families participate for the renewal
of the females. This possibility would allow a certain
genetic variability to be maintained.
Domestic animal diseases might represent a risk to
human health. Thus, identification of genes involved
in disease resistance is an important genetic research
subject. The scrapie case is an example where a national selection program for scrapie resistance was implemented as a precautionary measure. This dynamic
and deterministic model was applied here in the scrapie
case; however, it can be extended to other genes and selection schemes. For example, genes affecting important
performance traits such as adiposity or meat quality in
pigs (Le Roy et al., 1990; Larzul et al., 1997a; Hamilton et al., 2000) or prolificacy in sheep (Davis, 2005;
Fabre et al., 2006) could use this model. Moreover, this
deterministic model could evolve to a stochastic model
for a limited number of breeding animals in a small
population. Our model would need to be simplified for
use in stochastic simulation because stochastic simulations with a large number of choice variables can be
impractical.
LITERATURE CITED
Carroll, D. L. 2001. The genetic algorithm (ga) driver. http://
cuaerospace.com/carroll/ga.html Accessed Aug. 25, 2008.
Chakraborty, R., L. Moreau, and J. C. M. Dekkers. 2002. A method
to optimize selection on multiple identified quantitative trait
loci. Genet. Sel. Evol. 34:145170.
Davis, G. H. 2005. Major genes affecting ovulation rate in sheep.
Genet. Sel. Evol. 37(Suppl. 1):S11S23.
Dekkers, J. C. M., and R. Chakraborty. 2001. Potential gain from
optimizing multigeneration selection on an identified quantitative trait locus. J. Anim. Sci. 79:29752990.
Dekkers, J. C. M., and J. A. M. van Arendonk. 1998. Optimizing
selection for quantitative traits with information on an identified locus in outbred populations. Genet. Resour. Crop Evol.
71:257275.
Elsen, J. M., J. Vu Tien Khang, F. Schelcher, Y. Amigues, J. P.
Poivey, F. Eychenne, F. Lantier, and J. L. Laplanche. 1998.
A scrapie epidemic in a closed flock of Romanov. Proc. 6th
World Congr. Genet. Appl. Livest. Prod., Armidale, Australia.
27:269272.
Fabre, S., A. Pierre, P. Mulsant, L. Bodin, E. Di Pasquale, L. Persani, P. Monget, and D. Monniaux. 2006. Regulation of ovulation rate in mammals: Contribution of sheep genetic models.
Reprod. Biol. Endocrinol. 4:20.
Gibson, J. P. 1994. Short term gain at the expense of long term response with selection of identified loci. Proc. 5th World Congr.
Genet. Appl. Livest. Prod. Univ. Guelph, Ontario, Canada.
21:201204.
Goldberg, D. E. 1989. Genetics algorithms in search, optimization
and machine learning. Addison-Wesley Publ. Co. Inc., Boston,
MA.
Hamilton, D. N., M. Ellis, K. D. Miller, F. K. McKeith, and D. F.
Parrett. 2000. The effect of the Halothane and Rendement Napole genes on carcass and meat quality characteristics of pigs.
J. Anim. Sci. 78:28622867.
Holland, J. H. 1973. Genetic algorithms and the optimal allocation
of trials. SIAM J. Comput. 220:88105.
Lande, R., and R. Thompson. 1990. Efficiency of marker assisted
selection in the improvement of quantitative traits. Genetics
124:743756.
Larzul, C., L. Lefaucheur, P. Ecolan, J. Gogue, A. Talmant, P. Sellier, P. Le Roy, and G. Monin. 1997a. Phenotypic and genetic
parameters for longissimus muscle fiber characteristics in relation to growth, carcass, and meat quality traits in Large White
pigs. J. Anim. Sci. 75:31263137.
Larzul, R., E. Manfredi, and J. M. Elsen. 1997b. Potential gain from
including major gene information in breeding value estimation.
Genet. Sel. Evol. 29:161184.
Le Roy, P., J. Naveau, J. M. Elsen, and P. Sellier. 1990. Evidence
for a new major gene influencing meat quality in pigs. Genet.
Res. 55:3340.
Lewis, F. L. 1986. Optimal Control. Wiley, New York, NY.
Manfredi, E., M. Barbieri, F. Fournet, and J. M. Elsen. 1998. A
dynamic deterministic model to evaluate breeding strategies
under mixed inheritance. Genet. Sel. Evol. 30:127148.
Meuwissen, T. H., and J. A. van Arendonk. 1992. Potential improvements in rate of genetic gain from marker-assisted selection in
dairy cattle breeding schemes. J. Dairy Sci. 75:16511659.
Mongeau, M. 2003. Exploitation de la structure en optimisation
globale, Laboratoire Mathmatiques pour lIndustrie et la Physique. Handbook. UMR 5640. Toulouse, France.
Pong-Wong, R., and J. A. Woolliams. 1998. Response to mass selection when an identified major gene is segregating. Genet. Sel.
Evol. 30:313337.
APPENDIX: EQUATIONS
OF THE EXAMPLE
Two-year females:
f221tg = f211(t -1)g ;
m221tg = m211(t -1)g .
Three-year nonelite females:
f232tg = f221(t -1)g
w221tg (1 - w232tg )
W221(1 -W232 )
.
m232tg = m221(t -1)g
+
i(w221tg ) - w232tg i(w221tg w232tg )
(1 - w232tg )
r f s,
where i() is the selection intensity.
Three-year elite females:
f231tg = f221(t -1)g
w221tg w232tg
W221W232
;
m231tg = m221(t -1)g
+
i(w221tg ) - w221tg i(w221tg w232tg )
w232tg
Four- to six-year females:
Downloaded from jas.fass.org by guest on October 11, 2011
r f s.
894
Costard et al.
m111(t+1)g
m2a1tg = m2(a -1)1(t -1)g ;
1
1
2 f111(t+1)g
=
f2a1tg = f2(a -1)1(t -1)g ;
3
3
tghk ahkt
h =1 k =1
m2a 2tg = m2(a -1)2(t -1)g ; and
f2a
6
l
2 2tk a2
f2a 2tk la
a2 = 3
2
a2 = 3
f2a 2tg = f2(a -1)2(t -1)g .
f121tg = f111(t -1)g
W111
3
h =1 k =1
;
f2a
+
tghk (ahktW232
3
3
6
h =1 k =1
a2 = 3
Four-year nonelite males:
f2a
W131W142
6
+ f142th f2a
;
a2 = 3
m142tg = m131(t -1)g + i(w131tg w142tg )rm 3s.
w131tg (1 - w142tg )
f2a 2tk la
W131(1 -W142 )
m141tg = m131(t -1)g
i(w131tg ) - w142tg i(w131tg w142tg )
1 - w142tg
rm 2s.
a2 = 3
h =1 k =1
)
2
l
2 1tk a2
(1 - d )mne (1 -W142 )
(1 - d )m W + (1 - d )m (1 -W ) + dm
ne
142
T
e 142
(m
+m
)
141th
a2 = 3
3
2a2 2tk
l
+ f141th f2a
6
tghk ahkt ;
+m
2a2 1tk
+ f2a
Renewal.
Newborn males:
2
142th
2 1tk a2
6
;
(m
(1 - d )meW142
(1 - d )m W + (1 - d )m (1 -W ) + dm -W232
142
T
e 142
ne
(m
+m
)
142th
Four-year elite males:
l
2 2tk a2
6
a2 = 3
w131tg w142tg
l
2 1tk a2
1
1
2 f211(t +1)g
m131tg = m121(t 1)g .
3
2
)
m211(t +1)g =
f111(t +1)g =
).
a 2 = 3 (1 - d )meW142 + (1 - d )mne (1 -W142 ) + dmT
f
(1 - d )meW142
142th (
W
- f 2 ((1 - d )meW142 + (1 - d )mne (1 -W142 ) + dmT )
+ f141th (1 - d )mne (1 -W142 )+dmT f121th );
f131tg = f121(t 1)g ;
+
2a2 2tk
3
6
Two-year males after progeny testing:
f141tg = f131(t -1)g
+m
tghk (ahktW232
f211(t +1)g =
m121tg = m111(t -1)g + i(w111tg )rm1s.
f142tg = f131(t -1)g
142th
Newborn females:
Two-year males before progeny testing:
w111tg
(m
6
2a2 1tk
lf
2 1tk a2 121th
dmT
(1 - d )m W + (1 - d )m (1 -W ) + dm
T
e 142
ne
142
(m
+m
).
121th
2a2 1tk
Downloaded from jas.fass.org by guest on October 11, 2011
References
This article cites 16 articles, 4 of which you can access for free at:
http://jas.fass.org/content/87/3/885#BIBL
Downloaded from jas.fass.org by guest on October 11, 2011