ORIGINAL PAPER

Epistatic Regulation of Behavioural and Morphological Traitsin the Zebrafish (Danio rerio)

Dominic Wright Æ Roger K. Butlin Æ Orjan Carlborg

Received: 9 November 2005 / Accepted: 18 April 2006 / Published online: 3 June 2006

� Springer Science+Business Media, Inc. 2006

Abstract There is currently a lack of studies examining

epistasis in general, and specifically for behavioural traits

of evolutionary significance. The advent of more efficient

analytical methods for exploring epistasis in QTL studies

removes the computational restraint on this type of analysis

and suggests that performing further analyses of existing

datasets may reveal a more complete picture of the genetic

architecture of the traits. Here we report the results from an

epistatic QTL analysis of an F2 cross between a wild

population and a standard laboratory strain of zebrafish.

This further analysis was performed using a simultaneous

search to identify epistatically interacting QTL affecting

behavioural and morphological traits and uncovered sev-

eral novel epistatic interactions that reached either genome-

wide or suggestive significance levels as determined by a

randomisation testing approach. These results provide

novel insight into the genetic architecture of the regulation

of behavioural as well as morphological phenotypes and

call for more studies of epistasis for this group of traits.

Keywords Zebrafish Æ Epistasis Æ QTL Æ Boldness ÆGrowth

Introduction

There is a growing interest in exploring epistasis in the

genetic regulation of complex traits. A powerful approach

to explore epistatic regulation of complex traits is by QTL

mapping in experimental crosses and a number of recent

studies have shown that epistasis is now a measurable

phenomenon (Carlborg and Haley 2004). Investigations

into possible epistatic effects are, however, still rare given

the number of QTL studies performed, and there are many

experimental datasets available where epistasis might be

uncovered by reanalysis using epistatic QTL mapping

methods. Even though many of the existing experimental

datasets do not have the large sample sizes that are tradi-

tionally considered necessary for the estimation of epistatic

interactions, the efficiency of the new analytical tools

available for epistatic analysis suggests that performing

these analyses may be revealing.

To date, detection of epistasis has mainly been re-

stricted to morphological traits with high heritability

(Carlborg et al. 2003, 2005), though traits with low heri-

tability (often, but not solely, fitness related traits) are now

beginning to be analysed. For example, Civetta et al.

(2005) studied post-mating mortality in Drosophila lines

and Peripato et al. (2004) analysed litter size in mice:

despite the low heritabilities involved, significant epistatic

Edited by Jeanne Wehner

D. Wright

Institute for Integrative and Comparative Biology,

Faculty of Biological Sciences, University of Leeds,

Leeds LS2 9JT, UK

R. K. Butlin

Department of Animal and Plant Sciences,

The University of Sheffield, Sheffield S10 2TN, UK

O. Carlborg Æ D. Wright

Linneaus Centre for Bioinformatics,

Uppsala University, Uppsala SE 75124, Sweden

D. Wright (&)

Department of Animal Genetics, Uppsala University,

Uppsala, Sweden

e-mail: domwright@gmail.com

Behav Genet (2006) 36:914–922

DOI 10.1007/s10519-006-9080-9

123

QTL were identified in both studies. In both cases, rela-

tively small sample sizes were used: in the case of the

mouse study, 166 individuals were used with epistasis

explaining an additional 36% of the variation in the pop-

ulation, whilst in the Drosophila study 101 recombinant

inbred lines were sufficient to detect epistasis. As regards

behavioural epistatic analysis, studies are rare, mainly

being restricted to drug tolerance (for instance see Hood

et al. (2001), regarding pentobarbital withdrawal convul-

sions). As regards traits of a more evolutionary signifi-

cance, studies are restricted even further, with evidence of

the extent to which epistasis is prevalent being ambiguous.

Odour-guided behaviour in Drosophila has been analysed

using a combination of mutation screening and transcrip-

tional profiling, with an initial network of eight interacting

loci revealed (Anholt et al. 1996; Fedorowicz et al. 1998),

whilst analysis of particular genes in different backgrounds

also indicated the presence of epistasis (Anholt et al.

2003). Flint et al. (2004) found little evidence for epistasis

in a large F2 cross looking at fear-related behaviour in two

artificially selected strains of mice. Ruppell et al. (2004)

found some evidence for epistasis in foraging behaviour in

the honeybee, using four previously discovered QTL as the

basis for the analysis. In both of these studies, highly se-

lected lines as opposed to recently wild-derived individu-

als were used. Generally therefore the contribution to our

knowledge of the degree of epistasis present in natural

populations is somewhat limited. There remains a need to

study epistasis in behavioural traits with known effects on

fitness, in well studied ecological contexts, to develop a

more complete understanding of the types of genetic

architecture involved in adaptive evolution. When epistatic

effects are incorporated into the genetic model, not only

can more of the variation within a cross or population be

explained than with direct effects alone, but a more

complete picture of the architecture of a trait can be

obtained.

The anti-predator behaviour exhibited by fish is a much

studied phenomenon ecologically (Krause and Ruxton

2002; Pitcher and Parrish 1993), however analysis of the

underlying genetics is largely unexplored. The principal

manifestation of such defence behaviour is shoaling or

schooling (Magurran et al. 1995; Landeau and Terborgh

1986; Pitcher and Parrish 1993). This behaviour has the

effect of increasing vigilance, diluting individual risk and

causing predator confusion (Pitcher and Parrish 1993). As

well as such benefits, costs are also incurred by shoaling

behaviour, principally through an increase in resource

competition from shoal mates, although shoaling may also

lead to more efficient food location (Baird et al. 1991;

Ranta and Kaitala 1991). A further anti-predator response

is predator inspection. This involves a solitary fish or small

group approaching and then retreating from a predator

(Pitcher 1992). The benefits of inspection include increased

knowledge of the predator’s location or status, in terms of

the level of the potential hazard it represents (Csanyi

1985), whilst costs come from an increase in the proba-

bility of predation.

An F2 cross was set-up previously to examine the

architecture of these types of anti-predator behaviour using

the zebrafish, Danio rerio, as a model (Wright et al. 2006).

A wild population was crossed with a standard laboratory

strain in an F2 cross. These data has previously been

analysed using standard interval mapping and Genetic

Algorithm approaches (Nakamichi et al. 2001). Using an

established analytical technique (Carlborg et al. 2003,

2004, 2005), we extended the analysis of the existing

dataset to measure the effects of epistasis in both

behavioural and morphological traits. This enabled us to

analyse the effects of epistasis in traits of known evolu-

tionary significance, as well as to evaluate the use of such

analyses in a relatively small dataset.

Methods

F2 intercross

An F2 intercross was set-up between a population of wild-

derived zebrafish and a common laboratory population.

Wild fish (hereafter referred to as Santal after the nearest

village) were collected in 1997 from Bangladesh and

maintained for two generations under laboratory conditions

before the cross was conducted. Laboratory fish were from

the AB strain (for details of origin see http://www.zfin.org).

These two populations of fish differed strongly in both anti-

predation behaviour (with the Santal strain exhibiting both

a stronger tendency to shoal and increased inspection of a

novel object) as well as in morphology (the AB strain

showing greater growth rate than the Santal strain), see

Wright et al. (2006) for further details. These populations

were chosen due to the extreme differences both in terms of

behaviour and morphology as well as in genetic variation.

This enabled the genetic dissection of two populations that

had undergone strongly divergent natural and artificial

selection, as well as facilitated obtaining fully informative

markers for the cross. Two pairs of F1 fish were crossed to

yield a total of 184 F2 animals, with 166 derived from the

first pair and 18 from the second. Fish were housed in

standard aquaria (measuring 60 cm · 20 cm · 20 cm) and

fed with a commercial flake food. Phenotypes of fish re-

corded were two different anti-predator behaviours:

shoaling tendency and novel object inspection, with stan-

dard length (in mm) and weight (in grams) also recorded.

Each behavioural test was conducted twice, to give a total

of four trials.

Behav Genet (2006) 36:914–922 915

123

Phenotypic Measurements

Three different measures of ‘‘boldness’’ (in the form of

novel object inspection) were obtained: ‘‘time stimulus

zone first entered,’’ ‘‘total time in stimulus zone’’ and

‘‘number of entries to stimulus zone’’, whilst a composite

measure was also derived from these three observations

(the first axis of a Principal Component analysis). Shoaling

tendency was measured as the total time spent in the

stimulus zone adjacent to a shoal of conspecifics. Standard

length (the distance from the tip of the nose to the base of

the caudal fin) and weight were also recorded after the

behavioural tests. Further details of all phenotypes can be

found in Wright et al. (2006).

DNA Extraction and Genotyping

Methods of DNA extraction and genotyping are given in

Wright et al. (2006). The genetic markers used for this

study comprised all those used in the initial study, except

one which was dropped due to excessive segregation dis-

tortion (marker Z11781), giving a total of 65 markers

covering all 25 chromosomes of the zebrafish genome.

A full list of markers is presented in Wright et al. (2006).

QTL Analysis

In the previous interval mapping analysis of this dataset

(Wright et al. 2006), one significant and one suggestive

QTL (on chromosome 9 at 6 cM, and on chromosome

21 at 4 cM, respectively) were discovered for the ‘‘time

first entered stimulus zone’’ measure of boldness, whilst a

significant QTL was also discovered for the composite

measure (‘‘PC score’’) of boldness (once again on chro-

mosome 9, at 5 cM). A significant QTL was identified for

length (on chromosome 23 at 5 cM), whilst two suggestive

QTLs were identified for weight (on chromosomes 9 and

23). In this previous analysis, a Genetic algorithm tech-

nique was also used to simultaneously fit QTLs to the data,

with this approach supporting the above QTLs, as well as

detecting a number of other potential candidates.

The results presented here are based on a genetic model

incorporating epistasis, using a simultaneous search strat-

egy to discover QTLs (Carlborg and Andersson 2002;

Carlborg et al. 2000). This technique is based on a linear

model incorporating marginal effects for a pair of QTL as

well as the four possible pairwise interactions. For any

significant pairs, the genetic variance (r-squared value) due

to epistasis can be quantified through the linear model fitted.

The standard (marginal) genetic model for a QTL

incorporating additive and dominance effects is as follows:

y ¼ b0 þ FZ þ b1jaj þ b2jdj þ ej

where y is a vector of phenotypes, F is a vector of

regression coefficients containing fixed factors and cova-

riates in the model (in this case the fixed effect of rearing

tank for all traits and the covariate body length for the

behavioural traits), Z is a matrix of regression variables for

tank and any earlier detected QTL, aj and dj are regression

indicator variables for the additive and dominance effects

at location j whilst b1j and b2j are regression coefficients

for the additive and dominance effects at the same location.

Finally, ej is the error variable associated with the regres-

sion equation at this location. In the epistatic analysis, this

model is expanded to include the marginal (additive and

dominance) terms of the second QTL, as well as the four

pairwise interaction terms (additive by additive, additive

by dominance, dominance by additive and dominance by

dominance).

y ¼ b0 þ F Z þ b1jaj þ b2jdj þ b3kak þ b4kdk

þ b5jkaajk þ b6jkadjk þ b7jkdajk þ b8jkddjk þ ejk

Here b1 to b4 are the single QTL additive and dominance

effects for the QTL at locations j and k cM, respectively, b5

to b8 are the regression coefficients for the epistatic effects

between the QTLs and aajk, adjk, etc are the regression

indicator variables for these effects.

The parameters in the above equation (regression coef-

ficients) were estimated using a variation of the least

squares regression framework (Haley and Knott 1992;

Haley et al. 1994). Within this framework, QTL genotype

probabilities are estimated throughout the genome condi-

tional on the marker genotypes. The QTL genotypes are

then used to calculate the regression indicator variables for

the genetic effects of the QTL (with these effects estimated

by least squares regression). Here, the marker genotypes

are used to calculate the probability of every F2 offspring

being one of the four QTL genotypes (i.e., either QQ, Qq,

qQ or qq) at every location on the genome. These proba-

bilities then allow us to calculate the additive and domi-

nance effects of each QTL, every cM along the genome

(see Haley and Knott 1992). These are then used to cal-

culate the indicator regression variables for every point on

the two by two grid that is used by the epistatic model as

indicated by Haley and Knott (1992):

ai1aj2 ¼ ai1 � aj2

ai1dj2 ¼ ai1 � dj2

di1aj2 ¼ di1 � aj2

di1dj2 ¼ di1 � dj2

where a is the additive and d the dominance effect at locus

i and j, respectively for QTL 1 and QTL 2.

916 Behav Genet (2006) 36:914–922

123

Simultaneous mapping was used to detect epistatic QTL

pairs. This consists of calculating the QTL genotype

probabilities (as above) at every location, with the model

then being fitted to every possible pairwise combination of

loci (i.e., the statistical model is fitted exhaustively at every

location on the two dimensional grid). In each case the

model fit (residual sum of squares) was retained. For all

fitted pairs, significance was determined by one of three

methods:

(i) If both QTL were already significant by their marginal

effects, significance for the pair was declared with no

further testing.

(ii) When only one of the QTL in the pair had significant

marginal effects, a randomisation test was used to

gauge the significance of the second marginal QTL

and the interaction parameters of the pair. A 5% level

of genome-wide significance was used to indicate

significant pairs, whilst a 20% threshold was used to

indicate suggestive QTL pairs.

(iii) When neither of the QTL in the pair had significant

marginal effects, a randomisation test was used to

assess significance with 5% genome-wide signifi-

cance and 20% genome-wide suggestive thresholds.

In the analyses of each randomised dataset, a genetic

algorithm (Carlborg et al. 2000) was used to identify

the highest test-statistic, which was then retained in

the extreme-value distribution used to calculate the

empirical significance thresholds (1000 permutations

were used in each case).

The complete process of analysis therefore began with

running the standard least squares regression to detect

marginal QTL effects (i.e., a repeat of the initial analysis

performed in Wright et al. 2006). After this process, an

exhaustive two dimensional simultaneous search was per-

formed to detect epistasis. Significance of any epistatic

pairs was estimated as above, with the genetic algorithm

used to speed up the randomisation procedure that esti-

mated the significance level. Once epistasis had been

estimated, any significant or suggestive pairs were checked

for potential segregation distortion and the presence of any

outliers that could artificially distort the results (Knott et al.

1998). This is particularly important with small datasets

where individual extreme observations are more influential

for the results.

A 5% genome-wide threshold was set as significant,

whilst a 20% threshold was used to indicate suggestive

QTLs. The use of suggestive thresholds is always some-

what arbitrary. A commonly used threshold is that sug-

gested by Lander and Kruglyak (1995) of one false positive

QTL per genome scan. The 20% genome-wide significance

threshold used in this study is considerably more stringent

(depending on the dataset, an increase in threshold of ~0.6

LOD), and has been used in several other analyses previ-

ously (Carlborg et al. 2003, 2004, 2005) to further avoid

reporting false positive results, as well as being easier to

implement with this form of analysis.

Sex effects in this study have been previously found to be

non-significant in the marginal analysis and were therefore

omitted from the epistatic model. However, a general

problem with this type of epistatic analysis is that, by

pooling the sexes, certain effects that are in opposite

directions between the two sexes will be missed. One pos-

sible way to rectify this is by analysing the sexes separately.

However, in this instance the sample sizes then become so

small that they may no longer include each of the nine

genotypic classes, with any results therefore unreliable.

Therefore in this analysis the sexes were pooled.

Results

For the behavioural traits, one significant and two sugges-

tive epistatic pairs were detected for boldness (time first

entered) and a suggestive pair was also detected for bold-

ness (PC1 score) (Fig. 1 and Table 1). The significant pair

comprised of loci on chromosome 9 (hereafter referred to as

B9.10, with the B referring to the boldness trait, the nine

corresponding to the chromosome and the ten exact position

of the QTL in cM) and chromosome 12 (hereafter B12.39).

Both suggestive pairs involved a locus on chromosome 21

(B21.4), interacting with B9.10 and a locus on chromosome

18 (B18.12). The significant interaction for boldness (PC1

score) replicated the interaction between B9.10 and B12.39.

Two novel QTL (B12.39 and B18.12) were thus detected in

the epistatic analysis as only B9.10 (significant) and B21.4

(suggestive) had detectable marginal effects in both this and

the previous analysis. No QTL were discovered for shoaling

tendency that had not been identified in the previous anal-

ysis and no epistatic effects were identified.

For the morphological traits, one pair of interacting loci,

on chromosomes 23 (G23.8) and 6 (G6.10), was found to

have a genome-wide significant effect on length and a

suggestive effect on weight. Of these loci, only the QTL on

chromosome 23 had significant marginal effects and was

detected in the previous analysis.

Inclusion of epistatic effects in the genetic model con-

siderably increased the amount of variation explained for

both the behavioural and the morphological traits. For

boldness (time first entered) the variance explained in-

creased from 9% to 21% when only significant QTL were

included, rising to 44% when also including the suggestive

QTLs in the model (Fig. 2). For length, the variance ex-

plained increased from 10% to 23%.

Behav Genet (2006) 36:914–922 917

123

Although several different possible QTLs were identi-

fied using the genetic algorithm technique utilised in the

previous analysis (see Wright et al. 2006), none of the

QTLs detected by this method coincided with the novel

epistatic loci identified in the present analysis.

Interpretation of Epistasis

In the standard interval mapping analysis (Wright et al.

2006), B9.10 and B21.4 showed additive effects in different

directions, with B9.10 apparently showing transgressive

segregation: alleles from the low line (laboratory

strain—AA in Fig. 1) actually increasing the phenotype,

possibly with a degree of over-dominance (as compared to

the homozygous classes at this locus) also present. The

effects of B21.4 were more consistent with the parental

populations as the wild-type (high line) genotype increased

the phenotype.

Both of the epistatic pairs that affect boldness (time first

entered) and involve B21.4 indicate a dominant epistatic

mode of inheritance. In the case of the B21.4/B9.10 pair

(Fig. 1a), the decreasing effect of the B9.10 wild-type

genotype is masked by the presence of one or more of the

wild-type (S) B21.4 alleles, with all B21.4 SS or SA

genotypes giving a high phenotype, irrespective of the

B9.10 genotype. Only with the AA background are the

effects of the B9.10 genotype expressed, with the SS

homozygote having the lowest phenotype. The genotype–

phenotype map for B21.4 and B18.12 shows a similar

pattern (Fig. 1b, 1c). Once again, with either the SS or SA

150

200

250

300

350

400

450

500

550

600

SS SA AA

B12.39 genotype class

B9.10 SS B9.10 SAB9.10 AA

150

200

250

300

350

400

450

500

550

600

SS SA AA

B21.4 genotype class

B9.10 SSB9.10 SAB9.10 AA

a) c) *

0

5

10

15

20

25

SS SA AA

G23.8 genotype classes

G6.10 SS

G6.10 SA

G6.10 AA

150

200

250

300

350

400

450

500

550

600

SS SA AA

B21.4 genotype classes

B18.12 SSB18.12 SAB18.12 AA

d) * b)

Fig. 1 Genotype–phenotype plots for epistatic QTL pairs affecting

boldness (time first entered stimulus zone) (parts a–c) and length (part d).

Asterisks indicate epistatic pairs significant at the 5% genome-wide

level. (a) Boldness chromosome 21 locus (B21.4) and boldness

chromosome 9 locus (B9.10), (b) Boldness chromosome 21 locus

(B21.4) and boldness chromosome 18 locus (B18.12), (c) Boldness

chromosome 9 locus (B9.10) and boldness chromosome 12 locus

(B12.39) and (d) chromosome 23 locus (G23.8) and growth

chromosome 6 locus (G6.10). Genotypes at each locus are indicated

with SS (wild homozygote), SA (heterozygote) and AA (laboratory

homozygote), respectively

918 Behav Genet (2006) 36:914–922

123

B21.4 background all phenotypes are high. With the B21.4

AA background, the wild-type SS B18.12 homozygote has

effects in the expected direction, yielding a high phenotype

(though the A allele in this background appears to be

dominant, with both the AA and AS genotypes giving low

phenotypes).

The epistatic pair B12.39 and B9.10 (Fig. 1c) had a

significant effect on boldness (time first entered) and a

suggestive effect on boldness (PC1 score). Here, the two

homozygous wild-type loci (SS) interact to produce the

lowest phenotype, whilst the heterozygote genotype of the

B12.39 locus in the homozygous wild B9.10 background

produces the highest phenotype. This pattern was also

consistent with the epistatic pair for boldness (PC1 score).

The single epistatic QTL pair for length (mm) (G23.8

and G6.10) displays an interaction where the laboratory-

strain genotype at G6.10 decreases the phenotype in the

wild-type background of locus G23.8 (Fig. 1d). The

genotype–phenotype map indicates that in this case the two

wild-type (SS) homozygotes interact to produce a higher

phenotype than the other genotypes in the SS G23.8

background.

All of the anti-predator traits are strongly correlated with

each other (see Wright et al. 2006), whilst a similar cor-

relation can also be observed between the morphological

traits of length and weight. This is taken to indicate that the

observed epistatic QTLs for the different measures of

boldness reflect the same QTL pair being involved in

separate forms of the boldness phenotype. This lends

weight to the observed QTL actually being of relevance to

the behavioural or morphological trait in question.

Discussion

Reports of epistasis in QTL studies are still relatively rare.

However, this is because epistatic effects are rarely anal-

ysed: the importance of epistatic effects in populations can

only be truly ascertained when datasets are analysed

appropriately. Initially very large samples were thought to

be required for these analyses, and such datasets are still

desirable, but it is now clear that smaller sample sizes do

not preclude the detection of epistasis (Carlborg et al.

2003). In this study, epistasis was detected in an F2 pop-

ulation of less than 200, even with the use of stringent 5%

genome-wide significance levels. The results presented

here show that these analyses can reveal novel morpho-

logical and behavioural QTL and indicate high levels of

epistasis in evolutionarily important behavioural and mor-

phological traits.

Epistasis for boldness (time first entered) centres on a

network of four QTL (Fig. 3). Only two QTL were de-

tected through standard interval mapping techniques, thoseTa

ble

1O

ver

-vie

wo

fd

etec

ted

epis

tati

cp

airs

for

bo

ldn

ess

(B:

tim

efi

rst

ente

red

stim

ulu

szo

ne)

and

len

gth

(G).

*‘‘

Pre

vio

usl

yd

etec

ted

’’re

fers

tow

het

her

the

QT

Lin

qu

esti

on

was

det

ecte

db

y

stan

dar

din

terv

alm

app

ing

(see

Wri

gh

tet

al.

20

06

).C

orr

elat

ion

coef

fici

ents

±S

Ear

eg

iven

for

the

fou

rep

ista

tic

par

amet

ers

esti

mat

edin

the

full

mo

del

,w

hil

stth

eg

eno

me-

wid

esi

gn

ifica

nce

lev

elfo

rth

ep

air

inq

ues

tio

nis

also

giv

en

Ep

ista

tic

pai

rsM

od

elco

effi

cien

tsS

EG

eno

me-

wid

e

sig

nifi

can

ceQ

TL

1C

hro

mo

som

eL

oca

tio

n

(cM

)

Pre

vio

usl

y

det

ecte

d?*

QT

L2

Ch

rom

oso

me

Lo

cati

on

(cM

)

Pre

vio

usl

y

det

ecte

d?*

Ad

d·

add

Ad

d·

do

mD

om

·ad

dD

om

·d

om

B9

.10

91

0Y

esB

12

.39

12

39

No

–2

3.2

4±

22

.55

5.6

0±

29

.07

5.7

7±

27

.9–

10

4.5

3±

40

.1p

<0

.05

B9

.10

91

0Y

esB

21

.42

14

Yes

26

.88

±2

2.2

69

.02±

35

.1–

63

.58±

31

.1–

22

.27

±4

9.9

p<

0.2

0

B1

8.1

21

81

2N

oB

21

.42

14

Yes

–4

2.9

0±

24

.6–

77

.96

±3

9.2

14

4.7

7±

42

.78

6.4

0±

67

.7p

<0

.20

G6

.10

61

0N

oG

23

.82

38

Yes

–1

.83

±0

.8–

4.9

9±

1.2

0.5

5±

1.4

3–

2.4

0±

2.2

p<

0.0

5

Behav Genet (2006) 36:914–922 919

123

on chromosomes 9 and 21, with the other two QTL being

revealed by the epistatic analysis. The amount of variation

explained by the model also increased dramatically with

the inclusion of epistatic effects, from 16% to 44%, though

it must be noted that due to the relatively small sample size

used here, QTL effect sizes may be over-estimated (Beavis

1998). In the network, one QTL (B21.4) has dominant

epistatic interactions with two of the others (B9.10 and

B18.12), whilst the remaining pairwise interaction (B9 with

B12.39) is multiplicative in nature. While the wild-type

genotype at B9.10 and B12.39 will serve to reduce the

degree of boldness, these are counter-acted by the effects

of the B21.4 locus (effectively masking the effects of B9.10

if it contains at least one of the wild-type alleles), with the

low homozygote phenotype at this locus also being masked

by the wild-type genotype at B18.12. At a population level,

this is expected to lead to high boldness phenotypes in the

particular natural population from which the wild-type line

was derived.

The laboratory line has diverged recently, in evolu-

tionary terms, from a natural population (first being intro-

duced to the laboratory in the 1970s, and prior to this being

taken from a pet shop, with the population presumably

being collected from the wild at some point previous to

this—see http://www.zfin.org for further details).

Although, the exact point of origin of the AB strain is

impossible to detect, it nevertheless represents fairly sub-

stantial divergence in terms of selection pressures as

compared to the recently collected Santal strain. Therefore,

most differences between laboratory and wild-type lines

are likely to be due to high frequency, advantageous alleles

(for high boldness scores) in the wild-type line, due to the

directional or perhaps stabilising effect of natural selection,

and increase in frequency under relaxed or altered selection

pressures in the laboratory line of low boldness alleles. As

long as at least one high wild-type allele is present at

B21.4, any effects of polymorphisms at B9.10 and B18.12

will be effectively masked. Thus low boldness alleles at

these loci might be present in natural populations at

appreciable frequencies despite selection and contribute to

the evolution of reduced boldness following domestication

when the trait is neutral, or under weak selection for de-

creased boldness. Thus both the observed dominance and

epistatic interactions between loci are consistent with the

B9.10B21.4

B12.39B18.12

Fig. 3 A cartoon of the network of four epistatic QTL affecting

boldness (time first entered). Loci detected in the epistatic analysis are

connected by their significant or suggestive pairwise interactions. The

significant interaction is marked with an asterisk. The directions of

the arrows represent the principal directions of effects, whilst the sizes

of the arrows represent the strength of the epistasis (in terms of r-

squared accounted for). The size of the loci represents the strength of

their marginal effect

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

mar

gina

lef

fect

s on

ly

sim

ulat

aneo

usm

appi

ngm

argi

nal

effe

cts

only

sim

ulta

neou

sm

appi

ng w

ithep

ista

sis

mar

gina

lef

fect

s on

ly

sim

ulat

aneo

usm

appi

ngm

argi

nal

effe

cts

only

sim

ulta

neou

sm

appi

ng w

ithep

ista

sis

boldness(time first) length(mm)

tota

l r-s

qu

ared

suggestive

significant

Fig. 2 Variance explained by

epistatic interactions in the form

of r-squared values for traits

using models with loci selected

by standard interval mapping

only, models selected with

simultaneous interval mapping,

but without epistatic

interactions, and models

including all marginal and

epistatic effects of the loci

detected in the epistatic QTL

analysis

920 Behav Genet (2006) 36:914–922

123

likely selection experienced by this trait (Falconer and

Mackay 1996). However, when interpreting these results, it

would clearly be desirable to know more about genetic

variation within and among natural populations at puta-

tively neutral loci and to test whether the low boldness

alleles in the laboratory line are indeed derived from

natural polymorphism rather than de novo mutation.

Though there is some evidence for high variability both

within and among natural zebrafish populations (Gratton

et al. 2004; Wright 2004), no studies have yet looked at the

presence of specific alleles in natural zebrafish populations.

In another experimental cross utilising wild and

domesticated lines (in this case a wild · domestic fowl

cross (Carlborg et al. 2003; Kerje et al. 2003)), many

epistatic pairs were identified. The authors speculated that

the large degree of divergence between populations was a

potential reason for this observation, i.e., the length of time

since divergence might have increased the potential for co-

adaptation within lines (Carlborg et al. 2004). Although in

the study presented here there are also several large epi-

static interactions, the lines involved have not diverged to

the same degree as the above example. The AB laboratory

strain has been captive-bred for many years as compared to

the wild-derived Santal population (only two generations

from wild), however this difference is probably far less

than that between a chicken broiler line (with strong arti-

ficial selection for potentially hundreds of years, and spe-

cifically selected from the 19th century onwards as a layer

breed) and a wild-type red jungle fowl (Fumihito et al.

1994). If such strong differential selection had taken place,

the observed epistatic interactions may have been more

evenly spread throughout the traits observed.

The degree of epistasis that is expected to be present in a

trait depends on both the type of trait and the evolutionary

pressures acting upon it. Life history traits typically possess

low additive genetic variance, with larger dominance and

epistatic variance, morphological traits often possess high

additive variance with little epistasis or dominance, whilst

most behavioural traits fall somewhere between these two

extremes (Mousseau and Roff 1987; Roff and Mousseau

1987). Anti-predation behaviour is a trait strongly linked

with survival, depending on the environment (Krause and

Ruxton 2002; Seghers 1974; Magurran et al. 1995), and as

such is likely to be closely correlated with fitness. The

partitioning of the genetic variance for anti-predator

behaviour is therefore likely to be similar to a life-history

trait subjected to a degree of directional selection, chiefly

being comprised of dominance and epistatic variance

(Falconer and Mackay, 1996). The Boldness trait shows

this pattern while dominance and epistasis have less impact

on the morphological traits studied here (though potentially

there may also be some directional selection acting on

growth in zebrafish). The above view is also supported by

an early study by Kearsey and Kojima (1967), in which 12

traits were analysed in Drosophila. Major fitness compo-

nents showed epistatic interaction and strong directional

selection, whilst morphological traits (body size and bristle

number, etc) showed no dominance or interaction varia-

tion. It can also be pointed out that additive variance may

be present, but is less in comparison to the non-additive

and environmental variance present (Houle 1992), thereby

reducing its apparent effects. The failure to detect any ef-

fects of epistasis in shoaling tendency could indicate that

certain behavioural anti-predator traits do not exhibit a

similar pattern to boldness, however the weak power in the

case of this trait, with no significant QTL detected using

standard interval mapping, cautions against extrapolating

too much from this result.

In summary, these results demonstrate the importance of

epistasis in the genetic architecture of an evolutionarily

relevant behavioural trait, using a wild-derived population.

These results indicate not only that behavioural traits are

amenable to analysis of epistasis, but also that useful in-

sights can be gained using moderate sample sizes. This

study also contributes to the knowledge of the degree of

epistasis that exists within natural populations.

Acknowledgments DW was supported by a BBSRC studentship

and by an EU motility grant (HPRI-CT-2001-00153) to visit The

Linnaeus Centre for Bioinformatics. O.C. acknowledges the Knut and

Alice Wallenberg foundation for financial support.

References

Anholt RRH, Dilda CL, Chang S, Fanara J-J, Kulkarni NH, Ganguly I,

Rollman SM, Kamdar KP, Mackay TFC (2003) The genetic

architecture of odor-guided behaviour in Drosophila: epistasis and

the transcriptome. Nat Genet 35:180–184

Anholt RRH, Lyman RF, Mackay TFC (1996) Effects of single P-

element insertions on olfactory behaviour in Drosophila mela-

nogaster. Genetics 143:293–301

Baird TA, Ryer CH, Olla BL (1991) Social enhancement of foraging

on an ephemeral food source in juvenile wall-eye pollock,

Theragra chalcogramma. Environ Biol Fishes 31:307–311

Beavis WD (1998) QTL analyses: power, precision and accuracy. In:

49th Annual Corn and Sorghum Industry Research Conference,

Washington DC, ASTA, pp 250–266.

Carlborg O, Andersson L (2002) Use of randomization testing to

detect multiple epistatic QTLs. Genet Res 79:175–184

Carlborg O, Andersson L, Kinghorn B (2000) The use of a genetic

algorithm for simultaneous mapping of multiple interacting

quantitative trait loci. Genetics 155:2003–2010

Carlborg O, Brockmann GA, Haley CS (2005) Simultaneous mapping

of epistatic QTL in DU6i · DBA/2 mice. Mamm Gen 16:481–

494

Carlborg O, Haley CS (2004) Epistasis: too often neglected in com-

plex trait studies? Nat Rev Genet 5:618–625

Carlborg O, Hocking PM, Burt DW, Haley CS (2004) Simulta-

neous mapping of epistatic QTL in chickens reveals clusters

of QTL pairs with similar genetic effects on growth. Genet

Res 83:197–209

Behav Genet (2006) 36:914–922 921

123

Carlborg O, Kerje S, Schutz K, Jacobsson L, Jensen P, Andersson

L (2003) A global search reveals epistatic interaction between

QTL for early growth in the chicken. Genome Res 13:413–

421

Civetta A, Montooth KL, Mendelson M (2005) Quantitative trait loci

and interaction effects responsible for variation in female post-

mating mortality in Drosophila simulans and D. sechellia intg-

rogression lines. Heredity 94:94–100

Csanyi V (1985) Ethological analysis of predator avoidance by the

paradise fish (Macropodus opercularis L.) I. Recognition and

learning of predators. Behaviour 92:227–240

Falconer DS, Mackay TFC (1996) Introduction to quantitative

genetics. Prentice Hall

Fedorowicz GM, Fry JD, Anholt RRH, Mackay TFC (1998) Epistatic

interactions between smell-impairedloci in Drosophila mela-

nogaster. Genetics 148:1885–1891

Flint J, DeFries JC, Henderson ND (2004) Little epistasis for anxiety-

related measures in the DeFries strains of laboratory mice.

Mamm Gen 15:77–82

Fumihito A, Miyake T, Sumi SI, Ohno S, Kondo N (1994) One

species of the red jungelfowl (Gallus-gallus gallus) suffices as

the matriarchic ancestor of all domestic breeds. PNAS USA

91:12505–12509

Gratton P, Allegrucci G, Gallozzi M, Fortunato C, Ferreri F, Sbordoni

V (2004) Allozyme and microsatellite genetic variation in nat-

ural samples of zebrafsih, Danio rerio. J Zool System Evol Res

42:54–62

Haley CS, Knott SA (1992) A simple regression method for mapping

quantitative trait loci in line crosses using flanking markers.

Heredity 69:315–324

Haley CS, Knott SA, Elsen JM (1994) Mapping quantitative trait loci

in crosses between outbred lines using least squares. Genetics

136:1195–1207

Hood HM, Belknap JK, Crabbe JC, Buck KJ (2001) Genome-wide

search for epistasis in a complex trait: pentobarbital withdrawal

convulsions in mice. Behav Genet 31:93–100

Houle D (1992) Comparing evolvability and variability in quantita-

tive traits. Genetics 130:195–204

Kearsey MJ, Kojima K-I (1967) The genetic architecture of body

weight and egg hatchability in Drosophila melanogaster.

Genetics 56:23–37

Kerje S, Carlborg O, Jacobsson L, Schutz K, Hartmann C, Jensen P,

Andersson L (2003) The two-fold difference in adult size be-

tween the red junglefowl and white leghorn chickens is largely

explained by a limited number of QTLs. Anim Genet 34:264–

274

Knott SA, Marklund L, Haley CS, Andersson K, Davies W, Ellegren H,

Fredholm M, Hansson I, Hoyheim B, Lundstrom K, Moller M,

Andersson L (1998) Multiple marker mapping of Quantitative

Trait Loci in a cross between outbred wild boar and large white

pigs. Genetics 149:1069–1080

Krause J, Ruxton G (2002) Living in groups. Oxford University Press,

Oxford

Landeau L, Terborgh J (1986) Oddity and the confusion effect in

predation. Anim Behav 34:1372–1380

Lander ES, Kruglyak L (1995) Genetic dissection of complex traits:

guidelines for interpreting and reporting linkage results. Nat

Genet 11:241–247

Magurran AE, Seghers BH, Shaw PW, Carvalho GR (1995) The

behavioural diversity and evolution of guppy, Poecilia reticu-

lata, populations in Trinidad. Advances in the study of behavior.

Academic Press Inc., San Diego, pp 363–439

Mousseau TA, Roff DA (1987) Natural-selection and the heritability

of fitness components. Heredity 59:181–197

Nakamichi R, Ukai Y, Kishino H (2001) Detection of closely linked

multiple quantitative trait loci using a genetic algorithm.

Genetics 158:463–475

Peripato AC, De Brito RA, Matioloi SR, Pletscher LS, Vaughn TT,

Cheverud JM (2004) Epistasis affecting litter size in mice. J Evol

Biol 17:593–602

Pitcher TJ (1992) Who dares wins: the function and evolution of

predator inspection behaviour in fish shoals. Neth J Zool 42:371–

391

Pitcher TJ, Parrish JK (1993) Functions of shoaling behaviour in

teleosts. In: Pitcher TJ (ed) Behaviour of teleost fishes. Chapman

and Hall, London, pp 363–439

Ranta E, Kaitala V (1991) School size affects individual feeding

success in three-spined sticklebacks (Gasterosteus aculeatus).

J Fish Biol 39:733–737

Roff DA, Mousseau TA (1987) Quantitative genetics and fitness –

lessons from Drosophila. Heredity 58:103–118

Ruppell O, Pankiw T, Page Jr RE (2004) Pleiotropy, epistasis and

new QTL: the genetic architecture of honey bee foraging

behavior. J Hered 96:481–491

Seghers BH (1974) Schooling behaviour in the Guppy (Poecilia re-

ticulata): an evolutionary response to predation. Evolution

28:486–489

Wright D (2004) Identification of behavioural QTL in zebrafish(Danio rerio). PhD Thesis, University of Leeds, Leeds

Wright D, Nakamichi R, Krause J, Butlin RK (2006) QTL analysis of

behavioural and morphological differentiation between wild and

laboratory zebrafish (Danio rerio). Behav Genet epub Jan 12:1–14

922 Behav Genet (2006) 36:914–922

123