Evidence for adaptive phenotypic differentiation in Baltic Sea sticklebacks

Authors


  • Data deposited at Dryad: doi:10.5061/dryad.958f6

Correspondence: J. DeFaveri, Ecological Genetics Research Unit, Department of Biosciences, University of Helsinki, PO Box 65, FI-00014 Helsinki, Finland. Tel.: +358 9 191 57710; fax: +358 9 191 57694; e-mail: jacquelin.defaveri@helsinki.fi

Abstract

The evidence for adaptive phenotypic differentiation in mobile marine species remains scarce, partly due to the difficulty of obtaining quantitative genetic data to demonstrate the genetic basis of the observed phenotypic differentiation. Using a combination of phenotypic and molecular genetic approaches, we elucidated the relative roles of natural selection and genetic drift in explaining lateral plate number differentiation in threespine sticklebacks (Gasterosteus aculeatus) across the entire Baltic Sea basin (approximately 392 000 km2). We found that phenotypic differentiation (PST = 0.213) in plate number exceeded that in neutral markers (FST = 0.008), suggesting an adaptive basis for the observed differentiation. Because a close correspondence was found between plate phenotype and genotype at a quantitative trait loci (QTL; STN381) tightly linked to the gene (Ectodysplasin) underlying plate variation, the evidence for adaptive differentiation was confirmed by comparison of FST at the QTL (FSTQ = 0.089) with FST at neutral marker loci. Hence, the results provide a comprehensive demonstration of adaptive phenotypic differentiation in a high-gene-flow marine environment with direct, rather than inferred, verification for the genetic basis of this differentiation. In general, the results illustrate the utility of PSTFSTFSTQ comparisons in uncovering footprints of natural selection and evolution and add to the growing evidence for adaptive genetic differentiation in high-gene-flow marine environments, including that of the relatively young Baltic Sea.

Introduction

Differentiating between neutral and selective causes of divergence has fuelled a wide range of studies in both evolutionary (e.g. Lynch et al., 1999; André et al., 2010; Mariani et al., 2012; Whitlock & Glibert, 2012) and conservation (e.g. Crandall & Bininda-Emonds, 2000; Johansson et al., 2007; Hansen et al., 2012) biology. In particular, the ability to identify selection becomes important when populations exist in physically continuous, but environmentally heterogeneous systems. Under such circumstances, the homogenizing effects of gene flow may appear to constrain local adaptation (Slatkin, 1985; Räsänen & Hendry, 2008). Nevertheless, phenotypic (or genetic) divergence in adaptive traits can arise in response to spatially varying selective forces that favour different phenotypes in different environments (Slatkin, 1973; Hall et al., 2007; Savolainen et al., 2007). Therefore, understanding the distribution of adaptive traits – and their underlying genetic basis – in relation to environmental gradients is central for evaluations of local adaptation in the presence of gene flow. This is particularly relevant when the selective forces are correlated with geographical distance, because spurious correlations can arise from nonadaptive processes of isolation by distance (Storz, 2002; Hangartner et al., 2012; Meirmans, 2012). Therefore, after controlling for geographical distance, the contributions of selective and neutral processes in shaping the patterns of variation along environmental gradients may be more accurately assessed.

The marine environment provides a good example of such situations; due to the absence of obvious physical barriers, the opportunity for gene exchange among marine populations is typically high (Cano et al., 2008), and environmental factors acting as potential selective agents vary geographically (Conover et al., 2006). Although there have been many demonstrations of phenotypic differentiation among physically connected marine fish populations (e.g. Barlow, 1961; McQuinn, 1997; Cadrin, 2000; O'Reilly & Horn, 2004), evidence that the observed phenotypic divergence is adaptive and genetically based remains scarce (but see Billerbeck et al., 1997; Conover et al., 2006; Marcil et al., 2006; Cano et al., 2008; Hice et al., 2009). In the few situations where this has been demonstrated, the sampling schemes have typically not been sufficient enough to define the spatial scale of this divergence (Conover, 1998; Conover et al., 2006; but see: Hice et al., 2009). In general, determining whether the phenotypic patterns are adaptive and hence genetically based requires individuals originating from different environments to be reared under the similar environmental conditions (i.e. in a common-garden setting). However, common-garden experiments are rare in marine fish due to practical limitations (see Hutchings et al., 2007; Hice et al., 2009 for exceptions). Furthermore, identifying the specific genes that underlie the phenotypic patterns remains challenging, largely due to the fact that most phenotypes are the product of many underlying genes that interact with each other and with the environment (Naish & Hard, 2008; Mackay et al., 2009; Hill & Kirkpatrick, 2010). As such, the geography of adaptive phenotypic and genotypic variation in marine populations remains poorly understood (Hice et al., 2009).

One method that has been commonly applied towards identifying the forces driving phenotypic diversification involves the comparison of quantitative genetic divergence at putatively adaptive traits (QST; Spitze, 1993) with neutral genetic divergence (FST; Wright, 1951) among populations (Merilä & Crnokrak, 2001; McKay & Latta, 2002; Leinonen et al., 2008; Whitlock, 2008). If different trait values are favoured by natural (or sexual; Yu et al., 2011) selection in different localities/environments, local adaptation will occur and QST will exceed divergence at neutral genetic loci, which primarily reflect differentiation due to stochastic processes such as genetic drift (i.e. QST > FST). On the other hand, stronger neutral divergence as compared to that at quantitative traits (QST < FST) indicates that stabilizing or uniform selection is the main force acting to conserve similarity of trait values between populations (but see Whitlock, 1999; Goudet & Buchi, 2006; and Santure & Wang, 2009 for examples of QST suppression by epistasis or dominance). Under true neutrality or very weak selection, divergence at quantitative traits and neutral markers should be similar (QST = FST). Estimation of indices of quantitative trait divergence (QST) in their traditional sense is only possible through highly controlled breeding experiments that allow additive genetic components to be isolated from environmental components of divergence (e.g. Leinonen et al., 2008; Pujol et al., 2008). However, a strictly phenotypic measure of differentiation (PST; Leinonen et al., 2006) has also been used as a surrogate for QST, allowing direct investigations of phenotypic divergence in wild populations (reviewed in Brommer, 2011).

Although seemingly straightforward, QST (or PST)–FST comparisons are not without limitations or biases, most of which have been thoroughly addressed in several reviews and meta-analyses (Merilä & Crnokrak, 2001; McKay & Latta, 2002; Leinonen et al., 2008, 2013; Whitlock, 2008). Although the potential inaccuracies and violated assumptions in estimating QST (or PST) have been suggested to create an upward bias in the frequently reported incidents of QST > FST (Miller et al., 2008; Whitlock, 2008), recent re-evaluations of the QSTFST framework (Kronholm et al., 2010; Edelaar & Björklund, 2011; Edelaar et al., 2011; Meirmans & Hedrick, 2011) have highlighted how this upward bias may be further compounded by estimates of genetic divergence based on highly polymorphic markers that tend to deflate FST (Hedrick, 2005; Jost, 2008). An additional assumption of FST estimations is that under the island model at equilibrium, migration rates are exceedingly higher than mutation rates, and hence, FST is essentially determined independent of marker mutation rate (Storz, 2002; Kronholm et al., 2010; Edelaar & Björklund, 2011). Therefore, as mutation rates increase and/or migration rates decrease, the assumption of QST = FST under neutrality may be violated, particularly in systems of low gene flow (Hendry, 2002; LeCorre & Kremer, 2003, 2012). Consequently, several authors have called for caution to be exercised when interpreting results and drawing conclusions from QSTFST comparisons (e.g. Hendry, 2002; Whitlock, 2008; Santure & Wang, 2009; Brommer, 2011). Nevertheless, once critically applied, QST/PSTFST comparisons continue to be a useful and popular method for detecting local adaptation (e.g. Holand et al., 2011; Frei et al., 2012; Hangartner et al., 2012; Kronholm et al., 2012; Mariani et al., 2012; Leinonen et al., 2013). However, this approach has not often been applied towards marine fishes (but see: Leinonen et al., 2006; Mobley et al., 2011; McCairns & Bernatchez, 2012), due to logistical difficulties of breeding and rearing them in common-garden settings (Hutchings et al., 2007).

Although estimates of QST (or PST) are generally obtained from phenotypic data, in some cases the underlying loci – known as quantitative trait loci (QTLs) – have been identified. Hence, exploring the molecular mechanisms of adaptive evolution in the loci underlying adaptive traits can provide further insight towards the influence of natural selection on phenotypic divergence. Theoretical considerations (McKay & Latta, 2002; LeCorre & Kremer, 2003, 2012) suggest that discrepancy between QST and FST at the QTL (FSTQ) is likely to occur if (i) the trait expression is controlled by many QTLs, (ii) the strength of selection on each individual QTL is weak, (iii) gene flow is high, or (iv) a combination of these factors occurs. Yet, relatively few studies have empirically tested these predictions by directly comparing the allelic variation at QTL in relation to the corresponding trait variation in natural populations (e.g. Peichel et al., 2001; Rogers & Bernatchez, 2005; Hall et al., 2007; Raeymaekers et al., 2007; Mäkinen et al., 2008; Kronholm et al., 2012; Liedvogel et al., 2012). Most of these studies support the modelled expectations that differentiation at QTL is more accurately reflected in QST than in FSTQ due to covariance between QTLs (cf. McKay & Latta, 2002). Furthermore, because divergent selection is expected to reduce gene flow at adaptive loci, the genomic signature of increased differentiation at those loci can be detected through outlier analyses (Beaumont & Balding, 2004; Vasemägi & Primmer, 2005). However, these signals are difficult to detect when an adaptive trait is controlled by a large number of loci, also as a result of covariances among allelic effects (LeCorre & Kremer, 2003; Raeymaekers et al., 2007; Santure & Wang, 2009; Leinonen et al., 2013). Nonetheless, it should be possible to genetically confirm phenotypic signals of selection (as detected with QST or PST) for a particular trait if it is controlled by one major locus (e.g. Kronholm et al., 2012), thereby complementing our understanding of evolutionary responses to selection.

The purpose of this study was two-fold. Firstly, using the threespine stickleback (Gasterosteus aculeatus) as a model, we aimed to explore the utility of PSTFST comparisons in inferring adaptive population differentiation in a trait (lateral plate number) for which the development and expression are known to be largely unaffected by environmental and nonadditive genetic effects (Hermida et al., 2002; Colosimo et al., 2004; McGuigan et al., 2011). Because most of the variation in plate number is controlled by a major QTL tightly linked to the Ectodysplasin gene (Colosimo et al., 2004), we were able to further compare the divergence in this QTL (FSTQ) with the phenotypic estimate of divergence (PST) in the corresponding trait. Secondly, after establishing this relationship, we used FSTQ to study the genetic patterns of lateral plate differentiation in a large number of samples (n = 1288 individuals collected from 38 locations) covering the entire Baltic Sea basin (392 000 km2) – a high-gene-flow environment characterized by steep physiochemical gradients.

Materials and Methods

Study species and system

The threespine stickleback has proved to be an excellent system in which to explore the genetic basis of morphological evolution, as a wide array of phenotypic diversity has evolved over a very brief period of time (approximately 10 000 years; Bell & Foster, 1994). Despite the striking amount of divergence that exists between different populations inhabiting varying habitats (viz. marine, lakes, ponds, rivers), one trait that has been consistently conserved among marine and estuarine populations is heavy body armour. Specifically, although most marine/anadromous populations retain a complete row of lateral plates along the length of their entire body, a reduction in number of lateral plates has been repeatedly observed in many freshwater populations throughout their global distribution (Wootton, 1976; Klepaker, 1995; Colosimo et al., 2005). Although identification of the selective forces directly contributing to the parallel loss of body armour has remained a challenge, several hypotheses have suggested abiotic factors (e.g. temperature, salinity tolerance, external calcium availability) along with biotic forces (e.g. swimming performance and predation regime) as candidate selective agents (reviewed in: Reimchen, 1994; Barrett, 2010). Many of these factors vary also within the Baltic Sea and Kattegat/Skagerrak straits linking the Baltic Sea to the Atlantic Ocean, as does the number of lateral plates among stickleback populations within estuaries in these regions (Munzing, 1963; Wootton, 1976, 1984; Bańbura, 1994; Klepaker, 1996; Raeymaekers et al., 2007). In addition, because much of the variation in plate number appears to be under the control of a major quantitative trait locus (Ectodysplasin or ‘EDA’; Colosimo et al., 2005), investigations of the genetic variability at the locus underlying this trait are straightforward.

Sample collection

Sticklebacks were collected from 38 locations along the coast of the Baltic Sea and Kattegat/Skagerrak straits (Table 1; Fig. 1) during the breeding season (May–June) of 2009. Fish that were displaying male nuptial colouration (i.e. red throat and blue eyes; Rowland, 1982) or abdomens enlarged with ripe eggs were classified as adults and selected for the study. Samples were collected with seine nets and minnow traps (mesh size 6 mm), which were deployed along the shoreline at similar depths to avoid any potential depth- or distance-related sampling bias. Average annual salinities ranged from 0‰ to 30‰ across all sites (Table 1; see also Methods S1), thus encompassing an ecological gradient from nearly marine to nearly freshwater environments. A thermal gradient also exists along the Baltic Sea, with average annual temperatures ranging from 5.6 to 10.6 °C across all sites (Table 1; see Methods S1). All fish were killed on site with an overdose of tricaine methanesulphonate (MS-222) and transferred to 96% ethanol following the removal of pectoral fins, which were stored separately in 96% ethanol for subsequent molecular analyses.

Table 1. Sampling information of populations used in the study
CountryLocationCodeCoordinatesAverage annual salinity (ppt)Average annual temperature (°C) H E H O A R F IS N Mean (SD) standard length (mm)
  1. Basic population genetic parameters (HE, expected heterozygosity; HO, observed heterozygosity; AR, allelic richness; FIS, inbreeding coefficient) as calculated over 20 neutral microsatellite loci and mean (SD) standard lengths of the fish are reported. Populations in bold were phenotyped for number of lateral plates.

NorwayTonsbergTON

59°16′14″N

10°25′58″E

30.208.680.7260.69810.20.04735
NorwayKristiansandKRI

58°09′56″N

08°01′52″E

29.869.160.7360.71610.10.02834
DenmarkMariager FjordMAR

56°38′58″N

09°57′05″E

26.829.880.7330.70810.60.02736
SwedenFiskebäckskil FIS

58°14′05″N

11°24′06″E

23.529.890.7410.70710.30.0403549.5 (8.9)
DenmarkHvide SandeHVI

56°00′08″N

08°08′01″E

19.2610.150.7410.72110.90.03135
GermanyLindauKAP

54°35′29″N

09°49′57″E

17.0710.640.7510.73410.70.03436
SwedenVarbergVAR

57°06′01″N

12°14′25″E

16.639.860.7190.69710.00.02323
GermanyRostock ROS

54°05′33″N

12°09′12″E

11.2010.360.7120.79.70.0143656.4 (5.6)
DenmarkCopenhagen COP

55°42′05″N

12°35′55″E

10.769.850.7460.73610.40.0233648.5 (3.8)
DenmarkStege BugtSTE

54°58′48″N

12°17′13″E

9.1510.410.7570.7410.70.01736
SwedenKarlskronaKAR

56°10′13″N

15°24′34″E

7.559.430.7230.7139.00.01426
LatviaKolka KOL

57°45′53″N

22°35′27″E

7.548.740.7290.7439.5−0.0203654.7 (6.9)
SwedenVästlands shore VAS

56°59′53″N

18°11′53″E

7.219.390.7180.7329.70.0083558.7 (5.3)
PolandPuck Bay PUCK

54°45′37″N

18°30′07″E

7.199.990.7330.70410.20.0473551.2 (4.5)
PolandKarsiborKBOR

53°52′14″N

14°17′03″E

7.009.690.7190.7099.3034
SwedenKalmarVIK

57°24′18″N

16°38′33″E

6.828.320.7020.6979.80.00636
EstoniaSaaremaaVBB

58°25′05″N

22°06′33″E

6.598.810.7260.6949.50.04636
SwedenNyköping NYK

58°39′04″N

17°06′02″E

6.588.200.7250.69210.00.0083152.0 (6.6)
LatviaLenkupeLEN

56°39′06″N

21°03′14″E

6.379.270.7130.7169.60.01535
FinlandVuosnainenVUO

60°31′10″N

21°14′14″E

6.067.940.7350.72110.10.05936
FinlandEckerö ALA

60°11′36″N

19°37′06″E

6.038.160.7260.72610.10.0383156.3 (5.2)
FinlandTvärminneTVA

59°50′20″N

23°12′15″E

5.577.670.7280.7289.8−0.01029
SwedenHudiksvall HUD

61°43′03″N

17°06′26″E

5.327.300.7090.7039.40.0533656.3 (6.3)
EstoniaKäsmu BayEKAS

59°35′08″N

25°54′52″E

5.047.920.7310.70410.20.03834
SwedenAvikeAVI

62°30′31″N

17°43′29″E

4.826.360.7220.71110.00.0236
SwedenForsmarkFOR

60°24′09″N

18°11′05″E

4.737.820.7120.689.8036
EstoniaLetipeaLET

59°32′58″N

26°35′49″E

4.637.800.7260.71110.10.06134
FinlandMerirastila HEL

60°12′09″N

25°06′01″E

4.637.490.7010.65810.00.0233651.2 (6.4)
LatviaMeisa gravisMEI

57°42′13″N

24°21′35″E

4.478.750.7270.7359.5−0.03036
FinlandÖjaOJA

63°50′00″N

22°52′12″E

3.626.470.7070.6709.90.00934
FinlandPyhäjoki PJM

64°28′42″N

24°13′13″E

3.446.090.7290.7369.8−0.0103454.9 (2.7)
FinlandKaskinen FKAS

62°23′02″N

21°13′30″E

3.157.510.710.7099.60.0223657.7 (3.3)
SwedenSikea SIK

64°09′34″N

20°58′37″E

2.975.660.7170.6759.70.0023558.4 (5.2)
FinlandHaminaHAM

60°33′55″N

27°12′01″E

2.986.670.720.70310.30.04532
RussiaPrimorskPRI

60°21′02″N

28°37′04″E

2.447.700.6950.7137.8−0.02021
SwedenRaneaRAN

65°50′47″N

22°22′06″E

1.515.850.7110.6849.4−0.01036
RussiaPetergofa PET

60°03′14″N

29°58′16″E

0.007.420.7060.7009.70.0403556.7 (2.7)
RussiaReki NevaNEV

59°55′70″N

30°13′56″E

0.006.410.6990.7029.80.01335
Figure 1.

Map of sampling locations. Population codes follow those reported in Table 1. Bold, italic labels indicate populations that were phenotyped for number of lateral plates and used in PST comparisons. Inner pie graphs show the frequency of alleles associated with low (light grey) and high (dark grey) plate counts; outer pie graphs show the average number of lateral plates, expressed as the proportion out of 25.

DNA extraction and genotyping

Genomic DNA was extracted from the pectoral fins using a 10% Chelex®-100 resin (Bio-Rad Laboratories, Richmond, CA, USA). An average of 34 individuals per site were genotyped at 20 putatively neutral, unlinked microsatellite loci (Peichel et al., 2001; Table S1) – selected based on their distant positions from any predicted genes in the Ensembl Stickleback Genome browser version 66. An additional quantitative trait locus (QTL) tightly linked with the EDA gene controlling lateral plate variation (STN381; Colosimo et al., 2005) was also screened in the same individuals. Although this particular marker has been primarily used to indicate plate pattern rather than plate number (Colosimo et al., 2005), it nonetheless falls within the major chromosomal region that accounts for more than 75% of the variance in plate number (Colosimo et al., 2004; Kitano et al., 2008) and is in fact tightly linked with the EDA gene (Raeymaekers et al., 2007; Mäkinen et al., 2008). PCRs were carried out in a 10-μL reaction volume with a final concentration of 1 × Qiagen master mix, 0.5 × Q-solution, 2 pmol of each forward and reverse primer and approximately 5–20 ng template DNA. Forward primers were fluorescently labelled with either FAM, HEX or TET dyes, and panels were arranged to avoid overlapping size ranges. PCR cycling profile started with an initial activation step at 95 °C for 15 min, followed by 30 cycles of 94 °C for 30 s, 53 °C for 90 s and 72 °C for 60 s and concluded with an extension step at 60 °C for 5 min. PCR products were resolved using a MegaBace 1000 capillary sequencer (Amersham Biosciences, Piscataway, NJ, USA), and genotypes were scored with fragment profiler 1.2 software (Amersham Biosciences).

Quantification of lateral plates

A subset of 14 sites – including sites from the highly saline Kattegat/Skagerrak straits, brackish Baltic Proper and low-saline Gulf of Finland and Bay of Bothnia (Table 1; Fig. 1) – were selected for quantification of lateral plates. These samples, which were also used in the molecular analyses, were fixed in 10% formalin for at least 48 h, bleached in 3% H2O2/0.5% KOH and stained in a 2% KOH/alizarin red solution (modified from Potthoff, 1984) for ease of visualization of bony armour. The left side of each fish was photographed from a standard angle using a Nikon D60 (Nikon Inc., Melville, NY, USA) digital camera, and lateral plates were counted from the photographs. Because all individuals were keeled, the plates on the caudal peduncle (composing the keel; Reimchen, 2000) were not counted. Accordingly, the maximum number of plates observed was 25. Lateral plate counts were repeated for half of the samples by a second, independent source, and repeatability was confirmed to be high (R = 0.96, < 0.001).

Analysis of genetic markers

Basic population parameters (expected heterozygosity HE, observed heterozygosity HO, Wright's inbreeding coefficient FIS, deviations from Hardy–Weinberg equilibrium and linkage disequilibrium) were estimated with the program fstat 2.9.3 (Goudet, 2001). Allelic richness (AR) was calculated after being rarefied to 42 alleles (number of alleles in the population with the smallest sample size) using the program hp-rare (Kalinowski, 2005). Neutral genetic differentiation was estimated with θ (Weir & Cockerham, 1984) after removal of the QTL from the data set, following 1000 iterations and sequential Bonferroni corrections for multiple testing (Rice, 1989). Divergence at the QTL underlying plate number (FSTQ) was estimated as above, but independently. To account for mutation processes, locus-specific RST values were also calculated. All differentiation estimates were calculated with fstat 2.9.3 software. Nonmetric multidimensional scaling (NMDS) was performed to visualize pairwise distances as calculated with the neutral markers, using past 2.01 software (Hammer et al., 2001). The same analysis was performed with pairwise distances as calculated only with the QTL (marker STN381). Patterns of isolation by distance were tested for by comparing linearized FST values [FST/(1–FST)] with log-transformed geographical distances as measured by the shortest waterway (km) between sampling locations, using a Mantel test (Mantel, 1967).

Phenotypic differences in lateral plate number: PSTFST comparisons

We tested for population differences in square-root-transformed lateral plate numbers using one-way ANOVA and post hoc Tamhane tests (due to unequal variances) in spss 15.0 for Windows software (SPSS Inc., Chicago, IL, USA). Phenotypic divergence (PST) at this trait was calculated using the equation:

display math

where within (σ2W)- and among (σ2B)-population variance components were estimated between each pair of sites.

The posterior distribution of PST and each variance component was estimated with a Bayesian approach (Gelman et al., 2004) by running two chains of 10 000 iterations (following 5000 burn-in iterations) with a Gibb's sampler in winbugs 1.4 (Spiegelhalter et al., 2003). Every second iteration was drawn from each chain to obtain the posterior distribution. To determine whether the difference in lateral plate numbers was influenced by divergent selection, we followed the method of Whitlock & Guillaume (2009) by testing whether PST is in the tail of the distribution of neutral traits. This null distribution was simulated with FST values from the neutral marker data set and the within-population variance component, following Lind et al. (2011). The distribution of the test statistic PSTFST, which is the null hypothesis of a neutrally evolved trait, was simulated 10 000 times in R using the code provided by Lind et al. (2011). The observed PSTFST for lateral plate number was then compared with the simulated distribution. The expectation of this comparison is that PSTFST will exceed the neutral distribution if the number of lateral plates is under selection (Whitlock, 2008; Whitlock & Guillaume, 2009). Additionally, pairwise differences in both phenotypic and neutral genetic divergence were generated between all combinations of the 14 sites, resulting in PST and FST matrices with 91 pairwise comparisons. Data were square-root-transformed, and parameters were estimated by standard variance components analysis using the LME4 package in r 2.10. The resulting matrices were compared using a randomization test (Manly, 1991), with similar expectations that PST would exceed neutral FST if selection is acting on the trait. To further explore the relationship between PST and FST, a Mantel test for correlation was performed, with the expectation of congruence if both phenotypic and genetic differences have arisen from neutral evolutionary forces (e.g. Mariani et al., 2012).

Association between phenotype and genotype

The association between number of lateral plates and genotype at the corresponding QTL was tested in the subset of 14 sites using (i) a generalized linear model (GLM) and (ii) comparison of phenotypic divergence (PST) with divergence at the QTL (FSTQ). In the case of (i), the number of plates was a dependent variable, and the genotype at the QTL was a fixed effect. A log link function and Poisson-distributed errors were employed in the model, executed in r 2.10. In the case of (ii), the same procedure was followed as described above in the PSTFST comparisons, but with FSTQ used in the place of PST. Additionally, pairwise distance matrices for PST and FSTQ were compared after randomization in r 2.10. Correlation between PST and FSTQ matrices was first tested for with a Mantel test. To control for potentially confounding effects of neutral divergence, a partial Mantel test between PST and FSTQ matrices was performed using neutral FST as a covariate. Once the relationship between PST and FSTQ was confirmed, selection in lateral plates was investigated among all 38 sites with FSTQFST comparisons and correlations, with the similar expectation as PSTFST comparisons: elevated and uncorrelated levels of FSTQ if the QTL is under selection (Vasemägi & Primmer, 2005).

Signatures of selection: Outlier analysis

To confirm signatures of selection on the QTL, we performed an outlier analysis on the data set of combined neutral markers and the QTL, using FDist2 as implemented in LOSITAN (Antao et al., 2008). This analysis was first performed in a pairwise fashion on all pairs of the subset of 14 sites. A pairwise matrix was then generated whereby a value of one was assigned to any pair of sites in which a positive signal of selection was detected at the QTL, and a value of zero to those in which no signal was detected. This matrix was compared with the results of the post hoc Tamhane tests, for which a matrix was similarly constructed using 1 for significant differences in lateral plate numbers and 0 for nonsignificant differences (after P-values had been corrected for multiple testing), using a Mantel test. The outlier analysis was also performed on the data set containing all sites. In all cases, 50 000 simulations were performed under the stepwise mutation model at 95% confidence intervals and a false discovery rate of 0.1, after an expected FST was approximated. Preliminary analyses were also performed under the infinite alleles model, but no differences in results were observed.

Environmental correlations

To test whether lateral plate differentiation was related to environmental variables (viz. salinity and temperature), the linearized pairwise PST matrix was compared with matrices of log-transformed differences in salinity and temperature in the subset of 14 sites. To then account for neutral genetic processes, neutral FST was subtracted from linearized PST such that a matrix of (PST–FST) was generated. This matrix was then compared with matrices of differences in salinity and temperature with a Mantel test in R. In this comparison, a positive correlation would be expected if phenotypic differences were driven by selection stemming from these factors (Antoniazza et al., 2010; Hangartner et al., 2012). Because both of the environmental variables are correlated with geographical distance, a partial Mantel test between the (PST–FST) matrix and the matrices of differences in environmental variables was performed, using geographical distance as a covariate. The analyses were repeated with FSTQ in place of PST as the dependent variable, also using the subset of 14 sites. The analyses were performed again with FSTQ in all 38 sites. All Mantel tests were performed with the vegan package in r 2.10, and significance was tested with 1000 permutations.

Results

Differences in number of lateral plates

Most sites were polymorphic for the number of lateral plates, with the exception of PET, in which all individuals had 25 lateral plates. There was also a greater frequency of individuals with higher lateral plate counts in the Kattegat/Skagerrak straits and the southern coast of the Baltic Proper (Fig. 1; Fig. S1). Most of the remaining sites had a wide range of lateral plate counts, and the lowest plate counts were found in individuals from the Bothnian Bay (Fig. 1; Fig. S1). Accordingly, the mean number of lateral plates differed significantly among all 14 sites (F13,480 = 19.43, < 0.001; Fig. S1). Post hoc Tamhane tests revealed significant differences in 33 of 91 pairs of sites, most of which were between the Kattegat/Skagerrak and Baltic Proper (Fig. S1).

Genetic diversity and divergence

Site- and locus-specific diversity estimates as calculated with neutral markers can be found in Table 1 and Table S1, respectively. There was no difference between locus-specific FST and RST values (paired t20 = 1.081, = 0.29; Table S1). Global neutral differentiation (FST) over the subset of 14 sites was 0.006 (95% CI: 0.001–0.010), which increased to 0.008 (95% CI: 0.005–0.010) when all 38 sites were analysed. When analysed with neutral markers, most of the sites from the Baltic Proper (Fig. 1) formed a central cluster in the NMDS plot, from which the three Kattegat samples were separate (Fig. 2a). The site from the easternmost coast of the Gulf of Finland (PET; Fig. 1) was distant from the others (Fig. 2a). However, the NMDS plot of the sites as analysed with only the QTL formed a horseshoe distribution, indicative of a gradient (c.f. Podani & Miklós, 2002; Fig. 2b). There was a significant isolation-by-distance signal as calculated with neutral markers (14 sites: = 0.48, < 0.001; 38 sites: = 0.47, = 0.001; Table 2; Fig. S3). Global differentiation in the QTL (FSTQ) was 0.102 (95% CI: 0.069–0.135) over the subset of 14 sites and FSTQ = 0.089 (95% CI: 0.074–0.104) when all 38 sites were analysed.

Table 2. Matrix correlations (r) between phenotypic divergence (corrected for neutral divergence; PSTFST), divergence at the underlying QTL (corrected for neutral divergence; FSTQFST) and neutral divergence (FST), and between geographical distance and environmental variables (salinity or temperature), among (a) 14 sites with phenotypic data and (b) all 38 sites
Matrix 1Matrix 2 r p Matrix 3 r p
  1. Matrix 1 = dependent variable, matrix 2 = independent variable (corresponding statistics in columns 3 and 4), matrix 3 = independent variable upon which the dependent variable has been conditioned (corresponding statistics in columns 6 and 7). Bold text indicates significance (α < 0.05) after correction for multiple testing.

(a) 14 sites
F ST Distance 0.48 < 0.001    
P ST Distance 0.32 0.002    
F STQ Distance0.260.02   
(PSTFST)Distance0.340.01   
(PSTFST)Salinity0.310.006Distance0.130.12
(PSTFST)Temperature0.160.15Distance0.110.12
(FSTQFST)Distance0.250.03   
(FSTQFST)Salinity0.310.02Distance0.170.11
(FSTQFST)Temperature0.020.39Distance0.020.39
(b) All sites
F ST Distance 0.47 0.001    
F STQ Distance 0.17 0.001    
(FSTQFST)Distance0.170.006   
(FSTQFST)Salinity0.040.34Distance0.020.35
(FSTQFST)Temperature0.080.05Distance0.040.21
Figure 2.

Nonmetric multidimensional scaling plot based on pairwise FST values calculated with (a) 20 neutral markers and (b) the QTL underlying lateral plate number (STN381).

Divergence in number of lateral plates: PSTFST comparisons

Phenotypic divergence in plate numbers as estimated by PST was 0.213 (95% CI: 0.165–0.262). PST–FST values in lateral plate numbers were significantly higher than would be expected in a neutrally evolving trait (Fig. 3a). Similarly, pairwise estimates of PST were significantly higher than (randomization test, = 0.012) – and uncorrelated with (= 0.15, = 0.127) – pairwise neutral FST.

Figure 3.

Simulated distributions of PSTFST and of FSTQFST for a neutrally evolving trait and QTL, respectively, and the observed values of PSTFST in number of lateral plates (bold arrow) and FSTQFST in the QTL underlying number of lateral plates (thin arrow) as estimated from (a) the 14 sampling sites from which individuals were phenotyped (see Table 1) and (b) from all 38 sampling sites.

Association between genotype and phenotype

Much of the phenotypic variation in lateral plate number was explained by the genotype at the underlying QTL (R2 = 0.72, z4, 475 = 264.17, < 0.001). One allele (177) corresponded to higher plate counts, whereas the two alleles (196 and the rarer 190) corresponded to lower plate counts (Fig. S2). Pairwise estimates of PST were similar to (randomization test, = 0.20) – and strongly correlated with (= 0.86, < 0.001) – pairwise FSTQ. Controlling for neutral FST did not change this relationship (= 0.85, < 0.001). Accordingly, FSTQ was significantly higher than neutral FST when all 38 sites were analysed, as revealed by comparisons of FSTQ at the QTL underlying plate number with the distribution FSTQ of a neutrally evolving trait (Fig. 3b) and also by comparisons of pairwise distance matrices (randomization test, = 0.03). Divergence at the QTL was also uncorrelated with neutral divergence (= 0.09, = 0.138).

Divergence in number of lateral plates: Signatures of selection

When the outlier analysis was performed on each pair of the subset of 14 sites, signals of selection on the QTL were detected in 26 of 91 pairwise combinations (Table S2) – 25 of which also exhibited significant differences in lateral plate number (Table S2), resulting in a high correlation between matrices (= 0.781, = 0.001). When the outlier analysis was performed on the data set with all 38 sites, the QTL was detected as a clear – and only – outlier (< 0.001; Fig. 4).

Figure 4.

LOSITAN outlier detection in data set containing 20 neutral markers and the QTL underlying number of lateral plates (STN381). The FST for each locus is plotted against respective expected heterozygosity. Loci falling above the dashed line are classified as candidates under divergent selection, and those below the solid line are candidates under balancing selection. All loci between the two lines are classified as neutral.

Environmental correlations

After controlling for neutral differentiation and geographical distance, the pairwise matrix of lateral plate differentiation (PSTFST) was not correlated with salinity or temperature (Table 2). The same patterns were observed in the FSTQFST matrix comparison using the same 14 sites (Table 2). When all 38 sites were analysed, there was no correlation between the (FSTQFST) matrix and either of the environmental variables, particularly after controlling for geographical distance (Table 2). All other correlations can be found from Supplementary Table S3 and Fig. S3.

Discussion

The results of this study provide solid evidence for significant phenotypic and genotypic differentiation in an ecologically important quantitative trait across an environmentally heterogeneous seascape. Multiple analytical perspectives consistently supported the hypothesis that differentiation in lateral plate number in Baltic Sea sticklebacks is likely driven by local variation in selection, and highlight how selection can act on shaping the patterns of phenotypic and genotypic differentiation even in the face of high gene flow. Inferring natural selection and adaptive differentiation across wide geographical marine areas is challenging for a number of reasons, most of which can be attributed to the logistic difficulties in quantifying phenotypic differentiation and uncovering the genetic processes by which it is generated. Hence, there is need to devise and evaluate various methodological approaches that aim to assess the relative contributions of selection and drift on population structuring. In what follows, we discuss the implications of our findings in the context of recent debates around the shortcomings of QSTFST comparisons. We also discuss the potential biological explanations for the observed differentiation in Baltic Sea sticklebacks, and how our findings relate to other studies of adaptive differentiation in organisms living in high-gene-flow marine environments (c.f. Conover et al., 2006; Cano et al., 2008; Nielsen et al., 2009a).

QSTFST comparisons: Measuring QST/PST

The results from comparative studies of quantitative trait and neutral genetic differentiation (QST–FST) indicate that natural selection plays a predominant role in driving quantitative genetic differentiation (Merilä & Crnokrak, 2001; McKay & Latta, 2002; Leinonen et al., 2008; but see Edelaar et al., 2011). Our results accord with this view: estimates of lateral plate differentiation were an order of magnitude higher than differentiation in neutral loci. Several concerns have been raised regarding the precision of QST point estimates, most of which stem from the fact that they are typically based on a small number of populations (O'Hara & Merilä, 2005; Whitlock, 2008). Although this remains a valid point, it is important to note that the phenotypic divergence observed in our study was not only higher than neutral divergence – calculated from 14 populations – but also in pairwise differences. In these comparisons, it became clear that in addition to exceeding neutral divergence, phenotypic differentiation was also uncorrelated with neutral differentiation. Although the relationship between levels of differentiation in quantitative traits and neutral loci is generally inconsistent (Raeymaekers et al., 2007; Leinonen et al., 2008), congruence can be expected if both have arisen from the same processes (Merilä, 1997; Merilä & Crnokrak, 2001). For example, clinal variation in quantitative traits can be generated by stochastic processes of drift or gene flow restricted by geographical distance (Storz, 2002). By comparing the relative levels of neutral and phenotypic divergence, inferences about the contribution of selection to the observed variation have been strengthened in a number of studies (Hall et al., 2007; Hangartner et al., 2012; Mariani et al., 2012). Therefore, the lack of correlation between phenotypic and genetic distance in our study could indicate that differences in lateral plate number among Baltic Sea stickleback are a result of adaptive processes in response to environmental heterogeneity, rather than genetic drift.

More problematic is the fact that the accuracy of QST can be compromised by environmentally induced phenotypic variation (i.e. trait plasticity), particularly when estimated from data collected from the wild. Specifically, measures of PST are based on the contestable assumption that phenotypic divergence is equivalent to additive genetic divergence and not confounded by environmental effects (e.g. Merilä & Crnokrak, 2001; Pujol et al., 2008). To account for the unequal influence of environmental effects between different populations, Brommer (2011) suggested scaling the proportion of phenotypic variance to additive genetic variance across populations, allowing ‘population-level heritability’ (Mobley et al., 2011) to be accounted for with a constant c (Brommer, 2011). Some studies have applied this approach to explore the critical ratio of between-population additive genetic variance (c) to within-population additive variance relative to total variance to test the robustness of their PST estimates and confirm the conclusions drawn from PSTFST comparisons (e.g. Holand et al., 2011; Mobley et al., 2011; Mariani et al., 2012). Although most of these studies have found that environmental effects impact the estimates of trait divergence, it has been repeatedly demonstrated that the number of lateral plates in laboratory-reared sticklebacks is generally not influenced by rearing environment (Colosimo et al., 2004; McGuigan et al., 2011; but see: McCairns & Bernatchez, 2012). Therefore, because there is ample evidence that lateral plate number is a heritable trait subject to selection (Barrett, 2010), PST estimates of this particular trait are unlikely to be sensitive to inflation by environmental factors. However, nonadditive genetic effects such as dominance and epistasis, as well as maternal effects, are also known to impact the expression of quantitative traits (Whitlock, 1999; Storz, 2002; Santure & Wang, 2009). Although the latter two have been reported to only have a marginal influence on lateral plates in sticklebacks (Hermida et al., 2002), substantial dominance effects are possible (Colosimo et al., 2005). However, the effects of dominance mainly pertain to the suppression of QST, and hence, our finding of QST > FST should still remain robust evidence for population differentiation driven by natural selection.

QST –FST comparisons: Measuring FST

The uses of traditional measure of neutral population differentiation (FST) have recently received a fair amount of criticism, most notably in regard to the potential for highly polymorphic loci to bias estimates downwards (Hedrick, 2005; Jost, 2008; Meirmans & Hedrick, 2011). As such, a recent critique of QSTFST comparisons has arisen from the perspective that FST should only be calculated from markers with similar mutation rates as those of quantitative trait loci, such as single nucleotide polymorphisms (Edelaar & Björklund, 2011; Edelaar et al., 2011). Otherwise, the higher marker mutation rate produces large within-population variation, which could suppress FST if migration rate is relatively low (Hedrick, 2005; Kronholm et al., 2012; LeCorre & Kremer, 2012). Although we acknowledge that many microsatellite markers used in this study were highly polymorphic (cf. high levels of heterozygosity), the locus-specific RST values accounting for mutation rate were similar to their respective FSTs, indicating that – assuming stepwise mutation model – the high variability in these markers was not a result of their high mutation rate. Furthermore, simulation studies show that FST does not decline as a function of increasing heterozygosity when migration rate is high (e.g. Kronholm et al., 2010; Edelaar et al., 2011), and Baltic Sea sticklebacks are known to exhibit high migration rates (Jurvelius et al., 1996; Peltonen et al., 2004; DeFaveri et al., 2012). In addition, several of the loci screened in this study had reduced variability (e.g. STN38, STN79, STN199, STN178; HE < 0.50). Yet, differentiation in these loci – whether mutation rate was accounted for or not – was as low as in the other neutral loci. Therefore, it is unlikely that in our study, high mutation rates and heterozygosity would have inflated our neutral baseline estimates of differentiation.

PST –FSTQ comparisons and signals of selection

A lack of or weak correspondence between the degrees of molecular and quantitative trait differentiation across different species or populations is more of a rule than an exception (Merilä & Crnokrak, 2001; McKay & Latta, 2002; Leinonen et al., 2008, 2013). For example, Hall et al. (2007) found significant QST values in several phenology traits in European Aspen (Populus tremula), but differentiation in the underlying phenology candidate genes was an order of magnitude lower. Similarly, Raeymaekers et al. (2007) found strong signals of selection in several meristic traits in threespine sticklebacks through PSTFST comparisons. However, divergence at most underlying QTLs was not detected (Raeymaekers et al., 2007). These findings are not surprising in the light of results from analytical and simulation studies, which show that mean values of quantitative traits can change (i.e. QST > 0) without much change in allele frequencies across QTLs for which FST levels may remain similar to those observed in neutral loci, because the sum of small effects in many loci is enough to produce a large effect in a single quantitative trait (McKay & Latta, 2002; LeCorre & Kremer, 2003, 2012; Hall et al., 2007). This phenomenon can also interfere with the ability to identify signatures of selection through outlier analyses, which primarily detect divergent loci by elevated differentiation and/or reduced diversity. However, the loss of lateral plates in sticklebacks is controlled by a single locus of major effect (EDA; although several modifiers also contribute to the phenotypic variance in plate number); hence, most studies that have explored this trait have found a strong association between plate phenotype and EDA genotype (e.g. Colosimo et al., 2004, 2005; Kitano et al., 2008; Grøtan et al., 2012). As such, the correlation we observed between PST and FSTQ aligns with expectations and allowed us to explore the fine-scale spatial distribution of phenotypic divergence by means of the genetic variation in the underlying QTL. Furthermore, confirmation of the role of selection in shaping patterns of genetic variation at the QTL was achieved through detection of STN381 (an indel marker linked with EDA) as an extreme outlier when all populations were analysed, and also by comparing pairwise phenotypic divergence with detection of STN381 as an outlier in the respective population pair.

Environmental and geographical correlations with patterns of divergence

Isolation by distance (IBD) refers to the pattern of genetic structuring in neutral markers generated by limited gene flow between geographically distant populations (Slatkin, 1985). Within the past decade, there has also been considerable progress towards understanding how adaptive processes can also affect the genetic structuring of populations through ‘isolation by adaptation’ (Nosil et al., 2008; Edelaar et al., 2012). This has been complimented by incorporating correlations and comparisons of quantitative traits and ecological/environmental features that could be acting as selective agents (Storz, 2002; Hall et al., 2007; Hangartner et al., 2012). For example, when environmental selection gradients and geographical distances separating populations covary, differences in trait optima between populations will align with their spatial proximity. If those differences truly reflect adaptation to environmental conditions, phenotypic divergence should exceed divergence in neutral markers (Storz, 2002). In this study, we found evidence for IBD, indicating that nonadaptive processes have contributed to the patterns of genetic differentiation among Baltic Sea stickleback populations. However, divergence in lateral plate numbers – as measured by PST or FSTQ – followed the same pattern and exceeded that in neutral marker genes. This stronger IBD for lateral plates than neutral marker genes is indicative of selection acting on this trait and that this selection acts along a spatial gradient. This interpretation is further supported by the arched distribution of data points in the multidimensional scaling, which indicates an influence of ecological gradients (Podani & Miklós, 2002) – such as salinity and temperature. Although we observed a significant correlation between plate numbers and salinity, this relationship disappeared after controlling for geographical distance, which itself was strongly correlated with both of the environmental variables, suggesting that phenotypic divergence is likely driven by selective agents other than temperature or salinity. This is not entirely unexpected, as several recent studies of salinity-related effects on lateral plate expression and mortality have demonstrated a lack of influence induced by this variable (e.g. Marchinko & Schluter, 2007; Barrett et al., 2009; McCairns & Bernatchez, 2012; Grøtan et al., 2012). We also found that the highest number of plates and high-plate allele frequencies occurred in the lowest salinity regions of the Gulf of Finland (PET), lending further support to the hypothesis that alternative environmental and/or ecological forces have shaped the distribution of lateral plate numbers among populations. One likely explanation is local variation in predation regime, and its interactions with habitat characteristics (Leinonen et al., 2011), known to contribute to stickleback armour variation (see: Reimchen, 1994). In addition, plate morphs (and thus plate numbers) are known to be associated with certain behavioural traits, including reduced preference for environmental acclimation (Barrett, 2010) and benthic foraging (Bjærke et al., 2010). Thus, several factors might be acting together to shape the distribution of lateral plate numbers in the environmentally heterogeneous Baltic Sea.

Adaptive divergence in marine populations

Recent years have brought about a paradigm shift in marine population genetics: developments in molecular genetics have allowed researchers to gather more nuanced information about the extent and scale of adaptive population differentiation in marine populations, for which earlier studies based on neutral marker genes suggested no or little differentiation (reviews in: Conover et al., 2006; Cano et al., 2008; Hauser & Carvalho, 2008; Nielsen et al., 2009a). However, although these studies continue to uncover genomic patterns of differentiation (e.g. Knutsen et al., 2003; Nielsen et al., 2009b; André et al., 2010; Corander et al., 2012; Lamichhaney et al., 2012; Limborg et al., 2012), very few have been able to demonstrate genetic differentiation in phenotypic traits of ecological significance (but see: Luttikhuizen et al., 2003; Hice et al., 2009). To this end, our results provide a solid example of demonstrating adaptive divergence in a trait that has repeatedly been shown to be linked with various selection pressures (reviewed in Barrett, 2010), as well as in the major genetic locus coding its expression. Although several studies have described lateral plate variation among marine/estuarine populations of sticklebacks (e.g. Bańbura, 1994; Klepaker, 1996; Jones et al., 2006; Raeymaekers et al., 2007; McCairns & Bernatchez, 2012), to our knowledge, none have explicitly investigated the correlation with environmental gradients in these systems with such a fine-scale sampling scheme on a broad geographical scale or have explored the underlying genetic divergence in the QTL in relation to the environmental and phenotypic differences. Although lateral plate number in sticklebacks represents a quantitative trait having perhaps an exceptionally simple genetic architecture (but see e.g.: Haag et al., 2005; Johnston et al., 2011; Kronholm et al., 2012), the results provide an important and heuristically intuitive example of adaptive genetic differentiation in a conspicuous phenotypic trait. As such, it complements the short list of studies that have provided such evidence for marine organisms (e.g. Luttikhuizen et al., 2003) and vertebrates in particular (e.g. Conover et al., 2006; Hice et al., 2009). The importance of understanding the genetic basis of phenotypic variability in marine fish is underlined not only by recent debates about the quality of evidence for fisheries-induced evolution (Kuparinen & Merilä, 2007 Kuparinen et al., 2009; Audzijonyte et al., 2013), but also by the realization that cross-generational environmental and maternal effects have the potential to produce phenotypic effects of astonishing magnitude (e.g. Salinas & Munch, 2011).

Conclusion

Taken together, the results of this study demonstrate a strong geographical patterning of genetically based phenotypic variation in Baltic Sea sticklebacks, likely to be driven by spatially varying natural selection. As such, these results provide solid evidence for adaptive quantitative trait differentiation in the marine environment and, in particular, in the relatively young Baltic Sea seascape. The results also highlight the opportunities embedded into combined analyses of phenotypic and molecular genetic – both neutral and adaptive – variability towards inferring evolutionary processes in the wild.

Acknowledgments

We thank Janis Birzaks, Lotta Kvarnemo, Tuomas Leinonen, Hannu Mäkinen, Jouko Pokela, Anti Vasemägi, Josefin Sundin, Bertil Borg, Helmut Winkler, Sergey Titov and Dmitry Lajus for help in collecting the samples. Special thanks also to Ugo Szachnowski for staining and photographing the samples; to Scott McCairns, Tuomas Leinonen, Pekka Pamilo and Pim Edelaar for analytical assistance and/or helpful comments on the manuscript; and to Per Jonsson for modelling environmental estimates. The research leading to these results has received funding from the European Community's Seventh Framework Programme (FP/2007-2013) under grant agreement no. 217246 made with BONUS, the joint Baltic Sea Research and Development Programme (JM), Academy of Finland (no. 250435 and no. 134728; JM), and from LUOVA graduate school (JD).

Ancillary