Incipient habitat race formation in an amphibian


  • J. Van Buskirk

    Corresponding author
    1. Institute of Evolutionary Biology & Environmental Studies, University of Zürich, Zürich, Switzerland
    • Correspondence: Josh Van Buskirk, Evolutionary Biology & Environmental Studies, University of Zurich, CH-8057 Zurich, Switzerland. Tel.: +41 44 635 4983; fax: +41 44 635 6821;


    Search for more papers by this author


Theory defines conditions under which sympatric speciation may occur, and several possible examples of the process in action have been identified. In most cases, organisms specialize onto habitats that fall into discrete categories, such as host species used by herbivores and parasites. Ecological specialization within a continuous habitat gradient is theoretically possible, but becomes less likely with increasing gene flow among clinal habitat types. Here, I show that habitat race formation is underway in a frog, Rana temporaria, along a continuous and spatially mosaic habitat gradient. Tadpoles from 23 populations raised in an outdoor mesocosm experiment showed adaptive phenotypic variation correlated with the predator density in their pond of origin. A survey of microsatellite markers in 48 populations found that neutral genetic divergence was enhanced between ponds with very different densities of predators. This represents a new example of habitat specialization along a continuous habitat gradient with no spatial autocorrelation in habitat.


Theoretical studies describe conditions under which sympatric speciation may occur (Udovic, 1980; Kawecki, 1996; Dieckmann & Doebeli, 1999; Fry, 2003; Gavrilets, 2004; Thibert-Plante & Hendry, 2011), and possible examples of the process have been identified in nature (Emelianov et al., 2004; Barluenga et al., 2006; Fossoy et al., 2011). Speciation of initially undifferentiated forms can begin with the evolution of ecological specialization, a process known as host race or habitat race formation. Specialization onto hosts or habitats occurs most readily if dispersal among habitat types is restricted, or when individuals strongly prefer to mate with others originating from the same habitat (Futuyma & Moreno, 1988; Gavrilets, 2003; Bolnick & Fitzpatrick, 2007; Linnen & Farrell, 2010). This explains why specialization is usually found when discrete resources or intimate ecological relationships are involved (Schluter, 2000; Dres & Mallet, 2002; Savolainen et al., 2006). Perception of habitat types and subsequent habitat selection are relatively straightforward under these conditions.

When habitat variation is clinal, the ecological distinctions among habitat types are graded and often smaller. This implies that divergent selection in different habitats is weaker and that the rate of dispersal among habitats is potentially higher. Clinal habitats are therefore characterized by values of selection and gene flow that are unfavourable for the maintenance of genetic polymorphism, according to population genetic theory (Levins & MacArthur, 1966; Slatkin, 1973; Lenormand, 2002; Blanquart et al., 2012). Habitat race formation on a cline is still theoretically feasible but less likely (Haldane, 1948; Fisher, 1950; Slatkin, 1973, 1978; Lande, 1982). These models highlight two conditions that promote specialization in this context: (i) a steep cline, corresponding to a relatively rapid change along the gradient in the phenotypic value conferring highest fitness, and (ii) restricted gene flow along the cline, arising from habitat selection, positive assortative mating according to habitat of origin or natural selection against individuals that move to a habitat sufficiently different from their natal habitat. This leads to the hypothesis that ecological specialization and the outcome of sympatric speciation that may result from it are less relevant for species that experience continuously varying habitats.

I tested for genetic divergence among populations of an amphibian occupying a continuous gradient of habitat types. The study organism was the European common frog, Rana temporaria, which is an aquatic tadpole during the larval stage. Although larvae occupy spatially discrete wetlands in my study area, different populations are exposed to graded levels of predation risk spanning 2–3 orders of magnitude (Van Buskirk, 2005). This gradient represents an ecologically meaningful spectrum of habitats for amphibian larvae, because natural selection is highly divergent in ponds with many or few predators (Van Buskirk et al., 1997; Van Buskirk & Schmidt, 2000). This leads to the prediction that selection favours habitat specialization. However, spatially adjacent wetlands can harbour completely different predator densities, suggesting that dispersal and gene flow across the full habitat gradient is possible. Thus, R. temporaria represents the great many species that experience a continuously varying and spatially mosaic habitat structure. Theory mentioned above suggests that habitat specialization under these conditions should be weak.

Materials and methods

My approach was to estimate genetic divergence – in both quantitative traits and neutral markers – among populations of R. temporaria breeding in discrete wetlands within an 800-km2 region in northern Switzerland (Fig. S1, Table S1). The position of each wetland along the predation gradient was estimated by sampling predator density at the site during the larval period of R. temporaria. The pattern of quantitative genetic divergence of a subsample of the populations, measured in a common garden experiment, was interpreted to reflect adaptation to predation risk. Although the extent of adaptation can be revealed only with a reciprocal transplant experiment, the pattern of population differentiation is nevertheless informative about the identities of traits that contribute to performance along with their extent of divergence (Kawecki & Ebert, 2004; Van Buskirk & Arioli, 2005). Neutral marker data were used to infer the degree of gene flow across the habitat gradient relative to gene flow among sites with similar predation risk.

Estimating predation risk

I characterized the predation risk in 48 ponds using pipe sampling and, in some cases, dipnetting. Pipe sampling involved dropping a 0.1-m2 hollow pipe onto the bottom of the pond and removing all the captured animals with a small net (average of 22 samples per pond). Dipnetting involved sweeping a net along the substrate in a standard fashion so as to sample an area of 1.02 m2 per sweep (average of 34 net sweeps per pond). Samples were distributed among microhabitats in proportion to their occurrence in the pond. Pipe sampling and dipnetting were calibrated against each other using data from 114 sampling occasions on which I employed both methods (Van Buskirk & Arioli, 2005). Samples were conducted in early May, during the larval period of R. temporaria, and were repeated over 3–7 years per pond between 1997 and 2003. Predators included aeshnid dragonfly larvae, dytiscid beetle larvae, adult newts (Lissotriton, Mesotriton, Triturus), adult backswimmers (Notonecta glauca) and larval libellulid and corduliid dragonflies. Predation risk was the sum of the densities of all predators capable of killing tadpoles, weighted by a measure of their dangerousness. The weights were the mortality rates of tadpoles measured during exposure to each predator in outdoor mesocosm experiments (Van Buskirk & Arioli, 2005). Predation risk in the 48 ponds studied here was approximately log-normally distributed, ranging over > 3 orders of magnitude, with most ponds falling between 2 and 25 m−2 (Fig. S2). Spatial autocorrelation in predation risk was evaluated by plotting Moran's I against geographical distance (Sokal & Oden, 1978).

Common garden experiment

Tadpoles from a subsample of 23 R. temporaria populations (Table S1) were reared in a mesocosm experiment to discover whether phenotype assessed in a common environment was associated with predation risk in nature. The source ponds were selected to encompass a broad range of predator densities, but they also included ponds with different levels of canopy cover, vegetation structure and hydroperiod. The experimental mesocosms were 80-L plastic tubs filled with tap water and placed outdoors in a field at the University of Zurich. Two treatments, one with and one without a single-caged (nonlethal) Anax imperator dragonfly larva, were intended to represent wetlands with high and low predation risk. The predator cages were 11-cm sections of plastic pipe, 10 cm in diameter and capped on both ends with window screen. Mesocosms in the no-predator treatment contained an empty cage. Predators were each fed 300 mg of R. temporaria tadpoles three times per week. I repeated the experiment over 3 years (1999, 2001 and 2003) using nearly identical methodology, with four to ten replicates of each population–treatment combination, totalling 264 mesocosms (Van Buskirk & Arioli, 2005). Each mesocosm contained 15 individuals of nine sibships from a single population, collected as eggs within 1–3 days of oviposition in the source ponds and added to the mesocosms 4 days after hatching at Gosner (1960) stage 24. Mesocosms were stocked with 60 g leaf litter and a sample of water and zooplankton from a nearby wetland; developing larvae fed on naturally growing periphyton.

Tadpole behaviour was observed on two occasions, at about 28 and 40 days after hatching. The proportions of tadpoles swimming, resting inactively and invisible (hiding in leaves and detritus on the bottom of the mesocosm) were averaged from five to six instantaneous samples collected between 10:30 and 15:30. I measured tadpole morphology at about ages 27 and 41 days, by making photographs of a sample of five individuals per mesocosm on each day in a water-filled cuvet. The locations of 18 landmarks in lateral view were digitized from the photographs using image analysis software (Abramoff et al., 2004); the landmarks are described in detail in Van Buskirk (2009). Landmark configurations were subjected to geometric morphometric analysis, which consisted of scaling each form by its centroid size, Procrustes superimposition, projection onto Kendall shape space and principal components analysis to extract measures of shape known as relative warps (Zelditch et al., 2004). Here, I focus on two of the relative warps, RW4 and RW5 (6.4% and 6.0% of shape variation), because they showed significant plasticity induced by caged predators. The other RWs reflected shape variation due to unknown environmental factors, unexplained population differences, genetic variation within populations and measurement and presentation error.

Statistical analyses tested whether genetic variation among populations in behaviour and morphology was correlated with the predation gradient. I performed multivariate repeated-measures anovas (SAS Institute, 1990, p. 988), separately for the morphological and behavioural traits, testing for effects of the caged dragonfly treatment, predation risk in the source pond (averaged over five to seven annual samples for each pond) and their interaction. Next, I tested whether the slope of population divergence along the predation gradient was congruent with the direction of predator-induced plasticity, which is thought to be adaptive (Van Buskirk et al., 1997; Van Buskirk & Relyea, 1998). Observed trait values were randomly sampled 1000 times with replacement at the level of the mesocosm. For each randomization, I calculated the slopes of the five traits regressed against predation risk in the source pond, separately by treatment, and recorded the number of traits for which these slopes were parallel to plasticity. This number was then compared with the observed number to determine whether population divergence and plasticity were congruent more often than expected by chance.

Molecular genetic structure

I screened R. temporaria samples from all 48 ponds at eight variable microsatellite loci to characterize divergence among populations in neutral markers, and thereby indirectly estimate gene flow across the habitat gradient. In March 2000, I collected an average of 20.7 eggs per population, for a total of 996 samples (range 13–36, only three populations had < 19 samples). Insofar as possible, half-sibs sired by the same male were avoided by sampling one egg each from clutches of different ages and in different parts of the pond. Hatchling tadpoles were preserved in alcohol after they resorbed the yolk sac. DNA was extracted and previously described protocols were applied to genotype each specimen at each microsatellite marker (Garner et al., 2000). Sample sizes for each population and tests for linkage disequilibrium, null alleles and selection are described Van Buskirk (2012). Interpopulation differentiation, FST, was estimated by the allele identity method of Weir and Cockerham (1984) (Hardy & Vekemans, 2002). I tested for habitat-biased gene flow, termed ‘isolation by habitat’ (Nosil et al., 2008), using a partial Mantel test of the correlation between population differentiation [measured by FST/(1 − FST); Rousset, 1997] and the difference in habitat [measured by the absolute value of the difference in ln(predation risk) averaged across all years], controlling for ln(geographical distance).

I discarded one microsatellite locus prior to analysis because it showed evidence of selection, detected using the method of Beaumont & Nichols (1996). It was important to include only neutral markers because the aim was to estimate gene flow among pairs of populations, and FST correlates with gene flow only if divergence is more strongly influenced by drift than by selection or mutation (Slatkin, 1991; Whitlock, 2011). However, the results reported below did not substantially change if the discarded locus was included, which suggests that divergence at that locus was not associated with predation risk.


The common garden experiment revealed quantitative genetic divergence in morphology and behaviour along the predation gradient (Fig. 1); multivariate rmanova revealed significant effects of predation risk in the source pond for both kinds of traits (Table 1). Although tadpoles from all 23 populations were reared and assessed within the same environment, those originating from ponds with high predator density were phenotypically distinct: they had long and deep tails, were relatively inactive and spent more time hiding. Univariate tests of the separate traits on both sampling dates revealed significant associations between predation risk in the source pond and morphology (RW4: first sample F1,42 = 4.1, P = 0.0484, second sample F1,42 = 6.2, P = 0.0166; RW5: first sample, F1,42 = 4.4, P = 0.0425, second sample F1,42 = 5.9, P = 0.0199) and some behavioural traits (swimming: first sample F1,42 = 7.6, P = 0.0085, second sample F1,42 = 8.5, P = 0.0056, visibility: first sample F1,42 = 5.7, P = 0.0212, second sample F1,42 = 0.2, P = 0.67, proportion inactive: first sample F1,42 = 3.5, P = 0.0674, second sample F1,42 = 2.2, P = 0.14). Phenotypic plasticity induced by caged dragonflies was in the same direction as population divergence along the predation gradient, as in Fig. 1a, for nine of ten possible cases (five traits measured twice each). The resampling procedure found that plasticity and slopes were congruent for < 9 cases in 966 of 1000 randomizations (P = 0.034). This suggests that phenotypic variation among populations reflects in part adaptation to predation risk.

Table 1. Multivariate repeated-measures analyses testing for relationships between two types of traits, each sampled on two dates, and predation risk in the source pond. The morphological characters were RW4 and RW5 (Fig. 1); the three behavioural traits were the proportion of time swimming, resting inactively and hiding beneath the leaf litter. The treatment was the presence/absence of a single-caged dragonfly larva; predation risk was the summed density of all predator individuals in the source pond weighted by their ability to kill tadpoles. Entries in the table are Wilks' F and the P-value in parentheses; boldface highlights significant results.
SourceType of trait
Morphology (d.f. = 2,41)Behaviour (d.f. = 3,40)
Within-subjects effects
Date0.74 (0.4838)2.31 (0.0905)
Date × Treatment1.14 (0.3295) 5.80 (0.0022)
Date × Predation risk0.95 (0.3955)1.73 (0.1767)
Between-subjects effects
Treatment 24.71 (0.0000) 65.77 (0.0000)
Predation risk in source 3.81 (0.0304) 5.14 (0.0043)
Treatment × Predation risk0.73 (0.4889)1.92 (0.1418)
Figure 1.

Two measures of morphological shape (a and b) at about 41 days of age, and proportion of time spent swimming (c) and hiding in the substrate (d) at age 28 days, in tadpoles from 23 Rana temporaria populations reared in outdoor mesocosms. Grey circles represent the predator-free mesocosms; solid triangles are the caged dragonfly treatment. Error bars depict ± 1 SE. Predation risk in the source pond is the summed density of all predators weighted by their dangerousness to tadpoles. Drawings at the top illustrate shape changes represented by two relative warps. The grey/black outlines show specimens with scores 2 SD above/below the mean form, respectively. Tadpoles originating from ponds with many predators were deeper in head and tail, had longer tails and shorter heads, were less active and spent more time hiding than those from ponds with few predators. The data in panel d appeared in Van Buskirk and Arioli (2005).

Molecular population structure of the 48 R. temporaria populations revealed a bias in genetic divergence with respect to habitat. There was clear population structure over the spatial scale of the study area (46 km in longest dimension): average FST was only 0.0271, but there was strong isolation by distance [i.e. the relationship between FST/(1 − FST) and ln(geographical distance) was significant in a Mantel test; r = 0.24, P = 0.0002; Fig. S3]. These data imply that gene flow was somewhat spatially restricted at distances greater than about 10 km. Evidence for habitat specialization came from a significant positive relationship between the genetic divergence between pairs of populations and the difference between them in predator habitat, after accounting for their geographical separation (Fig. 2; r = 0.21, P = 0.0006; partial Mantel test). Ponds with very different densities of predators were more divergent than expected based on their spatial locations.

Figure 2.

Relationship between population differentiation at neutral markers and difference in habitat. Each point represents one pair of Rana temporaria populations (48 populations). The difference in predation risk was the absolute value of the difference in ln(predation risk) averaged over years. Population divergence was measured by FST/(1 − FST) after accounting for the geographical distance between the pair. Dotted lines depict the 95% confidence interval of a major axis regression through the 1128 points (population pairs). The positive correlation was significant in a partial Mantel test (r = 0.21, P = 0.0006; vegan package in r, R Core Development Team, 2013). Populations that occupied different habitats showed relatively high neutral divergence, suggesting that gene flow was restricted by habitat.

Predation risk showed no clear spatial autocorrelation on our study area, suggesting that predator density in nearby ponds was no more similar than that in ponds many km apart (Fig. 3). Morin's I was slightly, but not significantly, positive in ponds within a few hundred metres of one another.

Figure 3.

Spatial correlogram for predation risk measured during early May in 48 ponds in northern Switzerland. The heavy line indicates the observed value of Moran's I (a measure of association; Sokal & Oden, 1978); pairs of ponds are sorted into eight geographical distance bins arranged at equal intervals on a logarithmic scale. Sample size is the number of pond pairs contributing to the distance category. The shaded region represents the range of coefficients that do not differ significantly from the value of I expected by chance.


I describe evidence for incipient ecological specialization in a frog onto habitats defined by the density of predators. The evidence comes from two sources. First, tadpole behaviour and external morphology varied among populations in a manner consistent with adaptation to the density of predators in their pond of origin. When assessed in a common environment, animals originating from ponds with high predator densities in nature had deeper and longer tails and spent relatively little time swimming and more time hiding. Much of this variation is likely to be due to evolved population divergence. Although nongenetic maternal effects can affect growth and development of frog larvae (Kaplan, 1998), it is not obvious how the environment experienced by a tadpole might determine the behaviour and morphological shape of her offspring several years later. The second line of evidence was that populations occupying ponds characterized by very different predator densities were more divergent at neutral markers than were populations from ponds with similar predators. Divergence due to predators was not associated with a spatial habitat gradient, because there was no appreciable spatial autocorrelation in predation risk. These data reflect a balance among several processes acting together: strong natural selection that varies with habitat, fairly high dispersal among wetlands and gene flow that is more frequent between ponds with similar habitats.

Divergent selection across the predation gradient is suggested by estimates of survival and growth in experiments with and without predators. In the laboratory and outdoor mesocosms, predators impose high mortality on amphibian larvae that are especially active and have relatively shallow tail fins (Skelly, 1994; Van Buskirk et al., 1997; Van Buskirk & Relyea, 1998). An opposite combination of traits is associated with high performance in the absence of predators (Van Buskirk et al., 1997; Van Buskirk & Schmidt, 2000). Available functional data support these selection estimates (Anholt & Werner, 1995; Van Buskirk et al., 2003; Teplitsky et al., 2005; Johnson et al., 2008). Similar divergent selection is thought to operate in natural wetlands. That is, populations occupying ponds with many predators in nature presumably experience selection favouring genotypes that swim little and hide frequently, because these behaviours reduce encounter rates with predators (Werner & Anholt, 1993; Van Buskirk & McCollum, 2000). Tadpoles occurring in wetlands with few predators probably undergo selection for increased activity because resource intake is positively associated with foraging effort (Skelly & Werner, 1990). Similar arguments apply to morphological shape, although the functional mechanisms are less certain (Van Buskirk et al., 2003; Johnson et al., 2008). Perhaps the most convincing evidence of adaptation to predation risk in nature is the general agreement between the directions of selection estimated in experiments, plasticity induced by predators observed in the common garden experiment and the fixed population differences correlated with predation risk in nature.

Low population divergence at microsatellite markers means that dispersal among ponds is relatively high or that effective population sizes (Ne) are very large. Other studies suggest that Ne are low in amphibians – typically < 100 (reviewed in Ficetola et al., 2010) – and the pattern of isolation by distance argues in favour of high dispersal. A significant increase in FST with distance, at least beyond about 10 km (Fig. S3), indicates that dispersal limitation does indeed allow neutral divergence in this system. But for ponds within a few km of one another, dispersal and consequent gene flow is high enough to keep FST values in the range of 0.015.

In the absence of microsatellite data, the common garden experiment could be interpreted as supporting either local adaptation or habitat race formation. These two processes have different implications for population structure and potentially lead to different outcomes. Under local adaptation, demes adapt to local conditions because selection for alternative phenotypes in different sites is strong relative to dispersal (Kawecki & Ebert, 2004). This implies that the metapopulation is fairly viscous compared with the spatial scale of habitat heterogeneity, and it results in a spatial mosaic of adaptive genetic structure. The adapted units are local demes. Under habitat race formation, populations adapt to kinds of habitats rather than local sites; gene flow may be extensive, but it occurs mostly within habitat types. The adapted units are habitat races, each occupying many sites but sharing common adaptations to their habitats. This process can lead to ecological speciation (Via, 1999; Schluter, 2000; Räsänen & Hendry, 2008). A testable difference between local adaptation and habitat race formation is the pattern of gene flow: if populations are specializing onto habitat types, then gene flow must be biased by habitat (Nosil et al., 2008).

The microsatellite data imply habitat-biased gene flow. Divergence across the predation gradient visible in Fig. 2 is certainly not strong, but it demonstrates that gene flow among wetlands with very different predation habitats is somewhat restricted over longer periods of time. Two mechanisms could create this pattern. First, dispersing individuals may be more likely to settle and breed in habitats that resemble their natal habitat. This would be especially surprising because predation risk shows no spatial autocorrelation on the study area (Fig. 3). Consequently, a dispersing frog is likely to encounter a pond that is quite different from its natal site, and must occasionally show the temerity to abandon that pond in search of a more familiar habitat. Alternatively, there may be sexual or natural selection against frogs that reproduce in habitats that differ greatly from their natal habitat. I have no evidence to exclude either mechanism and other studies indicate that both can occur, sometimes simultaneously (Jaenike, 1986; Via, 1999; Via et al., 2000; Nosil et al., 2005).

The results are remarkable because wetlands in my study area do not fall into discrete categories with high and low predator density (Fig. S2). This habitat gradient is therefore typical of those experienced by many organisms, but it differs from habitats for which the classic examples of host or habitat races have been reported (Marchetti et al., 1998; Sorenson et al., 2003; Emelianov et al., 2004; Frantz et al., 2006; Räsänen & Hendry, 2008; Fossoy et al., 2011). There are good theoretical reasons to expect specialization onto discrete habitats or resources (Futuyma & Moreno, 1988; Smith & Skúlason, 1996), but several recent studies suggest that continuous habitat gradients can also favour ecological specialization (Aguilera et al., 2007; Mendez et al., 2010; De Luna et al., 2012; Richardson & Urban, 2013; Richter-Boix et al., 2013). Some studies report isolation of populations that breed or flower at different times (e.g. Gustafsson & Lonn, 2003; Møller et al., 2011; Richter-Boix et al., 2013), which may also reflect specialization along a continuous gradient, although breeding phenology is often correlated with use of discrete habitat types. The evidence is not entirely convincing in some of these cases because ecological and geographical isolation can be difficult to disentangle due to spatial separation of habitats, or because, when phenotypic data are available, the traits are not always measured under common environmental conditions.

The incipient habitat races described here will almost certainly never give rise to new species. European amphibians are not presently undergoing rapid radiation, and their geographical distributions suggest that speciation has occurred in allopatry (Nölert & Nölert, 1992; Hewitt, 1999). I propose that habitat generalists such as R. temporaria regularly experience selection favouring specialization along ecological gradients, but that specialization stalls at a relatively low selection/dispersal equilibrium or is constantly eroded by changes in landscape structure (Behm et al., 2010). These results nevertheless encourage a new look at evolutionary responses of habitat generalists to environmental variation and suggest the possibility for subtle habitat discrimination or mate choice in amphibians.


T.W.J. Garner developed the microsatellite markers and provided much useful advice. I thank more than 40 people who helped with the fieldwork and M. Arioli, D. Lang and especially E. Sabiote for help with genotyping. The study was supported by the Swiss NSF. Ethics permits were provided by the Veterinäramt of Kanton Zürich and collection permits by the nature conservation offices of Kantons Zürich and Thurgau.

Data deposition

The mesocosm-level phenotypic data and individual-level microsatellite data are available from the Dryad Digital Repository, doi:10.5061/dryad.f0824.