The effect of historical legacy on adaptation: do closely related species respond to the environment in the same way?

Authors


Rachel Prunier, Kellogg Biological Station, Michigan State University, 3700 E. Gull Lake Dr., Hickory Corners, MI 49060, USA.
Tel.: (269) 671 2350; fax: (269) 671 2165; e-mail: prunier@msu.edu

Abstract

The many documented examples of parallel and convergent evolution in similar environments are strong evidence for the role of natural selection in the evolution of trait variation. However, species may respond to selection in different ways; idiosyncrasies of their evolutionary history may affect how different species respond to the same selective pressure. To determine whether evolutionary history affects trait–environment associations in a recently diverged lineage, we investigated within-species trait–environment associations in the white proteas, a closely related monophyletic group. We first used manovas to determine the relative importance of shared response to selection, evolutionary history and unique responses to selection on trait variation. We found that on average, similar associations to the environment across species explained trait variation, but that the species had different mean trait values. We also detected species-specific associations of traits with the environmental gradients. To identify the traits associated uniquely with the environment, we used a structural equation model. Our analysis showed that the species differed in how their traits were associated with each of the environmental variables. Further, in the cases of two root traits (root mass and root length/mass ratio), two species differed in the direction of their associations (e.g. populations in one species had heavier roots in warmer areas, and populations in the other species had lighter roots in warmer areas). Our study shows that even in a closely related group of species, evolutionary history may have an effect on both the size and direction of adaptations to the environment.

Introduction

The many examples of convergence and parallel evolution in similar environments are some of the most striking examples of evolution by natural selection, suggesting that species respond similarly to the same environmental selection pressures. The repeated evolution of succulence in plants from arid environments, for example, reflects the consequences of natural selection for water storage in arid environments (Beard, 1976; Ogburn & Edwards, 2010). Similarly, the repeated evolution of benthic and limnetic forms of fish (sticklebacks, lake whitefish, pumpkinseed, and bluegill sunfish) in post-glacial lakes is the result of character displacement in response to competition in the newly created lake habitats (reviewed by Schluter & McPhail, 1993).

We recognize less often that species may respond to the same selection pressure in different ways. For example, within Drosophila melanogaster, Cohan & Hoffmann (1986) found similar responses to selection for ethanol tolerance in replicate experimental populations, but the genetic underpinnings of the ethanol tolerance differed from one replicate to another. The fitness associated with different genotypes at the alcohol dehydrogenase locus depended on the pre-existing genetic background of the individuals. Similarly, Young et al. (2007) showed that shrews with similar diets have functionally similar, but morphologically divergent, jaw structures. Thus, different morphologies may also be the result of similar selection (Young et al., 2007). Even in cases where convergent evolution is well documented, the unique evolutionary history of individual lineages may play a role in how each lineage adapts to its environment. For example, in the radiation of Anolis lizards on Caribbean islands, crown giant anoles on the island of Puerto Rico tend to have narrower pectoral width, deeper heads and wider pelvic widths than crown giant anoles on other islands (Langerhans et al., 2006). There are even examples where apparently adaptive responses to the same environmental selection arise from trait responses in opposite directions. For example, Nakazato et al. (2008) found that in two species of Andean tomatoes, the between-population associations of plant height with elevation and of days to wilting with mean annual precipitation were similar between species. In contrast, the association of leaf area with elevation depended on the species examined. In Solanum pimpinellifolium, plants tended to have larger leaves in the higher-elevation population, and in Solanum lycopersicum var. cerasiforme, plants tended to have smaller leaves in the higher-elevation population. Differential responses to apparently similar selection pressures could reflect lineage-specific differences in genetic variances and covariances between traits (Arnold et al., 2008; Hanson & Houle, 2008), the vagaries of genetic drift and independent mutational histories (Travisano et al., 1995), or subtle differences in the environment that were not measured. Different adaptive responses, although affected by nonadaptive forces, are ultimately driven by natural selection to a landscape that has multiple optima.

Large-scale trait–environment associations are often attributed to the effects of natural selection. For example, plants from drier and more nutrient-poor areas tend to have thicker or denser leaves (Cunningham et al., 1999; Fonseca et al., 2000; Niinemets, 2001), plant height is correlated with precipitation (Moles et al., 2009), and rooting depth increases with increasing drought (Canadell et al., 1996) and rainfall seasonality (Schenk & Jackson, 2002). All of these associations have been interpreted as adaptive responses to the environment, and these large-scale patterns suggest that species respond predictably to variation along these environmental gradients. Similarly, trait–environment associations between populations within species are often interpreted as adaptations to the local environment (Endler, 1986), but the extent to which these large-scale patterns are reflected within species is less explored.

In this study, we investigated a small evolutionary radiation to determine whether trait variation among populations within species is associated with similar environmental gradients across species and whether this trait variation reflects large-scale trait–environment associations (Table 1). Specifically, we investigated trait–environment associations within the white proteas (Protea sect. Exsertae), a group of six closely related species endemic to the mountains of South Africa that diverged recently (0.34–1.2 Myr; Valente et al., 2010). While the species are almost completely allopatric, they have broad environmental tolerances (Latimer et al., 2009) and do not cleanly partition environmental space (Fig. 1; see also Latimer et al., 2006). In fact, as Latimer et al. (2009) showed, the differences between species in their mean environments are not enough to account for the almost complete allopatry in this group, suggesting that adaptive differences between populations – in addition to differences between species – may contribute significantly to the diversity of the group (Carlson et al., 2011).

Table 1.   Expected trait–environment association, the corresponding trait and environment used to detect the association in this study, and the species and direction in which the pattern was found. + indicates that the trait–environment association was in the expected direction, and − indicates that the trait–environment association was in the opposite direction from the expectation.
Expected Trait–environment correlationCitationTraitEnvironmentSupported in our study?
Biomass decreases as elevation increasesClausen et al., (1940)Root mass, shoot mass, shoot height, leaves presentCOLDPCA+P. mundii, +P. punctata, + P. lacticolor, −P. aurea
Root/shoot ratio increases with decreased rainfallSchenk & Jackson (2002)Root mass/shoot massDRYPCA+ P. mundii
Leaves become larger and wider with increased rainfallCunningham et al. (1999)LWRDRYPCA+P. aurea and +P. mundii
Plant height increases with increased rainfallMoles et al. (2009)Shoot heightDRYPCANo
Roots length increases with increased droughtCanadell et al. (1996)Root lengthDRYPCANo
SLA increases with increased rainfallWright et al. (2004)SLADRYPCANo
Seeds are larger in lower-nutrient soilsBeadle (1966)Seed massFERTPCANo
Proteoid roots increase with decreased soil fertilityLamont (1982)Proteoid rootsFERTPCANo
SLA decreases with decreased soil fertilityCunningham et al. (1999)SLAFERTPCANo
Root branching increases with more concentrated rainfallNicotra et al. (2002)Root length/root massPPTCON+ P. mundii,P. aurea
Root depth increases with increasing rainfall seasonalitySchenk & Jackson (2002)Root lengthPPTCON+P. subvestita
Figure 1.

 White protea climate envelopes in three dimensions: COLDPCA (PCA summarizing coldness of winter), DRYPCA (PCA summarizing intensity of summer drought) and PPTCON (rainfall seasonality). Values were obtained by intersecting the locations of all observed populations of each species from the Protea Atlas (a database of 250k presence/absence observations of all species in the Proteaceae; Rebelo, 2006) with the South African Atlas of Agrohydrology (Schultze, 2007).

Our previous work on the white proteas found that among-population differences in leaf traits are adaptively associated with climate, for example thick or dense leaves appear to be adaptations to a harsh climate (Carlson et al., 2011). These results are consistent with expectations from large-scale surveys of trait–environment correlations across all flowering plant species (Wright et al., 2004). Differences in traits associated with access to nutrients and water and differences in seed size are also likely important foci of adaptation in most plants, including the white proteas. The environments in which the white proteas grow vary widely in the timing and amount of rainfall (Schultze, 2007), in hydrological conditions and in soil nutrients (Specht & Moll, 1983; Witkowski & Mitchell, 1987). In proteas, long tap roots are needed to reach the water table (Richards et al., 1995; Watt & Evans, 1999), and access to nutrients is facilitated by the production of proteoid roots, specialized cluster roots on lateral roots. Proteoid roots are especially important for uptake of phosphorous (Lamont, 1982; Neumann & Martinoia, 2002), but they also enhance uptake of nitrogen and micronutrients (Jeschke & Pate, 1995). Similarly, large seeds have been found to be adaptive in low-nutrient conditions (due to better provisioning; Beadle, 1966), both within (Bonfil & Kafkafi, 2000; Vaughton & Ramsey, 2001) and across species (Milberg et al., 1998; Shane et al., 2008).

Here we use leaf, shoot and root traits measured in a common environment to answer the following questions: (i) Does evolutionary history, similar evolutionary responses to the environment or lineage-specific evolutionary responses to the environment play a larger role in the trait diversification of the white proteas? (ii) What traits are associated with lineage-specific evolutionary responses to the environment? (iii) Do the species-specific trait–environment associations reflect large-scale trait–environment correlations?

Materials and methods

Study species

The white proteas are a well-supported, monophyletic clade within the genus Protea (Protea section Exertae, Valente et al., 2010). The current taxonomic treatment recognizes six species, P. aurea (Burm. f.) Rourke ssp. aurea (shuttlecock sugarbush), P. aurea ssp. potbergensis (Rourke) Rourke (potberg sugarbush), P. lacticolor Salisb. (hottentot sugarbush), P. mundii Klotzsch (forest sugarbush), P. punctata Meisn. (water sugarbush), P. subvestita N.E. Br. (waterlily sugarbush) and P. venusta Compton (creeping beauty). The clade is endemic to Southern Africa, and all of the species but Psubvestita are endemic to the Cape Floristic Region of south-western South Africa. Psubvestita is native to the Eastern Cape and Kwa-Zulu Natal provinces, as well as Lesotho (Fig. 2). The species of the white proteas radiated quickly after the origin of the group, and phylogenetic distance between populations and species is small (Prunier & Holsinger, 2010).

Figure 2.

 Sampling locations and ranges of the white protea species. Inset is a map of Africa, with enlarged area in the box.

All of the white proteas are sclerophyllous evergreen shrubs. P. venusta has a sprawling habit, but the remaining species are all ecologically similar. They grow upright and up to 4 m tall. They are killed by fire, regenerating from seeds stored in serotinous cones (Rebelo, 2001). They have largely allopatric ranges, although some populations of Ppunctata and Pvenusta are found in close proximity at the tops of the Swartberg and Kammannassie mountains. All are presumed to be pollinated by sugarbirds and sunbirds (Rebelo et al., 1984; Rebelo, 2006; de Swardt, 2008; Carlson & Holsinger, 2010). Because their flowering periods overlap and most species share bird pollinators, the white proteas often hybridize when grown together in cultivation and when they co-occur in the wild (Rourke, 1980; Prunier & Holsinger, 2010).

Sampling protocol

We collected seeds from wild populations in February–April 2008. We sampled 5–6 populations from each species (total populations = 30), spanning most of the geographical range of each species (Fig. 2, Table S1). We excluded P. aurea ssp. potbergensis from this study because we were able to sample only two populations of this geographically restricted subspecies. Within each population, we collected 5–8 seedheads from eight plants approximately 10 m apart along a linear transect through the population. We dried the seedheads in a cold room maintained at low humidity until they opened and the single-seeded fruits could be removed. We then selected fruits containing potentially viable seeds by identifying fruits in which the seed was filled and free of insect damage. We stored the fruits containing potentially viable seeds (henceforth seeds) at room temperature until planting.

Greenhouse experiment

We weighed, surface-sterilized with 10% bleach and planted one seed from eight maternal lines per population into each of six standard 288 plug flats (TLC Plastics or similar) filled with a standard nursery soil mix in a complete random block design in September 2008. We moved the seeds to a growth chamber (Conviron, Winnipeg Canada) programmed for short warm days and cool long nights (10 h: 20 °C, 14 h: 8 °C) to simulate autumn in the southern hemisphere (April-June), when the seeds germinate in the wild (Rebelo, 2001) and assessed germination every day until 47 days after planting, at which point germination had slowed to less than one plant every other day. One or two days after germination, we moved two seedlings from each maternal line to pots 10 cm × 10 cm wide at the top and 76.2 cm tall (Stuewe and Sons Inc., Tangent OR) in a greenhouse in Storrs, CT. The pots contained a soil mix with five parts peat/four parts sand/two parts fine perlite/one part charcoal. This soil mix is modified from a standard low-nutrient greenhouse mix by the addition of sand to provide better drainage. We randomly assigned seedlings to pots within the greenhouse. Supplemental light (80–150 μmols per m2) was added 30 min after sunrise and ended 30 min before sunset to ensure an even light environment across the greenhouse while maintaining natural red/far red ratios at dusk and dawn. This experiment was performed in the autumn and early winter in the northern hemisphere (September to December) at a latitude higher than the plants’ native ranges (41°N vs. 33°S). As a result, the days were shorter than they would experience in the wild, but by less than an hour per day through the duration of the experiment.

We kept the pots moist for the first month after transplanting; after that, we watered pots with tap water when they became dry. Due to uncertainty in optimal growing conditions, some plants (four per population) received no fertilizer. All other plants received 100 mL of no-phosphorous fertilizer (60 ppm 20-0-20; J. R. Peters, Inc., Allentown, PA, USA) once every other week. Plants in the no fertilizer treatment received 100 mL of tap water instead.

Harvesting and excavation

All of the plants were excavated in December 2008, 77–83 days after planting. Before the soil was disturbed, we removed the shoot at the soil surface, measured its height and removed a leaf for area and dry weight measurements. We measured fresh leaf area using a LiCor 3100 leaf area meter (LiCor, Lincoln, NE, USA) and dry leaf and shoot (stem and leaves) mass after drying leaves and shoots in an oven for 2 weeks at 60 °C. We extracted the roots by slicing the pots lengthwise and gently teasing the roots from the soil. Some fine roots were lost in extraction. Once the entire root was extracted from the pot, we washed off the remaining soil, measured the length of the longest root and noted the presence or absence of root rot and proteoid roots. Finally, we measured dry weight of the roots after drying them for 2 weeks at 60 °C. From these primary measurements, we calculated several derived traits: specific leaf area (SLA, leaf area/leaf dry mass), leaf length/width ratio (LWR, leaf length/leaf width), root/shoot biomass ratio (Rt:ShtM, root mass/shoot mass), root/shoot length ratio (Rt:ShtL, root length/shoot height) and root length/root mass (RtL:M, length of longest root/root mass). The leaf traits and the root/shoot ratios are widely used functional traits, and root length/mass ratio captures important variation in root allocation strategies. Plants that died or had substantial root rot were excluded from the analysis (< 8% of plants), resulting in a total of 384 plants (91 no fertilizer, 293 fertilizer).

Climate and soil data

To extract climate variables for populations used in the study, we intersected the locations of the sampled populations with layers from the South African Atlas of Agrohydrology, based on more than 30 years of weather data from more than 1000 weather stations (Schultze, 2007). We reduced the large number of available climate variables to a manageable number by choosing broad environmental categories that we expected to differ between species based on Latimer et al.’s (2009) work in the white proteas. We then performed separate principal components analyses on variables corresponding to these categories. This resulted in one unmanipulated and two composite variables that encapsulate variation in the categories that have been found to be important in delineating white protea distributions (Latimer et al., 2009): seasonal rainfall concentration (PPTCON), winter temperature (COLDPCA) and intensity of dry season drought (DRYPCA). PPTCON ranges from 0 (even rainfall) to 100 (all of the yearly rainfall falling in 1 month). PPTCON is positively correlated with total rainfall, which we did not include in our analysis (r2 = 0.26). COLDPCA is a measure of the coldness of winters and is strongly negatively correlated with elevation (r2 = 0.37). This axis was calculated as the first axis of a PCA of the average minimum daily temperature in the coldest month and the number of heat units in the coldest 3 months (the first axis explained 87% of variation in these two variables). Large values indicate warmer winters. DRYPCA is the first axis of a PCA of the number of days with < 2 mm rainfall in the driest 3 months and the total rainfall during those months (87% of variation in these two variables). Large values indicate milder, moister dry seasons. These environmental variables are the same as those used in the study of Carlson et al. (2011).

To gauge the degree to which plant traits are associated with the soil fertility of their home environments, we collected bulked soil samples at 15 and 30 cm depth from three locations at each seed collection site. Bulk samples were dried for 1 week at 60 °C and analysed at BEM Labs (Somerset West, South Africa) for percentage N, total P, total K, pH and cation exchange capacity. We constructed a final PCA, FERTPCA, from these NPK concentrations (variable loadings: percentage N 0.659, total P 0.496, total K 0.588, first axis explains 58.7% of variation). Values for the environmental variables for each population can be found in Table S1.

Does evolutionary history, similar responses to the environment or unique responses to the environment play a larger role in the trait diversification of the white proteas?

To determine the relative contributions of evolutionary history, shared responses to the environment, and unique responses to the environment to trait diversity, we conducted manovas in R (manova, R Core Development Team, 2011). The manova approach allows us to account for correlations between the response variables (all plant traits listed above) while testing their relationships with environmental gradients, the species effect and the interaction between species and environment. We standardized the environmental predictors so that each one had a mean of zero and a standard deviation of 1. This allows us to investigate the relative importance of evolutionary history (species effect), shared response to the environment (environment effects: FERTPCA, COLDPCA, DRYPCA and PPTCON) and unique response to the environment (all species × environment interactions). For this analysis, we included only the individuals that received fertilizer in the greenhouse, for a total of 293 individuals.

We estimated the relative importance of each factor by calculating the partial Wilks’η2 for each factor. Wilks’η2 is a measure of the partial variance explained by a factor and the multivariate approximation of SSeffect/(SSeffect + SSerror); (see Langerhans & DeWitt, 2004 for further discussion of partial η2). To create confidence intervals around the estimates of η2, we jackknifed the analysis by sequentially dropping one individual from the analysis (Davison & Hinkley, 1997). We estimated the marginal SS for each factor as the SS of the last factor added to the model, rotating the order of factors included so that each appeared in the final term. We used this marginal SS to estimate η2. Because interaction terms are always estimated after main effects, we removed the interaction terms from the model to estimate the η2 of each of the main factors. The main factor η2s reflect the variance in multivariate trait space that is due to the average response of species to the four environmental variables and the variance that is explained by the average trait differences between species.

Which traits drive any unique response to the environment?

We are interested in the relationships of many traits with the environment. However, these traits did not evolve independently of one another, and developmental responses within an individual are likely to be correlated across several traits. Separate multiple regressions would allow us to investigate the association between individual traits and environmental covariates (as in Nakazato et al., 2008; Carlson et al., 2011). However, they would not allow us to differentiate between direct associations between a trait and the environment and indirect associations that arise because a particular trait is developmentally or genetically correlated with another trait that has a direct association. We use a structural equation model of trait variation to account for trait–trait associations, allowing us to isolate direct associations of traits with environmental features. Structural equation modelling (Jöreskog, 1970; Jöreskog & Sörbom, 1996; Lee, 2007) provides a general approach to statistical analysis of unobservable, ‘latent’, variables by specifying their relationship to observable, ‘manifest’, variables. By including these latent variables, our model accounts for trait–trait associations and allows us to estimate only 28 relationships (lines in Fig. 3) rather than trying to examine all 105 pairwise relationships between 15 traits.

Figure 3.

 Relationships between latent variables, measured traits and outside factors estimated in the SEM analysis. Latent variables are represented by grey ovals and traits and outside factors (seedwt = seed weight, daystogerm = no. of days to germination, fert = fertilizer treatment) are represented by white boxes. All variables are standardized, and the relative size of each relationship is indicated by the width of the arrow that connects the variables. Black arrows indicate positive relationships and grey arrows indicate negative ones. Dotted arrows are nonsignificant relationships. Asterisks indicate the traits for which the regression coefficients were set to 1.

We analyse the coordinated response of whole plants in terms of four components: root, shoot, leaf and root/shoot. We are especially interested in these four components because together they represent the three fundamental components of the plant body (root, shoot and leaf), and root/shoot allocation has long been recognized as a fundamental axis associated with plant adaptation (Orians & Solbrig, 1977; Chapin, 1980; Lloret et al., 1999). These are the four latent variables (LV) identified in the centre of Fig. 3. Each of the manifest variables is related to one of the latent variables and to environmental covariates through a linear regression. For example, if rootlgi is the root length of the ith individual in the sample, then

image

where a is the species-specific intercept, βrootlg.root is the ‘loading’ of rootlg on the root latent variable, gspecies[i] is the species-specific vector of regression coefficients on the environment associated with individual i, envpop[i] is the vector of environmental covariates associated with the site of origin for individual i, εbin[i] is the random effect associated with the greenhouse bin in which individual i was grown and εi is the random error associated with individual i. Notice that the value of the root latent variable is defined only implicitly through its set of regression relationships with corresponding manifest variables; in the case of root, those regressions are the ones involving root mass, root length, proteoid roots, main root dead and root length/mass ratio. To ensure a unique value for each latent variable, the loading of one manifest variable is set to 1, in our case root length (root LV), the presence of leaves (shoot LV), SLA (leaf LV) and root/shoot weight (root mass/shoot mass, root/shoot LV). The choice of variable used for standardizing the latent variable relationships is arbitrary, because remaining loadings are estimated relative to the loading of 1 for the variable used for standardization. The loadings represent the extent to which each of the manifest variables is associated with the latent variable. Correlations between manifest variables are accounted for through their association with the latent variables. Two manifest variables with positive loadings on a latent variable are positively correlated with that latent variable. If that latent variable accounts for a large fraction of the variability in the traits it underlies (analogous to the first component of a PCA), then the traits will also be positively correlated.

Our data include a mixture of continuous and binary response variables. We used jags 2.1.0 (Plummer, 2003) to analyse the structural equation model in a Bayesian framework using a logistic link for binary response variables and diffuse normal priors on all loadings and regression coefficients. As suggested by Gelfand et al. (1995), we use hierarchical centring to improve convergence of the MCMC sampler. We report results from an analysis using four independent chains, each with a burn-in of 50 000 iterations, followed by samples taken at 160-step intervals for the next 200 000 iterations, for a total of 5000 samples from the posterior distribution. Standard convergence diagnostics (Gelman & Rubin, 1992) suggest that convergence was satisfactory. All potential scale reduction factor (Rhat) values were < 1.04, most were < 1.005, and the effective size of the sample from the posterior distribution of each parameter ranged between 100 and 5000. For more than 90% of the parameters, the effective sample size was > 500.

The model output includes the full posterior distribution for all parameters, which we summarize with the posterior means and 95% credible intervals. Estimates for which the 95% credible interval does not overlap zero are statistically distinguishable from zero. As an additional test of the importance of unique responses to the environment, we examined two models, one that imposes the same trait–environment association across all species (the common-trait model) and one that allows each species to exhibit a different trait–environment association (the varying-trait model). The common-trait model allows each species to have a different mean trait value but forces the regression coefficients describing trait–environment relationships to be the same for all species. This model can be summarized as

image

where αspecies[i] is a species-specific intercept and g is vector of regression coefficients shared by all species. The varying-trait model can be summarized as

image

where gspecies[i] is a vector of species-specific regression coefficients associated with the species to which individual i belongs.

We compared the adequacy of these models using DIC, an information criterion similar to AIC (Spiegelhalter et al., 2002). AIC cannot be used for model choice in a Bayesian context, because it depends on a maximum-likelihood estimate for the parameters. DIC is the Bayesian equivalent, including a component measuring the fit of the model to the data and a component describing the complexity of the model. A difference in DIC larger than 10 units indicates that one model is strongly preferred over the other. If the preferred model is the one in which the species are allowed to respond differently to the environmental gradients, we have evidence that species have different trait–environment relationships. We investigate those differences further by identifying the relationships for which the credible interval does not overlap zero. This is a conservative approach to detecting differences because the failure to detect an effect may reflect a lack of statistical power rather than the absence of an effect. Furthermore, because estimates for individual species trait–environment relationships emerge from a hierarchical model in which species means are drawn from a common distribution, a correction for multiple comparisons is not necessary (Gelman & Hill, 2007). To retain as large sample size as possible (N = 384), all of the plants were included in this analysis. The effect of the fertilizer was included in the model and does not affect the results reported.

Results

Does evolutionary history, similar responses to the environment or unique responses to the environment play a larger role in the trait diversification of the white proteas?

The manova revealed significant effects for all three classes of factors: shared response to the environment, unique responses to the environment and evolutionary history (Table 2). Much of the morphological variance among populations was explained by the model, as shown by the Wilks’η2 values for the species and direct environmental effects. Based on the estimates of partial variance explained (Wilks’η2) and the measures of significance, historical effects and shared response to the environment were similarly important in predicting trait variation. The factor that explained the most variance was shared response to rainfall seasonality, followed by the effect of history (species effect). The unique responses to selection were significant, but mostly smaller than the shared responses to selection. However, the effect of unique responses to winter temperature (η2 = 0.140) was nearly as large as those of shared responses to winter temperature (η2 = 0.188).

Table 2. manova table including Wilks’η2, the partial variance explained by each factor and 95% confidence intervals around the Wilks’η2, estimated by jackknifing.
Test forFactorWilks’λApprox. Fd.f.Wilks’η2P2.5% CI97.5% CI
Shared response to environmentCOLDPCA0.8124.46614, 2700.188< 0.00010.1830.195
DRYPCA0.7835.34614, 2700.217< 0.00010.2110.222
FERTPCA0.7496.45414, 2700.251< 0.00010.2450.257
PPTCON0.59113.36214, 2700.409< 0.00010.4030.416
HistorySpecies0.1329.74670, 1289.5450.333< 0.00010.3310.335
Unique response to environmentCOLDPCA × species0.4712.9270, 1194.3260.140< 0.00010.1370.143
DRYPCA × species0.6121.85670, 1194.3260.094< 0.00010.0920.096
FERTPCA × species0.6731.47970, 1194.3260.076< 0.010.0740.079
PPTCON × species0.6221.78670, 1194.3260.09< 0.0010.0890.093

Which traits drive the unique response to the environmental gradients?

The trait–environment correlations are underpinned by the structure of the structural equation model, which is detailed in Table 3 and Fig. 3. All of the manifest variables were significantly related to the latent variables. The fertilizer treatment had an effect only on the root latent variable.

Table 3.   Summary statistics of the relationships estimated in the structural equation model; mean, standard deviation and 95% credible intervals.
 Mean (β)SD2.50%97.50%
  1. The relationships between traits and latent variables are one-way, but the relationships between Root, Shoot and Leaf latent variables are reciprocal. The relationships between the RootShoot latent variable and Root and Shoot latent variables are also one-way with Root and Shoot informing RootShoot.

  2. * indicates relationships for which the 95% credible intervals do not overlap zero.

Trait–latent variable relationships
 SLA–Leaf−0.316*0.073−0.464−0.176
 Leaf LWR–Leaf−0.231*0.073−0.376−0.09
 Leaf Mass–Leaf1.054*0.0370.9841.13
 Root dead–Root−2.962*0.433−3.874−2.183
 Root length/mass–Root0.191*0.0440.1060.276
 Root weight–Root0.497*0.0320.4360.559
 Proteoid Roots–Root1.402*0.3940.672.188
 Root/Shoot Mass–RootShoot0.514*0.0530.4110.617
 Shoot Height–Shoot1.245*0.1141.0321.477
 Shoot Weight–Shoot1.8*0.1111.5922.027
External–latent variable relationships
 Daystogerm–Leaf−0.0350.038−0.1130.039
 Daystogerm–Root−0.11*0.056−0.223−0.002
 Daystogerm–RootShoot−0.0160.037−0.0870.056
 Daystogerm–Shoot−0.0230.016−0.0550.009
 Fertilizer–Leaf0.0930.077−0.0610.241
 Fertilizer–Root−0.277*0.11−0.496−0.056
 Fertilizer–RootShoot−0.0460.088−0.2160.131
 Fertilizer–Shoot0.0310.045−0.0590.121
 Seed weight–Leaf0.384*0.0770.2310.533
 Seed weight–Root0.218*0.1060.0130.42
 Seed weight–RootShoot0.0050.082−0.1520.172
 Seed weight–Shoot0.227*0.0350.1610.297
 Seed weight–Daystogerm−0.1880.108−0.4040.023
Relationships between latent variables
 Root–Shoot0.557*0.0470.460.645
 Leaf–Root0.602*0.040.5190.679
 Leaf–Shoot0.843*0.0220.7950.88
 Shoot–RootShoot−1.473*0.18−1.844−1.128
 Root–RootShoot1.104*0.0461.0181.198

With a delta DIC of −218.3, the model that allows trait–environment relationships to differ between species (DIC = 12152.0) is much more strongly supported than the model in which the species are forced to respond similarly to the environment (DIC = 12370.3). The significant species-specific trait–environment correlations show why this is the case (Fig. 4, Table S2). We detected species-specific trait–environment correlations in all species but P. venusta. We identified the most species-specific relationships in P. mundii, with 20 of the 51 significant trait–environment associations occurring between P. mundii populations. Winter temperature (COLDPCA) was the environmental axis most frequently involved in species-specific trait–environment associations with 21 significant trait–environment associations between traits and COLDPCA. Leaf traits varied the most with winter temperature (COLDPCA) and summer drought (DRYPCA). Root traits varied the most with winter temperature (COLDPCA) and rainfall concentration (PPTCON). Associations with shoot, root/shoot and seed traits were more evenly distributed across the environmental axes.

Figure 4.

 Correlations between measured traits and the local environment. The environment is summarized along four axes COLDPCA (coldness of winter), DRYPCA (intensity of summer drought) and PPTCON (rainfall seasonality). All traits and environments are standardized, and the size of each circle is proportional to the size of the effect. Black circles are positive relationships, and grey circles are negative. Only significant relationships are shown. Black arrows indicate reversals in trait–environment correlations. P. venusta is excluded from the figure because it had no significant trait–environment correlations.

When all (both significant and nonsignificant) trait–environment associations are considered, species only responded similarly in four of the 60 of the trait–environment combinations (Fig. S1). Across all species, seeds were always heavier in populations from wetter areas, leaves always had larger areas in wetter areas, shoots were always taller in populations from areas with less seasonal rainfall, and leaves were wider (lower LWR) in areas with more concentrated rainfall.

The dissimilarity between species in how their traits vary with their environment becomes even more apparent when only the significant associations are considered. For example, leaf traits in P. aurea, P. lacticolor, P. punctata and P. subvestita vary along a gradient in the intensity of summer drought (DRYPCA), but the identity of those traits differs between species. In P. aurea and P. lacticolor, leaf shape (LWR) varies across the gradient of summer drought, whereas in P. punctata and P. subvestita, leaf mass varies. In other cases, the same trait varies across an environmental gradient in more than one species, but the direction of the associations differs between species. For example, P. mundii and P. punctata individuals from warmer areas tend to have heavier root systems, but P. aurea individuals from warmer areas tend to have lighter root systems.

All of the cases in which species differed significantly in the direction of their relationships between a trait and the environment were root traits (arrows in Fig. 4). In addition to the differences in root biomass described above, P. mundii tends to have more branched root systems (low root-length-to-mass ratio) in warmer areas, whereas P. aurea tends to have less branched root systems (higher root-length-to-mass ratio) in warm areas. Also, P. mundii tends to have more branched roots in areas with more concentrated rainfall, but P. aurea has less branched roots in areas with more concentrated rainfall.

In addition to the opposite responses to COLDPCA and PPTCON listed above, other root traits were correlated with the environmental gradients. They were associated with rainfall seasonality, but the pattern of association varied markedly among species. P. mundii had the most varied root response to rainfall seasonality, with fewer proteoid roots, heavier roots and more branched roots in areas with more seasonal rainfall. P. subvestita individuals from areas with more seasonal rainfall had shorter roots. Root traits also varied in response to intensity of summer drought, but only in P. subvestita, in which individuals from wetter areas tended to have heavier, more branched root systems. Species also differed in the association between root traits and winter temperature. In addition to the conflicting trends in P. aurea and P. mundii described above, P. punctata had heavier root systems in warmer areas. We failed to detect significant associations between any root traits and soil fertility.

Seed traits did not vary predictably along the environmental gradients we studied. We expected to find heavier seeds in lower-nutrient areas (Bonfil & Kafkafi, 2000), but we detected no relationships between seed mass and soil fertility (FERTPCA). While seed mass did vary in association with winter temperature and rainfall, the nature of the association differed between species. In P. aurea, populations from warmer areas had smaller seeds. In P. mundii, populations from areas with less seasonal rainfall had smaller seeds. P. lacticolor and P. mundii populations from areas with warmer winters germinated faster.

Similarly, root/shoot ratios rarely varied with the environmental gradients in the direction we expected based on many studies of plant allocation strategies in other systems (Orians & Solbrig, 1977; Schenk & Jackson, 2002). Root/shoot ratios are expected to be higher when water availability is low (McCarthy & Enquist, 2007), but we detected such a relationship only in P. mundii. In areas with less intense droughts, P. mundii individuals invested more biomass in shoots than in roots. In P. aurea and P. lacticolor, root/shoot ratio was associated with winter temperature, but not with summer drought. Individuals in these species from areas with warmer winters tended to invest more in shoots than in roots.

Shoot traits were quite labile in response to the environmental gradients. P. mundii, P. lacticolor and P. punctata individuals from warmer areas tended to be bigger (more likely to have leaves, heavier shoots or taller shoots). P. mundii and P. subvestita individuals from more fertile areas tended to be larger as well. However, P. mundii, P. aurea and P. punctata individuals from areas with more concentrated rainfall tended to be smaller.

Many of the trait–environment relationships that we detected in leaves were as expected. For example, we found wider leaves in wetter areas (Cunningham et al., 1999) in P. aurea and P. lacticolor and heavier leaves in populations from wetter areas in P. mundii, P. punctata and P. subvestita. The one environmental association that we found with SLA, widely thought to be an important adaptation to many environmental stresses, was also in the direction that we expected. We found that in P. aurea, plants from colder sites tended to have lower SLA and thicker or more dense leaves. P. mundii and P. punctata populations from colder areas also tended to have smaller leaves.

Discussion

Large-scale trait–environment associations and the myriad examples of parallel evolution and convergence show that species often respond similarly to similar selection pressures. However, the vagaries of evolutionary history also play a role in both the mode and amount of adaptation to similar selection pressures. Here, we show that in a recent evolutionary radiation, similar responses to environmental gradients and species history were the largest predictors of trait variation, but that unique responses to selection were also important in trait diversification. These varied responses suggest that even in a closely related group, species are responding to similar selection pressures in different ways and that evolutionary history can have an effect on both the size and direction of the response to selection.

Evolutionary history and similar responses to the environment play the largest roles in trait diversification of the white proteas

The strongest predictor of multivariate trait variation is rainfall seasonality (Table 2), but the effect of history and other shared responses to the environment were nearly as large. Unique responses to the environment, while detectable, played a smaller role in trait differentiation. The large shared responses of traits to the environmental gradients show that even though the species differ in their mean trait values (significant species effect), many environmental responses are similar. This suggests that much trait evolution within species represents parallel evolution to similar environmental gradients. In particular, similar responses to rainfall seasonality (PPTCON) was the largest predictor of multivariate trait variation. However, the significant unique responses to the environment indicate that the species also respond differently to each environmental gradient.

The traits that drive the unique evolutionary response to the environmental gradients

In spite of broad similarity in multivariate trait–environment associations detected by the manova, the six species of the white proteas differ in the degree to which their traits are associated with the environment. Using the structural equation model, we detected 20 significant trait–environment associations in P. mundii, which might be due to the large environmental differences between the eastern and western ranges of P. mundii. In contrast, we did not detect any significant relationships between traits and the environment in P. venusta. P. venusta is found only near the tops of mountains and does not cover as much of the environmental gradients as do the other species. It also has smaller population sizes than the other species (Rebelo, 2006), so it might not harbour the genetic variability necessary to respond to selection.

In most of the cases in which more than one species had a significant trait–environment association, the differences were in the strength of the association, but in a few cases, there were differences in the direction of the association. In each of the 36 trait–environment combinations (e.g. COLDPCA and seed mass) in which we found at least one significant species-specific correlation, we found at least one species in which the relationship was different from that of other species (either not significant or in the opposite direction). Whereas others (Travisano et al., 1995; Ruzzante et al., 2003, Gomes & Monteiro, 2008 and Eroukhmanoff et al., 2009) have shown that evolutionary history affects the magnitude of a selection response, very few studies show closely related species responding to similar selection pressures in opposite directions.

All of the traits for which we detected trait–environment associations of opposite signs were root traits, and in all cases, P. aurea varied in the opposite direction of P. mundii or P. punctata. This result is especially surprising, because P. aurea’s environmental ranges are quite similar to that of P. mundii (Fig. 1) and its geographical distribution lies between two disjunct population segments of P. mundii. It appears that P. aurea and P. mundii populations are adapting to similar habitats in different ways. In two of the three cases when the two species have opposite trait–environment correlations, trait variation in P. mundii reflects the pattern expected from broad-scale comparisons. First, in areas with more concentrated rainfall, which in our data set is strongly correlated with total rainfall, P. mundii populations tend to have more branched root systems (lower root length/mass ratio) in areas where rainfall is highly concentrated (higher total), whereas P. aurea had more simple root systems. Nicotra et al. (2002) found that shrub species from drier areas of Australia invested more in the main axis of their roots than those from wetter areas, similar to the pattern that we detected in P. mundii. Second, P. mundii and P. punctata individuals from warmer areas have heavier root systems (and shoots), but P. aurea populations have lighter root systems (and nonsignificantly, shoots). Many studies have shown that plants at higher elevations tend to be smaller than those at lower elevations (Clausen et al., 1940; Woodward, 1986; Oleksyn et al., 2003; Byars et al., 2007), and in our system, winter temperature is strongly associated with elevation. In the third case, we have no expectation for the association between root length/mass ratio and temperature.

Trait–environment associations rarely reflect large-scale patterns

Some of the trait–environment associations that we detected were in the directions that we expected based on trends documented in many other studies, but some were not, and some expected relationships were not detected at all (Table 1). For example, large seed size has long been considered an adaptation to low-nutrient soils because of the additional resources that can be stored in a larger seed (Beadle, 1966; Stock et al., 1990; Milberg et al., 1998; Vaughton & Ramsey, 2001). Patterns consistent with this expectation have been found in studies of two other members of the Proteaceae, Banksia cunninghamii (Vaughton & Ramsey, 2001) and Protea compacta (Shane et al., 2008). In these studies, large seed size was shown to be an adaptive response to low-nutrient soils. In contrast to these expectations, we did not find any association between seed size and soil fertility. P. mundii individuals from areas with more concentrated (and more) rainfall did have larger seeds, but this might be a plastic response as a result of better maternal environment.

In root traits, we expected to find a negative association between soil fertility and the presence of proteoid roots, but we found no such association. Only in P. subvestita did we find the expected relationship between intensity of dry season drought and root weight and root length/mass ratio. Seedlings from populations that experience stronger droughts tended to have lighter, simpler root systems (for a given length, they weighed less). We found only one association between root/shoot biomass ratio and the intensity of dry season drought, even though this association is a textbook example of a resource investment trade-off (Larcher, 1995) and has often been detected in other studies (e.g. Blum, 1996; Lloret et al., 1999; Li & Wang, 2003). We detected this pattern in P. mundii, which has a large range of drought intensities. Our failure to detect this association in other species might reflect a smaller absolute range of drought intensities. Alternatively, rooting depth and root/shoot ratios might have little to do with drought intensity in our species. In Protea, rooting depth is often related to the depth of the water table (Richards et al., 1995; Watt & Evans, 1999), which depends on local geology. None of our environmental variables captures variation in water table depth.

The relationships that we found between above-ground traits and the local environment are consistent with those reported earlier for garden-grown plants by Carlson et al. (2011). However, in most cases, we only detected a significant association in one or two species. For example, both studies found that plants from areas with more concentrated precipitation have broader leaves (lower LWR), but in our study, it was only P. mundii in which we could detect the association. Carlson et al. (2011) ascribe the broader leaves in areas with more seasonal rainfall to the positive association between PPTCON and total rainfall. We may be detecting a relationship between leaf shape and total rainfall, a pattern seen across the Proteaceae (Thuiller et al., 2004; Yates et al., 2010). Narrow leaves lose heat through convection more efficiently than do wider leaves (Yates et al., 2010) and thus are likely to be favoured in drier areas where transpirational cooling is particularly costly. Carlson et al. (2011) found that the SLA of plant leaves grown in two common gardens was associated with differences in COLDPCA and DRYPCA of their source populations. In our study, only P. aurea populations varied in SLA along the temperature gradient (COLDPCA), and no species had a significant association between SLA and drought. Our failure to find strong associations between SLA and the environmental gradients in the greenhouse might be due to G × E interactions. The mild conditions in the greenhouse might have resulted in smaller differences in leaf morphology between populations (values of all traits for all populations are included in Table S3). Alternatively, the lack of significant associations that we detected in the greenhouse might be due to differences in power between the two analyses. In the study of Carlson et al. (2011) the trait–environment associations were estimated using all of the populations in all of the species, whereas in the study presented here, there were only a maximum of five populations per species to be used to detect an association between traits and the environment.

Conclusions

Our study shows that even in a closely related group of species, evolutionary history can play a role in trait diversification. Although on average the species responded similarly to the environment, they also had unique responses. In particular, the structural equation model revealed that in some cases, the root traits of P. aurea and P. mundii varied in opposite directions in response to the same environmental gradient. Evolutionary history can have an effect on adaptation at large and small evolutionary scales. Here, we document how evolutionary history can affect trait–environment associations even between very closely related species. We show that evolutionary outcome of selective pressures cannot always be predicted. Further, the lack of correspondence between large-scale trait–environment associations and those that we detect within the white proteas suggests that these gradients are unlikely to be driving much of the trait diversification in this group. This discrepancy between intra- and interspecific trait–environment associations might be a more general phenomenon and bears further investigation.

Acknowledgments

The authors thank A.G. Rebelo, M. McQuillian and G.F. Midgely for support in South Africa, A. M. Gawel for help collecting seeds and C. Morse for greenhouse support in CT. We thank C. Berdan and D. Fryxell for help in the CT greenhouse and C. Adams and many EEB graduate students for harvesting roots. A.M. Latimer and A.M. Wilson extracted the climate data. D. Weese gave helpful editorial comments. We thank Cape Nature, the Chief Directorate of Environmental Affairs of the Eastern Cape Province, the Eastern Cape Parks Board, the Department of Water Affairs and Forestry, Ezemvelo KZN, the many reserve managers and the private landowners for granting access to our wild study populations. This research was funded by NSF DDIG DEB 0709690 awarded to K. Holsinger and R. Prunier and DEB 0716622 awarded to K. Holsinger.

Ancillary