• Adaptation to divergent environments creates and maintains biological diversity, but we know little about the importance of different agents of ecological divergence. Coevolution in obligate mutualisms has been hypothesized to drive divergence, but this contention has rarely been tested against alternative ecological explanations. Here, we use a well-established example of coevolution in an obligate pollination mutualism, Yucca brevifolia and its two pollinating yucca moths, to test the hypothesis that divergence in this system is the result of mutualists adapting to different abiotic environments as opposed to coevolution between mutualists.
• We used a combination of principal component analyses and ecological niche modeling to determine whether varieties of Y. brevifolia associated with different pollinators specialize on different environments.
• Yucca brevifolia occupies a diverse range of climates. When the two varieties can disperse to similar environments, they occupy similar habitats.
• This suggests that the two varieties have not specialized on distinct habitats. In turn, this suggests that nonclimatic factors, such as the biotic interaction between Y. brevifolia and its pollinators, are responsible for evolutionary divergence in this system.
Adaptation to divergent environments creates and maintains biological diversity (Schluter, 2000; Funk et al., 2006), but we know little about the importance of different ecological factors, particularly the relative contributions of the abiotic environment and biotic interactions (Lexer & Fay, 2005; Johnson & Stinchcombe, 2007). It has been hypothesized that coevolution drives ecological divergence in many taxa (Ehrlich & Raven, 1964; Thompson, 2005). The pattern of co-phylogenesis observed at some taxonomic scales in putatively coevolving lineages, such as obligate pollination mutualisms, has offered some support for this contention (Weiblen & Bush, 2002). However, the hypothesis that coevolution produces diversification has rarely been tested against alternative explanations. In particular, there is a lack of tests for the role of ecological divergence caused by the abiotic environment in taxa that are hypothesized to coevolve (Johnson & Stinchcombe, 2007; Nuismer & Gandon, 2008; Toju, 2008). We know that abiotic stresses can have a significant effect on species in these systems (Segraves, 2003), and that many species’ boundaries are a result of environmental differences (Kruckeberg, 1957; Michalak et al., 2001; Lexer & Fay, 2005). It would thus be reasonable to suggest that diversification in these systems is the consequence of adaptation to divergent environments, rather than biotic interactions between mutualists. If this is the case, each taxon should possess different environmental requirements.
We address these questions in a classic example of an obligate mutualism, between Yucca brevifolia Engelm. (Joshua tree; Agavaceae) and its yucca moth pollinators (Lepidoptera: Prodoxidae). Yucca brevifolia is exclusively pollinated by two sister species of yucca moth, Tegeticula synthetica (Riley) and Tegeticula antithetica Pellmyr, which speciated whilst interacting with Y. brevifolia (Pellmyr & Segraves, 2003; Godsoe et al., 2008). The two moth species differ in overall body size and in ovipositor length, and are parapatrically distributed, with the larger T. synthetica occurring in the western half of Y. brevifolia's range and the smaller T. antithetica occurring in the eastern part. Interestingly, the two moth species co-occur in a 5-km-wide contact zone in south-central Nevada (Fig. 1). Previous work has suggested that the mutualism between Y. brevifolia and its pollinators is mediated by phenotype matching: the female moth uses a blade-like ovipositor to cut into the host flowers and lay her eggs inside the ovary before actively pollinating the flowers (Trelease, 1893), so that her larvae can consume a subset of the developing seeds. Because it must penetrate deeply enough to lay eggs on top of the ovules, but not so deep as to damage the flower, the ovipositor should match the length of the route by which it reaches into the floral ovary (Pellmyr, 2003).
Previous work has documented two varieties of Y. brevifolia based on tree architecture and height to first branching (Rowlands, 1978). These taxa are associated with different species of pollinator (Pellmyr & Segraves, 2003, Godsoe et al., 2008), and 12–13 morphological traits examined are significantly different between these varieties (Godsoe et al., 2008). With these data, it is possible to assign an individual tree to the correct variety with 95% accuracy. The varieties are sister taxa (Pellmyr et al., 2007) and have been described as separate species (Lenz, 2007) but, in the absence of an estimate of gene flow based on nuclear markers, we follow previous nomenclatures that treat these entities as varieties (Rowlands, 1978; Baldwin et al., 2002; USDA, 2008).
There is a complex history of co-divergence between tree varieties and moth species. The ancestral host plant of the two moth species is Y. brevifolia (Godsoe et al., 2008). The distributions of these two varieties parallel those of their pollinators, with Y. brevifolia var. brevifolia (hereafter var. brevifolia) in the west and Y. brevifolia var. jaegeriana (hereafter var. jaegeriana) in the east. Molecular analyses of chloroplast markers indicate that these two varieties are polyphyletic, but diverged significantly earlier than their pollinators (Smith et al., 2008). However, in spite of this history of ancient divergence in the plants, there are indications of ongoing gene flow between the two varieties (Smith et al., 2008).
The matching between moth and floral phenotypes suggests that the differences between var. brevifolia and var. jaegeriana may have arisen as a result of coevolutionary interaction with the divergent pollinator species (Godsoe et al., 2008). However, it is also possible that natural selection mediated by the abiotic environment could have produced these differences. To test this alternative hypothesis, we examined whether the two varieties have different climatic requirements. The distribution of Y. brevifolia crosses a number of major climatic divisions in the south-western USA. This species occurs primarily in the Mojave desert of the south-western USA, with some populations reaching the Sonoran and Great Basin deserts, the Colorado Plateau semi-desert, as well as montane forest habitats (Fig. 2a). Much of the ecology of plants in this region has been shaped by their ability to survive extreme climates (Bailey, 1995). Moreover, the east–west split between the two varieties of tree has previously been ascribed to a known trend in precipitation across North American deserts (Rowlands, 1978). The Mojave desert receives most of its precipitation in the winter, whereas eastward deserts receive an increasing proportion of summer precipitation, approximating a 50 : 50 summer to winter ratio near the south-eastern distribution limit of Y. brevifolia (Brown, 1982). It is thus reasonable to examine whether adaptations to different climates could explain why there are two distinct varieties of Y. brevifolia with distinct distributions.
Most plants involved in reciprocally obligate pollination mutualisms are large and long lived (Janzen, 1979; Pellmyr, 2003). For this reason, traditional tests for the role of environmental variation on species’ distributions, such as reciprocal transplant experiments, are not typically feasible. This system is, however, an ideal candidate for applying Geographic Information System (GIS)-based methods to test whether these varieties have different climatic requirements (Kozak et al., 2008; Swenson, 2008), because Y. brevifolia's distribution is exceptionally well characterized.
The most popular GIS-based approach is to use ecological niche modeling to test whether the predicted environmental requirements of the taxa are identical or similar (Warren et al., 2008). However, simulation work has shown that some such comparisons can erroneously infer that species have different niches when one species disperses to environments unavailable to the other (W. Godsoe, unpublished). For this reason, we have not focused our investigations on the determination of whether different taxa live in identical environments. Instead, we test whether the two varieties of Y. brevifolia have a similar response (pattern of occurrence given important environmental variables) in the environments that both varieties encounter. This is performed in three steps. First, principal component analysis (PCA) is used to determine which environments are available to both varieties. Second, the ability of ecological niche models to predict the presence of each variety of plant is assessed. If these models can accurately determine where Y. brevifolia occurs, it would suggest that climate is an important determinant of whether environments are suitable for this species. Third, we compared the ability of these distribution models to predict presences in new regions. If there has been ecological divergence caused by physical environmental factors, we should expect a model developed from trees of one variety to be a poor predictor for the presence of the other variety.
Materials and Methods
Distribution data assembly
We combined existing survey datasets with extensive, complementary field surveys of nearly all known populations of Y. brevifolia to produce a dataset of 5765 presence points [Fig. 1; see also Table S1 (Supporting Information)]. To compensate for variation in sampling density between different areas, we created a stratified subsample of our dataset by overlaying a 5 km × 5 km grid and randomly sampling one presence point from every cell that contained any trees. This produced a final dataset of 319 locations where var. brevifolia was present and 417 locations where var. jaegeriana was present.
Niche models were based on climate data from 19 WORLDCLIM bioclimatic variables (Hijmans et al., 2005; http://www.worldclim.org/) at a spatial resolution of 30 arc seconds (~1 km). We believe that these are the most important abiotic variables to study for Y. brevifolia because previous authors have hypothesized that the two varieties specialize on different climatic conditions (Rowlands, 1978). Moreover, climate is believed to be a particularly important determinant of species distributions in the American South-west (Bailey, 1995). Individual populations of Y. brevifolia are separated from each other by as much as 50 km. We interpret this as an upper bound on how far away we should expect to find disjunct populations. We thus restricted our analyses to a minimum convex polygon around the range of Y. brevifolia plus a 50-km buffer (Fig. 1). We divided our study area into a western region containing presences of var. brevifolia and an eastern region containing presences of var. jaegeriana. In the south, this division traverses a low-elevation region (Fig. 2b), inhabited by neither variety (Fig. 1). Yucca brevifolia disperses over short distances. Seeds are typically dispersed less than 30 m by rodents (Vander Wall et al., 2006), and Yucca moths typically disperse pollen over distances of less than 50 m (Marr et al., 2000). We are thus reasonably confident that each variety cannot disperse across this boundary. In the north, there is a great deal of suitable habitat for Y. brevifolia and the two varieties come into contact. We are less confident that our division represents a natural barrier in this region but, in the north, similar environments are available on both sides of this boundary. We thus interpret our division as a reasonable approximation of the environments to which each variety can disperse.
Niche modeling algorithms can produce unreliable extrapolations in environments to which the focal organism cannot disperse (W. Godsoe, unpublished). Thus, we identified regions in which any climatic variable in one portion of the Mojave is more extreme than any value of that variable in the other portion of our study region. These regions represent environments to which one variety cannot disperse, and so we treat inferences of niche models in these regions with caution (Elith & Graham, 2009). It is difficult to visualize all 19 WORLDCLIM variables, and so we have summarized this variation by calculating principal components from 2000 randomly selected locations from the eastern and western Mojave, respectively. Maps generated from these principal components provide a simple way to explore the climates available to both varieties (Fig. 2). In particular, when the color of such a map in one region is more extreme than any shade found in the other region (see, for example, Fig. 2c,d), it indicates that a climate in the first region is more extreme than any climate available in the second. These principal components also provide a convenient way to link the climate data used to model the distribution of Y. brevifolia to existing descriptions of environmental variation in the American South-west.
We developed niche models using boosted regression trees (BRT) and maximum entropy (Maxent; Friedman et al., 2000; Phillips et al., 2006), two algorithms that performed well in an extensive comparison of available methods on empirical data (Elith et al., 2006). Unless otherwise noted, we created models by using all of the presences for a particular variety.
BRT is a multi-stepped machine-learning algorithm. The method starts by constructing a decision tree that produces a crude classification of observations into presences and absences. The algorithm builds subsequent trees to explain observations that were poorly predicted in previous steps, and then builds a composite model from a large number of these decision trees, allowing for nonlinear terms, complex interactions and an evaluation of the relative importance of different parameters. We implemented BRT using Elith et al.'s (2008) modifications of the gbm package in R (R Development Core Team, 2006; Ridgeway, 2006). Our models used presence points for a given variety and an equal number of pseudoabsences selected at random from the appropriate region of the Mojave desert. BRT requires three user-defined parameters that adjust the complexity and stochasticity of the resulting model. We compared the cross-validated area under the receiver operating curve (AUC) scores of models using 27 parameter combinations (three different bagging fractions, three different tree complexities and three different learning rates). AUC scores varied little between different parameter combinations (0.85–0.89). A relatively fast learning rate of 0.01 and tree complexity of eight produced slightly higher AUC scores for both varieties, and so we used these parameter values in subsequent analyses. A bagging fraction of 0.25 produced the best model for var. jaegeriana; a bagging fraction of 0.75 produced the best model for var. brevifolia. As a compromise, we used a fraction of 0.5 for subsequent analyses. We estimated the number of trees in the final model with a 10-fold cross-validation procedure (Elith et al., 2008). This resulted in models from a composite of approximately 400 decision trees.
Maxent is a machine-learning algorithm that estimates the probability of occurrence by constraining this probability distribution to resemble the distribution of observed presences and by minimizing the information contained in the residuals (Phillips et al., 2006). We sampled 10 000 background points for analyses of var. brevifolia from the western portion of our study area, and 10 000 background points for analyses of var. jaegeriana from the eastern portion of our study area. We used the default parameter settings for Maxent version 3.1.0 (logscale output, regularization multiplier = 1, convergence threshold = 0.00001).
Assessment of model accuracy
We used AUC statistics as our primary measure of model accuracy (Freeman, 2007). This is a nonparametric estimate of a model's ability to distinguish between presences and absences. The AUC metric varies from zero to unity, with a score of unity representing perfect discrimination and a score of 0.5 representing a model that performs no better than random. This metric is threshold independent and so is an appropriate choice for comparing Maxent and BRT, two methods with continuous outputs that vary from zero to unity. We computed cross-validated AUC scores to measure the accuracy of niche models for each variety within its home range. In Maxent, we obtained cross-validated scores by randomly selecting 90% of the observations, and then calculated AUC scores on the remaining 10%. For BRT, we used the R scripts of Elith et al. (2008) to create 10 datasets, each using 90% of the available data to create the model and 10% to calculate AUC scores; the final AUC score is the average over all 10 data partitions.
Comparing predictions of niche models
To test whether the two varieties have similar environmental requirements, we must find a way to compare the predictions of two niche models developed in different regions. Unfortunately, AUC scores will not necessarily reflect the ability of the models to predict locations in new regions (Randin et al., 2006; Peterson et al., 2007). For this reason, we specifically designed a measure of niche similarity that compares the ability of niche models of each variety to predict to a new region. Using the maps from our PCAs, we determined that comparable climates are available approximately 300 km to the east or west of the contact zone. We set aside subsets of this region near the contact zone for tests of the ability of models developed for each variety to predict presences in adjacent regions with similar climates. For var. brevifolia, we created a niche model for the 225 presences further than 200 km from the contact zone. We measured the ability of this model to predict the 94 var. brevifolia presences within 200 km of the contact zone. We then used a model of var. jaegeriana trees to predict to the same set of presences (Fig. S1A, see Supporting Information), and compared the accuracy of both models using AUC scores. Similarly, we produced a model of the 138 var. jaegeriana presences further than 200 km from the contact zone, and then applied it to the 279 var. jaegeriana presences within 200 km from the contact zone. Finally, we compared the accuracy of these predictions to that of a model developed for var. brevifolia trees (Fig. S1B). Although sample size varied from one model to another, a recent analysis has found that BRT can produce reasonably robust niche models with only 30 presences (Guisan et al., 2007). Our models used at least four times more presences than this minimum.
For these comparisons, we present AUC scores with standard errors computed in the PresenceAbsence package in R (Freeman, 2007). However, we note that this standard error estimate is technically only valid for comparing models against random prediction, and for the particular set of presence and pseudoabsence points used.
This approach for measuring niche similarity has three advantages over existing methods (Peterson et al., 1999; Warren et al., 2008). First, there is some concern that a good niche model in one region may be less accurate in another (Peterson et al., 2007). In the absence of ecological divergence, we might still expect a model developed in one region to produce more accurate predictions in that region than would a model derived elsewhere. Our test mitigates this problem by forcing the models of both varieties to predict in a new region.
Second, this design permits us to specifically compare our models in one region. Niche models produce higher AUC scores in study areas with a great deal of unsuitable habitat than they do in study areas with a great deal of suitable habitat (Lobo et al., 2008). Thus, comparisons of AUC scores from different regions may give the illusion that one model is more accurate than another, even though they may produce equally good predictions when applied to the same study region. We eliminate this problem by restricting our model comparisons to a single study area per test.
Third, our method also has a useful ecological interpretation. If the sharp distribution limits that we observe in sympatry are a consequence of different climatic requirements, a model of one variety should predict very few of the presences for the other variety near the contact zone. For this reason, a model of the wrong variety would be far less accurate than a model of the correct variety. However, if the two varieties have different distributions for some other reason, such as biotic interactions or dispersal limitation, we should expect the distribution models to produce predictions of a comparable quality.
We present two metrics of the ability of niche models of each variety to predict the distribution of the other variety. First, we use the cross-validated AUC score as a measure of the ability of a model to make predictions to the home range of a variety. We compare this score with the ability of the same model to make predictions to the range of the other variety (hereafter referred to as AUC accuracy away from the home range). Second, we test whether BRT models of one variety correctly predict presences of the other variety significantly better than chance (Fig. 3), a frequently reported measure of niche similarity (Peterson et al., 1999; Warren et al., 2008). As our models were fitted to a dataset of 50% presences and 50% pseudoabsences, we identified presences as observations with a probability of occurrence of more than 50%. We then used a χ2 test following Peterson et al. (1999), with Yates’ correction, to determine whether the model predicted a greater number of presences than would be expected by chance.
The first three principal components summarized 88% per cent of the environmental variation. PC1, representing 49% of the variation, contrasts between temperature and precipitation (Table S2, see Supporting Information) and is strongly associated with elevation. PC2, representing 30% of the variation, is negatively correlated with precipitation in the coldest quarter, precipitation in the wettest month and several other traits that are high in regions near the Pacific coast (i.e. the west portion of our study region), and is positively correlated with temperature seasonality and precipitation in the driest month, all of which are properties of interior climates. For this reason, it may be thought of as a contrast between coastal (low values of PC2) and interior (high values of PC2) climates. PC3, representing 9% of the variation, is positively associated with temperature variation and summer precipitation, which are characteristics of the Sonoran desert and Colorado Plateau at the eastern edge of the range of Y. brevifolia (Fig. 2a,d). There is much more variation in PC2 in the western Mojave than in the east, and much more variation in PC3 in the eastern Mojave than in the west (Fig. 2c–f). Together, PC2 and PC3 partially represent the previously described east–west gradient from strong winter precipitation in the Mojave desert to strong summer precipitation in the Sonoran desert (Bailey, 1995), although neither shows a steep east–west gradient further north, near the contact zone between var. brevifolia and var. jaegeriana (Fig. 2c,d). In this region, PC1 exhibits the most variation, but both varieties of Y. brevifolia occupy a comparable elevation range in allopatry (Fig. 4b,d). It is in sympatry that we see a sharp deviation from this pattern, where var. brevifolia occurs in a region with low scores on this PC, and var. jaegeriana occurs in regions with high scores (Fig. 4c).
Niche model accuracy and comparisons
Maxent and BRT each produced models with AUC scores above 0.8 when niche models of each variety were used to predict their own ranges (Table 1). When niche models of one variety were applied to the range of the other, they predicted many presences near the contact zone, but were less accurate in more distant locations (Fig. 3d,e). This pattern was most pronounced when we applied predictions derived from the southern portion of the range of var. brevifolia to the prediction of var. jaegeriana (AUC = 0.611 and AUC = 0.645 for Maxent and BRT, respectively; Figs 3, S2; Table 1). In that case, the Maxent and BRT models predicted essentially no presences in the southern portion of the range of var. brevifolia, producing the lowest AUC scores in the study. The climate in this region has low loadings on PC2; moreover there is no region with an equivalent climate in the eastern Mojave (Figs 2, S2).
Table 1. Area under the receiver operating curve (AUC) scores for ecological niche models of each variety, projected within their range (Yucca brevifolia var. brevifolia in the western portion of our study area; Y. brevifolia var. jaegeriana in the eastern portion) and onto the range of the other variety
BRT, boosted regression trees; Maxent, maximum entropy.
Yucca brevifolia var. jaegeriana trees near the contact zone were predicted nearly as well by a model derived from var. brevifolia as they were by var. jaegeriana away from the contact zone (AUC ± SE: 0.718 ± 0.02 vs 0.739 ± 0.02). The distribution of var. brevifolia near the contact zone was actually predicted better by a model of the distribution of var. jaegeriana than it was by a model of var. brevifolia away from the contact zone (AUC = 0.748 ± 0.04 vs 0.658 ± 0.04). Across the entire range, BRT produced niche models that performed much better than chance (prediction based on var. brevifolia to var. jaegeriana, P = 1.1e-51; prediction based on var. jaegeriana to var. brevifolia, P = 0.0000578).
Our combined analyses indicate that Y. brevifolia is exposed to a diverse array of climates, and that these climates affect which environments are suitable to this species. There is little evidence, however, that trees associated with different pollinators have different ecological requirements. Both the eastern and western partitions of our study area contain locations with climates absent in the other partition (Fig. 2). Yucca brevifolia var. brevifolia occurs along the western edge of the Mojave desert, a region with low loadings on PC2 associated with a more coastal climate. This region includes the transition from desert to montane woodlands (Fig. 2a,c,e; Bailey, 1995). The BRT niche model for var. jaegeriana predicts few presences in this region, and the Maxent model predicts none (Fig. S2). Conversely, a few populations of var. jaegeriana reach into the edge of the Sonoran desert, a region characterized by summer precipitation; this climate has no equivalent further west. This region has high scores on PC3 (Fig. 2d,f), and niche models of var. jaegeriana predict few presences in these regions.
Our niche models indicate that, when eastern and western trees encounter similar climates, they have a similar response. Not only do models of one variety predict many presences in regions with comparable climates, they actually produce predictions of equal, if not greater, accuracy than models of the correct variety. BRT niche models for each tree variety even produced acceptable predictions across the range of the other variety, and were far more accurate than would be expected by chance. Projections of the niche model of one variety across the entire distribution of the other must be interpreted with caution, but this result at least suggests that, in many places, the climatic requirements of each variety are very similar.
A simple explanation for this pattern would be that moth and tree lineages diverged into allopatric or parapatric populations at some time in the past. We know from PCA that different climates suitable to Y. brevifolia are available in the eastern and western portions of the Mojave. Given this, allopatric lineages should occupy different environments, even in the absence of ecological divergence. Unfortunately, the data available do not allow us to compare the ecological requirements of Y. brevifolia across its entire range. Therefore, it is still possible that the tree varieties would perform differently if they were both to encounter environments with high values for PC2 or PC3. To illustrate this, consider an experimental design that corresponds to the extant distribution of Y. brevifolia. Imagine if we had planted both varieties in the environments near their contact zone in Nevada, but only one variety in a study site to the west and one in a study site to the east. We could compare the requirements of each variety in Nevada, but it would be impossible to compare their requirements in the other two study sites. By analogy, the extant distribution of the two varieties is similar to a single common garden ‘planted’ in locations in which the two varieties could encounter each other, but with missing or incomplete information about other locations. This interpretation does, however, depend on our assessment of the environments to which the two varieties of Y. brevifolia can disperse. If the two tree varieties really can disperse across the southern Mojave, the absence of var. jaegeriana in the west and var. brevifolia in the east constitutes strong evidence of ecological differentiation. Nevertheless, we are reasonably confident in our assertion that the trees cannot disperse across this barrier because our distribution models indicate that this region contains a broad area of habitat unsuitable to both varieties, and we know that Y. brevifolia disperses over short distances (Vander Wall et al., 2006).
The main conclusion of our study is thus that the two varieties of Y. brevifolia have similar climatic requirements across the many environments that they both encounter. This demonstrates that the two varieties have not specialized on distinct climates. In particular, we know that the abrupt boundary between the two varieties in sympatry cannot be explained by adaptation to different climates, as niche models of one variety are reasonably good at predicting presence for the other variety. It is possible that recent contact, or unmeasured abiotic variables, drive this boundary, although the strong performance of our niche models in allopatry suggests that climate is indeed a very important determinant of Y. brevifolia's distribution.
Comparing the predictions of distribution models is a relatively new technique. We have only a limited understanding of the properties of these methods, and so it is important to consider their potential weaknesses. To our knowledge, there are no analyses of the statistical power of these methods. Our comparisons of niche models indicate that the performance of the two varieties is similar, but with current analytical tools we cannot rule out the possibility that there is a biologically significant difference in the performance of the two tree varieties that is too small to be detected with distribution-based methods. We have followed existing conventions for these methods that test for similarity using null models, but recognize that more theoretical work is needed to develop a nuanced interpretation of distribution-based data.
It is possible that the incorporation of more data, such as Joshua tree population density at different locations, could improve the power of our tests. If the tree varieties have different environmental requirements, we might expect var. brevifolia trees to be at low densities in environments with high values of PC3, and var. jaegeriana trees to have low densities in regions with low values of PC2. We were unable to collect sufficient density data to test this idea, but the limited information at hand is inconsistent with this hypothesis. Rowlands (1978) analyzed the percentage cover data from c. 30 populations. He found that south-eastern populations of var. jaegeriana were unusually dense, but that across the rest of the range the density of both varieties increased gradually with elevation. Moreover, the highest percentage cover of var. brevifolia was in the south-eastern portion of the range of that variety, a part of the western Mojave with particularly high values of PC3.
Another limitation is that we have not explicitly partitioned out the effects of mutualist-mediated fitness from the abiotic environment. This is an important problem, and one that can now be addressed rigorously, albeit with complex experimental designs that require reciprocal transplants and manipulations of the available mutualists (Nuismer & Gandon, 2008; Piculell et al., 2008). In comparison, our approach is necessarily crude. We found that the two tree varieties had similar climatic requirements, particularly in the northern portion of their range where they encounter one another. This indicates that climatic requirements are insufficient to explain the observed differences in the two taxa. This is an interesting qualitative result, but extensive additional work will be needed to quantify the relative influence of the biotic and abiotic environment.
Aside from the statistical issues, many inferences about the causes of ecological divergence are hindered by our lack of historical information. We know that agents of evolutionary change may have been important in the past, but that subsequent genetic or environmental changes have obscured this fact (Coyne & Orr, 2004). Although there is little evidence of current ecological differentiation, it is possible that the two varieties specialized on different environments in the past, but that these environments do not currently exist (Jackson & Overpeck, 2000). The best way to deal with this problem is to place our observations in a phylogenetic context; in other words, to test whether speciation is associated with pollinator divergence and not with changes in an organism's abiotic requirements. This is not possible in our study because we have only investigated a single example of speciation in the moth lineage and in the plant lineage. The similarity in environmental requirements suggests that adaptation to different climates has not driven ecological divergence in this system. To our knowledge, this represents one of the first tests of this important hypothesis in a coevolving system (but see Toju, 2008). Nonetheless, this result must be interpreted with caution, as a mechanism of divergence may have been very important in the past, but may be largely irrelevant now (Coyne & Orr, 2004).
We do not yet understand why vegetative traits, such as leaf length and height to first branching, are different between the two varieties. We propose two possible explanations. First, tree architecture may be developmentally linked to floral traits. In particular, branching occurs at the site of previous inflorescences (Rowlands, 1978), and so selection on floral development might indirectly alter other aspects of tree morphology. Second, isolation between the two varieties may have provided the opportunity for genetic divergence. These changes could have resulted from either drift or selection, but they did not alter the ability of the trees to survive in the climates currently available to both varieties. These are important questions which will require a careful investigation of the genetic and physiological basis of these traits.
Obligate mutualisms have offered valuable insights into how ecology shapes the evolution of biological diversity (Weiblen, 2002; Pellmyr, 2003). In particular, we now know that strong interspecific interactions over ecological time-scales are linked with macroevolutionary patterns of diversification. This has been interpreted as evidence that coevolution between mutualists creates and maintains biological diversity, but it has been difficult to determine whether the macroevolutionary pattern is caused by coevolution, or is a side-effect of other ecological or evolutionary forces. Recent work (Godsoe et al., 2008) indicates that the traits that show the greatest degree of divergence between the two varieties of Y.brevifolia are tightly associated with pollination, and match the morphology of their respective pollinator. The evidence presented here suggests that the abiotic environment does not explain the pattern of co-diversification we see in this obligate mutualism, which, together with the results of Godsoe et al. (2008), provides support for the hypothesis that co-divergence is the result of coevolution.
We thank the US National Park Service, US Fish and Wildlife Service, US Department of Energy, Joshua Tree National Park, Death Valley National Park, Mojave National Preserve, Desert National Wildlife Refuge and The Nevada Test Site for providing access to study sites. W. Kent Ostler helped arrange access to some populations and assisted with data collection. Paul DePrey, Rob Fulton, Ashley Cushman, Tasha LaDoux, Mark Hagan and David Reznick provided logistical support. We thank Gary Reese, Benjamin A. Waitman and Trent M. Draper for technical assistance. Jorge Soberon, Luke Harmon, Jeffrey Evans, Melanie Murphy, Niko Balkenhol, Chris Drummond, Ben Ridenhour and Raymond Dezzani provided valuable comments and suggestions. Jane Elith generously provided her R scripts, and advice on using boosted regression trees. Financial support was provided by National Science Foundation grant DEB-0516841, and the Cooperative Ecosystems Studies Units Network (CESU) of the US National Park System. Bioinformatics facilities were supported by NIH/NCRR P20RR16448 and P20RR016454. W.G. was supported by a pre-doctoral fellowship from the Canadian National Science and Engineering Research Council and the Foundation for Orchid Research and Conservation. J.B.Y. received support from the Joshua Tree National Park Graduate Student Research Program. Any use of trade, product, or firm names in this publication is for descriptive purposes only and does not imply endorsement by the US government.