Adaptive Species Differentiation and Population Uniformity in Viola Species Sharing Similar Geographical Distribution but Differing Habitat Preferences

Selection of favorable alleles is sufficient to maintain species cohesion even with low levels of gene flow. Here, in the study of two sister species growing in contrasting restricted ecological areas, we assess whether their divergent traits have been subject to uniform selection among populations. Two different analyses were performed to evaluate the relative importance of selection and drift. First, we compared FST and PST indexes, analogous to QST, in eleven populations each of Viola eizanensis and V. chaerophylloides. PST was computed for two reproductive traits, two leaf traits and two allocation traits. Second, we compared the observed PST with the expected distribution of PST under neutrality generated by phylogeny-based simulation assuming the Brownian motion model for trait evolution. These analyses indicated that spring leaf mass per area and seed number per capsule were divergent between species and uniform among populations. We suggest that these candidate traits are associated with ecological speciation. We also show that uniform, not divergent, selection among populations is probably common in both species. This result suggests that uniform selection maintaining similarity of traits among populations is a more dominant force than divergent selection in species growing in restricted ecological areas.


Introduction
The maintenance of species cohesion has been of continued interest in evolutionary biology.In the classic view, evolutionary biologists regard species as evolutionary units that are defined as groups of interbreeding individuals isolated from other such groups (Mayr, 1963).Ehrlich and Raven (1969) challenged this view by reviewing evidence of gene flow in animal and plant populations that were generally too restricted to maintain species integration.Since this seminal paper, many efforts have been made to estimate gene flow in the wild and evaluate its role in species cohesion (e.g.Ellstrand & Elam, 1993;Levin, 2000;Slatkin, 1985).By reviewing both recent estimates of gene flow and analyzing the strength of selection on phenotypic traits, Rieseberg and Burke (2001) and Morjan and Rieseberg (2004) reached two key conclusions.First, there are numerous species that lack sufficient gene flow to prevent divergence, although gene flow is often considerably greater than suspected from earlier studies.Second, most selection coefficients for leading QTLs associated with interspecific differences are high enough for advantageous alleles to spread across the range of a species.Based on this evidence, they argued that advantageous QTLs are likely to contribute to species cohesion, even with a very low level of gene flow.However, their calculations of selection coefficients for QTLs were crude, probably resulting in an overestimation of the strength of selection.Further studies are required to demonstrate how selection on QTLs can be effective to maintain species integration under low levels of gene flow.divergent natural selection favoring different phenotypes in different lineages is most likely to have been involved to achieve this differentiation.Second, if F ST ≈ Q ST , this implies that the observed degree of differentiation in quantitative traits can be explained by genetic drift alone.The third possible outcome is F ST > Q ST .This implies that observed degree of lineage differentiation is actually smaller than what would be expected under genetic drift alone.In other words, uniform natural selection favoring the fittest phenotype probably acted on the traits.
Most comparative studies of F ST and Q ST have shown that the degree of differentiation in quantitative traits (Q ST ) typically exceeds that observed in neutral marker genes (F ST ) (e.g.Leinonen, O'Hara, Cano, & Merila, 2008;McKay & Latta, 2002;Merilä & Crnokrak, 2001).These studies suggest a prominent role for natural selection in accounting for patterns of quantitative trait differentiation among contemporary populations.In several other studies, Q ST and F ST values did not differ significantly, suggesting that genetic drift alone was a sufficient explanation for the differentiation observed (e.g.Jorgensen, Richardson, & Andersson, 2006;Lascoux, Thorsen, & Gullberg, 1996;Widén, Andersson, Rao, & Widén, 2002).Finally, other studies in which F ST exceeded Q ST (e.g.Baruch, Nassar, & Bubis, 2004;Edmands & Harrison, 2003) suggested the footprints of uniform selection.Among this latter group particularly lie some plant species that have restricted ecological niches.Examples include Scabiosa canescens (Waldmann & Andersson, 1998), Liatris scariosa var.novae-angliae (Gravuer, von Wettberg, & Schmitt, 2005), Cedrela odorata (Navarro et al., 2005), Thlaspi caerulescens (Jimenez-Ambriz et al., 2007) and Brassica insularis and Centaurea corymbosa (Petit et al., 2001).Petit et al. (2001) suggested that restricted ecological niches could cause species to experience homogenous selection pressures, resulting in a lower Q ST .However, all of those studies compared Q ST and F ST within a single plant species.Here, we studied a closely-related species pair, Viola eizanensis (Makino) Makino that prefers more shaded habitats, and V. chaerophylloides (Regel) W. Becker var.sieboldiana (Maxim.),which prefers more open habitats.We tested the prediction that species differences apparently adaptive to habitat differences were subject to divergent selection between species and uniform selection among populations within each species.To test this prediction, we compared F ST and P ST indexes, analogous to Q ST .This approach would be error-prone because of the downward bias of P ST , but P ST based on measuring phenotypes grown in the common garden would be a suitable approach for finding candidate traits that are most likely under heterogeneous selection (Whitlock, 2008).
In addition to the analysis of F ST and P ST , as described above, we performed a phylogeny-based simulation to generate the distribution of P ST expected under neutrality.As the comparison between F ST and Q ST is assumed to be non-biased by maternal effects, environmental effects, migration rates, mutation rates and selection on marker genes (Edelaar & Bjorklund, 2011;Edelaar, Burraco, & Gomez-Mestre, 2011;Hendry, 2002;Merilä & Crnokrak, 2001;Whitlock, 2008), phylogeny-based analysis permits more accurate assessment of whether the neutral hypothesis can be rejected for particular traits.To detect non-Brownian traits, we applied the Brownian motion model for trait evolution following Freckleton and Harvey (2006).There are three possible outcomes from the comparison of observed P ST with its expected neutral distribution.First, if P ST exceeds the 95% confidence interval of its neutral distribution, divergent natural selection would be suggested.Second, if P ST is included within such a confidence interval, the genetic drift hypothesis is not rejected.Third, if P ST is smaller than the confidence interval, uniform selection would be suggested.
Viola eizanensis and V. chaerophylloides var.sieboldiana provide an ideal opportunity for revealing the role of selection in species cohesion and differentiation.First, their sister relationship is supported by an unpublished ITS tree of Viola (Inoue, personal communication).Second, they are endemic to Japan, both having similar geographical distributions and altitudinal ranges (Hama, 2002;Igari, 1996).Third, their genetic structure and phylogenetic relationships were determined in our previous study (Toyama & Yahara, 2009).Fourthly, they are similar in most other ecological traits being small perennial herbs that produce chasmogamous (CH) flowers in spring and cleistogamous (CL) flowers in summer and autumn, and propagate only by elaiosome-bearing seeds dispersed by ants.They do, however, differ in habitat preference, V. eizanensis prefers more shaded places in temperate forests, and V. chaerophylloides var.sieboldiana prefers more open areas in temperate grasslands (Hama, 2002;Igari, 1996).Reflecting this difference in habitat preference, apparent adaptive phenotypic differences between two species have been described (Toyama and Yahara, in prep.).For example, the dry weight of CH flowers relative to the above-ground plant weight was larger in V. chaerophylloides than in V. eizanensis, a difference that appears to be adaptive as pollinator visitations are more frequent in V. chaerophylloides.The two species also differ in the display size and dry mass of CH flowers, and weight and area of spring leaves, differences that appear to be adaptive to shaded or open habitats.To our knowledge, this is the first report to compare results of both an analysis of F ST and P ST and a phylogeny-based analysis, and also to examine the relative importance of natural selection and genetic drift to apparently adaptive traits by comparing sister species.
The present study aims to test our hypothesis that divergent traits between two sister species growing in contrasting restricted ecological areas have been subjected to uniform selection among populations in each species.We address the following questions: 1) What traits are divergent between species? 2) What traits are uniform among populations?3) Is uniform selection among populations common in species having a restricted ecological niche?4) Is there any difference between the two analyses?

Study Species and Populations
The study species, populations and individuals were the same ones used in our previous study (Toyama & Yahara, 2009, Supplementary Table and Supplementary Figure).In July 2004, we collected about five individuals from each of the 22 populations that cover the distribution range of the two taxa in Japan.We collected individuals growing at least three meters apart to avoid any potential vegetatively reproduced individuals (although neither of these species are known to do this under ordinary environmental conditions).Only five individuals were sampled per population, which is an insufficient sampling, but individuals from each population clustered together (Toyama & Yahara, 2009).From our preliminary analyses, we judged that the number of populations examined was more important than the number of individuals per population.This decision was supported by a simulation study by Goudet and Büchi (2006) showing that a better estimate of Q ST was obtained from many populations (> 20) with few families (5) rather than few populations with many families.Modified from Toyama and Yahara (2009).

Greenhouse Experiment
This experiment was carried out in the Kyushu University nursery.After sampling, plants were planted in pots (12 cm diameter) and grown in the 50% shaded greenhouse.Pots were randomized every three days during the study.Quantitative traits were measured after the plants were grown in this common environment for at least nine months (from July 2004 to late March 2005) and measurements were performed from late March 2005 to June 2005.
We measured eight traits.CH and CL flower numbers were counted every three days and each flower was marked with colored tape (1 mm × 2 mm) to identify new flowers.For measurement of display size of CH flowers (DSCH, cm 2 ), a picture of each flower was taken and processed together with a visual scale (1 cm 2 ), which enabled measurement to the nearest 0.01 cm 2 using the Scion Image software ver.4.0.2.The dry mass of CH flowers (DMCH, mg) was recorded using the first flowers produced, which were dried at 60°C and weighed with an electronic balance to the nearest 0.01 mg.Seed number per capsule of CL flowers (SN/CL) were counted for each flower.For weekly counting of leaf number, each leaf was marked with colored tape (1 mm × 2 mm) to enable the identification of new leaves.For the measurement of spring leaf area (SPLA, cm 2 ), a measurement to the nearest 0.01 cm was obtained using the same technique as with the display size of CH flowers.To calculate spring leaf mass per area (SPLM/A, mg/cm 2 ), a leaf with a known area was harvested from each individual plant, dried and weighed as previously described.We then calculated total investment in spring leaves (LEAF, mg) by multiplying SPLM/A with total SPLA for each individual.Total investment in CH flowers (W CH ) or CL flowers (W CL ) was calculated by multiplying the total flower number with the dried flower mass of each individual plant.The proportion of CH flower investment to the above-ground part of each plant (CH/AG) was calculated as ( ) and the proportion of CL flower investment to the above-ground part of each plant (CL/AG) was calculated as ( ) (2)

Quantitative Trait Analysis
Generally, there is a correlation among flower traits and a tradeoff among allocation traits.A principal component analysis was performed to identify uncorrelated components in the chasmogamous (DSCH and DMCH) and allocation traits (LEAF, CH/AG and CL/AG) using R v2.3.1 (R Development Core Team, 2006).We analyzed the variation of the first principal component for chasmogamous traits and the first and second principal components for allocation traits.
A variance component analysis with species, populations nested within a species and individuals nested within a population as random effects was performed for each trait using the restricted maximum likelihood (REML) method.The estimates of σ S 2 (genetic variance between species), σ P 2 (genetic variance among populations within species) and σ I 2 (genetic variance within populations) were used to quantify the level of species differentiation (P CT ) and population differentiation (P SC ) for each trait.P CT and P SC were calculated as follows, assuming the differences within and between populations or species were strictly genetic; The 95% confidence interval for P CT and P SC was estimated by 10 000 bootstrap replications using R v2.3.1 (R Development Core Team, 2006).
To test the difference between P CT (or P SC ) and neutral, we compared them with 95% confidence intervals for F CT (or F SC ) estimated by AFLP (Toyama & Yahara, 2009).The use of five AFLP primer combinations on 121 individuals amplified a total of 533 fragments.Replications proved that the AFLP data was reliable (repeatability 99.2 % in V. eizanensis, 99.8 % in V. chaerophylloides).Analysis of molecular variance (Excoffier, Smouse, & Quattro, 1992) was used to investigate the partitioning of genetic variation with the HIERFSTAT package in R (Goudet, 2005).The confidence interval for each F-statistic was estimated by bootstrap over loci with 10 000 repeated samplings of the data.All analyses were performed in R v2.3.1 (R Development Core Team, 2006).
To generate the expected distribution of P CT and P SC under neutrality, we applied the phylogeny-based simulation technique of Freckleton and Harvey (2006) using R v2.3.1 (R Development Core Team, 2006).In this simulation, we first calculated P CT and P SC using the original data.Traits that had two or more data sets from one individual were analyzed using mean phenotypic values.Second, common ancestral trait values of X k were reconstructed using a pair of adjacent tips (individual i and j) calculated as ( ) ( ) where v i and v j were the lengths of the branches leading to nodes i and j respectively.To account for statistical error (Felsenstein, 1985), the node below k was increased from v k to ( ) Third, the standardized contrasts were calculated independently by their position on the phylogeny as follows;   Fourth, the contrasts were randomized on the tree.Fifth, using contrasts and the above two formulae, a random value for the ancestral trait was computed at each node.The sign (positive or negative) of the second equation was assigned at random.This procedure was repeated until randomized values had been assigned to all nodes.Fifth, we calculated the simulated P CT and P SC .Finally, steps two to five above were repeated 10 000 times to estimate the empirical confidence interval of simulated P CT and P SC .
In the above simulation, we used a maximum parsimony (MP) tree instead of the neighbor joining (NJ) tree of Toyama and Yahara (2009) with the number of substitutions per branch as branch length.The MP tree showed an almost identical topology to the NJ tree.
This simulation is associated with calculation errors when there are polytomies on the tree.We made some assumptions about the unknown branching patterns of multiple nodes to permit analysis.Because V. eizanensis and V. chaerophylloides do not reproduce vegetatively under ordinary environmental conditions, it is unlikely that polytomies would be due to clonal individuals.Therefore, we set the virtual node to having only one branch length.We randomly selected a branching pattern from all possible topologies in each repeat simulation.

Results
Principal component analysis indicated that 77.1% of variation in CH flower traits could be explained by the first principal component (PC1).PC1 and PC2 could also explain 51.7% and 33.3%, respectively, of the variation in allocation traits (Table 1, Figure 1).In the allocation traits, PC1 separates plants with larger vegetative organs from plants with larger reproductive outputs, and PC2 discriminates plants with a larger allocation of CL flowers from plants with a larger allocation of CH flowers.Phenotypic variation in each quantitative trait is shown in Figure 2.     Estimates of P CT ranged from 3.28 × 10 -9 (PC2 for allocation traits) to 0.631 (SN/CL) (Figure 3a).By comparing the 95% CI of P CT with that of F CT , it was shown that P CT was significantly higher in two traits: SN/CL and SPLM/A, and significantly lower in four traits: PC1 for CH flower traits, SPLA, and PC1 and PC2 for allocation traits.Estimates of P SC ranged from 0.121 (SPLM/A) to 0.704 (SPLA) (Figure 3b).By comparing the 95% CI of P SC with that of F SC , it was shown that P SC was significantly lower in three traits; SN/CL, SPLM/A and PC2 for allocation traits.
The 95% confidence intervals of P CT expected under neutrality are shown in Figure 4a.A significantly higher P CT value than the neutral expectation was observed in SN/CL and SPLM/A, and a significantly lower P CT value was observed in PC2 for allocation traits.The 95% confidence intervals of P SC expected under neutrality are shown in Figure 4b.A significantly lower P SC value than the neutral expectation was observed in SPLM/A and PC2 for allocation traits.

Adaptive Divergence between Species and Phenotypic Uniformity among Populations
Ecological speciation is defined as the evolution of reproductive isolation between populations by divergent natural selection arising from differences between ecological environments (Coyne & Orr, 2004;Schluter, 2001).
In this study, we were able to identify candidate traits associated with ecological speciation.Viola eizanensis and V. chaerophylloides are separated by their differing light preference, and comparative analysis between F ST and P ST values and phylogeny-based simulation analysis supported the observation that two of six traits (SN/CL and SPLM/A) were subject to divergent selection between species and uniform selection among populations (Figures 3, 4).
Spring leaf dry mass per area values (SPLM/A) were larger for V. eizanensis than V. chaerophylloides (Figure 1d).This interspecific differentiation may be related to a difference in leaf longevity, as leaves having larger mass per area are expected to have greater longevity and are seen as adaptive for shaded habitats.This positive correlation between leaf mass per area and leaf longevity has been reported in many studies (e.g.Reich, Walters, & Ellsworth, 1992, 1997;Reich et al., 2003; I. J. Wright et al., 2005).Generally, in high light conditions, photoinhibition decreases plant productivity so that natural selection favors the production of new leaves.
Conversely, in dark conditions the photosynthesis rate is low but stable so that natural selection favors the maintenance of leaves.This correlation between light conditions and leaf longevity has been previously reported (e.g.Sterck, 1999).However, further studies are needed to test whether the two species differ in leaf longevity as we predict.
For the number of CL seeds per capsule (SN/CL), V. chaerophylloides tends to have more seeds than V. eizanensis (Figure 2b).This interspecific differentiation may be related to the trade-off between seed number and size.In general, species producing many small seeds are considered to be superior colonizers, whereas species producing fewer large seeds are considered to be superior competitors (Coomes & Grubb, 2003).V.
chaerophylloides grows in open and more disturbed grassland habitats and has more CL seeds over the growing season, a characteristic that may be advantageous for seed dispersal into newly available open habitats.On the other hand, V. eizanensis, grows in closed and more stable woodland habitats and produces fewer CL seeds, which are potentially better able to survive in shaded environments with correspondingly fewer resources, and withstand unfavorable conditions for seed germination and seedling growth.Such association between light conditions and seed number/size has been observed in some plant species (Foster & Janson, 1985;Grubb & Metcalfe, 1996;Hewitt, 1998).
Phenotypic plasticity in the above two traits could affect their ecological speciation.Recently, an individual-based simulation study showed that phenotypic plasticity could either enhance or degrade the process of ecological speciation depending on the timing (Thibert-Plante & Hendry, 2011).If the evolution of plasticity occurs before dispersal to different environments, it can increase the progress of ecological speciation.If plasticity occurs after dispersal, progress can be slowed because migrants are better suited to their new environments.Considering our findings, the SPLM/A and SN/CL traits, subject to divergent selection between species and uniform selection among populations, might represent evolved plasticity before the speciation process.There is no evidence to support this hypothesis, but it might be possible to confirm by examining the plastic response of the two species and/or common ancestral species in different light conditions.In summary, the two species we studied differ in two functional traits, spring leaf dry mass per area and the number of CL seeds per capsule, and these differences may have evolved as a result of their individual light requirements.Further studies are needed to reveal how these differences play a role in their respective speciation processes.
In contrast to the above findings, the other four traits were subject to uniform selection or genetic drift (Figure 3, 4).Particularly, PC2 for allocation traits was subject to uniform selection within and between species in both analyses (Figures 3,4).This result suggests that a common plasticity mechanism has evolved for the allocation ratio of CH and CL flowers in the ancestor of both species.Theoretically, the reproductive system of CH and CL flowers evolved as a result of adaptation to variation in biotic and abiotic environments (Schoen & Lloyd, 1984).Many previous studies have shown that the proportions of CH and CL flower production are variable in response to light availability (Cheplick, 2006;Lecorff, 1993;Masuda & Yahara, 1994;Masuda, Yahara, & Maki, 2004;Mattila & Salonen, 1995;Schemske, 1978;Wilken, 1982), soil nutrient availability (Lecorff, 1993), soil moisture (Brown, 1952) and plant density (Wilken, 1982).However, further studies are required to elucidate whether the ratio of CH to CL flower investment is regulated by the same mechanism in many other cleistogamous species.For the other three traits (CHPC1, SPLA, AlloPC1), we couldn't detect a strong selection within or between species, although the comparative analysis between F ST and P ST was significant (Figures 3 and  4).These results suggest that genetic drift might have played the primary role in driving phenotypic differentiation.

Low Values of P SC in the Restricted Ecological Niche of Species
The comparative analysis between F SC and P SC values and the phylogeny-based simulation showed that in both species P SC was often lower than F SC and its neutral distribution (Figures 3b,4b).In no cases was the observed P SC higher than the observed F SC or expected P SC under neutrality.These results imply that observed phenotypic differentiation among populations of the two species can be explained by random genetic drift or uniform selection rather than by divergent selection.This finding contrasts with previous reports where Q ST was larger than F ST (Leinonen et al., 2008;McKay & Latta, 2002;Merilä & Crnokrak, 2001).
This contrast may be due to the restricted ecological niches of the two species; adaptation to woodland and grassland habitats may be associated with homogenous selection pressures for phenotypic variation.Our lower P ST result is consistent with findings in other ecologically restricted plant species (Grubb & Metcalfe, 1996) such as Scabiosa canescens (Waldmann & Andersson, 1998), Liatris scariosa var.novae-angliae (Gravuer et al., 2005), Cedrela odorata (Navarro et al., 2005), Thlaspi caerulescens (Jimenez-Ambriz et al., 2007) and Brassica insularis and Centaurea corymbosa (Petit et al., 2001).Thus, we suggest that the niche requirement of restricted species could act as uniform selection among populations.

Comparison between P ST and Phylogeny-based Analyses
We applied two different methods to elucidate the relative importance of selection and drift.In all cases, phylogeny-based analysis was more conservative than P ST analysis (Figures 3, 4).In Figures 3a and 4a, uniform selection was suggested for four traits by P CT analysis, but only one trait in the phylogeny-based analysis.
Similarly, in Figures 3b and 4b, uniform selection was suggested for three traits by P SC analysis, but only two traits in the phylogeny-based analysis.There are three possibilities to explain this discrepancy.First, in comparative analysis between F ST and P ST , underestimated P ST values could more easily result in apparent support for uniform selection than in phylogeny-based analysis.This study was not performed with progeny of experimental crosses, but with plants sampled from natural populations.In this case, non-additive genetic effects (e.g.dominance, maternal) could cause overestimation of genetic variance within populations and underestimation of Q ST (Merilä & Crnokrak, 2001;Whitlock, 2008).However, dominance effects might be small because of the predominance of inbreeding in these two species (Goudet & Buchi, 2006).Second, linkage disequilibrium due to being highly selfing may also cause overestimation of F ST .Our study assumed that the AFLP molecular markers are selectively neutral and unlinked to loci under selection.However, in partially selfing organisms, variation at neutral loci can be greatly influenced by selection acting on linked and even unlinked loci via genetic hitchhiking (Hedrick, 1980).Charlesworth et al. (1997) also demonstrated that, for subdivided populations, heterogeneous selection enhances F ST values even at neutral loci distant to the selected locus.The AFLP mutation rate could also bias F ST estimation.To our knowledge, there is no general mutational model for AFLP markers, but Edelaar et al. (2011) demonstrated that estimates of the differences between Q ST and F ST was biased depending on the mutation rate of the genetic markers used.We should await further studies on this topic before drawing conclusions.Thus, we must be careful in interpreting the results of comparative analysis between F ST and P ST .Third, a conservative policy in phylogeny-based analysis could also cause this discrepancy.As the distribution of P ST expected under neutrality was performed under maximum parsimony criteria, it might be more difficult to detect uniform selection in the comparison with observed P ST .Thus, there are some limitations in this study, but in areas where the results in two analyses are consistent sufficient data is presented for some initial conclusions.

Conclusion
Traits divergent between species and unified among populations are considered to be of particular importance in understanding the speciation process.In the present study, we detected adaptation in two traits in response to light preferences.In V. eizanensis, having larger leaf mass per area is expected to lead to greater leaf longevity and is a trait considered adaptive to shaded habitats.Likewise, the production of fewer seeds is expected to result in a higher seedling survival rate under restricted light conditions.Conversely, in V. chaerophylloides, smaller leaf mass per area is expected to ensure effective photosynthesis under higher light conditions, and having more CL seeds may be advantageous for seed dispersal into newly available open habitats.We have also demonstrated that the ratio of CH to CL flower investment is uniform both within and between species.To our knowledge, this is the first report to reveal the selection of allocation ratio between CH and CL flowers.Additionally, we have shown that low P SC values are consistent with previous studies performed on species having a restricted ecological niche (Gravuer et al., 2005;Jimenez-Ambriz et al., 2007;Navarro et al., 2005;Petit et al., 2001;Waldmann & Andersson, 1998).Therefore, we suggest that uniform selection is a more dominant force in these species maintaining uniformity of traits than has previously been suggested in recent reviews.Our finding supports the view that low gene flow species are held together primarily by the spread of advantageous alleles (Rieseberg aud Burke 2001;Morjan and Rieseberg 2004).However, our calculations of P ST based on measuring phenotypes grown in a common garden was biased downwards (Whitlock, 2008); therefore our analysis should be considered a first screen.To further test this view, mapping QTLs for the two traits divergent between species and unified among populations is a potential approach (e.g.Raeymaekers, Van Houdt, Larmuseau, Geldof, & Volckaert, 2007).V. eizanensis, and V. chaerophylloides, are inter-fertile and our on-going work with F2 hybrids of the two species will hopefully lead to a better understanding of the role of natural selection upon cohesion and differentiation of these species.
Supplementary Figure.Geographical locations of the 22 populations used in this study Filled circles represent V. eizanensis (A-K), and open circles represent V. chaerophylloides var.sieboldiana (L-V).Letters refer to population symbols in Supplementary Table.

Figure 1 .
Figure 1.Principal component analysis of CH flower and allocation traits

Figure 2 .
Figure 2. Phenotypic variation of six traits in V. eizanensis (populations A-K) and V. chaerophylloides (populations L-V) The mean ± S.D. of (a) PC1 for CH flower traits, (b) seed number per capsule, (c) spring leaf area, d) spring leaf mass per area, (e) PC1 and (f) PC2 for allocation traits is shown.Letters indicate population symbols referred to in Toyama and Yahara (2009) and Figure S1.

Figure 3 .
Figure 3. Confidence intervals for (a) F CT and P CT , (b) F SC and P SC Each dot represents the observed P CT or P SC value and each bar represents the 95% confidence interval.The upper and lower horizontal dotted lines indicate the 95% confidence intervals of F CT and F SC .Asterisks indicate that the 95% CIs do not overlap.

Figure 4 .
Figure 4. Comparison between (a) observed P CT and the distribution of P CT expected under neutrality and (b) observed P SC and expected P SC Each dot represents an observed value and each bar represents the 95% of confidence interval of its neutral distribution.Asterisks indicate that the observed value does not overlap with the 95% CI for neutrality.
Table.Sampled population codes and locations of two Viola species

Table 1 .
Contribution of CH flower and allocation traits to the principal components of phenotypic variation