Mating systems are among the most labile characteristics of flowering plants, with transitions frequently occurring among populations or in association with speciation. The frequency of mating system shifts has made it difficult to reconstruct historical evolutionary dynamics unless transitions have been very recent. Here, we examine molecular and phenotypic variation to determine the polarity, timescale, and causes of a transition between outcrossing and self-fertilization in sister subspecies of Clarkia xantiana. Phylogenetic analyses and coalescent-based estimates of the time to most recent common ancestor indicated that outcrossing is ancestral to selfing and that there has been a single origin of selfing. Estimates of divergence time between outcrossing and selfing subspecies were 10,000 (95% CI [credible interval]: 3169–66,889) and 65,000 years ago (95% CI: 33,035–151,448) based on two different methods, suggesting a recent and rapid evolutionary transition. Population genetic data indicated that the transition to selfing was associated with a 80% reduction in molecular diversity, which is much greater than the 50% reduction expected under a shift from obligate outcrossing to obligate self-fertilization alone. Our data also suggest that this severe loss of diversity was caused by colonization bottlenecks. Together with previous studies, evidence for reproductive assurance in C. xantiana now connects variation in plant–pollinator interactions in the field to phenotypic and molecular evolution.

Transitions between outcrossing and self-fertilization have occurred repeatedly throughout the evolutionary history of angiosperms (Stebbins 1974; Holsinger 2000). Selfing has arisen in some cases through the loss of a self-incompatibility system (reviewed in Igic et al. 2008) but frequently occurs when self-compatible populations transition between outcrossing and autonomous selfing as primary forms of reproduction (Richards 1999). Although phylogenetic studies have revealed important insights into the macroevolution of mating systems (e.g., Schoen et al. 1997; Goodwillie 1999; Barrett 2002; Igic et al. 2006; Foxe et al. 2009), details of the historical dynamics of transitions have been challenging to reconstruct because mating systems are highly labile. The evolutionary history of conspecific populations and recently diverged taxa that differ in mating system has been of particular interest for understanding the polarity of transitions, the timing of divergence, and the mechanisms responsible for mating system shifts (e.g., Foxe et al. 2010; Ness et al. 2010).

The direction of evolutionary change has typically been assumed to occur from outcrossing to selfing. However, it has been debated as to whether mating system transitions are unidirectional and whether selfing is an evolutionary dead end (Stebbins 1957; Takebayashi and Morrell 2001; Escobar et al. 2010). In higher level phylogenies, it is difficult to infer whether and how often a character state is gained or lost, particularly when that character evolves rapidly and when diversification and extinction rates of lineages are associated with character states (Goldberg and Igic 2008). In the past, genomic and computational resources limited the feasibility of studies investigating the historical dynamics of mating system transitions at the population level or among closely related species. However, given the increasing ease of obtaining sequence data and the development of divergence population genetic approaches based on the coalescent (Hey and Nielsen 2004; Drummond and Rambaut 2007), it is now feasible to examine the recent evolutionary history of shifts in reproductive strategies.

Two general models have been proposed to explain the evolution of self-fertilization (mating system models reviewed in Uyenoyama et al. 1993; Goodwillie et al. 2005). One model suggests that selfing can evolve due to an inherent 50% transmission advantage of selfers over obligate outcrossers (Fisher 1941; Jain 1976). This automatic transmission model rests on the condition that selfing individuals can contribute genes to the next generation not only by selfing but also by siring offspring on other individuals, resulting in a fitness advantage over obligately outcrossing individuals. A second model of mating system evolution suggests that selfing evolves as a form of reproductive assurance in environments where outcrossing is unlikely to occur because pollinators or mates are unreliable (Stebbins 1957; Lloyd 1979, 1992; Cheptou 2004; Morgan and Wilson 2005). This model has received increasing support from field studies showing that selfing elevates seed production (Herlihy and Eckert 2002; Elle and Carney 2003; Kalisz et al. 2004; Busch 2005) and that the selfing phenotype is favored by natural selection when limited mates and effective pollinators cause pollen limitation (Moeller and Geber 2005; Moeller 2006). It is important to note that these two models are not necessarily mutually exclusive; however, they provide substantially different predictions about the process of mating system evolution.

Although field studies provide insights into the adaptive significance of alternative mating strategies, they are less successful in identifying the original causes of mating system transitions. It has been proposed that the two different causes of the evolution of selfing leave different signatures at the molecular level and that population genetic analyses can provide information regarding the history and mode of mating system divergence (Schoen et al. 1996). Under both models, selection on mating system modifier loci (and hitchhiking effects on linked loci) should cause losses of molecular variation. In addition, high rates of selfing should cause as much as a 50% reduction in neutral diversity across the genome (Charlesworth et al. 1993; Nordborg 2000); for transitions from mixed mating (partial selfers) to higher rates of selfing, the expected loss of neutral variation should be less than 50%. The key difference between models is that an advantage of selfing under reproductive assurance is the potential for populations to colonize new sites void of pollinators or mates (Baker 1955, 1967; Pannell and Barrett 1998). Frequent population bottlenecks associated with colonization (i.e., founder effects) under the reproductive assurance model are expected to result in strong genome-wide reductions in neutral diversity, greater than the 50% reduction in molecular diversity that is expected for a shift from obligate outcrossing to obligate self-fertilization alone (Schoen et al. 1996). Population bottlenecks are also expected to result in high variance in the frequency distribution of polymorphisms among loci due to random genetic drift (Charlesworth et al. 2003; Wright and Gaut 2005).

Under the automatic transmission model, by contrast, the advantage of selfing is dependent on the availability of mates (i.e., relatively large population sizes) and pollen vectors for outcrossing (Jain 1976). Frequent population bottlenecks are not predicted to accompany the evolution of selfing via automatic selection and therefore populations should maintain higher levels of neutral genetic variation than under reproductive assurance (Schoen et al. 1996). One problem for differentiating between models is that the transition from outcrossing to selfing must have occurred recently. When mating system transitions have occurred in the distant past, it is difficult to distinguish whether extant patterns of genetic diversity were affected by mating system divergence versus other demographic and evolutionary phenomena that may have occurred after the speciation process (i.e., complete reproductive isolation and genealogical exclusivity). Therefore, examining the signature of mating system evolution using molecular population genetic data is most powerful when taxa have only recently begun to diverge (Schoen et al. 1996).

In this study, we examine the evolutionary history of mating system evolution in Clarkia xantiana A. Gray (Onagraceae), which consists of a predominantly selfing subspecies, parviflora, and a mixed mating but primarily outcrossing subspecies, xantiana. The geographic distribution, pollination ecology, and floral biology of this species are well documented (Eckhart and Geber 1999; Runions and Geber 2000; Moeller 2006; Eckhart et al. 2011), which has allowed us to comprehensively sample populations and floral variation across the species’ range. Both taxa are self-compatible and preliminary studies indicate that they are cross-compatible to some extent, producing viable and fertile F1 hybrids in controlled environments (D.A. Moeller, unpubl. data). The two subspecies can be distinguished primarily by their flowers and flowering time (Eckhart and Geber 1999). Where the subspecies are sympatric and co-occur, they are distinct and hybrids are not commonly observed.

Compared to other investigations of mating system evolution using molecular population genetics, this system is unusual in that a series of field experiments have supported the reproductive assurance model. Specifically, selfing provides reproductive assurance and the selfing phenotype is favored by natural selection in environments where few mates and effective pollinators limit opportunities for outcrossing (Fausto et al. 2001; Geber and Eckhart 2005; Moeller and Geber 2005). The pattern of mating system divergence among populations is also correlated with the distribution and abundance of specialized pollinators (Moeller 2006). In this article, we examine the evolutionary history of outcrossing and selfing taxa of C. xantiana using DNA sequence data (eight nuclear and three chloroplast loci) and microsatellite variation (four loci) collected from 154 individuals representing 31 geographic populations. We used phylogenetic- and coalescentbased methods to determine the number of transitions between mating strategies, the ancestral mating system, and the time of divergence. We then used taxon-wide and population-specific patterns of sequence polymorphism to test theoretical predictions for patterns of molecular variation under the reproductive assurance and automatic transmission models of mating system evolution.



Clarkia xantiana A. Gray ssp. xantiana and C. xantiana A. Gray ssp. parviflora (Eastw.) Harlan Lewis & P. H. Raven are winter annuals that are endemic to the southern Sierra Nevada foothills of Kern and Tulare Counties, California. Their distributions extend south through the Tehachapi Mountains to the Transverse Ranges (Liebre and San Gabriel Mountains), where populations are less common (Fig. 1). The subspecies are parapatric with a narrow zone of sympatry (∼5 km) at xantiana's eastern and parviflora's western range margin (Eckhart and Geber 1999; Eckhart et al. 2010). In the zone of sympatry, subspecies often occur within meters to tens of meters of one another in the same site. Their distributions occupy different sections of a dominantly west-to-east environmental gradient where the eastern section, occupied by parviflora, receives less and more variable precipitation compared to the western section, occupied by xantiana (Eckhart et al. 2010, 2011).

Figure 1.

Map showing the location of the 16 C. xantiana ssp. parviflora and 15 C. xantiana ssp. xantiana populations sampled in this study. The numbers that label each population correspond to those in Table S1. Color shading shows elevations in 500-m intervals from deepest canyons (pale yellow, 0–500 m) to highest peaks (white, >2500 m).

In xantiana, outcrossing is effected mainly by solitary bees, some of which are specialized on the genus Clarkia (Moeller 2005; Eckhart et al. 2006; Moeller 2006). Outcrossing is also promoted by physical separation of the anthers and receptive stigma (herkogamy) and the temporal precedence of anther dehiscence over stigma receptivity (protandry). Floral manipulation experiments in the field have demonstrated that autonomous selfing does not contribute significantly to seed production in xantiana (Moeller et al. 2012). However, geitonogamous selfing (between flowers on the same individual) and biparental inbreeding are likely common due to multiflowered inflorescences and limited dispersal. In parviflora populations, visits by pollinators are uncommon and the specialist solitary bee pollinators of Clarkia are not present (Moeller 2006). Herkogamy and protandry are strongly reduced and autonomous selfing occurs readily just prior to, coincident with, or shortly after flower opening (Eckhart and Geber 1999; Runions and Geber 2000; Moeller 2006). Rates of autonomous self-fertilization in pollinator-free greenhouse environments typically exceed 90% for parviflora and less than 10% for xantiana (Moeller 2006). Within-population studies indicate outcrossing rates range from 0.62—0.73 (95% CI [credible interval]: 0.58–0.76) for six xantiana populations and 0.12 and 0.13 (95% CI: 0.10–0.15) for two parviflora populations (Moeller et al. 2012).


We sampled 31 populations (15 xantiana, 16 parviflora) spanning the species’ distribution except for a section of the Tehachapi Mountains that is privately owned and inaccessible (Fig. 1, Table S1). Populations of the two subspecies that share the same number as a label (e.g., 5x and 5p) co-occur in the narrow zone of sympatry (Fig. 1, Table S1). We planted one seed from each of five haphazardly chosen field-collected maternal families from each population (four individuals from 68p). Plants were initially grown in environmental chambers, where leaves were collected for DNA extraction, and subsequently grown in a greenhouse, where floral traits were measured. Therefore, molecular data and floral phenotypes were ascertained from the same individuals. Floral traits were measured on the first two flowers that opened on each plant; we used the average of these two measurements in our analyses (see also Moeller 2006). We measured petal length, petal width, herkogamy, and dichogamy. Herkogamy was measured as the distance from the receptive stigma to the nearest anther and was measured just as the stigma became receptive. Dichogamy was measured as the amount of time between long-anther dehiscence and stigma receptivity, which was determined by repeatedly observing flowers throughout floral development. We estimated individual autofertility (autonomous selfing rates) by determining fruit set across all flowers on each plant (mean of 35.5 flowers/plant). Plants were spaced widely and not disturbed during flowering to prevent contamination.


DNA was extracted using Qiagen DNeasy plant mini kits (Valencia, CA). We used PCR to amplify eight nuclear genomic regions (nDNA) and three chloroplast regions (cpDNA) from the 154 individuals representing the 31 populations. The PCR primers used to amplify single-copy nDNA loci were designed from EST sequences isolated from C. breweri flower buds (see Moeller et al. 2011). The three cpDNA (psbA-trnH, Kress and Erickson 2007; trnT-trnL; and trnL-trnF, Taberlet et al. 1991) regions were chosen because they are among the most variable regions of the chloroplast (Shaw et al. 2005).

PCR products were sequenced directly except for g3pdh and a subset of other loci. For these exceptions, PCR products were cloned into pGem-T Easy vectors (Promega Corp., Madison, WI) and one to five cloned products were sequenced. Sanger sequencing was conducted using the ABI BigDye version 3.1 (Applied Biosystems, Foster City, CA) chemistry in conjunction with an ABI 3730xl DNA sequencer. All chromatograms were manually inspected using Sequencher version 4.8 (Gene Codes Corp., Ann Arbor, MI) to determine polymorphic sites, which were coded according to IUPAC ambiguity codes, and assemble consensus sequences from bidirectional sequence reads. We used the program Phase version 2.1 (Stephens et al. 2001; Stephens and Donnelly 2003) to infer the haplotypes treating cloned fragments as known haplotypes. To avoid biasing results due to nucleotide misincorporation into cloned products, rare polymorphisms detected from cloned loci were confirmed by directly sequencing PCR products. Alignments were inferred using the program Muscle (Edgar 2004) with the default settings and were minimally edited to reduce erroneous homology statements. The total length of reliable sequence was used for the phylogenetic analyses (see Table 1 for GenBank accession numbers).

Table 1.  Summary statistics for 11 sequenced loci: number of sequences obtained (N), length of fragment used in the phylogenetic analyses (Lp), length of fragment for phased coalescent methods (Lc), segregating sites (S), haplotype diversity (Hd), population structure (Fst), genealogical sorting index (gsiT), average posterior probability of phased haplotypes (Phase), nucleotide substitution model (NST), and GenBank accession numbers.
LocusNxantiana Fstparviflora Fstxantiana gsiTparviflora gsiTPhaseNSTGenbank accessions
  1. 1Violated the assumptions of a molecular clock.

  2. 2Sums are provided for all metrics except for gsiT values, which are estimates based on 100 bootstrap replicates of the concatenated dataset, and the phase posterior probabilities, which represent the mean.

  3. 3Sequences were <200 bp and could not be deposited in GenBank. They are available upon request from the corresponding author.

  4. *Significant at P < 0.05

 a23  79 19 60 473 473 250.880.420.70.05*0.30*0.86HKY + IJQ009828−JQ009844, JQ010748−JQ010836
 d13 100 30 70 839 334 210.430.430.760.06*0.25*0.98HKY + G + IJQ009731−JQ009827
 d5 143 65 78 392 391 370.940.150.590.11*0.16*0.89GTR + G + IJQ009584−JQ009730
 f 9 152 73 74 543 542 800.880.330.440.06*0.40*0.93HKY + GJQ010595−JQ010747
 g3pdh1 138 61 72 504 5031000.910.380.540.040.45*0.9GTR + G + IJQ009845−JQ009982
 i11 154 75 79 424 424 480.840.140.460.21*0.23*0.83HKY + GJQ010441−JQ010594
 ipi2 145 70 75 914 913 250.690.150.830.13*0.010.9GTR + IJQ010291−JQ010440
 k221 153 75 78 649 648 830.950.190.740.19*0.34*0.9GTR + G + IJQ010137−JQ010290
 psbA− trnH 14566 79 228  67 120.040.480.730.010.2*K81uf + IJQ009448−JQ009583
 trnT− trnL 12053 67 122  41  10.450.130.13K81uf3
 trnL− trnF 15475 79 416 415  40.340.140.12HKYJQ009983−JQ010136

Genomic DNA from the same 154 individuals was used to genotype four dinucleotide repeat microsatellite loci. Primer sequences and amplification conditions can be found in Moeller et al. (2011). PCR was conducted separately for each locus with the 6-FAM and NED dyes; amplified products were combined for fragment separation on an ABI 3730xl using LIZ-500 as a size standard. A subset of individuals was re-run to confirm alleles. All fragment sizes were determined by directly examining each electropherogram.


We estimated the phylogenetic relationships among individuals using both the maximum-likelihood approach of Garli version 0.951 (Genetic Algorithm for Rapid Likelihood Inference; Zwickl 2006) and the Bayesian method implemented in MrBayes version 3.1 (Huelsenbeck and Ronquist 2001; Ronquist and Huelsenbeck 2003). To evaluate the range of topologies and associated likelihood scores across independent runs, we conducted 1000 analyses of the original nonbootstrapped dataset with Garli using the default parameter settings. Statistical support for topological relationships based on the maximum-likelihood method was assessed through 1000 bootstrap replicates. All analyses with Garli were performed using the computing resources associated with The Lattice Project (Bazinet and Cummings 2008). We used the default settings within MrBayes except for the following: GTR +Γ+ I nucleotide substitution model, a heating parameter of 0.03, and 0.01 as the mean of the prior on branch length. Bayesian Markov chain Monte Carlo (MCMC) posterior probabilities were estimated after removing 25% of the 4.1 × 106 generations as burnin. To provide polarity among nucleotide states and infer ancestral relationships, we used six congeneric outgroup taxa (C. concinna, C. davyi, C. arcuata, two accessions of C. amoena ssp. huntiana, and C. amoena ssp. caurina).

Given that the phylogenetic analyses cannot reconstruct reticulate relationships that result from recombination or hybridization, we also estimated relationships using the phylogenetic network method NeighborNet within SplitsTree4 (Huson and Bryant 2006). Outgroups were excluded from the analyses and the network was constructed based on Kimura-2-parameter distances (Kimura 1980). Support for edges were estimated from 1000 bootstrap replicates.

To quantify the evolutionary distinctiveness of the two taxa based on the degree of genealogical exclusivity, we calculated the ensemble genealogical sorting index (gsiT; Cummings et al. 2008) using 100 bootstrap replicates from the analyses with Garli. The gsiT is a normalized index under which 0 represents a random arrangement of conspecifics on the tree and 1 represents monophyly. Significance of the gsiT index was assessed through 5000 permutations. We also estimated the degree of incomplete lineage sorting among the loci by calculating the gsiT index from 100 bootstrap replicates of Garli analyses run for each gene separately.

We used the program InStruct (Gao et al. 2007) to further examine the distinctiveness of the subspecies and investigate patterns of introgression. InStruct groups individuals into subpopulations and accounts for inbreeding by relaxing the assumption of Hardy-Weinberg equilibrium within subpopulations (Gao et al. 2007). Using both the phased haplotype and microsatellite data, we ran two independent MCMC chains of 2.0 × 106 generations with 1.0 × 106 generations as burnin under a model allowing for admixture. The number of populations (i.e., k) was fixed at two. Convergence between the chains was assessed based on the Gelman–Rubin statistic and we present the results from the chain with the lowest Deviance Information Criterion.


We used the *Beast method within Beast version 1.6.1 (Drummond and Rambaut 2007) to estimate the time to most recent common ancestor (TMRCA) of each taxon and the time of divergence between the two taxa (i.e., the time after which there has not been significant gene flow; Heled and Drummond 2010). We defined two taxon sets corresponding to the two subspecies but did not enforce monophyly. Outgroups were not included in the analysis as the multispecies coalescent can be used to infer root position (Heled and Drummond 2010). To avoid biased results due to recombination (Strasburg and Rieseberg 2009), we used the largest nonrecombining segment within each nDNA locus as identified by the program IMgc (Woerner et al. 2007). Loci were partitioned separately and assigned the best-fitting nucleotide substitution model based on the results from ModelTest (Posada and Crandall 1998) and using Akaike's Information Criterion (Akaike 1974). Loci that violated the assumption of a molecular clock, based on analyses run using Paup* (Swofford 2003), were assigned a relaxed lognormal molecular clock. Trees were unlinked, except for the nonrecombining cpDNA, such that parameter estimates are based on integrating across loci. Using a Yule model of speciation, analyses were run for 1.0 × 108 generations removing the initial 1.0 × 107 as burnin. We present the mode of time since divergence and the average across all loci of the mode of TMRCA for each taxon. Sufficient sampling of parameter space throughout the MCMC sampler was evaluated using Tracer version 1.5 (Rambaut and Drummond 2009) and replicate runs were conducted to ensure accuracy of results.

We also estimated time of divergence using the isolation-with-migration model implemented in IMa2 (Hey and Nielsen 2007). Unlike *Beast, IMa2 allows for gene flow during taxon divergence and does not assume a Yule model. We ran a full model treating each taxon as a separate group. As with the *Beast analyses, only the nDNA loci were treated as independent and the largest nonrecombining block was used. The DNA sequences were assigned the HKY nucleotide substitution model (Hasegawa et al. 1985) and the appropriate inheritance scalar (i.e., 1 for autosomal nDNA and 0.5 for cpDNA loci). Analyses with IMa2 also included the microsatellite loci, which were assigned the stepwise mutation model and an inheritance scalar of 1. We ran 20 chains under a geometric heating scheme with appropriate priors for the migration, effective population sizes, and divergence time parameters that were identified through preliminary exploratory analyses. Runs consisted of 3.0 × 105 steps as burnin followed by 3.0 × 106 steps. Marginal histograms were evaluated and compared to ensure consistent parameter estimates.

To convert parameter estimates from *Beast and IMa2 to demographic timescales (e.g., years instead of mutations) we used the following mutation rates: 1.5 × 10−8 (Koch et al. 2001) and 0.9 × 10−9 (Berry et al. 2004) substitutions per site per year for the nDNA and cpDNA loci, respectively, and 0.00024 mutations per locus per generation for microsatellites (Thuillet et al. 2002). To convert the microsatellite mutation rate from a per generation to per year timescale, we used a generation time of 1.77 years based on data from a demographic study of 20 xantiana populations, which included study of seed germination and dormancy (Eckhart et al. 2011). Generation time was inferred from matrix population models using the method of Cochran and Ellner (1992).


For nDNA, we identified synonymous and nonsynonymous sites by aligning each haplotype to a reference coding sequence from the EST library or one identified using Blast (Altschul et al. 1990). Using the program Sites (Hey and Wakeley 1997), we calculated the average number of segregating nucleotides per site, θw (Watterson 1975), the degree of linkage disequilibrium, r2, the population recombination rate, ρ, and Tajima's D (Tajima 1989). If populations of the selfing taxon have been affected commonly by population bottlenecks, as predicted under the reproductive assurance hypothesis, then we expect elevated variance in Tajima's D relative to neutral equilibrium expectations (Wright and Gaut 2005). Because geographic sampling strategy can affect inferences about the site frequency spectrum (Ptak and Przeworski 2002; De and Durrett 2007), we calculted θw and Tajima's D for each population separately and at the subspecies level. We were unable to estimate r2 and ρ for many local population samples due to small sample sizes and low levels of polymorphism. Therefore, we only present results based on estimates calculated at the subspecies level. We used DnaSP version 5.0 (Librado and Rozas 2009) to estimate the degree of population structure, Fst, within each subspecies (Hudson et al. 1992).

We used a two-tailed t-test to test for significant differences between means of each statistic. We also tested whether the variance among estimates of Tajima's D in each subspecies was elevated relative to expectations under mutation-drift equilibrium using the program HKA (Hey and Nielsen 2004). We excluded cpDNA from these analyses because they contained very low levels of polymorphism.

For microsatellite loci, we used the program Genodive version 2.0b21(Meirmans and Van Tienderen 2004) to calculate the number of alleles, allele size ranges within each subspecies, observed heterozygosity, Ho, and expected heterozygosity, He, within populations for each locus, inbreeding coefficients, Fis; Weir and Cockerham 1984, within populations for each locus, and population structure among populations within subspecies, Fst; Weir and Cockerham 1984. The program Fstat version (Goudet 1995) was used to estimate rarefied allelic richness. We also conducted two-tailed t-tests to determine whether allelic richness, Fis and Fst differed significantly between the two subspecies.



Phylogenetic analyses were based on 1483 sequences and 5504 bp of which 425 were variable. Because of the few hybrid individuals in our sample and recent divergence, the two subspecies did not harbor any fixed differences between them; however, xantiana and parviflora averaged 30 and 5 unique polymorphisms per locus, respectively. The average number of polymorphisms shared between the subspecies per locus was six. Subspecies were not reciprocally monophyletic on the best topology estimated under a maximum-likelihood or Bayesian method, but there were only two samples of xantiana and one sample of parviflora that caused the taxa to be polyphyletic (Fig. 2). One clade with a high posterior probability contains all parviflora samples and only a few apparently admixed xantiana individuals (see below), which suggests that parviflora is a distinct group and that xantiana paraphyletic.

Figure 2.

(A) The phylogenetic tree with the highest maximum-likelihood from the 1000 replicate analyses of the nonbootstrapped dataset with Garli. Numbers associated with branches represent the percent of 1000 bootstrap replicates supporting that clade; only bootstrap values >80 are shown. Branches in red are those with greater than 0.80 posterior probability from the Bayesian analysis. Asterisks at tips denote clades consisting of all samples from the same population. (B) InStruct results showing the fractional assignment of individuals to clusters at k= 2. Individuals are shown in the same order as they appear on the phylogeny and color designates the proportion of an individual's multilocus genotype that belonged to one of the two clusters. (C) Observed autofertility and standardized values of flower size, herkogamy, and protandry for each individual.

The SplitsTree network showed two distinct groups corresponding to each subspecies, with considerable genetic distance between them, and a few individuals from sympatric populations found among heterospecifics. Similar to the phylogenetic analyses, we found that statistical support was greatest within parviflora where individuals from the same population often form a strongly supported group (Figs. 2 and S1).

The results from InStruct illustrated a sharp genetic difference between the two taxa. At k = 2, individuals were clustered into groups according to subspecies designation, which contrasts with the phylogenetic results that placed several xantiana individuals within a clade that included all parviflora samples. InStruct results also showed little indication of widespread admixture or hybridization. What admixture is present is found mainly within xantiana samples from sympatric populations (populations 5x, 6x, and 22x; Figs. 1 and 2).

The genealogical sorting index (gsiT) suggested a high degree of genealogical exclusivity for each taxon; gsiT values were 0.75 (P < 0.01) and 0.58 (P < 0.01) for parviflora and xantiana, respectively (Table 1). The gsiT values for each taxon based on individual gene trees were substantially less than observed for the concatenated dataset. Values for xantiana were also lower than parviflora for nearly all loci, which is likely the result of xantiana being paraphyletic (i.e., the node uniting all xantiana samples includes all parviflora samples and, thus, drives the gsiT values toward 0).


The estimates of TMRCA for each of the subspecies corroborated the phylogenetic results in indicating that xantiana is the ancestral taxon (Fig. 3). The mode of the posterior distribution of coalescence time to the ancestor uniting all xantiana alleles was 170,852 years ago (95% CI: 49,877–920,373) and of all parviflora alleles was 95,071 years ago (95% CI: 32,112–588,610).

Figure 3.

Density plot of the average across all loci of TMRCA for each subspecies estimated with *BEAST

The time since divergence (i.e., the time at which the subspecies began to diverge and evolve along separate evolutionary trajectories) between the two subspecies from *Beast was 9420 years (95% CI: 3169–66,889; Fig. 4). The estimate of the time since divergence from IMa2 was 66,152 years ago (95% CI: 33,035–151,448; Fig. 4). Estimates of Ne and population migration rates, 2Nem, from IMa2 can be found in Figure S2.

Figure 4.

Distributions of the divergence time between the two subspecies based on two different methods, IMa2 and *BEAST


Estimates of population genetic parameters based on all sites and a dataset with only noncoding sites were strongly correlated (θw, r= 0.85; Tajima's D, r= 0.95) and produced similar results. Therefore, we only present results based on analyses of all sites. Nucleotide polymorphism within populations (θw) was approximately fivefold greater in xantiana than parviflora (t=−11.80, df = 237, P < 0.001; Fig. 5; Table S2). A similar pattern was present when estimates of θw were calculated from taxon-wide samples, where individuals within a subspecies were pooled (Fig. 5; Table S3). Coalescent-based estimates of Ne from IMa2 were similarly fourfold greater in xantiana than parviflora (Figure S2). Genetic variation was strongly partitioned among populations in parviflora, with significantly greater differentiation among populations of parviflora (inline image) than xantiana (inline image; t= 5.28, df = 16, P < 0.01; Fig. 5).

Figure 5.

Comparison of the two subspecies in (A) θw, (B) Tajima's D, (C) linkage disequilibrium (r2), and (D) population structure (Fst). Boxes show the interquartile range, bars illustrate the median, and the whiskers extend out to 1.5 times the interquartile range.

Tajima's D did not differ significantly between the two subspecies regardless of whether it was calculated on taxon-wide or population-specific samples (t= 1.50, df = 14, P= 0.156 and t=−0.99, df = 163, P= 0.323, respectively; Fig. 5; Tables S2 and S3). The mean estimate of D from population-specific samples was close to zero for both subspecies; however, pooling individuals within subspecies caused the mean estimate of D to become negative (Fig. 5). The variance among population-specific estimates of D was significantly greater than equilibrium expectations for parviflora (inline image, inline image; P < 0.05) but not xantiana (inline image, inline image) suggesting that population bottlenecks have affected parviflora populations.

We also observed strong phylogenetic structure among parviflora populations, which also suggests that population bottlenecks have occurred. Individuals from eight of the 16 parviflora populations clustered together in the same clade (1p, 5p, 20p, 23p, 27p, 48p, 68p, and 83p), often with appreciable bootstrap support and high posterior probabilities (Fig. 2). By contrast, there was only one clade that included all individuals from the same population of xantiana (36x).

Linkage disequilibrium (r2) was significantly greater in parviflora (inline image) than xantiana (inline image; t= 2.51, df = 14, P= 0.03; Fig. 5). A similar difference was found using the population recombination rate, ρ, which was correlated with r2 (r= 0.75).

Patterns of microsatellite variation were similar to those from sequence data. Allelic richness and observed heterozygosity were greater in xantiana (71 alleles; Ho= 0.40) than parviflora (37 alleles; Ho= 0.06; Table 2); however, rarefied allelic richness was not significantly different (t=−1.3583, df = 6, P= 0.22). As expected, we also found significantly higher inbreeding coefficients in parviflora (inline image) than xantiana populations (inline image; t= 5.17, df = 81, P < 0.01; Table 2) and greater population structure in parviflora (inline image) than xantiana (inline image; Table 2); however, this difference was not significant (t= 0.80, df = 6, P= 0.45).

Table 2.  Summary statistics for four microsatellite loci: percent successful amplification, allelic ranges, number of alleles per locus (A), allelic richness (AR), observed heterozygosity (Ho), expected heterozygosity (He), inbreeding coefficients (Fis), and population structure (Fst).
Locus% AmplificationAllelic range (bp)AARHoHeFisFst
cx30.99244–268249–2551312.9 30.4400.680.120.3210.160.1
cx70.99113–131107–12510 8.9 4.80.330.040.440.140.20.740.170
cx90.97154–220194–2033029 60.3200.830.330.6110.120.34
cx110.96 88–152122–164372022.


Our results provide evidence that the predominantly outcrossing xantiana is the ancestral taxon from which the predominantly selfing parviflora evolved. Although there are multiple lines of evidence that support the distinctiveness of the two taxa (i.e., phylogenetic, genotypic clustering, and morphological), we also found that divergence began recently and that there has been an insufficient amount of time for the sorting of ancestral polymorphism. Given that the two taxa have not completed the speciation process (i.e., acquired complete reproductive isolation and genealogical exclusivity), inferences about the mode of mating system evolution are not likely confounded by events that occur postspeciation. Molecular population genetic analyses indicated substantially reduced molecular diversity and strong genetic bottlenecks in parviflora, which are consistent with theoretical predictions for the reproductive assurance model. This is the first system that we are aware of where evidence for reproductive assurance connects plant–pollinator interactions in the field to patterns of natural selection on phenotypes to patterns of sequence variation.


Studies of mating system transitions are often framed around the assumption that selfing is derived from outcrossing. In some cases, phylogenetic studies have shown that selfing has arisen multiple times from outcrossing progenitors (Schoen et al. 1997; Goodwillie 1999) but in many other cases, it remains unclear as to the polarity of mating system shifts. In this study, we found support that xantiana is the ancestral progenitor from which parviflora arose based on two different types of analyses. Specifically, an outgroup rooted phylogeny placed xantiana basal to parviflora. In addition, coalescent analyses showed that the TMRCA was approximately 100,000 years earlier for xantiana than parviflora.

The two methods we used to estimate time of divergence suggested that xantiana and parviflora began to diverge approximately 66,152 (IMa2; 95% CI: 33,035–151,448) or 9420 (*Beast; 95% CI: 3169–66,889) years ago during the upper Pleistocene. The discrepancy between the two methods can be explained by considering the different assumptions they make about the divergence process. The estimate of divergence time from *Beast represents the point at which significant gene flow ceases to occur between taxa (Heled and Drummond 2010). The estimate of divergence time from IMa2 explicitly allows for migration during the divergence process. Given that the InStruct analyses suggest that there has been recent but limited hybridization between the two taxa, our samples likely violate the assumption of no recent gene flow in *Beast. Recent introgression following secondary contact also violates the assumption of constant gene flow throughout the divergence process in IMa2, which may result in inflated estimates of time of divergence (Becquet and Przeworski 2009). As a result of these violations, the two methods will produce estimates of time since divergence that are biased in opposite directions (i.e., *Beast will produce shallower estimates and IMa2 will produce deeper estimates).

Our results also suggest that the limited introgression between subspecies is asymmetric, primarily from parviflora to xantiana. Specifically, the results from InStruct (Fig. 2) show multiple instances where individuals with xantiana-like phenotypes have admixed genomes, but we did not observe this for individuals with parviflora-like phenotypes. We also found that asymmetric migration rates from IMa2 were greater in the direction of parviflora to xantiana than the opposite direction (Fig. S2). Asymmetric introgression is also supported by how samples from sympatric sites (i.e., areas where introgression is most likely to occur) are grouped in the phylogenetic analyses. Xantiana samples from some sympatric sites (5x and 6x) are embedded within the strongly supported clade that includes all parviflora samples (Fig. 2). By contrast, parviflora samples from the same sympatric populations (5p and four individuals from 6p) comprise strongly supported clades that are nested among other parviflora individuals (Fig. 2). The same pattern of asymmetric introgression has also been observed between sister selfing and outcrossing Mimulus species (Sweigart and Willis 2003). One possibility is that hybrids are formed when outcrossers sire seeds on selfers and the resulting F1 hybrids are more likely to backcross to the outcrossing parent. In C. xantiana, we have observed that pollinators prefer and frequently visit both parental xantiana and F1 hybrids, but discriminate against the smaller flowered parviflora (D.A. Moeller, unpubl. data). This pattern of pollinator behavior increases the likelihood for backcrossing to xantiana rather than parviflora.


The TMRCA predated the time of divergence between xantiana and parviflora by approximately 100,000 years. Although TMRCA is expected to predate the time of divergence, this very broad difference further emphasizes how recently they began to diverge. Such shallow species tree divergence also makes it particularly difficult to evaluate the evolutionary distinctiveness of the two taxa under the criterion of genealogical exclusivity (Baum and Shaw 1995). The substantial amount of DNA polymorphism, principally within xantiana, suggests that the lack of reciprocal monophyly is due to retention of shared ancestral polymorphism rather than a lack of nucleotide variation. Based on simulation studies, the time necessary to achieve genealogical exclusivity for a majority of the genome is 4–7 Ne (Hudson and Coyne 2002). Because of the high levels of polymorphism in xantiana, the time necessary for xantiana and parviflora to achieve genealogical exclusivity is on the order of hundreds of thousands of years.

Despite the expected time necessary for xantiana and parviflora to each achieve complete genealogical exclusivity, there are multiple lines of evidence to suggest that the two taxa are incipient species and represent independent evolutionary lineages. For example, gsiT values for each taxon based on a concatenated matrix are significant and close to one. Additionally, the lack of reciprocal monophyly is not caused by individuals from throughout each taxon's range but rather is restricted to only a few samples from sympatric populations. Overall, the distinctiveness of the two taxa is supported by a combination of morphological differences (Figs. 2 and S3) and the formation of distinct genotypic clusters with no evidence for widespread admixture (Fig. 2; Mallett 1995). The general paraphyletic relationship of xantiana with respect to a nearly monophyletic parviflora is also expected of taxa at an intermediate state on the continuum of divergence (i.e., polyphyly to paraphyly to monophlyly; Rosenberg 2003).


Our estimate of the time since divergence suggests that the rate of evolution to a selfing phenotype has been quite rapid. The phylogenetic placement of xantiana as ancestral to parviflora suggests that the selfing phenotype of parviflora likely evolved from a phenotype characterized by larger flowers and higher levels of protandry and herkogamy (Fig. 2; see also ancestral character state reconstructions in Fig. S4). Consequently, the distinct morphological differences between the two subspecies (Fig. 2) are primarily the result of phenotypic changes over the past 65,000 years along the lineage leading to parviflora. Such a rapid rate of phenotypic evolution contrasts with the divergence of the selfing Arabidopsis thaliana from the outcrossing A. lyrata, which may have begun one million years ago (Tang et al. 2007). Other instances of rapid evolution of selfing (i.e., within the last 100,000 years) include the divergence of Capsella rubella from its self-incompatible progenitor, C. grandiflora (Foxe et al. 2009; Guo et al. 2009), and of selfing populations of Leavenworthia alabamica from self-incompatible populations (Busch et al. 2011). Additional studies estimating the temporal dynamics of divergence between closely related taxa with different reproductive strategies will help to clarify the rate at which such phenotypic changes occur.


The transition from outcrossing to selfing, alone, has dramatic effects on the genome, including reductions of up to 50% of genetic diversity (reviewed in Charlesworth and Wright 2001; Glemin et al. 2006). For transitions from partial selfers, such as xantiana, to more predominant selfers, such as parviflora, the reductions in genetic variation are expected to be less than 50%. Theoretical models have predicted that the reproductive assurance and automatic transmission modes of mating system evolution leave different signatures at the molecular level (Schoen et al. 1996). Under the former, the reduction in genetic diversity should be much greater than 50% owing to colonization bottlenecks (i.e., founder effects), whereas under the latter, extreme losses of diversity are not expected. We found that nucleotide diversity within parviflora was one fifth of that in xantiana and that the variance of Tajima's D in parviflora was significantly elevated relative to neutral equilibrium expectations. Both results suggest strong bottleneck events following the divergence of parviflora from xantiana (Charlesworth et al. 2003) and provide support for the reproductive assurance hypothesis.

Founder effects and population bottlenecks can also explain the greater degree of population structure (Fst) within parviflora relative to xantiana (Tables 2 and 3; Fig. 5). This result is consistent with previous studies showing greater population structure in selfers than outcrossers (Hamrick and Godt 1996; Nybom 2004). Population structure may also explain the discrepancy between estimates of Tajima's D based on taxon-wide versus population-specific samples. Taxon-wide samples resulted in somewhat negative values of D for both taxa (an excess of rare polymorphisms) whereas population-specific samples resulted in mean values of D near zero. This effect of geographic sampling strategy has been commonly found in recent molecular population genetic studies (Pool and Aquadro 2006; Arunyawat et al. 2007; Moeller et al. 2007) and simulations suggest that pooling across local populations can skew the site frequency spectrum (Ptak and Przeworski 2002; De and Durrett 2007) particularly in nonequilibrium populations (Städler et al. 2009).


The reproductive assurance model predicts that selfing in plants evolves under chronic pollen limitation of reproduction, which can arise when pollinators or mates are unreliable (Lloyd 1992; Schoen et al. 1996; Morgan and Wilson 2005). However, processes that are independent of mating system evolution (e.g., genetic bottlenecks associated with the speciation process) could also result in the genetic signature similar to that predicted for the reproductive assurance model. For example, automatic selection could have driven the evolution of selfing, with population bottlenecks occurring subsequently for reasons unrelated to the mating system transition. Field studies, which provide direct insight into the adaptive significance of selfing versus outcrossing, serve as an independent functional assessment of mating system variation and can be used to validate the results of molecular population genetic studies.

In C. xantiana, the results of field experiments provide additional support that the reproductive assurance model best explains the shift in mating system between xantiana and parviflora. Outcross pollination in xantiana is effected by solitary bees, particularly those specialized on the genus Clarkia (Moeller 2005; Eckhart et al. 2006), whereas in parviflora populations, pollinator visits are uncommon and specialist pollinators have not been found (Fausto et al. 2001; Moeller 2006). Moreover, patterns of population differentiation in mating system mirror geographic patterns of pollinator abundance and community composition (Moeller 2006). Reciprocal transplant experiments between the geographic ranges of xantiana and parviflora have also shown that xantiana suffers strong pollen limitation in parviflora's range but neither subspecies was strongly pollen limited in xantiana's range (Geber and Eckhart 2005). Emasculation experiments have verified that selfing elevates female fertility in field populations of parviflora, suggesting that selfing provides reproductive assurance (Moeller 2006). Finally, manipulative field experiments have shown that under pollinator and mate limitation, natural selection favors the selfing phenotype (reduced herkogamy and protandry), whereas selection maintains the outcrossing phenotype when pollinators and mates are reliable (Moeller and Geber 2005). In sum, patterns of DNA sequence variation are consistent with field studies in suggesting that selfing arose as a mechanism of reproductive assurance in environments where outcrossing is unlikely.

Associate Editor: J. Kelly


We are indebted to M. Geber and V. Eckhart for sharing their extensive knowledge of this system, P. Tiffin for assisting with the initiation of this project, and B. Morris for estimating generation time from demographic data. We also thank E. Beckman, R. Bier, R. Carter, J. Iverson, C. Kelly, and J. Reese for help collecting and processing the molecular data. Computational resources and assistance were provided by the Minnesota Supercomputing Institute (MSI) and A. Bazinet and M. P. Cummings. J. Kelly and two anonymous reviewers provided helpful comments that improved the manuscript. Funding for this project was provided by the National Science Foundation (DEB-1025004 to D. A. Moeller), the University of Minnesota, and the University of Georgia.