The unpalatable and warning-patterned butterflies Heliconius erato and Heliconius melpomene provide the best studied example of mutualistic Müllerian mimicry, thought–but rarely demonstrated–to promote coevolution. Some of the strongest available evidence for coevolution comes from phylogenetic codivergence, the parallel divergence of ecologically associated lineages. Early evolutionary reconstructions suggested codivergence between mimetic populations of H. erato and H. melpomene, and this was initially hailed as one of the most striking known cases of coevolution. However, subsequent molecular phylogenetic analyses found discrepancies in phylogenetic branching patterns and timing (topological and temporal incongruence) that argued against codivergence. We present the first explicit cophylogenetic test of codivergence between mimetic populations of H. erato and H. melpomene, and re-examine the timing of these radiations. We find statistically significant topological congruence between multilocus coalescent population phylogenies of H. erato and H. melpomene. Cophylogenetic historical reconstructions support repeated codivergence of mimetic populations, from the base of the sampled radiations. Pairwise distance correlation tests, based on our coalescent analyses plus recently published AFLP and wing colour pattern gene data, also suggest that the phylogenies of H. erato and H. melpomene show significant topological congruence. Divergence time estimates, based on a Bayesian coalescent model, suggest that the evolutionary radiations of H. erato and H. melpomene occurred over the same time period, and are compatible with a series of temporally congruent codivergence events. Our results suggest that differences in within-species genetic divergence are the result of a greater overall effective population size for H. erato relative to H. melpomene and do not imply incongruence in the timing of their phylogenetic radiations. Repeated codivergence between Müllerian co-mimics, predicted to exert mutual selection pressures, strongly suggests coevolution. Our results therefore support a history of reciprocal coevolution between Müllerian co-mimics characterised by phylogenetic codivergence and parallel phenotypic change.
Citation: Hoyal Cuthill J, Charleston M (2012) Phylogenetic Codivergence Supports Coevolution of Mimetic Heliconius Butterflies. PLoS ONE 7(5): e36464. doi:10.1371/journal.pone.0036464
Editor: Corrie S. Moreau, Field Museum of Natural History, United States of America
Received: January 24, 2012; Accepted: April 6, 2012; Published: May 7, 2012
Copyright: © 2012 Hoyal Cuthill and Charleston. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This research was supported by a Discovery Project grant (number DP1094891) from the Australian Research Council. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
The Neotropical butterfly genus Heliconius (Nymphalidae, Heliconiinae) is highly diverse, with 39 species , , many of which can be subdivided into multiple wing pattern morphs, or races . These unpalatable butterflies ,  have diversified to form regional Müllerian  mimicry complexes , each involving multiple species with a convergently evolved  predator-warning pattern , . For almost 150 years, biologists have debated whether the remarkable adaptive radiation of the Heliconius was driven by reciprocal ecological associations, a process now known as coevolution . Unpalatable Müllerian  co-mimics share the cost of educating inexperienced predators  (unlike palatable Batesian  mimics, which may “parasitize” their unpalatable models ). According to Müller's original model , two unpalatable co-mimics will both gain in fitness by their resemblance, though the ratio of these fitness gains will be proportionate to the square of the ratio of their population sizes. Where population sizes differ, this predicts greater fitness benefits for the rarer population because the more abundant co-mimic is expected to lose a greater number of individuals to encounters with inexperienced predators . It has been suggested that mimicry (and, particularly, mutualistic Müllerian mimicry ) may provide some of the most favourable conditions for coevolution, which has been defined (in the strict sense) as reciprocal evolutionary change  under mutualistic or competitive selection . Therefore, mimetic wing pattern evolution among Heliconius butterflies may provide key evidence regarding the importance of coevolution in adaptive radiation , .
In particular, the parallel wing pattern radiations of H. erato and H. melpomene have been a primary case study in the debate over coevolution between Müllerian co-mimics , , , . Both H. erato and H. melpomene are unpalatable and have wing colour patterns that deter potential predators , , . Across the Americas, each of these two species exhibits approximately 30 warning pattern morphs . With few exceptions, the wing patterns of H. melpomene and H. erato match in every region where they co-occur  (Figure 1), suggesting that they are Müllerian co-mimics , , . The two species are reciprocally monophyletic  and do not hybridise . Therefore mimicry between them has involved convergence at the genetic and phenotypic level , . Purifying selection against intra-species hybrids with unusual wing patterns  acts with assortative mating  to generate partial reproductive isolation between the parapatric morphs , potentially showing speciation in action , , .
Figure 1. Wing patterns morphs and geographic distributions of the Müllerian co-mimics H. erato and H. melpomene.
Mimetic morphs , ,  (aligned rows), from H. erato (left) and H. melpomene (right), included in this study (photographs show the type specimens). Coloured boundaries on a satellite image of Central and South America indicate the geographic range of each morph . Numbers indicate countries where morphs were sampled (by , ), West of the Andes: 1 Costa Rica, 2 Panama, 3 West Ecuador; East of the Andes: 4 Colombia, 5 French Guiana, 6 Trinidad, 7 Peru and 8 East Ecuador.doi:10.1371/journal.pone.0036464.g001
Codivergence is the parallel divergence of ecologically associated lineages within two distinct phylogenies , and is one predicted outcome of coevolution , . Codivergence may not, in itself, prove coevolution in the strict sense . However, codivergence can be considered some of the strongest available evidence for coevolution , , since, as Page  states, “it is difficult to imagine that [codivergence] can occur without at least some degree of coevolution”. Topological and temporal congruence (similarity of branching pattern and timing, respectively) between the phylogenies of H. erato and H. melpomene, compatible with a history of codivergence, would therefore support their coevolution , , ,  (in reference to their Pleistocene refugium hypothesis, Brown et al.  and Sheppard et al.  suggested that coevolution between H. erato and H. melpomene may have been aided by population isolation, but see  for a critical review). In contrast, a lack of topological or temporal congruence would suggest that coevolution did not occur (as previously suggested , , ).
Despite considerable discussion of phylogenetic branching patterns , ,  and timing , , , and an early biogeographic character-based analysis , there has been no previous test of codivergence between the mimetic populations of H. erato and H. melpomene using methods from cophylogenetic analysis (reviewed in ), which were developed specifically for this purpose. Cophylogenetic analysis tests for topological congruence between a pair of phylogenies that represent associated entities (such as genes and species, parasites and their hosts, populations and biogeographic regions , or mimics and models ) to determine whether there is statistically significant evidence for codivergence between these associated phylogenies (as described below). We present the first explicit cophylogenetic test for codivergence between mimetic populations of H. erato and H. melpomene, using multilocus coalescent ,  phylogenies. We then re-examine divergence times based on the Bayesian multilocus coalescence model , using an established  – and recently corroborated  – fossil-calibrated molecular clock, as well as a root-node calibration based on the most comprehensive published butterfly phylogeny .
One of the greatest challenges in reconstructing phylogenetic branching patterns and divergence times for recent radiations with low sequence divergence, such as that of the Heliconius, is the incomplete sorting of ancestral polymorphisms among divergent populations . Incomplete lineage sorting can cause individual gene trees to conflict with each other and with the true population tree. A related problem is that individuals sampled from divergent lineages may not form reciprocally monophyletic clades on an individual gene tree, or on a population tree built using gene sequence concatenation or gene tree consensus methods . Coalescent methods are designed to take individual gene histories into account by modelling the processes of mutation and inheritance, specifically, the coalescence of sampled genes, back through a gene tree, to their most recent common ancestor (reviewed by Rosenberg and Nordborg ). Coalescent phylogenetic methods reconstruct the relationships between divergent populations, that are partially or completely genetically-isolated , , by optimally reconciling the histories of multiple gene loci within one population-level tree , . Coalescent methods have rarely been applied to heliconian population genetics , and have not previously been used to reconstruct the phylogenies of H. erato and H melpomene.
In this study, we use coalescent  and character support  methods to delimit monophyletic populations of H. erato and H. melpomene, among individuals sampled at the level of country, biogeographic region, and morph (Table S1). Phylogenetic relationships ,  and divergence times  for these populations are reconstructed using Bayesian  and parsimony based (Minimise Deep Coalescence, MDC ) coalescent methods. These phylogenies provide the basis for cophylogenetic tests of topological congruence, conducted across the set of corresponding phylogenetic estimates returned by the coalescent analyses. Estimated branching patterns, cophylogenetic histories, and divergence times are presented and discussed, based primarily on phylogenies of the country level mimetic populations, which received the highest support for monophyly from the character based  and Bayesian coalescent  analyses. To ensure that the conclusions are not exclusive to our coalescent population phylogenies, we also conduct cophylogenetic analyses on phylogenies reconstructed from recently published amplified fragment length polymorphism (AFLP)  and colour pattern gene  data.
Materials and Methods
Molecular data, phylogenetic and population genetic analysis
The primary dataset of 290 multilocus DNA sequences (Table S1) included eight pairs of mimetic wing pattern morphs from H. erato and H. melpomene (as well as the morphs H. erato chestertonii and H. melpomene plesseni, whose co-mimics were not included in this study), sampling populations from the major Neotropical biogeographic regions (East and West of the Andes) and seven countries ,  (Figure 1). Twenty-two related species , , ,  were also included, in order to estimate positions and divergence times, for the population radiations of H. erato and H. melpomene, within the wider radiation of genus Heliconius. These species, and morphs of H. erato and H. melpomene, were selected to maximise coverage of the four included gene loci.
Table 1. Significance of topological congruence between H. erato and H. melpomene phylogenies.doi:10.1371/journal.pone.0036464.t001
This primary dataset comprised 3,533 base pairs sampled from four gene loci, namely the mitochondrial loci cytochrome c oxidase COI and COII, and the nuclear loci mannose 6-phosphate isomerase (Mpi) and triose-phosphate isomerase (Tpi). DNA sequences were downloaded from GenBank (accession numbers and source studies Table S1). Sequences of each gene locus were aligned using MUSCLE . Alignments for individual gene loci were used in Maximum Likelihood (ML) phylogenetic analyses, implemented in Treefinder , to produce individual gene trees as input for the MDC  coalescent phylogenetic analyses (described below). The ML phylogenetic analyses were conducted using the best-fit substitution model for each gene, selected by jModelTest 0.1.1  under the Akaike information criterion (AIC) . Sequences for the four included gene loci were also concatenated to produce a multilocus data matrix (Table S3).
Population genetic statistics were estimated based on the multilocus data matrix, using the program SITES . These statistics included the average pairwise genetic divergence within a population and an effective population size parameter, θ, estimated as the product of effective population size and mutation rate (θ = 4 Neμ where Ne = effective population size and μ is the neutral mutation rate , ). Two tests of population monophyly were performed, on the multilocus data matrix, for specimens of H. erato and H. melpomene grouped at the level of country, biogeographic region, morph or species. First, monophyly was assessed with Shimodaira-Hasegawa (SH) tests  on monophyly-constrained ML trees, using Treefinder. This test compared the support for each population level using the AIC, which evaluates the fit of a statistical model to the data against the number of parameters imposed by that model – in this case, the number of constraints required for monophyly at the given population level. Since the AIC is a measure of information loss, the preferred phylogenetic hypothesis will be the one with the lowest AIC value. Incongruence length difference (ILD) tests, conducted using PAUP* 4.0b10 , indicated significant incongruence between the nuclear loci for both H. erato and H. melpomene (p = 0.01 in each case), so the nucleotide substitution model was partitioned by gene locus (COI, GTR+I+G; COII, HKY+I+G; Mpi, GTR+G; Tpi, HKY+G). The second test of population monophyly was based on an explicit Bayesian multilocus coalescent model  and conducted using Bayesian phylogenies reconstructed for each geographical sampling level (further details below).
Coalescent population phylogenies were reconstructed by minimising the number of deep gene coalescences  in Mesquite  (conducted separately for the clades of H. erato and H. melpomene and for the combined taxon set), and using a Bayesian multi-population coalescent model in *BEAST , part of the BEAST 1.6.1 package  (for the combined taxon set). Such methods, which are based on an explicit model of gene lineage coalescence, have been found to accurately reconstruct population level phylogenies and are robust to low levels of gene flow , . MDC ,  phylogenies were each reconstructed using a heuristic, population level tree search, which incorporated the branch lengths of the four gene trees, did not auto-resolve gene tree polytomies, used subtree pruning and regraft branch-swapping, and stored up to 100 equally good trees at each search step. Bayesian coalescent analyses were based on partitioned nucleotide substitution models, selected under the Bayesian information criterion (COI, HKY+I+G; COII, HKY+I+G; Mpi, HKY+G; Tpi, HKY+G). The focal *BEAST analyses were run using a relaxed, uncorrelated log-normal molecular clock (selected based on Bayes factor comparisons against an, otherwise identical, analysis run with a strict molecular clock), allowing the mutation rate to vary within the phylogeny . A Yule prior was specified for the branching process of the population tree. Since the two mitochondrial gene loci (COI and COII) are non-recombining, a linked tree was specified for these loci in the *BEAST analyses . Each *BEAST analysis was run with a Markov chain Monte Carlo (MCMC) chain length of 108 steps, parameter sampling every 104 steps, and a conservative burn-in of 25%. Effective sample size (ESS) values, for the posterior distribution of each parameter, were assessed to check chain convergence in each *BEAST run. Output from *BEAST was analysed using Tracer  and visualised using FigTree 1.3.1 (A. Rambaut, http://tree.bio.ed.ac.uk/software/figtree/). The MDC and Bayesian coalescent population phylogenies were used as input for the cophylogenetic analyses (described below).
The Bayesian phylogenies also provided the basis for the second test of population monophyly. In this test, hypotheses of population divergence (at the level of country or morph versus species) were tested by comparing the coalescent likelihood  and population tree posterior  calculated under the Bayesian multi-population coalescent model . The coalescent likelihood calculates the probability of each gene tree g given the population tree S, as follows:
where b denotes the branches of the population tree S, is the implied history of g over b, and is the function for effective population size through time . The population tree posterior is the sum of that tree's log likelihood and log prior probability, plus the log prior probability densities for any other included priors . For each of these parameters, Tracer was used to calculate the mean value across the MCMC samples (excluding the burn-in) as well as the 95% Bayesian confidence interval (CI), which is the shortest interval containing 95% of the sampled values. The preferred phylogenetic hypothesis, in this test, is the one with the highest coalescent likelihood and tree posterior. Comparing these parameter values between population trees allowed us to evaluate independent phylogenetic estimates for each geographical sampling level (and so did not require nested hypotheses of population monophyly, as does the Bayesian coalescent method for population delimitation of Yang and Rannala  for example). The Bayesian coalescent analysis was based on a reduced dataset consisting of those country level populations which were sampled at all four gene loci, according to the requirements of *BEAST. Biogeographic region was not included as a population level in the Bayesian coalescent analyses, due to the unavailability of gene sequences with sufficient coverage of the four included loci.
Population divergence times were estimated in *BEAST under the Bayesian coalescent model , which estimates and incorporates both phylogenetic branching patterns and effective population sizes. Such methods, which explicitly model gene lineage coalescence, are expected to give relatively accurate estimates of divergence times compared, for example, to estimates from individual or concatenated gene loci . Two time-calibration methods were used. First, a relaxed, uncorrelated log-normal molecular clock was used to estimate divergence times based on a rate of 0.01909 substitutions per site per million years . This rate was taken from a recently published fossil-calibrated molecular clock estimate among 7 loci (including both mitochondrial and nuclear genes) for a butterfly clade (Papilioninae, Papilionidae)  and is very close to Brower's  widely used arthropod molecular clock rate for mitochondrial loci of 0.02 substitutions per site per million years . This substitution rate was set for one reference locus (COI) and specified as a prior for the 3 remaining loci (COII, Mpi, Tpi), after Heled and Drummond . This analysis produced tree topologies identical to those of the *BEAST analyses (described above), in which no time-calibrated substitution rate was specified (there, the reference locus rate and priors for the other loci were set to 1, giving branch lengths in units of substitutions per site). A second analysis used a time calibration interval for the divergence of genus Heliconius from genus Eueides, based on a fossil-calibrated phylogeny of the butterflies . This interval, from 12–23 Mya, was specified as a uniform prior on the age of the phylogenetic root node, based on the published 95% Bayesian confidence interval for the Heliconius-Eueides divergence time estimated at 18 Mya .
The MDC and Bayesian coalescent population phylogenies of H. erato and H. melpomene were used to conduct cophylogenetic analyses, using TreeMap 3  and Jane 3 . These analyses tested for statistically significant topological congruence between two given phylogenies (here, corresponding phylogenetic estimates for H. erato and H. melpomene, as described below), as would be compatible with a history of codivergence between mimetic populations. Cophylogeny mapping reconstructs histories that explain the similarities and differences between associated phylogenies given a cost regime for the recoverable historical events , . This is achieved by mapping current ecological associations (e.g. between mimics and models) back into the internal nodes of one phylogeny (e.g. that of the model) to reconstruct a cophylogenetic history (e.g. the history of mimicry between two species). In our context, the recoverable historical events are codivergence (parallel divergence of mimetic lineages), duplication (divergence of mimic lineages without model divergence), model switch (divergence of a mimic lineage onto an additional model lineage), and loss (absence of a mimic on a model lineage where it would otherwise be expected).
Müllerian co-mimics may benefit from a shared warning pattern to different degrees , . Heliconius erato has several characteristics, independent of hypothetical divergence times, which suggest that it has had the dominant role in its mimicry relationship with H. melpomene (as suggested by Eltringham in 1916 , ). These include generally greater current , and possibly historical , abundance, greater gregariousness, a wider geographic distribution, and pupal mating . Therefore, we treated H. erato as the model and H. melpomene as the mimic in our main cophylogeny mapping analyses, conducted using TreeMap and Jane. For comparison, these analyses were also repeated with a reversed model-mimic relationship.
Cophylogeny mapping in Jane uses heuristics to find solutions that minimise the overall cost of a historical reconstruction given a cost regime. The default event costs are zero for a codivergence event, one for duplication and model switch events, and two for loss events. TreeMap 3 attempts to find a Pareto set of solutions, that is, all the histories that could be optimal, given the input phylogenies and associations, under a range of event cost regimes. This range is very permissive: codivergence is set at zero cost and all the other costs are assumed to be positive, but do not need to be specified. Statistical analysis can then be performed (in both programs) to test whether the cost of a historical reconstruction is significantly lower than expected by chance, by generating a pseudo-random sample of minimal costs from a null distribution of problem instances with the same model phylogeny. The null distribution is generated by repeatedly randomising either the associations between the leaves (also known as the terminal taxa) or the branching order of the associate (mimic) tree. Thus there are two null hypotheses we might test: either (a) the current associations between model and mimic are not a consequence of a history of coevolution with the model phylogeny, or (b) the branching order of the mimic tree is not dependent on the branching order of the model tree. We prefer the latter test as it accounts for differences in probability of different tree shapes (by randomising the mimic phylogeny), but we conducted both tests for completeness and comparability with other studies.
An additional pairwise distance correlation test of topological congruence was performed in TreeMap 3. This test compares the significance of correlations of pairwise distances between leaves, for associated clades in the two phylogenies, against a distribution of such measures estimated by randomising subtrees of the mimic phylogeny. Each non-trivial internal node of each phylogeny is tested as follows. The leaves descended from the node are determined and their associated leaves in the other phylogeny are found, then the most recent common ancestor of those associated leaves is found. Thus we determine associated subtrees, which are then tested by randomising the mimic subtree. The test statistic is the Spearman correlation of distances between corresponding pairs of leaves. This provides a general test of the topological congruence between two associated phylogenies that does not require an explicit estimate of the history of associations between them. Rather, it tests against a null hypothesis that the mimic phylogeny is independent of the model phylogeny.
To ensure that the results of the cophylogenetic analyses were not exclusive to our coalescent population phylogenies, a cophylogenetic pairwise distance correlation test (as described above) was performed (i) on recent, genome-wide, AFLP phylogenies of H. erato (including 85 specimens) and H. melpomene (including 78 specimens) ; and (ii) on phylogenies reconstructed from a recently published 10-gene dataset (with 127 specimens), which included 5 linked genes involved in heliconian wing colour pattern determination . To produce the input for these cophylogenetic analyses, the topologies of the published AFLP phylogenies  were replicated and the 10-gene dataset  was downloaded from GenBank (accession numbers from Table S1 in Hines et al. ) and re-analysed. For the 10-gene dataset, MUSCLE was used to produce separate gene locus alignments for H. erato (plus its relatives H. himera and H. clysonymus) and H. melpomene (plus its relatives H. cydno, H. ismenius and H. numata). Alignments for each gene locus were then concatenated to produce two multilocus data matrices: one including all 10 genes and the other containing only the 5 colour pattern genes. A ML phylogeny was then reconstructed for each multilocus data matrix in Treefinder, using a partitioned nucleotide substitution model with the best-fit substitution model for each gene selected by jModelTest under the AIC. For each of these phylogenies, monophyletic clades of each wing pattern morph were then collapsed to a single leaf, to avoid pseudo-replication of mimicry associations.
Phylogenetics and population genetics
Shimodaira-Hasegawa character support tests  on monophyly-constrained ML trees could not reject population monophyly at the level of country, biogeographic region, or morph for H. erato or H. melpomene (p>0.4 in all cases). Taking into account the number of parameters imposed by each monophyly constraint, using the AIC  difference from the unconstrained ML tree (corresponding to consistent species level monophyly only), country was the favoured monophyly level for the sampled within-species populations (AIC scores: ML tree = 44,033, countries = 47,500, regions = 47,780, morphs = 47,502; AIC difference: countries = 3,468, regions = 3,747, morphs = 3,470). Similarly, the highest coalescent likelihood mean (clm) and population tree posterior (ptp) values were shown by the Bayesian phylogenies of the country level populations, indicating that the Bayesian multi-population coalescent model  also favoured population divergence at the level of country (corresponding phylogenies Figures S1 and S2) over divergence at the level of species or morph (corresponding phylogenies Figures S3 and S4): country level clm = 1,774 [1,685 to 1,862]; ptp = −17,478 [−17,598 to −17,359]; species level clm = 1,220 [1,130 to 1,310], ptp = −18,287 [−18,399 to −18,179]; morph level clm = 1,701 [1,613 to 1,788], ptp = −17,617 [−17,732 to −17,503]. The Bayesian CI values for the species level phylogeny do not overlap with those for the other population levels. However, the overlap in 95% confidence intervals for the population levels of country and morph indicates that there is relatively little information by which to choose between these hypotheses.
Monophyly of sampled morph populations (at least at the level of country) is supported by the gene sequence data and provides the most probable coalescent history for the sampled gene loci. This concurs with the greater clustering of individuals into monophyletic country level populations observed on recent genome-wide AFLP phylogenies, relative to phylogenetic estimates based on three concatenated mitochondrial loci . These results (see also , ) do suggest that the wing pattern morphs sampled from multiple countries (here H. erato hydara, H. erato petiverana, H. melpomene melpomene and H. melpomene rosina) may be non-monophyletic. However, neutral markers for recently diverged populations that can experience ongoing low-level gene flow (including those used in this study) may show relatively low levels of phylogenetic structure  and we note that the character support analyses were unable to reject monophyly of the higher population levels of biogeographic region or morph. Based on the character support and Bayesian coalescent analyses, we therefore focussed our cophylogenetic analyses on the country level populations, which received the highest monophyly support. However, phylogenies for region and morph level populations, which received lesser support, were also analysed.
Topological congruence between the population phylogenies of H. erato and H. melpomene
To account for phylogenetic uncertainty, we conducted cophylogenetic analyses across the set of phylogenies returned by the coalescent analyses (listed in Table 1). Across these MDC and Bayesian coalescent population phylogenies, the overwhelming indication is of significant topological congruence (Table 1). The cophylogenetic analyses suggest that, in almost all cases, there are more codivergence events between the mimetic populations of H. erato and H. melpomene than would be expected by chance if their phylogenies were independent. Figure 2 shows an example pair of phylogenies from this set of phylogenetic estimates with similarly high congruence (see Table 1). Reconciling the phylogeny of H. melpomene with that of H. erato indicates remarkable topological congruence, with 8 codivergence events out of out of a possible 11, two duplications followed by model switches, and one loss (Figure 3). Interestingly, we still obtain highly significant congruence between the phylogenies when the mimic-model relationship is reversed (Table S2).
Figure 2. Phylogenies of H. erato and H. melpomene illustrating branching orders of co-mimetic country level populations within each species.
Example phylogenies independently estimated for H. erato (black, left) and H. melpomene (orange, right) using the Minimise Deep Coalescence (MDC) method . These correspond to cophylogenetic analysis “separate MDC countries 1” in Table 1. H. erato/H. melpomene co-mimics (see Figure 1) are indicated by grey lines. This is one of several possible phylogeny pairs with similarly high congruence (see Table 1). Taxon labels indicate the sampled biogeographic region (East or West of the Andes), and country (abbreviations are: CR Costa Rica, Pa Panama, E Ecuador, C Colombia, FG French Guiana, T Trinidad and Pe Peru).doi:10.1371/journal.pone.0036464.g002
Figure 3. Cophylogenetic reconstruction of the history of mimicry between country level populations of H. erato and H. melpomene.
Example cophylogenetic history of the mimicry relationship between H. erato and H. melpomene based on MDC phylogenetic estimates (shown in Figure 2), and reconstructed using TreeMap 3. Bars indicate 95% Bayesian confidence intervals for divergence times, corresponding to Figure S1. Grey-filled dots correspond to reconstructed codivergence events; white-filled dots represent duplication events which, in this case, are both followed by model switch events – one for the colonisation of H. erato etylus in East Ecuador by H. melpomene ecuadoriensis, here the sister to H. melpomene malleti (also from East Ecuador), the other for the colonisation of the H. erato hydara population from Trinidad by a population of H. melpomene melpomene, here the sister of a lineage from French Guiana (one of the sampled mainland countries closest to Trinidad); the only mimicry loss event is indicated at the most recent common ancestor of H. erato hydara populations from Trinidad and Panama. Taxon labels correspond to those in Figure 2.doi:10.1371/journal.pone.0036464.g003
The pairwise distance correlation test also showed significant congruence at the roots of the H. erato and H. melpomene phylogenies, for the majority of our phylogenetic estimates (Table 1), suggesting that the phylogenies have been highly dependent on each other throughout their history.
Pairwise distance correlation tests conducted on the recent AFLP phylogenies  (Figure S5) and on the 10-gene dataset phylogenies  (Figures S6 and S7) also suggest significant topological congruence between the radiations of H. erato and H. melpomene (Table 1), contrary to the conclusions of these authors , .
Temporal congruence between the population phylogenies of H. erato and H. melpomene
Using the program SITES , we estimated average uncorrected genetic divergence (average pairwise substitutions per site, excluding gaps and indels) at 0.0245 within H. erato and 0.0153 within H. melpomene (excluding other, putatively incipient, species). Thus, the average uncorrected sequence divergence among individuals sampled from H. melpomene is considerably lower than (62% of) that estimated for H. erato, as previously suggested . However, the effective population size parameter for H. melpomene, measured as the product of effective population size and mutation rate (θ = 4 Neμ), was estimated at 41% to 63% of that for H. erato, see also . Specifically, the value of θ calculated using the method of Watterson  was 0.1058 for H. erato and 0.0429 for H. melpomene, and the θ value calculated based on pairwise nucleotide diversity (see ) was 0.0245 for H. erato and 0.0153 for H. melpomene, with each estimate of θ calculated as an average per base pair, with gaps, indels and sequences for putatively incipient species excluded . Effective population size is known to be positively correlated with average genetic diversity. Therefore, the lower average pairwise genetic diversity of H. melpomene, relative to H. erato, is an expected consequence of a lower effective populations size (e.g. see ), and is compatible with similar origination dates for the sampled clades of H. melpomene and H. erato, as discussed below.
In results similar to those of Flanagan et al. , average uncorrected sequence divergence between H. erato and the closely related species H. hecalesia (0.0605) was greater than that between H. melpomene and its close relative H. cydno (0.0303), or that between H. melpomene and the silvaniforms (0.0512), which form the outgroup to H. melpomene plus the H. cydno group . Divergence times estimated on our Bayesian coalescent phylogenies of the Heliconius (Figures S1, S2, S3 S4), using both substitution rate ,  and node age calibrations , also suggest that the split between H. hecalesia and H. erato (country population phylogenies: rate calibrated age 2.8 Mya [1.9, 3.76], node calibrated age 4.57 Mya [2.89, 6.3]; morph phylogenies: rate calibrated age 2.77 Mya [1.82, 3.79], node calibrated age 2.88 Mya [1.41, 4.63]) is older than that between the H. cydno species group and its H. melpomene sister clade (country populations: rate calibrated age 0.35 Mya [0.23, 0.48], node calibrated age 0.6 Mya; morphs: rate calibrated age 0.4 Mya [0.23, 0.58], node calibrated age 0.43 Mya [0.17, 0.73]), as previously suggested , as well as that between H. melpomene – plus the H. cydno group – and the silvaniforms (country populations: rate calibrated age 1.93 Mya [1.29, 2.64], node calibrated age 2.46 Mya; morphs: rate calibrated age 1.9 Mya [1.29, 2.6], node calibrated age 1.66 Mya [0.8, 2.63]). However, like the population genetic results discussed above, the Bayesian divergence estimates suggest temporal congruence between the population radiations of H. erato and H. melpomene (Figure 3), contrary to . First, sampled H. erato and H. melpomene were estimated to have begun their radiations at similar times. Specifically, the age of the most recent common ancestor of the included populations was estimated at 200,000 to 500,000 years for H. erato and 200,000 to 800,000 years for H. melpomene (H. erato: country populations, rate calibrated age 0.35 Mya [0.24, 0.47], node calibrated age 0.49 Mya [0.3, 0.7], morphs, rate calibrated age 0.25 Mya [0.2, 0.34], node calibrated age 0.22 Mya [0.11, 0.36]; H. melpomene: country populations, rate calibrated age 0.47 Mya [0.31, 0.66], node calibrated age 0.75 Mya [0.43, 1.11], morphs, rate calibrated age 0.2 Mya [0.11, 0.32], node calibrated age 0.21 Mya [0.08, 0.36]). Second, estimated ages of mimetic populations of H. erato and H. melpomene were often similar. Of the codivergence events reconstructed for the MDC phylogenies of Figure 3, for example, 95% Bayesian confidence intervals overlap where they are available. A historical reconstruction incorporating the estimated divergence times on the Bayesian country level phylogeny finds 5 codivergence events out of a maximum of 8 (Figure S8). Ages for the first divergence of eastern and western populations of H. erato and H. melpomene also show overlapping confidence intervals, although the point estimate for H. erato was estimated at approximately twice that for H. melpomene, as discussed below (H. erato: country populations, rate calibrated age 0.2 Mya [0.13, 0.27], node calibrated age 0.32 Mya [0.2, 0.45], morphs, rate calibrated age 0.18 Mya [0.13, 0.24], node calibrated age 0.16 Mya [0.07, 0.25]; H. melpomene: country populations, rate calibrated age 0.1 Mya [0.04, 0.17], node calibrated age 0.13 Mya [0.04, 0.23], morphs, rate calibrated age 0.1 Mya [0.05, 0.16], node calibrated age 0.11 [0.04, 0.19]).
Evidence for codivergence
Our coalescent population phylogenies for H. erato and H. melpomene (e.g. Figure 2; Figures S1, S2, S3 S4) have many features in common with previous phylogenetic estimates , , including a strong signal from biogeographic region (East or West of the Andes). The MDC coalescent phylogenies represented in Figure 2, for example, share major topological features with recent, genome-wide, AFLP phylogenies of H. erato and H. melpomene  (Figure S5). These features include a relatively basal split, within each species, between two clades; one clade containing eastern and western populations of, mimetic, H. erato hydara/H. melpomene melpomene plus the populations of the other western morphs (mimetic, H. erato petiverana/H. melpomene rosina and H. erato cyrbia/H. melpomene cythera), the other clade containing the remaining eastern populations of H. erato hydara/H. melpomene melpomene plus populations of the other eastern mimetic morphs. Within the solely eastern clades, a basal split between two major sub-clades is also shared with the recent AFLP phylogenies; one sub-clade containing H. erato hydara/H. melpomene melpomene and the French Guianan population of, mimetic, H. erato erato/H. melpomene thelxiopeia, the other sub-clade containing populations of, mimetic: H. erato lativitta/H. melpomene malleti, H. erato emma/H. melpomene aglaope, H. erato etylus/H. melpomene ecuadoriensis and H. erato favorinus/H. melpomene amaryllis.
Cophylogenetic analyses, conducted across the set of coalescent phylogenetic reconstructions, give an overall picture of statistically significant topological congruence between the evolutionary radiations of H. erato and H. melpomene co-mimics (contrary to previous suggestions , ) (Table 1). In particular, all phylogenetic estimates for the, best supported, country level populations are compatible with a history of repeated codivergence between mimetic populations.
In the interpretation of their AFLP phylogenies, Quek et al.  emphasised elements of incongruence between the topologies for H. erato and H. melpomene. They  noted, specifically, that the earliest branching lineages within each species did not represent co-mimetic morphs (these were H. erato etylus sampled from East Ecuador, which instead falls within the eastern clade of our figure 2, and H. melpomene nanna sampled from Brazil, which was not included in our coalescent analyses).
However, a cophylogenetic analysis conducted on these recent AFLP phylogenies  also indicates that patterns of evolutionary branching among co-mimics are significantly more similar than expected by chance (Table 1; Figure S5), despite elements of incongruence such as those described above. This suggests that an early lack of phylogenetic resolution  and as well as the complexity of more recent estimates of phylogenetic branching patterns  have previously concealed significant topological congruence between the phylogenies of H. erato and H. melpomene, which is revealed by quantitative cophylogenetic analysis.
Phylogenies of H. erato and H. melpomene based on a reanalysis of the complete 10-gene dataset of Hines et al.  also show significant topological congruence (Table 1; Figure S6), and biogeographic clustering patterns complementary to those of the coalescent population phylogenies illustrated in Figures 2. Similar reanalyses of only the 5 colour pattern genes from this dataset  are less able to cluster individuals of the same morph and show reduced biogeographic signal but also indicate statistically significant topological congruence between phylogenies of H. erato and H. melpomene co-mimics (Table 1; Figure S7).
As expected for recent evolutionary radiations of populations that still experience low-level gene flow, phylogenetic reconstructions for H. erato and H. melpomene are subject to some uncertainty, and there are differences between phylogenetic estimates based on different gene partitions (e.g. Figures S6 and S7), taxon partitions, and methodologies (e.g. the MDC and Bayesian phylogenetic estimates shown in Figures 2 and S1, respectively). However, several topological features are common to phylogenetic estimates based on different methodologies and data partitions (as discussed above) and the consistent result that emerges when we consider these various phylogenies is one of statistically significant topological congruence in the branching patterns of co-mimetic populations within these two species.
To be compatible with codivergence, ecologically associated phylogenies must be both topologically and temporally congruent . For example, the phylogenies of H. erato and H. melpomene might show topological but not temporal congruence if wing patterns arising from an earlier radiation (previously suggested to be that of H. erato , ) were secondarily converged upon during a later, but topologically similar, radiation by a mimic (previously suggested to be H. melpomene) . Previous studies have generally suggested that the phylogenies of H. erato and H. melpomene were temporally incongruent. In 1996 , Brower estimated that two eastern clades within H. erato and H. melpomene were of similar ages (150,000 – 200,000 years old), based on uncorrected average within-clade genetic divergence. However, his estimation that a key divergence between populations East and West of the Andes occurred earlier in H. erato (1.5–2 Mya) than in H. melpomene (65,000 years ago)  has been taken as evidence against codivergence of the two species , . In the same vein, Flanagan et al.  suggested that H. erato was approximately twice as old as H. melpomene, based on corrected genetic divergences from their nearest relatives (thought to be Heliconius hecalesia and Heliconius cydno, respectively). These apparent discrepancies in divergence times have previously been taken as evidence against coevolution of H. erato and H. melpomene , .
As in previous studies , we find that H. melpomene shows lower genetic diversity than H. erato, as measured by the average uncorrected pairwise divergence between individuals. However, population genetic comparisons indicate that the overall effective population size (estimated as θ, the product of effective population size and mutation rate) of H. melpomene is smaller than that of H. erato (observed here and also by Flanagan et al. ). The effective population size is the size of an idealized breeding population that would experience the same effects of random mutation as a real population under study . Effective population size is generally positively related to, but less than, the census population size . Therefore field observations suggesting that H. melpomene generally has a census population size approximately half that of H. erato (e.g. see ) are compatible with the difference in effective population size estimated from sampled genetic variation. Within-species genetic diversity is positively correlated with effective population size , , , . Indeed, the population size parameter θ determines the average genetic diversity of the population, because it takes into account both effective population size and mutation rate (these parameters can be separated using independent estimates of the mutation rate, for example from fossil calibrated divergence times , however such information is not available for H. erato or H. melpomene). The estimated difference in effective population size predicts that average genetic variation could be lower within the less abundant species H. melpomene, even if its radiation was temporally congruent with that of H. erato (e.g. see ).
Like Flanagan et al. , we estimated splits between H. melpomene and closely related clades (the H. cydno clade or the silvaniforms) to be younger than the split between H. erato and its close relative H. hecalesia (Figures S1, S2, S3 S4). This finding was supported both by differences in average sequence divergence (here and also Flanagan et al. ) and by the Bayesian coalescent phylogenetic analyses (which are more robust, since they estimate and take into account effective population sizes ). However, the crucial test for codivergence is not whether H. erato and H. melpomene first diverged from respective outgroups at similar times but whether their internal population radiations were temporally congruent.
Our Bayesian coalescent reconstructions indicate similar ages for the most recent common ancestor of the sampled populations of H. erato (200,000 to 500,000 years) and H. melpomene (200,000 to 800,000 years), with overlapping 95% Bayesian confidence intervals (e.g. Figure 3). This is compatible with a contemporaneous codivergence event at the start of the sampled radiations of these species (e.g. Figure 3). Thus we concur with Brower's  hypothesis that much of the phenotypic diversity within H. erato and H. melpomene evolved relatively recently, but estimate the origin of the sampled morphs of H. erato to be more recent than his estimate of 1.5–2 MY, and contemporaneous with that of H. melpomene, contrary to his conclusions . The age of the first divergence between eastern and western populations (see Figure 3; and Figures S1, S2, S3 S4) was estimated at 160,000 to 320,000 years for H. erato and at 100,000 to 130,000 years for H. melpomene (though a smaller sample of 2 western populations were included in the Bayesian analyses for H. melpomene, compared to 3 western populations of H. erato). However, the 95% Bayesian confidence intervals overlap, suggesting that contemporaneous codivergence of western and eastern populations, within the two species, cannot be rejected, contrary to previous suggestions .
Overall, the population genetic and Bayesian coalescent divergence time estimates strongly suggest that the parallel phenotypic radiations of H. erato and H. melpomene occurred during an overlapping time period, contrary to previous suggestions , , . The phylogenetic reconstructions and divergence time estimates are compatible with a series of contemporaneous codivergence events, occurring during a Müllerian mimicry relationship sustained over at least 350,000 years.
Codivergence and coevolution
Congruent phylogenies are often considered necessary to sustain hypotheses of (strictly reciprocal) coevolution , . Here, we find significant topological and temporal congruence between the phylogenies of H. erato and H. melpomene, which demonstrates that coevolution between the two species was possible (contrary to some previous suggestions , ). Furthermore, codivergence can be considered some of the strongest evidence that coevolution did occur , . In the case of H. erato and H. melpomene, the codivergent populations identified by the cophylogenetic reconstructions frequently represent distinct mimetic wing patterns (e.g. Figure 1). Thus, population codivergence is correlated with parallel genetic ,  and phenotypic  variation. When sustained codivergence is accompanied by multiple examples of parallel phenotypic change (as in the co-mimetic morphs of H. erato and H. melpomene illustrated in Figure 1), reciprocal coevolution can be considered a more probable mechanism than, for example, secondary, one-sided evolutionary change by one species (previously suggested to be the less abundant H. melpomene) to match its co-mimic (previously suggested to be the more abundant H. erato) , as follows.
As noted previously , evidence for topological and temporal congruence between associated phylogenies, such as that presented here, cannot completely reject the possibility that one of the taxa in question (such as H. erato) radiated first (within the limits of the overlapping Bayesian confidence intervals) and prompted a second parallel radiation by the other taxon (such as H. melpomene). It has been suggested to be unlikely that, in every case, mutations associated with parallel wing pattern change would occur first in H. erato and second in H. melpomene over multiple codivergence events . A generally larger population size (as estimated here and suggested previously , ) may have made such wing pattern mutation in H. erato more probable (e.g. ). However, some parallel wing pattern changes between the co-mimics may have been prompted by an initial mutation in H. melpomene, although uncertainty in divergence time estimates make the precise order of divergences extremely difficult to determine.
Perhaps more relevant to the argument for coevolution, sensu stricto, of H. erato and H. melpomene ,  is whether their evolution was reciprocal, in that change in one species exerted a selection pressure on the other and vice versa. Müllerian mimicry theory  predicts that two unpalatable co-mimics will both benefit from a shared warning pattern, although a less abundant species will receive greater fitness gains than a more abundant co-mimic (as introduced above). For example, based on our estimate that the effective population size for H. erato is, overall, approximately twice that for H. melpomene, the classical Müllerian model would predict mutual fitness benefits with a, respective, ratio of 1:4 (e.g. see ). While this supports previous suggestions that H. erato has generally played the dominant role in its mimicry relationship with H. melpomene (e.g. ), relative abundances at approximately these levels have been predicted to promote reciprocal coevolution involving convergent evolutionary change (e.g. see Figure 2 of  and Figure 7 of ). Furthermore, field studies demonstrate that Heliconius abundance can fluctuate both regionally and seasonally, and the local population size of H. melpomene can equal or exceed that of H. erato . Potentially this would create mutual peaks in selection for Müllerian mimicry during periods of similar abundance, for example at shared population bottlenecks . Codivergence may not prove coevolution in the strict sense . However, where there is a history of codivergence , such as that suggested by this study – particularly between ecological associates predicted to exert reciprocal selection pressures, such as these Müllerian co-mimics – at least some degree of coevolution is strongly suggested. Biogeographic comparisons suggesting that the dominant model H. erato has sometimes converged towards H. melpomene  offer another compatible line of evidence for reciprocal evolutionary change. While theory suggests that the mutual fitness benefits of Müllerian mimicry will promote coevolution , , , evidence for this has previously been rare . Therefore, our evidence for sustained codivergence between mimetic populations of H. erato and H. melpomene represents an important empirical case for the study of coevolution.
Coevolution is a powerful concept because it describes a mechanism for the coordination of evolutionary change in genetically separate populations . Consequently, evidence for coevolution has fundamental implications for ecology, population genetics and wider evolutionary theory . Here, we have presented evidence for phylogenetic codivergence between mimetic populations of H. erato and H. melpomene. Such codivergence represents some of the strongest evidence for coevolution , . Therefore, the parallel radiations of H. erato and H. melpomene support a hypothesis of reciprocal coevolution between Müllerian co-mimics characterised by population codivergence and parallel phenotypic change . Consequently, we suggest that these parallel radiations deserve to be reinstated (after , ) as one of the most striking known examples of coevolution.
Bayesian coalescent phylogeny for the Heliconius with country level populations of H. erato and H. melpomene. Branch labels give posterior probabilities, the axis indicates time (Mya) based on substitution rate calibration, and scale bars show 95% Bayesian confidence intervals for the mean node age.
Bayesian coalescent phylogeny for the Heliconius with country level populations of H. erato and H. melpomene, labelled as for Figure S1 and showing a time axis based on node age calibration.
Bayesian coalescent phylogeny for the Heliconius with morph level populations of H. erato and H. melpomene, labelled as for Figure S1 and showing a time axis based on substitution rate calibration.
Bayesian coalescent phylogeny for the Heliconius with morph level populations of H. erato and H. melpomene, labelled as for Figure S1 and showing a time axis based on node age calibration.
Phylogenies for H. erato (left) and H. melpomene (right) reproduced from  and corresponding to cophylogenetic analysis “Quek et al., 2010 AFLP” in Table 1. H. erato/H. melpomene co-mimics sampled from the same country are indicated by grey lines. Taxon labels indicate the sampled biogeographic region (East or West of the Andes), and country (abbreviations correspond to Figure 2). Shaded circles indicate the significance of a pairwise distance correlation test conducted for the shaded node (with p values corresponding to the key).
Maximum likelihood phylogenies independently estimated for H. erato (left) and H. melpomene (right) based on the 10-gene dataset of , corresponding to cophylogenetic analysis “Hines et al. 2011, 10-genes” in Table 1. Taxon labels indicate the sampled biogeographic region (abbreviations are: Am Amazon, Ca Caribbean, Ch Chocoan-Parana), and country (abbreviations correspond to Figure 2 with additional abbreviation: B Brazil). Further annotation corresponds to Figure S5.
Phylogenies reconstructed as for those of Figure S6 except based on only the 5 colour pattern genes of  and corresponding to cophylogenetic analysis “Hines et al. 2011 colour pattern genes” in Table 1. Further annotation corresponds to Figure S5.
History of mimicry between H. erato (black phylogeny) and H. melpomene (blue phylogeny), reconstructed using Jane 3, based on the phylogeny shown in Figure S1: white-filled circles represent codivergence, solid circles represent duplications, arrows represent model switches, and dashed lines represent losses.
Sampling information for the study system, including accession numbers.
Significance of congruence between phylogenies with a model (here H. melpomene) to mimic (here H. erato) relationship, reversed relative to Table 1.
DNA sequences used in this study: aligned data matrix in Nexus format.
Satellite image of Central and South America courtesy of NASA Earth Observatory. Photographs of Heliconius type specimens courtesy of Butterflies of America (specimens held by The Natural History Museum, London) and The Linnean Society of London. We would also like to thank the reviewers of this article for their very constructive comments.
Conceived and designed the experiments: JHC MC. Analyzed the data: JHC MC. Contributed reagents/materials/analysis tools: MC. Wrote the paper: JHC MC.
- 1. G. Lamas ed 2004 Atlas of Neotropical Lepidoptera. Checklist Part 4A Hesperiodea-Papilionoidea. Gainesville: Scientific Publishers
- 2. Brown JKS (1981) The biology of Heliconius and related genera. Ann Rev Entomol 26: 427–456.
- 3. Sheppard PM, Brown KS, Benson WW, Singer MC (1985) Genetics and the evolution of Müllerian mimicry in Heliconius butterflies. Phil Trans Roy Soc Lond 308: 433–613.
- 4. Engler H, Spencer KC, Gilbert LE (2000) Preventing cyanide release from leaves. Nature 406: 144–145.
- 5. Kapan DD (2001) Three-butterfly system provides a field test of Müllerian mimicry. Nature 409: 338–340.
- 6. Müller F (1879) Ituna and Thyridia; a remarkable case of mimicry in butterflies. Trans Entomol Soc Lond 1879: xx– xxix:
- 7. Bates HW (1862) Contributions to an insect fauna of the Amazon valley (Lepidoptera: Heliconidae). Trans Linnean Soc 23: 495–556.
- 8. Reed RD, Papa R, Martin A, Hines HM, Counterman BA, et al. (2011) optix drives the repeated convergent evolution of butterfly wing pattern mimicry. Science 333: DOI:10.1126/science.1208227.
- 9. Benson WW (1972) Natural selection for Müllerian mimicry in Heliconius erato in Costa Rica. Science 176: 936–939.
- 10. Thompson JN (1994) The coevolutionary process. Chicago: University of Chicago Press.
- 11. Joron M, Mallet JL (1998) Diversity in mimicry: paradox or paradigm? Trends Ecol Evol 13: 461–466.
- 12. Janzen DH (1980) When is it coevolution? Evolution 34: 611–612.
- 13. Thompson JN (1989) Concepts of coevolution. Trends Ecol Evol 4: 179–183.
- 14. Gilbert LE (1983) Coevolution and mimicry. Futuyma DM, Slatkin M, eds. Coevolution. Sunderland: Sinauer Associates Inc. pp. 263–281.
- 15. Brower AVZ (1996) Parallel race formation and the evolution of mimicry in Heliconius butterflies: a phylogenetic hypothesis from mitochondrial DNA sequences. Evolution 50: 195–221.
- 16. Mallet J (1999) Causes and consequences of a lack of coevolution in Müllerian mimicry. Evol Ecol 13: 777–806.
- 17. Chai P (1986) Field observations and feeding experiments on the responses of rufous-tailed jacamars (Galbula ruficauda) to free-flying butterflies in a tropical rainforest. Biol J Linn Soc 29: 161–189.
- 18. Pinheiro CEG (2003) Does Müllerian mimicry work in nature? Experiments with butterflies and birds (Tyrannidae). Biotropica 35: 356–364.
- 19. Brown KS, Sheppard PM, Turner JRG (1974) Quaternary refugia in tropical America: evidence from race formation in Heliconius butterflies. Proc Roy Soc Lond B 187: 369–378.
- 20. Beltrán M, Jiggins CD, Brower AVZ, Bermingham E, Mallet J (2007) Do pollen feeding, pupal-mating and larval gregariousness have a single origin in Heliconius butterflies? Inferences from multilocus DNA sequence data. Biol J Linn Soc 92: 221–239.
- 21. Counterman BA, Araujo-Perez F, Hines HM, Baxter SW, Morrison CM, et al. (2010) Genomic hotspots for adaptation: the population genetics of Müllerian mimicry in Heliconius erato. PLoS Genet 6: DOI:10.1371/journal.pgen.1000796.
- 22. Hines HM, Counterman BA, Papa R, de Moura PA, Cardoso MZ, et al. (2011) Wing patterning gene redefines the mimetic history of Heliconius butterflies. Proc Natl Acad Sci USA 49: 19666–19671.
- 23. Kronforst MR, Young LG, Kapan DD, McNeely C, ONeill RJ, et al. (2006) Linkage of butterfly mate preference and wing color preference cue at the genomic location of wingless. Proc Natl Acad Sci USA 103: 6575–6580.
- 24. Arias CF, Muñoz AG, Jiggins CD, Mavárez J, Bermingham E, et al. (2008) A hybrid zone provides evidence for incipient ecological speciation in Heliconius butterflies. Mol Ecol 17: 4699–4712.
- 25. Page , RDM (2003) Introduction. Page RDM, ed. Tangled trees: Phylogeny, cospeciation, and coevolution. Chicago: The University of Chicago Press. pp. 1–21.
- 26. Futyma DJ, Slatkin M (1983) Introduction. Coevolution. Sunderland: Sinauer Associates Inc. pp. 1–13.
- 27. Mallet J, Jiggins CD, McMillan WO (1996) Mimicry meets the mitochodrian. Curr Biol 6: 937–940.
- 28. Knapp S, Mallet J (2003) Refuting refugia? Science 300: 71–72.
- 29. Flanagan NS, Tobler A, Davison A, Pybus OG, Kapan DD, et al. (2004) Historical demography of Müllerian mimicry in the neotropical Heliconius butterflies. Proc Natl Acad Sci USA 101: 9704–9709.
- 30. Quek S-P, Counterman BA, de Moura PA, Cardoso MZ, Marshall CR, et al. (2010) Dissecting comimetic radiations in Heliconius reveals divergent histories of convergent butterflies. Proc Natl Acad Sci USA 107: 7365–7370.
- 31. Ceccarelli FS, Crozier RH (2007) Dynamics of the evolution of Batesian mimicry: moelcular phylogenetic analysis of ant-mimicking Myrmarachne (Aranae: Salticidae) species and their ant models. J Evolution Biol 20: 286–295.
- 32. Maddison WP (1997) Gene trees in species trees. Syst Biol 46: 523–536.
- 33. Heled J, Drummond AJ (2010) Bayesian inference of species trees from multilocus data. Mol Biol Evol 27: 570–580.
- 34. Brower AVZ (1994) Rapid morphological radiation and convergence among races of the butterfly Heliconius erato inferred from patterns of mitochondrial DNA evolution. Proc Natl Acad Sci USA 91: 6491–6495.
- 35. Simonsen TJ, Zakharov EV, Djernaes M, Cotton AM, Vane-Wright RI, et al. (2011) Phylogenetics and divergence times of Papiloninae (Lepidoptera) with special reference to the enigmatic genera Teinopalpus and Meandrusa. Cladistics 27: 113–137.
- 36. Heikkilä M, Kaila L, Mutanen M, Peña C, Wahlberg N (2011) Cretaceous origin and repeated Tertiary diversification of the redefined butterflies. Proc Roy Soc Lond B. DOI:10.1098/rspb.2011.1430.
- 37. Rosenberg NA, Nordborg M (2002) Genealogical trees, coalescent theory and the analysis of genetic polymorphisms. Nat Rev Genet 3: 380–390.
- 38. Maddison WP, Knowles LL (2006) Inferring phylogeny despite incomplete lineage sorting. Syst Biol 55: 21–30.
- 39. Eckert AJ, Carstens BC (2008) Does gene flow destroy phylogenetic signal? The performance of three methods for estimating species phylogenies in the presence of gene flow. Mol Phylogenet Evol 49: 832–842.
- 40. Zhang C, Zhang D-X, Zhu T, Yang Z (2011) Evaluation of a Bayesian coalescent method of species delimitation. Syst Biol 60: 747–761.
- 41. Shimodaira H, Hasegawa M (1999) Multiple comparisons of log-likelihoods with applications to phylogenetic inference. Mol Biol Evol 16: 1114–1116.
- 42. Beltrán M, Jiggins CD, Bull V, Linares M, Mallet J, et al. (2002) Phylogenetic discordance at the species boundary: comparative gene genealogies among rapidly radiating Heliconius butterflies. Mol Biol Evol 19: 2176–2190.
- 43. Edgar RC (2004) MUSCLE: multiple sequence alignment with high accuracy and high throughput. Nucleic Acids Res 32: 1792–1797.
- 44. Jobb G, Haeseler AV, Strimmer K (2004) TREEFINDER: a powerful graphical analysis environment for molecular phylogenetics. BMC Evol Biol 4: 18.
- 45. Posada D (2008) jModelTest: phylogenetic model averaging. Mol Biol Evol 25: 1253–1256.
- 46. Akaike H (1974) A new look at statistical model identification. IEEE T Automat Contr 19: 716–723.
- 47. Hey J, Wakely J (1997) A coalescent estimator of the population recombination rate. Genetics 145: 833–846.
- 48. Watterson GA (1975) On the number of segregating sites in genetical models without recombination. Theor Popul Biol 7: 256–276.
- 49. Wakely J, Hey J (1997) Estimating ancestral population parameters. Genetics 145: 847–855.
- 50. Swofford DL (2003) PAUP* Phylogenetic analysis using parsimony (*and other methods) version 4. Sunderland: Sinauer Associates.
- 51. Maddison WP, Maddison DR (2010) Mesquite: a modular system for evolutionary analysis version 2.73. http://mesquiteproject.org.
- 52. Drummond AJ, Rambaut A (2007) BEAST: Bayesian evolutionary analysis by sampling trees. BMC Evol Biol 7: 214.
- 53. Yang Z, Rannala B (2010) Bayesian species delimitation using multilocus sequence data. Proc Natl Acad Sci USA 107: 9264–9269.
- 54. McCormack JE, Heled J, Delaney KS, Townsend Peterson A, Lacey Knowles L (2010) Calibrating divergence times on species trees versus gene trees: implications for speciation history of Aphelocoma jays. Evolution 65: 184–202.
- 55. Charleston MA, Robertson DL (2002) Preferential host switching by primate lentiviruses can account for phylogenetic similarity with the primate phylogeny. Syst Biol 51: 528–535.
- 56. Conow C, Fielder D, Ovadia Y, Libeskind-Hadas R (2010) Jane: a new tool for the cophylogeny reconstruction problem. Algorithm Mol Biol 5: 16.
- 57. Charleston MA (1998) Jungles: a new solution to the host/parasite phylogeny reconciliation problem. Math Biosci 149: 191–223.
- 58. Eltringham H (1916) On specific and mimetic relationships in the genus Heliconius. Trans Ent Soc Lond 1916: 101-148. pp. 101–148.
- 59. Wright S (1931) Evolution in Mendelian populations. Genetics 16: 97–159.
- 60. Frankham R (1996) Relationship of genetic variation to population size in wildlife. Conserv Biol 10: 1500–1508.
- 61. Charlesworth B (2009) Effective population size and patterns of molecular evolution and variation. Nat Rev Genet 10: 195–205.
- 62. Kronforst MR, Gilbert LE (2008) The population genetics of mimetic diversity in Heliconius butterflies. P Roy Soc B 275: 493–500.
- 63. Baxter SW, Nadeau NJ, Maroja LS, Wilkinson P, Counterman BA, et al. (2010) Genomic hotspots for adaptation: the population genetics of Müllerian mimicry in the Heliconius melpomene clade. PLoS Genet 6: 1–12.
- 64. Gandon S, Michalakis Y (2002) Local adaptation, evolutionary potential and host-parasite coevolution: interactions between migration, mutation, population size and generation time. J Evol Biol 15: 451–462.
- 65. Turner JRG (1987) The evolutionary dynamics of Batesian and Muellerian mimicry: similarities and differences. Ecol Entomol 12: 81–95.
- 66. Gilbert LE (1984) The biology of butterfly communities. Vane-Wright R, Ackery P, eds. The biology of butterflies. New York: Academic Press. pp. 41–54.
- 67. Futuyma DJ (1986) Evolutionary biology: Second edition. Sunderland: Sinauer Associates Inc.
- 68. Wilson EB (1927) Probable inference, the law of succession, and statistical inference. J Am Stat Assoc 22: 209–212.