The evolution and genetics of sexually dimorphic ‘dual’ mimicry in the butterfly Elymnias hypermnestra

Abstract

Sexual dimorphism is a major component of morphological variation across the tree of life, but the mechanisms underlying phenotypic differences between sexes of a single species are poorly understood. We examined the population genomics and biogeography of the common palmfly Elymnias hypermnestra, a dual mimic in which female wing colour patterns are either dark brown (melanic) or bright orange, mimicking toxic Euploea and Danaus species, respectively. As males always have a melanic wing colour pattern, this makes E. hypermnestra a fascinating model organism in which populations vary in sexual dimorphism. Population structure analysis revealed that there were three genetically distinct E. hypermnestra populations, which we further validated by creating a phylogenomic species tree and inferring historical barriers to gene flow. This species tree demonstrated that multiple lineages with orange females do not form a monophyletic group, and the same is true of clades with melanic females. We identified two single nucleotide polymorphisms (SNPs) near the colour patterning gene WntA that were significantly associated with the female colour pattern polymorphism, suggesting that this gene affects sexual dimorphism. Given WntA's role in colour patterning across Nymphalidae, E. hypermnestra females demonstrate the repeatability of the evolution of sexual dimorphism.

1. Introduction

Understanding the relationship between genetic variability and the many levels of biological diversity is a central aim of genomics. Single genes of large effect are often found to be responsible for striking examples of adaptive variation [1,2]. Thus, much morphological diversity is derived from genetic variation at a relatively small number of genetic loci [3–6]. Mimetic butterflies are models for studying the relationship between exceptional phenotypic diversity resulting from limited genetic diversity for a number of reasons, including the manifest adaptive value of mimetic phenotypes, the fecundity and ease of rearing butterflies, and the incredible morphological diversity of butterflies [7,8]. Unravelling the genomic and developmental basis of butterfly phenotypes has advanced understanding of the evolution of sexual dimorphism [9], mimicry [5] and evolvability [10].

The Batesian mimetic butterfly genus Elymnias (Lepidoptera: Nymphalidae: Satyrinae) lends itself to the study of mimicry and sexual dimorphism because its 53 recognized species can vary dramatically in colour, pattern and wing size to mimic a variety of different model species in the families Nymphalidae, Pieridae, Papilionidae, Erebidae (Arctiinae) and Zygaenidae throughout tropical and subtropical Asia [11,12]. Moreover, only dorsal or both dorsal and ventral wing surfaces may be mimetic, and individual species can mimic multiple models via morphological differences that vary between sexes, locales or syntopic forms [13,14]. Within the genus Elymnias, there are several examples of allopatrically distributed species mimicking the same widespread model, thereby resembling each other, and of different populations or forms of a single species mimicking different models [13,14]. The most widespread and locally abundant species in this genus is the common palmfly, Elymnias hypermnestra [11]. This species is a ‘dual mimic’ [15]: it is sexually dimorphic and each sex resembles a dramatically different model species. All males of this palm-feeding species resemble melanic, unpalatable models in the genus Euploea [13,16] (figure 2). However, female mimicry is geographically variable: some disjunct populations are sexually dimorphic with orange females that mimic Danaus, while other populations are monomorphic, and melanic females mimic Euploea models along with the males (figures 1 and 2). Orange and melanic females do not co-occur. Naive, captive insectivorous birds (Pycnonotus sinensis formosae, Zosterops japonicus simplex and Copsychus malabaricus) with no prior exposure to the model or mimic readily consume adult males, orange females and melanic females representing four E. hypermnestra subspecies, indicating that the species is a palatable Batesian mimic (S.-H.Y. 2016, unpublished data). This species provides a unique opportunity to study the genomic basis of dual mimicry to assess whether the trait is controlled by loci known to control sexual dimorphism [2,17], mimicry [6,18,19] or both. In addition, the experimental advantages of this variable and widespread species might allow identification of loci that play an important role in the tremendous morphological diversity of its congeners.

Figure 2.

An ASTRAL species tree of Elymnias hypermnestra based on 3000 random autosomal 10 kb windows infers multiple clades of melanic and orange female forms. Branch colour indicates quartet score branch support. The sample IDs correspond to the same numbers in figure 1, and their colour indicates subspecies affiliation. Orange or dark backgrounds indicate the female colour pattern of the lineage, and representative images of females of the same subspecies as each sample are shown around the periphery. Images of the putative model species mimicked by orange and melanic females are provided at the top. Representative males of four subspecies are shown at the bottom. (Online version in colour.)

Figure 1.

Elymnias hypermnestra comprises three genetically and geographically distinct populations. (a) The geographic distribution of 48 E. hypermnestra populations representing 15 subspecies. Orange females and melanic females are indicated with background colours, demonstrating disjunct distributions of each colour pattern. Collection locations of each specimen used in this study are indicated with its sample ID (electronic supplementary material, table S1), which is coloured to indicate its subspecies. The dark outlines on the map indicate genetically distinct populations, as inferred by (b) principal component analysis. The points in this plot indicate sample ID and colour pattern. The same three populations are indicated by an (c) ADMIXTURE plot. The sample ID and colour pattern are indicated below each bar. (Online version in colour.)

Here, we examine the evolution and biogeography of sexually dimorphic dual mimicry in E. hypermnestra. Orange females of E. hypermnestra tinctoria (Thailand) and E. hypermnestra baliensis (Bali) produce orange patterns using different combinations of ommochrome pigments, suggesting independent evolution of orange morphs in these two geographically distant populations [20]. However, the evolutionary history and current population structure of E. hypermnestra were unknown, making it impossible to distinguish between single- and multi-origin scenarios. Moreover, while researchers have identified many genes that control the development of mimetic colour patterning in butterflies [5], including doublesex, responsible for female-limited polymorphic mimicry in Papilio polytes [2,21], the genes controlling sexually dimorphic dual mimicry are not understood. Since E. hypermnestra is dimorphic in some regions and monomorphic in others, this species has the potential to elucidate how sex-specific effects emerge and contribute to phenotypic variation. We assembled a high-quality reference genome and then resequenced low-coverage reads from 45 individuals representing 18 subspecies across the species's range. This allowed us to address the following three questions: (i) What is the population history and current population genetic structure of E. hypermnestra? (ii) Does the orange female colour pattern have a single evolutionary origin? (iii) What gene(s) are responsible for whether a population is dimorphic with orange females or monomorphic with melanic females?

2. Results

(a). Genetic structure of E. hypermnestra populations

We first assembled a reference genome for E. hypermnestra baliensis to facilitate downstream analyses. Using k-mer analysis [22] and SCO content evaluation [23,24], we found that the E. hypermnestra reference genome presented here is among the best assembled, most complete and least redundant nymphalid genomes available (electronic supplementary material, table S1). To better understand natural variation in E. hypermnestra across its large distribution spanning approximately 55 longitudinal degrees from western India to eastern Indonesia, we resequenced the genomes of 45 samples with at least approximately 20× coverage representing 18 subspecies across Asia (figure 1a; electronic supplementary material, table S2). We called SNPs in our resequenced data relative to the reference genome. This genome-wide SNP data indicated substantial genetic structure. The samples formed three distinct clusters in a principal component analysis of these data (figure 1b). We calculated fixation (F_ST) indices between each pair of subspecies and found the same three populations (electronic supplementary material, figure S1). The same three groups were also identified by ADMIXTURE [25] (figure 1c). Increasing the number of putative populations increased the likelihood of the admixture model, but the results assuming 2–5 populations all had comparable cross-validation errors (electronic supplementary material, figure S2).

(b). Repeated evolution of the Danaus mimetic colour patterns in E. hypermnestra

A 6-locus intraspecific phylogeny of E. hypermnestra suggested that neither orange nor melanic females were monophyletic, but support values on this tree were low (electronic supplementary material, appendix S1). We therefore inferred a species tree with ASTRAL using gene trees from 3000 unlinked, autosomal 10 kb windows. This tree was also inconsistent with either orange or melanic female morphs forming a monophyletic group (figure 2), as there were five melanic and four orange lineages. Trees inferred from Z-linked windows or complete mtDNA genomes (electronic supplementary material, figure S3) were topologically similar to the species tree inferred from autosomal loci (figure 2). The E. hypermnestra hainana subspecies/genetic population was distinctive in the PCA (figure 1b) and in the species tree, where all five samples form a strongly supported branch (figure 2). However, samples of this subspecies were not monophyletic in the 6-locus tree (electronic supplementary material, appendix S1), underscoring the potential bias of inferring intraspecific phylogeny using only few protein-coding markers [26].

We developed a coalescent model using the phylogeny of E. hypermnestra to examine when a given number of gene flow events are likely to have occurred during the evolutionary history of E. hypermnestra. Our species tree suggested that most subspecies are monophyletic (figure 2), so, while we recognize that subspecies are not necessarily monophyletic groups [27], we treated each subspecies as a group for computational ease. We found residual covariance between taxa in our model that was best explained by gene flow (electronic supplementary material, figure S4a). Some gene flow events were between melanic clades in different regions, but others were between melanic and orange clades, suggesting that gene flow was not only between subspecies with the same female morphs, consistent with results from Papilio polytes [28].

Finally, we used EEMS [29] to visualize estimated relative migration rates across geographic space (electronic supplementary material, figure S4b). Our results were consistent with biogeographic patterns evident in the species tree (figure 2). EEMS predicted several strong barriers to gene flow. The strongest barrier coincides with Wallace's Line, a well-known biogeographic demarcation that separates Bali (orange females) from Lombok (melanic females), and extends northward between Borneo and Sulawesi [30]. A second barrier separates Sumatra (melanic females) from Java (orange females), and the third barrier separates melanic E. hypermnestra hainana from all other populations. Intraspecific genetic diversity was highest on the Asian mainland and decreased from west to east along the Indo-Australian Archipelago (electronic supplementary material, figure S5).

(c). A genome-wide association study of the orange Danaus-like colour pattern suggests reuse of WntA

To identify the genetic locus or loci associated with orange and melanic female colour patterns in E. hypermnestra, we performed genome-wide association mapping of female colour patterns using the full SNP call set from the 45 re-sequenced samples. If male butterflies were sequenced, their collection locality was used to infer the female colour pattern from that area. We performed a genome-wide association study (GWAS) using GEMMA [31] because it incorporates the population structure and relatedness among samples. We saw no peaks in the unaligned genome without an equivalent in the aligned genome (electronic supplementary material, figure S5).

While many sites fell above the 1% false discovery rate (FDR), correcting for multiple testing, these sites had relatively little linkage disequilibrium (LD). Importantly, the most strongly associated sites had no other neighbours (figure 3). This was consistent with our gene tree of the 200 bp region surrounding these SNPs in which neither orange nor melanic colour patterns were monophyletic (electronic supplementary material, figure S3c).

Figure 3.

(a) Association between Elymnias hypermnestra female colour pattern and genetic variation. p-values are from SNP-wise Wald tests. Blue and red dashed lines represent the 10% and 1% false discovery rates (FDR), respectively. The full GWA results (with unplaced scaffolds) are shown in electronic supplementary material, figure S6. (b) An enlargement of chromosome 22 in (a); depicting the region of the two SNPs most significantly associated with female colour pattern. The locations of three nearby genes in the Melitaea cinxia reference genome are shown below the plot. (Online version in colour.)

The two most strongly associated sites were 3 bp apart; both exceed 1% FDR (figure 3a). Adding the population structure (as measured by the first principal component from the PCA) as a covariate removed neither of these sites (electronic supplementary material, figure S7). Looking at the genotype of the two sites on the scaffold (figure 3b; electronic supplementary material, figure S7), we observed that they predicted wing pattern almost perfectly (electronic supplementary material, table S6). These sites were 150 kb away from WntA, a patterning gene that has repeatedly been shown to be involved in melanization across the family Nymphalidae [32].

3. Discussion

(a). Repeated evolution of a mimetic colour pattern

Females of the dual mimic E. hypermnestra either resemble Euploea with a melanic colour pattern similar to males, or have an orange colour pattern mimicking Danaus. Our analyses shed light on the evolutionary and genetic mechanisms responsible for the geographic mosaic of female colour pattern in this facultatively sexually dimorphic species.

Analysis of the population structure suggested the presence of three genetic populations in E. hypermnestra. The first group represented the described subspecies E. hypermnestra hainana found in Taiwan, southern China including Hainan, northern Vietnam and central Laos (figure 1). The second genetic population comprised E. hypermnestra found on Java, the Lesser Sunda Islands and Seram. The third included individuals from South Asia including Sri Lanka, Indochina south of hainana and Sumatra (figure 1). The geographic border between E. hypermnestra hainana and the rest of E. hypermnestra's range coincides with the Hoang Lien Son Range and surrounding high elevation areas in the ‘Tail of the Himalayas’. The other border between genetic populations lies between Java and Sumatra. While these are currently separate land masses, the two islands were conjoined during Pleistocene low sea stands together with Borneo and the Thai–Malay peninsula to form a single land mass along the edge of the continental shelf, Sundaland [30]. Thus, the border of these populations lacks an obvious barrier to dispersal, though this area is frequently associated with genetic discontinuities within and between other butterfly species (D.J.L. 2015, unpublished data). While all E. hypermnestra hainana females are melanic, the other two populations include areas with orange females and areas with melanic females, which could be explained by the convergent evolution of colour patterns in disjunct locales.

We were able to trace the evolutionary history of the orange/melanic transition using phylogenetic analysis. As suggested by our species tree, the orange and melanic morphs of E. hypermnestra did not form monophyletic groups. This is not uncommon in butterflies—for instance, a single morph of Heliconius may prevail in a given region, but actually comprise distinct Heliconius species that are only monophyletic at colour pattern loci [33]. It is still unclear why variability between melanic and orange morphs of E. hypermnestra evolved and how it is maintained. The lack of monophyletic female colour patterns in E. hypermnestra may result from a geographic mosaic of selection to mimic the most common unpalatable model in a region. While differences in Danaus and Euploea local abundance have not been demonstrated, they tend to live in different habitats [11]. Characterization of the host plants, predators and butterfly communities where different female forms live may shed light on this issue, including assessment of model species abundance. Moreover, studying geographic variability in the chemical ecology of the mimicry ring may provide insight on the relationship between mimetic morphs and their models [34].

(b). WntA and the orange/melanic shift

To identify genetic factors underlying the shift between orange and melanic colour patterns in E. hypermnestra, we performed a GWAS of female colour pattern. In GWAS analyses of similar systems, there are usually large peaks of many linked sites [2]. This raises the question of why there is apparently little LD in this system. One possibility is that LD is lost because of filtering. On average, we identified one polymorphic site for every 100 bp. Another possible explanation is that, unlike most previous functional genomics studies on Lepidoptera, this study sampled butterflies across a wide geographic range with the strong population structure. Most other work was done within a narrower geographic range. For instance, all butterflies sampled by Kunte et al. [2] were from a single F3 generation. When we compared our GWAS (figure 3) to results of other studies with geographically extensive sampling (such as those on Arabidopsis), we found similarly rapid linkage decay resulting in narrow peaks [35,36].

Many previous studies demonstrate that WntA is associated with colour patterning in other nymphalid butterflies. In Heliconius, WntA is related to colour pattern transitions among different species and is typically expressed in regions of the butterfly wing that are melanic in mature adults [37]. Moreover, linkage mapping has shown that WntA is associated with a similar transition in Limenitis arthemis; in this case, an ancient cis-regulatory element mediates a transition from a mimetic white banded to a non-mimetic, unbanded form [38,39]. These data on these two SNPs in E. hypermnestra were consistent with them being cis-regulatory elements regulating WntA 150 kb downstream. While this is an unusually long distance between a regulatory element and its target, it is not unprecedented. Regulatory elements have been found megabases away from the genes they regulate [40], and optix enhancers are up to 220 kb away in Heliconius [41,42]. We found pronounced similarities between WntA's known effects on wing patterning in butterflies and the phenotype observed in E. hypermnestra. For instance, Mazo-Vargas et al. [32] created CRISPR WntA knockouts for a variety of nymphalids and found two conserved characteristics of WntA. First, WntA typically acts on the basalis (B), the central symmetry system (CSS) and the marginal band system (MBS), three regions of butterfly wings which are conserved across nymphalids. Moreover, WntA is typically expressed in melanic regions, likely because it is associated with upregulation of melanin. Both traits were found in the orange/melanic switch in E. hypermnestra (figure 2), further suggesting that WntA is involved in this transition of female colour pattern.

The potential involvement of WntA in E. hypermnestra mimicry polymorphism suggests that the gene functions somewhat differently than in Heliconius or Limenitis. Mimicry in E. hypermnestra is sexually dimorphic: while females may be orange, males are always melanic [13]. This implies that polymorphism affects females differently than males. Several mechanisms are plausible: by upregulating WntA in melanic females; downregulating WntA in orange females or changing the spatial pattern of WntA expression. This is an unusual example of a single gene involved in both sexually dimorphic and non-sexually dimorphic mimicry. This suggests a slightly different role for WntA in this system than in others, where WntA affects both sexes. Future functional genomics work can elucidate the specific nature of WntA on this variation. Two other peaks in our GWAS stood out, one on chromosome 20 and one on chromosome 6. Many of the genes have unknown functions, suggesting an angle for further research (electronic supplementary material, table S6).

(c). Predictability of evolution

Studies on wing patterns in Nymphalidae have revealed that a common toolkit of genes, including optix, cortex and WntA, underlie wing patterning and support the hypothesis that evolutionary outcomes can be predictable [2,10,37,38]. This study complements work on the predictability of evolution in two critical ways. For one, Elymnias diverged from the clade with Limenitis and Heliconius over 80 million years ago [43], making this one of the oldest cases of gene reuse in Nymphalidae that has been studied. Moreover, this demonstrates how sexual dimorphism can create variation with a single component of the toolkit: the same gene, WntA, seems to underlie sexually monomorphic variation and sexually dimorphic variation. This variation, in turn, allows for a greater phenotypic diversity than single genes of large effect would establish alone. The seemingly adaptive variability between sexes and among populations of E. hypermnestra has provided a fascinating natural experiment to study the genomic basis and evolution of a novel sexually dimorphic trait.

4. Methods

(a). Reference genome assembly and quality

The E. hypermnestra reference genome was generated from two E. hypermnestra baliensis females from Bali. We isolated DNA from the thorax tissue using a phenol–chloroform extraction method and constructed Illumina paired-end (PE) libraries with insert sizes 250 and 500 bp using the KAPA Hyper Prep Kit (KR0961, v. 1.14) from 2 µg genomic DNA [44]. We constructed mate pair (MP) libraries with insert sizes of 2, 6 and 15 kb using the Nextera Mate Pair Library Prep kit (FC-132-1001) and 4 µg genomic DNA (electronic supplementary material, table S3). The five, unique barcoded libraries were pooled in a ratio of 59 : 30 : 6 : 3 : 2 and sequenced 2 × 100 bp on a single lane of Illumina HiSeq 4000 (electronic supplementary material, table S2). We trimmed low-quality regions and adapters from raw PE reads using Trimmomatic v. 0.36 [45] where bases in the reads that were below a quality score of 15 were trimmed using a sliding window of 4 bp and all reads less than 36 bp in length were discarded. We used Platanus v. 1.2.4 [44] to trim adapter sequences and low-quality regions from MP reads. Trimmed libraries were assembled using the default settings of Platanus v. 1.2.4 and the assembly was polished using Redundans v. 0.13a (default settings; [46]). We removed scaffolds < 5 kb from this assembly, generated a species-specific repeat library and masked repeats using RepeatScout 1.0.5 and RepeatMasker 4.0.8 [47,48], respectively, to produce the final assembly. We estimated genome size and heterozygosity using 21-mer frequencies in the raw 250 bp PE library using GenomeScope [22].

We assessed the quality of our assemblies and other well-assembled nymphalid genomes using BUSCO v. 3 and the endopterygota gene set (2440 single-copy orthologs) from OrthoDB v. 9 [23,24]. The accessions of the assemblies tested are in the electronic supplementary material, table. We assigned E. hypermnestra scaffolds to Melitaea cinxia chromosomes using RaGOO [49,50]. This pipeline assigned 206/947 scaffolds (542 Mb/566 Mb) to chromosomes.

Finally, we generated a preliminary gene annotation set for the E. hypermnestra genome using MAKER v. 3.01.02 [51,52]. We used de novo transcripts from Bicyclus anynana (NCBI BioProject) as evidence for transcription, as no transcriptome data exist for Elymnias. We downloaded raw reads from BioProject PRJEB10924 using the SRA toolkit, trimmed remaining adapters using Trimmomatic, and assembled transcripts using Trinity v. 2.8.0 [53] with default settings. Furthermore, we used protein sequences from the UniProt/SwissProt protein database [54], and RefSeq protein models for Danaus plexippus, Papilio xuthus, Bombyx mori, Vanessa tameamea, Pieris rapae, and Drosophila melanogaster as evidence for protein-coding regions. We trained SNAP using this evidence, then used SNAP, Augustus v. 3.2 with Heliconius melpomene parameters, and GeneMark-ES 4 with MAKER to generate the final gene models [55]. We functionally annotated predicted proteins using BLASTp against the Uniprot/SwissProt database and combined that information using scripts included in MAKER.

(b). Whole genome resequencing and quality control

Adult E. hypermnestra were collected in the wild and preserved in ethanol and/or by freezing at −80°C (electronic supplementary material, table S1) before genomic DNA was extracted from the thorax tissue using a phenol–chloroform DNA extraction protocol. We constructed approximately 250 bp PE libraries using the KAPA Hyper Prep Kit (KAPA Biosystems) and sequenced them to approximately 20× coverage using 2 × 80 bp Illumina NextSeq 500 (electronic supplementary material, table S1). We trimmed adapters and low-quality regions from raw resequencing reads using TrimGalore 0.6.1 and cutadapt v. 1.18 [56], then removed reads containing overrepresented sequences (identified using FastQC). We mapped reads to the E. hypermnestra reference genome using Bowtie2 v. 2.3.0-beta7 with parameter ‘--very-sensitive-local’ [57]. We marked duplicate reads using PicardTools v. 2.8.1 and realigned around indels using the Genome Analysis ToolKit's (GATK, v. 3.8) RealignerTargetCreator and IndelRealigner. Finally, we called SNPs using the GATK UnifiedGenotyper with default settings except for the following values: heterozygosity prior = 0.02; minimum allowable base quality score = 30 and minimum mapping quality = 20 [58]. We removed genotypes with phred-scaled quality less than 10. We then produced FASTA formatted genome sequences for each individual using the GATK FastaAlternateReferenceMaker [59]. Our data reached approximately 20× coverage on average and had average mapping rates of 94.76% (electronic supplementary material, table S4).

(c). Population structure analyses

We inferred E. hypermnestra population structure with ADMIXTURE 1.3.0 [25]. We first performed linkage-disequilibrium-based pruning on our SNP dataset using plink v. 1.90, including only SNPs with r² < 0.10 in 50 bp sliding windows with 10 bp steps according to plink's --indep-pairwise utility. This yielded 108 189 SNPs. We ran ADMIXTURE with 10-fold cross-validation for parameters k = 2 through 10. We looked at the cross-validation error and the value of k that minimized the residuals (electronic supplementary material, figure S2) [60]. We performed principal component analysis on the same filtered dataset using plink [61].

(d). Phylogenetic analyses

Since LD returns to background levels over approximately 50 kb in Heliconius [62], we split the E. hypermnestra genome into non-overlapping 10 kb windows, kept every fifth window, then extracted alignments of sequences for each window from individual fastas with GATK. We tested for recombination within each alignment using PhiPack [63], then filtered out windows with recombination p-values > 1 × 10⁻¹⁰ and at least 100 informative sites. PhiPack uses patterns of polymorphism to infer the probability of past recombination events; as p-values decrease, the probability of recombination in the tested window increases. We randomly selected 3000 autosomal alignments that passed these filters and inferred an unpartitioned gene tree from each using IQ-TREE, which selected the best model with ModelFinder and estimated branch support using 1000 ultrafast bootstraps [64–66]. Finally, we inferred a species tree using the default settings of ASTRAL-III [67], which computed a consensus topology with support values derived from the fraction of gene trees that support a particular four-taxon topology (quartet scores).

(e). Genome-wide association for colour

We filtered out SNPs with 10 or more missing alleles or minor allele frequency < 0.10 from the unpruned dataset for a total of 5.4 million SNPs. We assigned phenotypes to each sample based on that population's female wing colour pattern (electronic supplementary material, table S2), then computed site-wise Wald χ² test p-values using GEMMA v. 0.98, including GEMMA's centred kinship matrix as a covariate [31]. We calculated genome-wide cut-off scores using the FDR method [68]. GWA results were plotted by ordering E. hypermnestra scaffolds to the Melitaea cinxia chromosome-level assembly [49].

Data accessibility

The reference genome and sequence data generated for this study are publicly available at NCBI under BioProject accessions PRJNA660054 and PRJNA660057.

Authors' contributions

D.J.L. and M.R.K. conceived and designed the study; D.M.R., N.W.V. and S.N. performed analyses and collected data; S.-H.Y., D.P. and D.J.L. collected specimens; N.W.V., D.J.L. and M.R.K. directed the project; D.M.R., N.W.V. and D.J.L. wrote the manuscript with input from all co-authors.

Competing interests

The authors declare they have no competing interests.

Funding

Fieldwork was funded by grants 9285-13 and WW-227R-17 from the Committee for Exploration and Research of the National Geographic Society to D.J.L. This work was funded by NSF grant nos DEB-1120380 and DEB-1541557 to D.J.L., MOST grant nos 108-2621-B-110-004-MY3 to S.-H.Y. and NIH grant no. GM131828 to M.R.K.