Refining mimicry: phenotypic variation tracks the local optimum
Summary
- Müllerian mimicry between chemically defended preys is a textbook example of natural selection favouring phenotypic convergence onto a shared warning signal. Studies of mimicry have concentrated on deciphering the ecological and genetic underpinnings of dramatic switches in mimicry association, producing a well-known mosaic distribution of mimicry patterns across geography. However, little is known about the accuracy of resemblance between natural comimics when the local phenotypic optimum varies.
- In this study, using analyses of wing shape, pattern and hue, we quantify multimodal phenotypic similarity between butterfly comimics sharing the so-called postman pattern in different localities with varying species composition.
- We show that subtle but consistent variation between populations of the localized species, Heliconius timareta thelxinoe, enhance resemblance to the abundant comimics which drive the mimicry in each locality.
- Those results suggest that rarer comimics track the changes in the phenotypic optimum caused by gradual changes in the composition of the mimicry community, providing insights into the process by which intraspecific diversity of mimetic pattern may arise. Furthermore, our results suggest a multimodal evolution of similarity, with coordinated convergence in different features of the phenotype such as wing outline, pattern and hue.
- Finally, multilocus genotyping allows estimating local hybridization rates between H. timareta and comimic H. melpomene in different populations, raising the hypothesis that mimicry refinement between closely related comimics may be enhanced by adaptive introgression at loci modifying the accuracy of resemblance.
Introduction
Chemically defended animal species often show striking convergence in their colour patterns with other prey coexisting in the same habitat. This convergence may be explained, since the work of Müller (1879), by natural selection favouring superficial resemblance and operated by visual predators, a phenomenon called Müllerian mimicry. Theory proposes that predators learn upon experience the association of prey distastefulness and prey visual appearance, generally distinctive warning patterns. Mimicking a locally abundant warning signal, well known by local predators, constitutes a benefit to a defended prey species by decreasing predation risk. Mimicry benefits associated with a given warning signal depend on the relative numbers of prey sampled vs. available to learning predators and are usually driven by the most toxic or the most abundant prey species (Mallet & Joron 1999).
Müller's general principle has been largely supported by theoretical models and empirical experiments (Turner 1977; Sheppard et al. 1985; Turner 1987; Ruxton, Sherratt & Speed 2004). Field transplant experiments have confirmed selection favouring local patterns (Mallet & Barton 1989; Kapan 2001; Chouteau & Angers 2011; Merrill et al. 2012) and the strong frequency-dependent selection acting on warning signals in diverse communities (Chouteau, Arias & Joron 2016). Natural selection for local mimicry explains local polymorphism of distinct colour patterns (Kapan 2001; Joron & Iwasa 2005), as well as the maintenance of geographical races with sharply distinct patterns, for instance the so-called postman vs. rayed patterns of Heliconius erato and H. melpomene (Mallet & Barton 1989). Most of those studies investigated the mimicry benefits associated with alternative warning strategies using completely distinct patterns, corresponding to mimetic optima and describing an adaptive landscape with adaptive peaks separated by valleys of low fitness (Leimar, Tullberg & Mallet 2012). However, for a given morph, variations may be found between individuals, between sexes and between localities. Fewer studies have addressed the significance of mimicry variation around a given adaptive peak, or the underpinnings of precise resemblance within a given mimicry ring. Mimetic communities often involve assemblages of species which differ between localities, and species indeed vary in the level of mimicry precision to others (Penney et al. 2012), but it is unclear what determines the level of mimicry accuracy in those communities or to which extent the mimicry optimum may vary through space or time.
The ability of predators to generalize the signals of defended prey is an important determinant of selection on resemblance in a mimicry system (Rowe, Lindstrom & Lyytinen 2004; Ihalainen et al. 2012). Sharing key components of a warning signal with comimics can sometimes enhance protection and may allow crossing a valley of low fitness (Beatty, Beirinckx & Sherratt 2004; Balogh et al. 2010). For instance, jacamars trained to avoid butterflies with an orange patch were less likely to attack butterflies with a similar pattern painted red than those painted black (Langham 2004), suggesting that a red patch may be sufficient to reduce predation regardless of the differences in hue. Nevertheless, in that example, red butterflies still received higher predation than orange controls (Langham 2004), suggesting that small deviations of colour hue are, to a certain extent, perceived by predators and translate into fitness differences. Increased predation against imperfect mimics has also been shown in laboratory experiments (Ihalainen et al. 2008), meaning that even if coarse resemblance is attained, selection may still favour the improvement of mimicry. Those findings support the classical scenario for the evolution of mimicry, first through a major phenotypic change allowing coarse resemblance for certain key warning feature, followed by the gradual improvement of mimicry under selection by narrowly generalizing predators (Sheppard et al. 1985; Turner 1987; Franks & Sherratt 2007; Balogh et al. 2010; Leimar, Tullberg & Mallet 2012).
The strength of selection for resemblance is affected by the complexity and diversity of the prey community. For instance, discrimination against imperfect mimics is less accurate when the community of prey is complex, that is composed of several distinct warning signals (Ihalainen et al. 2012). However, whether variations of pattern within a mimicry ring itself affect the intensity of selection for resemblance has rarely been tested, since experimental predators usually are only trained on a single prey type. Quantifying variations in resemblance of wild individuals within and between species of a given mimicry ring allows investigating if and how the accuracy of resemblance evolves.
Phenotypic similarity is influenced not only by natural selection favouring accurate mimicry but also by the genetic architecture underlying variation in phenotype. For instance, phenotypic variation remaining within the generalization range of predators might be little influenced by mimicry selection and more by genetic correlations or environmental trade-offs. Drift, mutations, environmental plasticity or hybridization are different sources of variation which can affect the accuracy of similarity and dissimilarity of phenotypes within and between species of the mimicry ring. In addition, specific genetic architectures can favour convergence and a good level of resemblance. In Heliconius butterflies, variations in mimetic colour patterns are largely controlled by a few Mendelian loci of large effect, often coined the ‘colour pattern toolbox’, an architecture which may facilitate secondary convergence in wing patterns (Reed et al. 2011). The most striking example of mimicry achieved through a shared architecture is when alleles at colour-patterning loci are shared among comimics via adaptive introgression. This was documented, for instance, in the pair of species H. timareta and H. melpomene (Heliconius Genome Consortium 2012; Pardo-Diaz et al. 2012). In this case, colour pattern resemblance reflects the shared origin of adaptive alleles in both species, but selection for mimicry may also play a role, first by facilitating the invasion of introgressed alleles in the receiving population and second by favouring further refinement of resemblance in the new genome. Describing variations in resemblance between species is therefore required to disentangle the relative importance of shared genetics and mimicry selection in the evolution of accurate resemblance.
In this study, we investigate selection for mimicry perfection by quantifying phenotypic similarity among multiple species forming the so-called postman mimicry ring in Northern Peru.
The ‘postman’ wing pattern is a very strong warning signal working as a major (and possibly ancient) mimicry attractor in this region (Hines et al. 2011). However, subtle geographic variation in colour pattern was reported for one of the comimics (Heliconius timareta thelxinoe), between localities separated by about 175 km (Mérot et al. 2013). The two localities display some differences in the assemblage of species participating in the postman mimicry. In the Alto Mayo, H. timareta thelxinoe co-occurs mostly with H. telesiphe while, in the Cordillera Escalera, the ‘postman’ community is dominated by H. erato favorinus and H. melpomene amaryllis (Fig. 1). Theoretical simulations suggest that the most abundant species of a mimicry ring generally drives the evolution of phenotypic resemblance in other species (Turner 1977; Mallet 1999; Franks & Sherratt 2007; Ruxton et al. 2008). We therefore hypothesize that the subtle variation in colour pattern found in certain species participating in the mimicry ring might be the footprint of selection for different mimetic optima, reflecting spatial changes in the phenotypic composition of the different communities.
To test whether the local mimetic community may influence the strength and the nature of selection for mimicry, we investigate whether geographic variations for various modalities of the warning signals (colour pattern, reflectance of the colour patch, wing shape) within participating species appear to track variations in composition of the mimetic community. We use colour pattern modelling (Le Poul et al. 2014), geometric morphometrics and spectral colour measurements to quantify phenotypic similarity between comimics and analyse it in the light of neutral molecular variation, hybridization rates and species composition in the distinct communities.
Materials and methods
Species studied, specimen collection and density
The four species of the ‘postman’ mimicry ring (Fig. 1, H. melpomene amaryllis Feder & Feder, H. timareta thelxinoe Lamas & Mérot, H. erato favorinus Hopffer and H. telesiphe Doubleday, further abbreviated with their species name only) were sampled in two tropical montane areas separated by 175 km, the Escalera and the Alto Mayo (San Martín, Peru). Sampling localities were chosen along an altitudinal continuum, which ranges from 400 m to 1300 m in the Escalera (‘E’, 06°27′28″S; 76°17′53″W) and from 1100 m to 1800 m in the Alto Mayo (‘A’, 05°39′58″S; 77°44′35″W).
To estimate the relative frequencies of each species, we used collection data corrected by the number of collecting days. On collecting days, all butterflies from the four species encountered were caught with entomological nets. A subset of this sample was used for genetic and phenotypic analysis (Table S1 and S2, Supporting information). We considered two populations of H. melpomene in the Escalera (‘Low E’, below 1000 m; and ‘Escalera’, above 1000 m, sympatric with H. timareta). We also included an additional population of H. melpomene from Moyobamba (06°05′13″S; 76°59′36″W, Peru) to investigate geographic variation in H. melpomene between area of sympatry and allopatry with H. timareta.
Phenotypic description and analyses
Data acquisition
Images of ventral (v) and dorsal (d) forewings (FWs) and hindwings (HWs) were captured in normalized light conditions (CIE standard illuminant D50) using a Nikon D90 digital camera with a Nikon micro 105/2.8GEDVR lens.
Measurements of wing reflectance were done with a spectrometer (AvaSpec-3648, Avantes) and a deuterium–halogen light source (DH-200, Avantes) connected to a 1·5 mm diameter sensor (FCR-7UV200-2-1·5×100, Avantes) inserted in a miniature black chamber (an opaque black plastic tube surrounding the reflectance probe to exclude ambient light from the measurement). Reflectance spectra were taken at 90° incidence relative to a 99% reflectance standard (300–700 nm spectralon) and to dark current. Spectra were recorded with the software Avasoft 7.0 using an average of five measures with an integration time of 23 ms. On all wing surface (FWv, FWd, HWv, HWd), we recorded reflectance of colour patches and black area at the same location for each specimen.
Wing colour pattern
Colour pattern was analysed using colour pattern modelling (Le Poul et al. 2014), which allows quantifying colour pattern variation and similarity across the entire wing. Detailed methods are given in supplementary material. Briefly, wing outline was extracted individually from the background. Within this area, the RGB colours are categorized into three colour classes (black, red or yellow). We called ‘yellow’ the bar on all species hindwing so that pattern itself would be comparable, despite some hue variation between species and between individuals of different ages or wear. All individual wings were aligned by rotation, translation, scaling and normalization based on an iterative process. Variations in patch boundaries can then be fully described and compared between individuals.
Each wing surface (FWd, FWv, HWd, HWv) was characterized by a set of pixels with homologous position across specimens. Each pixel was associated with three presence/absence binary values for black, red and yellow, which allows colours to be treated separately. Variations in patch boundaries were then analysed for each surface separately with a principal component analysis applied to the set of pixels. The resulting components describe a morphological space used for subsequent analyses after the dimensionality reduction proposed by Baylac & Friess (2005).
Small colour elements located in the basal part of the wing were scored following Mérot et al. (2013) with an index describing variation at two qualitative characters: the red line on the ventral forewing (0 = absent to 5 = a well-marked line, Fig. S1a) and the extension and the number of basal red spots on the ventral hindwing (0 = no dots to 5 = large spots, Fig. S1b).
Wing outline
Wing outline, extracted and aligned through the first step CPM procedure, was further analysed with elliptical Fourier analysis using custom scripts developed in Matlab (Jones et al. 2013), following the directions of (Neto et al. 2006). A PCA was applied to the first twenty Fourier harmonics and allowed describing a morphological space for wing outline.
Wing venation: analysis by geometric morphometrics
Wing venation was described using 15 (FW) and 14 (HW) landmarks, placed at vein intersections and vein termini on the ventral side, as described in Mérot et al. (2013). Standard tests of repeatability were done by taking the landmarks five times per wing on subsamples of five butterflies from a single species, population and sex. Landmark coordinates were digitalized using TpsDig2 (Rohlf 2010) and superimposed using a general Procrustes analysis (Bookstein 1991; Zelditch et al. 2004). For each set of landmarks, superposition includes all samples in a multidimensional Procrustes space whose tangent space is a Euclidian morphological space for wing venation. A principal component analysis was applied on this wing venation data followed by the dimensionality reduction proposed by Baylac & Friess (2005). Wing size was measured using log-transformed centroid size (Bookstein 1991).
Colour spectra and visual models
Colour spectra were analysed using Avicol V.6 (Gomez 2006). Colour spectra obtained between 300 and 700 nm were smoothed using a local Fourier correction at 650 nm (Fig S5 and S7). A mean colour spectrum was calculated for each colour patch of each individual by averaging measurements on the left and the right wing. Each kind of colour patch was further analysed separately.
Analyses of spectra were carried out using models of animal vision, which take into account the observer's vision and the illuminating light. Each photoreceptor is characterized by a sensitivity function, which determines the wavelength of reflected light perceived by the eye. Analyses in the main text are with a tetrachromatic V-type bird visual system [Peafowl, Pavo cristatus, (Hart 2002)] and a light environment corresponding to large sunny gaps in a tropical forest (Théry, Pincebourde & Feer 2008). Analysis with other visual system and other incident lights gave consistent results and is presented in supplementary materials (Fig. S9). A physiological model of Endler & Mielke (2005) was applied to visualize colour distribution in an unconstrained space, a tetrahedron whose vertices corresponds to the four photoreceptors (Fig. S6). Within this colour space, the relative location of each individual's patch colour can be compared and the overlap between two clouds of points (Table S6) can be calculated. Relative excitation of each photoreceptor was treated as multivariate data and analysed with PCA analyses.
Statistical analysis
Within each species, differences in colour pattern, colour spectra, wing outline and venation between geographic populations were tested by a one-way manova on each subset of PCs with geographic origin as a factor. Then, within each species and for each trait, we tested discrimination between groups defined by geographical populations with a linear discriminant analysis. We compared the cross-validation values to discrimination between simulated populations of similar size following the methods described in Evin et al. (2013).
Phenotypic similarity between each species and each population of H. timareta was quantified with several methods within each morphological space. First, to measure the magnitude of phenotypic similarity between individuals, Euclidian distances in the morphological space were computed between all pairs of specimens. For colour spectra, we used perceptual chromatic distances between all pairs of specimen, expressed in JNDs (just noticeable differences) in a perception model of Vorobyev & Osorio (1998). This model takes into account the noise due to errors in photoreceptor response and estimates more accurately whether the discriminability between colours could really be perceived by predators. Perceptual and Euclidian distances between a mimic and each population of H. timareta (Alto Mayo vs. Escalera) were analysed with a mixed model with geographic population as factor and identity of the compared specimen as random factors. Secondly, to compute an indicator of similarity between groups, Mahalanobis distances were calculated between the three species and the two populations of H. timareta. Contrary to Euclidian distances, this measure applies in a transformed space and is scaled by intragroup variation. Thirdly, to confirm our similarity estimates, we calculated an index of cross-classification by performing a linear discriminant analysis on a subset of two groups [e.g. H. melpomene/H. timareta (A); H. melpomene/H. timerata (E)]. Since the latter two analyses are sensitive to unbalanced sample sizes (Kovarovic et al. 2011), we performed them on subsets of similar sizes, by randomly drawing 30 specimens from each group, repeating the procedure 1000 times and averaging over the repetitions.
All analyses were performed in r 3.0.3. (R Core Team 2014) using the packages ade4, nlme, Mass and Rmorph (Baylac 2012).
Multilocus microsatellite analysis
The multilocus microsatellite genotype, performed on H. melpomene and H. timareta, is based on 11 loci and is an extension of the genotyping performed in Mérot et al. (2013) with the addition of 31 H. melpomene amaryllis and 69 H. timareta thelxinoe. Detailed methods can be found in supplementary materials; genetic diversity and statistics in Table S3. F-statistics were calculated using genetix 4.05 (Belkhir et al. 1996-2004). We used STRUCTURE (Pritchard, Stevens & Donnelly 2000), a multilocus Bayesian clustering method, to determine population structure, to confirm the assignment of individuals to species and to detect admixed genotypes. STRUCTURE was run with 500 000 updates of the Markov chain after an initial burn-in of 50 000 updates, to achieve chain convergence for a set of models with different numbers of clusters (K = 1–6). Since H. melpomene and H. timareta can hybridize and disperse, we used the ‘admixture model’ and ‘correlated allele frequencies’. The most likely K was chosen using the maximum likelihood value in structure and the ΔK-method (Evanno, Regnaut & Goudet 2005) and then evaluated following recommendations from STRUCTURE documentation. Over the entire set of specimens included, the likelihood reaches a maximum plateau between K = 2 and K = 3. When K = 2, the two clusters correspond to the two species (identified phenotypically), so K = 2 bears a clear biological meaning and was retained. K = 3 further splits H. timareta into two clusters with no obvious association with biological variables (such as geographic populations), nor with any identifiable experimental bias. Further structure analyses run on each species separately support the absence of intraspecific clustering. Posterior probabilities of being member of a cluster were estimated and allowed detecting potential hybrids. Hybrid detection was also run separately with NewHybrids 1.0 (Anderson & Thompson 2002). Relying on the results from simulated hybrids (Mérot et al. 2013) and hybrids from controlled crosses (Mérot et al. 2015), individuals were considered as ‘pure’ if the posterior probability was above 0·9 in structure and above 0·7 in NewHybrids.
Results
Geographical variation within species
Populations of H. timareta from the Escalera and the Alto Mayo exhibit significant differences in colour spectra, pattern, wing outline and venation (Table 1). This result was consistent whether males and females were treated together or separately. A linear discriminant analysis on H. timareta allows discriminating the two populations, with a low level of misclassifications, better than the range of misclassification obtained from randomly simulated population (Table 1). Centroid sizes did not exhibit significant differences between the two populations of H. timareta (anova FW P = 0·12, HW P = 0·77).
| d.f. | F | Pillai | P | Misclassification (LDA) (%) | |
|---|---|---|---|---|---|
| Colour spectra | |||||
| Red FWd | 1,132 | 13·5 | 0·24 | < 0·001 | 28 |
| Red FWv | 1,133 | 8·9 | 0·17 | < 0·001 | 39 |
| Yellow HWd | 1,134 | 8·7 | 0·17 | < 0·001 | 29 |
| Yelow HWv | 1,135 | 7·9 | 0·15 | < 0·001 | 34 |
| Colour pattern | |||||
| Red patch FWd | 1,136 | 16·5 | 0·66 | < 0·001 | 13 |
| Yellow bar HWd | 1,137 | 18·9 | 0·70 | < 0·001 | 11 |
| Wing venation | |||||
| FW | 1,138 | 7·1 | 0·68 | < 0·001 | 30 |
| HW | 1,139 | 7·0 | 0·69 | < 0·001 | 27 |
| Wing outline | |||||
| FW | 1,133 | 5·7 | 0·29 | < 0·001 | 23 |
| HW | 1,120 | 5·3 | 0·30 | < 0·001 | 21 |
Within H. melpomene and within H. erato, neither colour spectra, wing outline nor size displayed any significant geographical phenotypic variations (Table S4 and S5). For pattern and venation, slightly significant differences were observed between populations of H. melpomene and H. erato. However, discrimination between geographic populations reaches 41–71% of misclassifications, which is within the misclassification range from randomly simulated populations of similar sample size.
Relative frequency of postman comimics
The Escalera and the Alto Mayo areas display differences in their geography and topography, their community composition and the altitudinal ranges of local species (Fig. 1c). The Escalera is a relatively thin, mid-elevation Eastern Cordillera jutting from adjacent Amazonian lowlands. On its slopes below 1000 m, H. erato and H. melpomene are the only species with a postman pattern and are generally very abundant, at our collection places and elsewhere in the foothills (> 200 specimen of each species in 65 days, approximately 1 : 1 ratio, Table S2). In the highest areas, between 1000 and 1300 m, H. erato becomes less abundant while H. melpomene co-occurs with H. timareta. The Escalera is the southernmost tip of the known distribution of H. timareta and the latter can be as abundant as H. melpomene, but only locally, as observed at our collection places (1 : 1 ratio, Table S2). At a finer scale, H. timareta gradually increases in frequency relative to H. melpomene (Fig. 1c). Despite intensive collection (83 days, Table S2), only two specimens of H. telesiphe were found in the Escalera, and only at the highest collection station (1300 m).
The Alto Mayo represents a wider, higher and more continuous area of montane forest surrounded by Andean areas. There, H. melpomene and H. timareta display less overlap since the transition is more abrupt and occurs at higher elevation (1300 m, Fig. 1c). H. timareta was never found below 1200 m and was locally abundant at one collection station (1400–1600 m). Over 1300 m, H. telesiphe is an abundant species in the Alto Mayo and the commonest postman-patterned species. H. erato was not found in our standardized collection stations of the Alto Mayo although it is occasional up to 1200 m, and thus parapatric with H. timareta.
Variation in wing phenotype in the postman mimetic community
Wing colour pattern
For forewing colour pattern, the first axis of the PCA (Fig. 2a) displays variation from the large rounded red patch of H. erato and H. melpomene, to the dislocated, slender patches H. telesiphe. H. timareta is intermediate on this axis with the Escalera population closer to H. melpomene and H. erato. H. telesiphe is sitting in a quite distant position overall, because it exhibits a small additional subapical red patch on the forewing, missing in all other species. Nevertheless, the Alto Mayo population of H. timareta is the closest population to H. telesiphe in the colour pattern space because of its somewhat dislocated, zigzagging red patch shape, reminiscent of the main red patch of H. telesiphe.
On the hindwing, all four species display a yellow/white pattern barring both sides of the hindwing longitudinally. The first PC shows variation in the width of the bar. H. melpomene and H. erato have a rather wide and long yellow bar while H. telesiphe has a thinner and shorter bar (Fig. 2b). Again, H. timareta appears in intermediate position along this direction of variation, with the Escalera population having a wide yellow bar like H. melpomene, while Alto Mayo populations show a thinner bar like H. telesiphe. The second PC shows a continuous gradient from a long and pointed bar seen in H. erato to the thin bar characteristic of H. timareta from the Alto Mayo, straight and wide in the proximal part and curved posteriorly in the distal part. H. melpomene and H. timareta from the Escalera sit in an intermediate position.
Results were consistent on the ventral colour pattern (Fig. S2) and whether males and females were treated together or separately (Fig. S3).
Small elements of pattern
All specimens of H. telesiphe exhibit bright and large red spots on the ventral hindwing and a well-marked red line on the ventral forewing (index = 5) (Fig. S1). By contrast, H. erato nearly never displays red dots or lines except in a few specimens (4–18%, depending on population) that possess small faded spots or a slight line (index = 0–1). Along this gradient between H. erato and H. telesiphe, most H. melpomene samples are closer to the H. erato phenotype, with no red line (89–100% depending on the population) and small faded spots (89–96%). Then, 45 to 47% of H. timareta samples from the Escalera exhibit small red spots (index = 2–3) and no or a tiny red line (index = 0–3), making them closer to the H. erato and H. melpomene phenotypes. On the contrary, a well-marked red line and large red dots (index = 4–5) are observed in more than 93% of H. timareta samples from the Alto Mayo (vs. only 53–55% from the Escalera population).
Wing venation and outline
For forewing shape, results were congruent whether described by venation or by wing outline. For both venation and outline, the first PCs display continuous variations from the rather rounded forewings (Fig. 3a,c) of H. melpomene, shared by H. timareta from the Escalera, towards the thinner and more elongated wings of H. telesiphe. By contrast, H. timareta from the Alto Mayo display resemblance to H. telesiphe through its more elongated wing shape. In the venation analysis, which takes into account the discal cell and the distal veins, the second PC recapitulates phylogenetic relationships, separating H. erato and H. telesiphe from H. melpomene and H. timareta.
For the hindwing shape, the first PCs of both the venation and outline analysis are mostly associated with sexual dimorphism (presence/absence of male androconia), but the second PCs display the same variation as the forewing, from the rounded wings of H. melpomene to the thin, elongated wings of H. telesiphe (Fig. 3b,d). Both H. timareta populations are in intermediate position.
Colour spectra
For both forewing and hindwing, the reflectance of colour patches also follows a continuum between the four species from H. erato and H. melpomene to H. telesiphe (Fig. 4). The two populations of H. timareta occupy different positions along this continuum in the colour space. On the one hand, the Escalera population displays similar colours to H. melpomene and H. erato (red–orange and bright yellow), with which it overlaps in the three-dimensional colour space (Fig. S6, Table S6). On the other hand, the position of H. timareta from the Alto Mayo comes closer to H. telesiphe for the red forewing patches (with a deeper red) and, to a lesser extent, for the whitish hindwing patch.
It is notable that colours are generally lighter on the ventral side than on the dorsal side and exhibit more intra- and interspecific variability. On the ventral side, both red and yellow patches of H. timareta and H. telesiphe reflect more short wavelengths and UV than H. erato and H. melpomene. The higher variability in colour on the ventral side is also shown by the analyses of perceptual distances, since discrimination was higher for the ventral side than for the dorsal side. On the ventral side, visual contrast within species was smaller than between species, for both red and yellow and reaches 2–6 JND (Fig. 5i,j). On the dorsal side, within-species and between-species visual distances spanned a similar range (2 JND, Fig. 5g,h), except between H. telesiphe and all others for the white dorsal hindwing patch.
Similarities in wing phenotype in the postman mimetic community
For forewing and hindwing colour patterns, forewing shape and hue of all patches, resemblance indicators suggest that the Escalera population of H. timareta is more similar to the most abundant local comimics of the Escalera (H. erato, H. melpomene) than the Alto Mayo population is. Euclidian and perceptual distances as well as Mahalanobis distances between H. timareta and H. erato or H. melpomene are smaller for the Escalera population than for the Alto Mayo population (Table S7, Fig. 5). The reverse was found for distances with H. telesiphe, suggesting that the Alto Mayo population of H. timareta is more similar to H. telesiphe than the Escalera population is. Those results are supported by the differences in cross-classification rates from the discriminant analysis (Table S8). Results were consistent for males and females taken independently and together.
For hindwing shape, similarity estimates do not show any consistent trend differentiating the two geographic populations of H. timareta.
Genetic divergence and population structure
No genetic differentiation was found between the two geographic populations of H. timareta. The FST value over all loci was estimated at 0·025, and the two populations do not split into different genetic clusters. For H. melpomene, no genetic differentiation was found either: FST reaches only 0·009 between the four populations, and they do not display any genetic clustering.
Genetic differentiation between H. melpomene and H. timareta was significant in the two geographic areas and of similar magnitude (A: FST = 0·148; E: FST = 0·157).
Hybrid detection
The level of admixture follows the expected trend, with slightly higher rates in the Escalera, where H. timareta and H. melpomene populations show more altitudinal overlap and are the most phenotypically similar. However, this difference was not significant, possibly due to our limited samples sizes (test of equal proportions: P = 0·67).
Bayesian clustering analyses with structure detected 4·2% of admixed individuals in the Escalera (Fig. S11, two back-crosses to melpomene, two back-crosses to timareta and 3 F1 hybrids out of 167 specimens), and 3·2% in the Alto Mayo (one back-cross to melpomene, 1 F1 or back-cross to timareta out of 63 specimens). Surprisingly, given that H. timareta does not occur below 1000 m, one admixed specimen was detected in the low Escalera population at about 700 m of elevation. Assignment tests with NewHybrids 1·0 found the same admixed individuals except in one case, but also pointed out six more putative back-crosses, raising the proportion of admixed individuals to 8·7% in the Escalera and 6·3% in the Alto Mayo.
Discussion
Tracking variations in the local prey environment
Our results show that despite an absence of genetic differentiation at neutral markers, geographic populations in one of the species, H. timareta, display subtle but consistent variations in wing phenotype, associated with enhanced similarity to locally abundant comimics. Our results therefore denote a geographic shift in mimicry association consistent with quantitative changes in the composition of the mimicry community.
Geographic mosaics of covarying warning pattern are commonly observed in animal clades involved in mimicry associations, from butterflies to frogs (Thompson 2005; Twomey et al. 2013). An iconic case is the faithful covariation of Heliconius melpomene and H. erato throughout their shared range, the two butterflies switching in concert from red postman pattern (in certain Andean valleys, south-eastern Brazil and Carribean area) to orange-rayed Amazonian patterns (Sheppard et al. 1985; Reed et al. 2011). Our results suggest that comimetic species not only exhibit adaptive shifts of colour pattern but can also track one another geographically for more subtle variations within a type of pattern. Quantitative variations in pattern and hue appear adaptive and match local specificities of the surrounding communities of comimetic prey. Such partial phenotypic divergence represents an intermediate situation between undifferentiated populations and colour pattern races participating in disjoint mimicry assemblages, opening a window on the process by which diversity within species may evolve. This also brings support to the hypothesis that the variation in abundance of certain key species can influence the local mimicry optimum in a way that translates into changes in the selected phenotypes of other species, a process which was proposed to contribute to mimicry diversification in the face of selection for resemblance (Turner & Mallet 1996).
Geographic variations were not found, at our scale, within H. melpomene amaryllis nor H. erato favorinus, two species which are generally abundant and widespread compared to H. timareta. This observation fits the prediction that the commoner species should be less influenced by selection for perfect resemblance in a Müllerian association (Ruxton, Sherratt & Speed 2004; Ruxton et al. 2008). However, in other cases, where H. melpomene is not the dominant species, such as in the Amazon lowlands, subtle geographic variations are reported where they match the local butterfly communities. For instance, among the orange-rayed Amazonian forms, local populations from the Marañon valley exhibit a partial lack of orange rays, which presumably enhance their resemblance to the commonest local species, H. himera (J. Mallet, pers. com.). Similarly, the two parapatric races named H. melpomene aglaope and H. m. malleti represent a geographic continuum of populations whose wing pattern varies quantitatively in the width and shape of the yellow forewing patch, matching similar variations in the patterns of local H. erato populations and likely other comimetic species. Together with those examples, our results suggest that coordinated quantitative covariations of the elements of the warning signal are adaptive and, despite receiving little quantitative attention, reflect the tracking of a geographically changing optimum defined by a community of multiple mimetic species.
The multimodality of traits involved in mimicry
Resemblance between populations of H. timareta and the local comimetic species appear congruent over the different traits measured in our study: colour pattern, hue and wing shape, suggesting a multimodal and coordinated evolution of resemblance. This is consistent with recent studies showing that, although colour might have a greater influence on predator decision, pattern features also play a significant role in signalling (Finkbeiner, Briscoe & Reed 2014).
Whether wing shape is also under selection for mimicry is unclear. Wing venation itself cannot presumably be seen by predators, but it may be indirectly under selection since wing motion and wing outline are associated with mimicry (Srygley 1994; Jones et al. 2013). Long and narrow forewings (high aspect ratio) are generally associated with fast and extended flight while broader wings may be better suited for slow and more fluttery flights (Betts & Wootton 1988), and H. telesiphe displays a notably faster and more elusive flight than H. erato and H. melpomene (pers. obs.). Flight behaviours may be perceived by bird predators as informative characters for prey identification and may participate in the definition of the local mimicry optimum (Srygley & Ellington 1999), so the similarity in shape of H. timareta to H. telesiphe in the Alto Mayo and to H. melpomene in the Escalera may represent an adaptive response to selection on flight patterns.
A striking result from the spectral analysis is that, although both wing sides show a signal of local mimicry adaptation, hue on the dorsal side is less variable (within and between species) than on the ventral side. In butterflies, ventral and dorsal patterns can be genetically correlated as shown by experimental evolution lines of Bicyclus anynana (Beldade, Koops & Brakefield 2002) but also strikingly different, as in Morpho for instance, which exhibit a bright blue dorsal side and a camouflaged ventral side. Pattern differences between the ventral and dorsal sides then generally result from different selective pressures. Here, the lower variance measured on dorsal hue presumably reflects the fact that dorsal patterns in Heliconius are more frequently visible to predators in full light during sunny hours, when the butterflies are flying or basking with wings spread out, translating into stronger natural selection on hue resemblance. By contrast, ventral patterns are exhibited when resting in the shadow or roosting at night or at dusk or dawn (Mallet 1986; Finkbeiner, Briscoe & Reed 2012), in a light environment where colour differences are less noticeable than in sunlight (Théry, Pincebourde & Feer 2008). Interspecific differences in ventral hue, especially in UV reflectance, might also be associated with intraspecific social and sexual communication (Silberglied & Taylor 1978).
The evolution of accurate mimicry
Most variations in resemblance quantified between comimics here are quite subtle. All populations belong to the general ‘postman’ wing pattern, but the precision of mimicry is enhanced between taxa from the same locality, as expected if predators select for mimicry refinement. A long-standing debate is whether evolution takes place by small gradual or large changes (Punnett 1915; Fisher 1927; Nicholson 1927; Fisher 1930). Mimicry evolution is classically suggested to occur first via a large phenotypic change, which allows coarse resemblance, then by a phase of more subtle refinement (Sheppard et al. 1985; Turner 1987; Franks & Sherratt 2007; Balogh et al. 2010; Leimar, Tullberg & Mallet 2012). Introgression of alleles of the optix locus from H. melpomene, controlling the presence of the red forewing patch, is thought to have been a major step allowing H. timareta to join the postman mimicry ring (Heliconius Genome Consortium 2012; Pardo-Diaz et al. 2012). Our results describe the second phase of mimicry evolution in which small variations in shape, colour and pattern improve the match to the local optimum. Those mimicry improvements may have evolved through the selection of allelic variants at optix and other unlinked loci modifying wing morphology and pattern. In H. melpomene, the dislocated vs. rounded shape variation of the red forewing patch is associated with a quantitative trait locus (QTL) unlinked to optix (Baxter, Johnston & Jiggins 2008). Differential selection towards different mimetic optima determined by changes in the local species communities would then be predicted to affect such ‘secondary’ mimicry-refining loci, which act epistatically with the major mimicry switches already identified.
A role for gene flow?
The process of gradual evolution is particularly efficient in populations with a rich supply of genetic variation for traits contributing to the warning signal (Ruxton et al. 2008). Apart from mutation and intraspecific migration, gene flow from closely related species through hybridization may be an important source of variation (Grant & Grant 1994). Hybridization with local species may bring new, locally adapted alleles, and our data show that a certain percentage of F1 hybrids and admixed individuals between H. melpomene and H. timareta are found in natural populations. Genomewide signatures of past and ongoing gene flow are documented between H. melpomene and H. timareta, and introgressed alleles of the red-patterning gene optix from H. melpomene determine the general postman wing pattern in H. timareta of Northern Peru (Heliconius Genome Consortium 2012; Pardo-Diaz et al. 2012; Martin et al. 2013). The occurrence and timing of the introgression need to be clarified more finely to assess its role in the differentiation we report between the two locations. For instance, one hypothesis is that resemblance may have had more time to evolve in one population if the postman pattern is older there. Similarly, recurrent interspecific gene flow for adaptive alleles at optix and other loci readily favoured by selection may participate in the refinement of mimicry in places where H. melpomene determines the position of the optimum and overlaps with H. timareta.
One intriguing result here is that traits which might not be readily perceived by predators (small red spots, costal line, forewing venation) also display a higher similarity between H. timareta and H. melpomene in the Escalera than in the Alto Mayo. If those traits are indeed subject to weaker selection by predators (cf. the stronger resemblance on dorsal phenotypic traits), their similarity might reflect local gene flow between hybridizing H. melpomene and H. timareta. Our genetic analysis does not support higher gene flow in the Escalera than in the Alto Mayo, but the resolution is limited by the small number of markers. We suggest it would be worth exploring the contribution to resemblance of interspecific gene flow in this and similar systems.
Conclusion
Here, we showed that subtle variations in wing pattern across short geographic distances allow populations of comimics to track gradual changes in the mimicry optimum, which suggest that predators do not generalize widely and that selection is operating to maximize mimicry efficiency even between relatively similar warning signals. This demonstrates that fitness peaks in the morphological (signalling) space are sharp and that the quantitative movements of the position of the optimum are of strong adaptive significance for certain species of the community. Effectively, this amounts to H. timareta switching mimetic association and to our knowledge, this is the first quantification of such a mimicry switch in a community context. All traits quantified here display coordinated variations, suggesting multimodal and gradual evolution of multiple traits improving the level of resemblance and describing the second phase of colonization of a new fitness peak. Our study also raises the possibility that the observed geographic pattern of adaptation and the accuracy of mimicry might partly be enhanced by interspecific gene flow providing adaptive alleles readily favoured by selection.
Acknowledgments
We thank V. Llaurens and D. Gomez for their help with colour analysis, A. Evin and V. Debat for their help with geometric morphometrics and people involved in fieldwork. We are grateful to Jim Mallet for introducing us to this study system and for helpful comments on this manuscript. We are grateful to Evan Twomey and one anonymous reviewer for their valuable comments on this manuscript. We thank the Peruvian Ministerio de la Agricultura (SERFOR), the SERNANP-BPAM and PEHCBM-ACR-CE and the Museo de Historia Natural for research and export permits in Peru. This work was supported by ERC starting grant MimEvol and ANR-JCJC HybEvol to MJ.
Data accessibility
Data available from the Dryad Digital Repository: http://dx.doi.org/10.5061/dryad.h4j6c (Mérot et al. 2016).
