The diversification of Heliconius butterflies: what have we learned in 150 years?

What is a (Heliconius) species?

Although arguments over the ‘correct’ species concept may be sterile, examining the nature of species can help us more carefully consider the evolution of diversity. Species are often considered the fundamental unit of biodiversity, but distinguishing between species and lower taxonomic levels has proved problematic. The widely adopted Biological Species Concept (BSC) (Mayr, 1942) is useful because it explicitly addresses the mechanisms (i.e. reproductive isolation) that result in the discontinuities we observe in nature. However, the strength of reproductive isolation observed between taxa, including among species and subspecies of Heliconius, is often broadly continuous (Mallet et al., 2007; Merrill et al., 2011a; Nosil, 2012). Even if we disregard a strict BSC, where genetic barriers must be absolute, the degree of reproductive isolation required for species status is arbitrary. In addition, it has been noted that if ‘gene exchange were widespread and substantial between sympatric taxa’, it would present a serious problem for the BSC (Coyne & Orr, 2004; p. 41). Recent genomic analysis of Heliconius species exposes just such a challenge (Martin et al., 2013) (Fig. 2).

Studies of hybridization in Heliconius twenty years ago led to an alternative to the BSC – termed the Genotypic Cluster Definition (Mallet, 1995). In contrast to the BSC, Mallet's (1995) Genotypic Cluster Definition takes a descriptive approach to species delimitation using sympatric coexistence of distinct multilocus genotypes as the defining character of a species, without prejudice as to the processes that maintain species. This genotypic cluster approach is now broadly applied in Heliconius, with a conservative approach to the elevation of taxa to species status. In other words, many populations with different wing-pattern phenotypes are considered subspecies, or geographic races, because they form transition zones with abundant immediate forms where they meet. The ~30 named races of Heliconius erato, for example, have distinctive colour-pattern phenotypes, but belong to the same species. In contrast, the geographic replacement of H. erato found in south eastern Ecuador, H. himera, is considered a distinct species because, where the two species co-occur, hybrids are rare (but still account for ~10% of individuals) (Jiggins et al., 1996). In other words, the Genotypic Cluster Definition implies that where they co-occur, ‘species’ are characterized by a bimodal distribution of traits, even if gene flow persists. By emphasizing the importance of multilocus genotypes, the Genotypic Cluster Definition is a useful tool for investigating gene flow and the maintenance of distinct species, and the genomic architecture of gene flow and divergence.

A genomewide view of porous species

The sequencing of a Heliconius reference genome (The Heliconius Genome Consortium, 2012), and resequencing of additional taxa (Kronforst et al., 2013; Martin et al., 2013), makes it possible to infer genomewide patterns of divergence. The patterns observed broadly support the adoption of a genotypic clustering approach towards species delimitation. Below the species level, low levels of divergence are observed between colour-pattern races across most of the genome, but strong differentiation is seen at a handful of loci known to be under divergent selection (The Heliconius Genome Consortium, 2012; Kronforst et al., 2013; Martin et al., 2013). Between-species comparisons, on the other hand, reveal much higher levels of genomewide divergence despite convincing evidence that large-scale gene flow persists due to occasional hybridization (Kronforst et al., 2013; Martin et al., 2013). Significant phylogenetic discordance (i.e. where different parts of the genome imply different evolutionary histories) has been frequently observed in Heliconius (Beltran et al., 2002; Bull et al., 2006; Kronforst et al., 2006a; Martin et al., 2013). Although some level of discordance is expected due to the stochastic nature of gene coalescence, there is now strong evidence for porous species boundaries that permit gene flow for millions of years after initial divergence. For example, across 40% of its genome, H. melpomene from Panama is more closely related to the population of H. cydno with which it co-occurs than to an allopatric population of H. melpomene in French Guiana (Martin et al., 2013) (Fig. 2).

Genomic divergence may begin at a few narrow regions containing key loci under selection, and these islands of divergence may grow as speciation proceeds, eventually expanding to encompass the whole genome (Wu, 2001; Feder et al., 2012a). The narrow peaks of differentiation observed between recently diverged Heliconius races, and the widespread genomic divergence between species, are consistent with the early- and late-stage expectations of this islands-of-divergence model. Whether this is pertinent to our understanding of speciation depends on whether populations divergent at just a handful of narrow genomic regions do in fact represent incipient species. It has been argued that Heliconius wing-pattern races, which are structured at just a few loci of large phenotypic effect in otherwise genetically mixed populations, do not (Cruickshank & Hahn, 2014). However, population divergence at a few loci under strong selection is a form of reproductive isolation, albeit localized within the genome. More significant is the mass of evidence that divergence in colour pattern plays a key role – alongside the evolution of additional traits – during Heliconius speciation. Regardless, studies in Heliconius have demonstrated that selection may initially limit gene flow at discrete regions of the genome involved in adaptive divergence.

Elsewhere, divergence between forms of threespine sticklebacks and European crows resembles the proposed initial stages of speciation (Hohenlohe et al., 2010; Jones et al., 2012; Poelstra et al., 2014), whereas it could be argued that divergence observed between host races of Rhagoletis and sister species of Anopheles gambiae and Ficedula flycatchers looks more like the later stages (Lawniczak et al., 2010; Michel et al., 2010; Ellegren et al., 2012; Hahn et al., 2012). Thus, the ability to sample Heliconius genomes across a continuum of divergence greatly informs our understanding of speciation genomics. Nevertheless, there are arguably no empirical examples of the transition between these two states either from Heliconius or from other taxa. The lack of intermediate stages may be expected if a rapid and unstable transition occurs at the species boundary (Feder et al., 2012a). This is perhaps the critical transition in the evolution of species. Selection maintaining initial localized divergence may later provide opportunities for further differentiation in physically linked genomic regions (‘divergence hitchhiking’; Via & West, 2008; Via, 2012; Feder et al., 2012a) or across the genome more generally (‘genome hitchhiking’; Feder & Nosil, 2012). Examples of the transition between the two states would therefore be important for understanding the role of these two forms of hitchhiking (Via & West, 2008; Feder & Nosil, 2012; Feder et al., 2012b; Via, 2012). As more genomes are sequenced, this apparent gap in our sampling may well be overcome.

Mimicry: can selection lead to a predictable genome?

The spectacular mimicry in Heliconius has become a textbook example of natural selection (e.g. Barton et al., 2007). Cyanogenic glycosides, either synthesized de novo or sequestered from host plants as larvae, render Heliconius butterflies unpalatable to vertebrate predators (Engler et al., 2000; Cardoso & Gilbert, 2007; Engler-Chaouat & Gilbert, 2007; Hay-Roe & Nation, 2007; Chauhan et al., 2013). Predators include insectivorous birds that learn the wing patterns of distasteful prey and subsequently avoid them (Brower et al., 1963; Chai, 1986; Chai, 1988; Pinheiro, 1996; Langham, 2004). Unpalatable prey therefore benefit from protection against predators by displaying colour patterns similar to other unpalatable species, a widespread phenomenon known as Müllerian mimicry (Müller, 1879). Despite the difficulty of observing predation in the wild, three lines of evidence confirm the importance of mimicry as an adaptation in Heliconius: (i) evidence for learning and sight rejection based on previous experience from focal predators studied in cages (Chai & Srygley, 1990; Merrill et al., 2012) or in the wild (Langham, 2004), (ii) higher recapture rates of released butterflies that match the local mimicry community (Benson, 1972; Mallet & Barton, 1989; Kapan, 2001) and (iii) lower attack rates on artificial butterflies matching local comimics (Merrill et al., 2012; Finkbeiner et al., 2014).

Estimates of selection coefficients favouring mimicry are high, whether calculated directly from the recapture rates of released butterflies [overall s = 0.52 in H. erato (Mallet & Barton, 1989); s = 0.64 in H. cydno (Kapan, 2001)] or indirectly from cline width and linkage disequilibrium measurements across hybrid zones (per locus s = 0.13–0.40 in H. erato and H. melpomene) (Mallet et al., 1990; Rosser et al., 2014). Each mimetic wing pattern therefore represents a towering fitness peak in the adaptive landscape, driving convergence across a wide diversity of prey species sharing a habitat. However, despite strong evidence for intense selection for mimicry in Heliconius, we still have limited knowledge about the actual communities of predators, and there are difficulties in obtaining direct estimates of predation rates in the wild. This has hindered our understanding of how variation at individual loci controlling wing patterns directly affects fitness in the wild, a problem common to many systems (Barrett & Hoekstra, 2011).

What factors influence the fitness landscape?

Fitness landscapes are often simplified to a single isolated trait, but in reality they must include multiple interacting traits. Heliconius warning signals provide a tractable example of a phenotype comprising a composite of multiple characters, some of which may have additional functions (Sherratt & Beatty, 2003). In Heliconius, warning signals can involve the spatial arrangement of wing-pattern elements (see Brown, 1981; for a review), colour hue (Crane, 1954; Sweeney et al., 2002, 2003; Bybee et al., 2012; Llaurens et al., 2014), wing shape (Jones et al., 2013) and behaviour (Srygley, 1999, 2004, 2007; Finkbeiner et al., 2012). The addition of new characters to warning signals can enhance their efficiency (Sherratt & Beatty, 2003; Rowe & Halpin, 2013). However, we also expect trade-offs between ecological functions; for example, mimetic flight behaviours and wing shape carry aerodynamic costs (Srygley, 1999, 2004). Wing patterns are also involved in intraspecific communication and mate recognition, resulting in conflict between mimicry and intraspecific communication (Estrada & Jiggins, 2008). The relative fitness associated with butterfly warning patterns therefore results from the resolution of synergies and trade-offs between functions (Mallet & Gilbert, 1995; Salcedo, 2010), although little is known about how these interactions affect the diversity of patterns or the coexistence of species.

Despite this, local abundance is clearly central to the fitness of specific warning patterns. Positive number dependence (Müller, 1879; Turner, 1984; Mallet & Joron, 1999) underpins mimicry and leads to convergence among coexisting taxa. An individual's risk of predation decreases as the local density of its warning pattern increases: as numbers increase, local predators will have had greater opportunity to associate particular patterns with unpleasant experiences and so avoid them. Heliconius butterflies participate in diverse but often coexisting mimicry rings (groups of species that converge on the same warning phenotype). Hence, there may be several, distinct mimetic phenotypes in any given area, all of which impose strong stabilizing selection on participating taxa. This complexity, alongside variation in local predator and prey communities, shapes a rugged adaptive landscape with distinct fitness peaks (Turner, 1984). A given Heliconius species may join many distinct mimicry rings according to local context (Brown, 1979), leading to the well-documented spatial mosaic of mimetic races observed across the Neotropics in many widespread species, the most prominent being the parallel geographic radiations of H. erato and H. melpomene (Brown, 1976; Linares, 1997) (Fig. 1). A few species, such as H. doris, H. numata, and some populations of H. hecale, H. cydno and H. timareta, maintain local polymorphisms (Brown & Benson, 1974; Brown, 1976; Kapan, 2001; Chamberlain et al., 2009). In H. numata, for example, multiple forms participate in distinct mimicry rings dominated by butterflies in the tribe Ithomiini (Nymphalidae) (Figs 1 and 3), and the persistence of polymorphism appears to be maintained by a balance between migration and selection across mimicry communities (Joron et al., 2001; Joron & Iwasa, 2005). Nonetheless, there are also many widespread Heliconius species that are monomorphic across their range, and the reasons for extreme variability in some but not all species remain unclear.

Spatial variation in prey communities is therefore crucial to the diversification of warning signals. Nevertheless, the establishment of entirely new phenotypes is paradoxical under strict number-dependent selection. One further possibility is that during periods of relaxed selection, drift may allow new variants to rise above a threshold density until mimicry selection takes over (Mallet, 1993, 2010; Mallet & Joron, 1999; Sherratt, 2006; Chouteau & Angers, 2012). Stochastic changes in the abundance of butterflies carrying warning signals, leading to turnover in the composition of prey communities, may also cause fluctuations in the direction and intensity of selection (Turner & Mallet, 1996). Heliconius warning patterns provide one of the most convincing examples of Wright's shifting balance (Mallet, 2010).

Fitness valleys and genetic leaps

Evolving a new mimetic resemblance implies crossing a substantial fitness valley. This presents an additional challenge to our understanding of colour-pattern diversity within Heliconius (Fisher, 1930; Turner, 1977, 1984), but also an excellent opportunity to test hypotheses on how fitness valleys can be bridged. The ‘two-step’ theory (Poulton, 1913; Nicholson, 1927) predicts that crossing a valley can be achieved with a mutation of major effect, followed by mutations of smaller effect that refine mimicry. A long tradition of laboratory rearing and crossing experiments between mimetic forms within species has revealed that (i) phenotypic plasticity is not involved in major shifts between warning pattern phenotypes and (ii) that major colour-pattern elements (Beebe, 1955; Turner & Crane, 1962; Emsley, 1964; Sheppard et al., 1985; Mallet, 1989; Linares, 1996; Jiggins & McMillan, 1997; Gilbert, 2003; Joron et al., 2006; Kronforst et al., 2006b) or even entire wing-pattern polymorphisms (Brown & Benson, 1974; Joron et al., 2006) are controlled by a relatively small number of Mendelian loci (Fig. 4). The distinct phenotypes resulting from alternative alleles at these loci may represent the large-effect mutations hypothesized to bridge troughs in the fitness landscape. Modifier loci, and loci of small effect, which refine wing-pattern phenotypes, have also been documented in several Heliconius species (Sheppard et al., 1985; Nijhout, 1991; Baxter et al., 2008; Jones et al., 2013; Papa et al., 2013; Huber et al., 2015) and perhaps represent the second step in the ‘two-step’ theory.

The difficulty of estimating selection coefficients for individual loci in nature, however, makes it challenging to measure the effect sizes of individual mutations in terms of fitness. In addition, we still lack information about the precise mutations at patterning loci underlying phenotypic shifts, which is required to test whether the distribution of effect sizes of adaptive mutations reflects theoretical expectations. Advances in the functional characterization of causative loci will perhaps reveal the number, timing and effects of mutational steps involved in adaptive evolution in Heliconius. Differentiation of individual colour-pattern elements likely involves multiple, sequential mutations targeting the same gene(s) (McGregor et al., 2007; Baxter et al., 2008; Martin & Orgogozo, 2013). These may build up in separate populations and be later brought together through hybridization. As a consequence, ‘ready-made’ alleles of large phenotypic effect, capable of crossing deep adaptive valleys, can be made available through adaptive introgression (Gilbert, 2003; Mallet et al., 2007; Pardo-Diaz et al., 2012; The Heliconius Genome Consortium, 2012). Indeed, recent research has shown that mimicry in Heliconius can be achieved through hybridization and subsequent sharing of key loci between closely related species (Pardo-Diaz et al., 2012; The Heliconius Genome Consortium, 2012). In this way, phenotypic evolution can occur through selection on extant genetic variation, rather than large-effect novel mutations. It seems likely that adaptive introgression and homoploid hybrid speciation are considerably more common than formerly believed.

The role of genetic architecture in navigating fitness landscapes

Adaptive shifts in one phenotypic axis often require correlated or compensatory change in others. Shifts between mimetic fitness peaks, for example, often require correlated changes in multiple wing-pattern elements. Where selection favours rapid change, or acts to maintain complex polymorphisms, recombination will break down the key associations between alleles underlying co-adapted traits if left unchecked. How then, are these complex, composite phenotypes maintained? In species with well-defined geographic races, the correct assortment of alleles affecting warning patterns does not require linkage between loci as this is ensured by spatial segregation [e.g. the four major wing-pattern genomic regions of H. erato and H. melpomene are fixed for different alleles in different geographic regions (Mallet, 1989; Supple et al., 2013)]. In contrast, in those species where local polymorphism is maintained (such as H. numata), tight linkage between loci, or strong assortative mating, is required to facilitate the coexistence of multiple combinations of congruous alleles, which together produce mimetic phenotypes. ‘Supergenes’, which allow multiple functional elements to segregate as a single Mendelian locus despite recombination elsewhere, are classically associated with polymorphic mimicry (Brown & Benson, 1974; Charlesworth & Charlesworth, 1975; Turner, 1977; Joron et al., 2006; Thompson & Jiggins, 2014). In H. numata, coexisting mimetic phenotypes are determined by the supergene P, at which polymorphic chromosomal inversions maintain linkage disequilibrium and protect co-adapted allele combinations (Joron et al., 2011) (Fig. 3). Dominance relationships among alleles at the P supergene locus limit the expression of intermediate nonmimetic phenotypes when sympatric morphs interbreed (Joron et al., 2006; Le Poul et al., 2014). In contrast, similar supergene architectures are not observed in the sister species of H. numata, namely H. ismenius and H. hecale, which do not maintain local polymorphisms (Huber et al., 2015). Variation in selection regime between taxa, as well as the age, tempo and genealogy of adaptation, therefore leads to different genetic architectures. For example, the contrasted phylogeographic history of the parallel radiations in H. erato and H. melpomene suggests that the tempo of adaptation varied markedly, H. melpomene having recently colonized an older diversification in H. erato (Flanagan et al., 2004; Quek et al., 2010; Hines et al., 2011; but see Cuthill & Charleston, 2012). However, clear predictions of genetic architectures resulting from different adaptive scenarios are lacking.

Is evolution ‘predictable’?

Mimetic warning patterns are a classic example of convergent evolution, where similar adaptive traits appear in distantly related taxa. For example, phenotypic divergence within both H. erato and H. melpomene has led to near-perfect convergence in warning pattern between geographic races of these two species, whose lineages separated ~12 mya (Fig. 1). Although very similar mimicry switches in the two lineages map to homeologous genomic regions (Fig. 4), there is no evidence for sequence homology at the cis-regulatory regions involved in mimicry so far studied (Supple et al., 2013). Thus, the evidence points towards independent evolution of the colour patterns between the lineages, albeit using similar genetic machinery. The repeated use of specific genes has been frequently observed in convergent evolution (Mundy, 2005; Coyle et al., 2007), even among distantly related taxa (Arendt & Reznick, 2008). Particular features of some genes, for example their position in regulatory networks, may make them repeated targets of natural selection, and so in this sense ‘predictable’ (Stern & Orgogozo, 2009). This seems to be the case in Heliconius, where the recurrent evolution of mimetic phenotypes has largely been driven by evolutionary change in the same set of ‘toolkit’ genes (Fig. 4). Many of the loci that control convergent wing-pattern elements in distantly related species map to similar positions in the genome (Joron et al., 2006; Baxter et al., 2008; Reed et al., 2011), although phylogenetic analyses imply the causative mutations are often independently derived (Hines et al., 2011). Notably, the P locus ‘supergene’, controlling essentially all colour-pattern variation in H. numata, maps to the same genomic region as one of these major switch genes in both H. erato and H. melpomene, where it controls the expression of (at least superficially) very different colour-pattern elements (Joron et al., 2006). Remarkably, the homeologous genomic region in Biston betularia (the peppered moth) controls the switch between carbonaria and typica morphs (Van't Hof et al., 2013; see Gallant et al., 2014, for another example). It appears that these genes are responsible for both convergent and divergent wing-pattern phenotypes not only within Heliconius, but also across the Lepidoptera.

Candidate genes underlying colour-pattern shifts have been identified from patterns of gene expression. Microarray, immunohistochemistry and in situ hybridization data have shown that expression of a homologue to the Drosophila homeotic gene optix perfectly prefigures areas of the wing fated to be red across the Heliconius radiation (Reed et al., 2011; Martin et al., 2014). A second unlinked gene, WntA, is associated with variation in the size and position of melanic ‘shutter’ elements, which often define the shape of yellow and white forewing areas (Martin et al., 2012). To date, however, functional evidence is restricted to WntA, obtained by heparin injection, which affects the wnt signalling pathway (Martin et al., 2012). A lack of experimental of methods for full functional verification remains a major challenge for Heliconius research. Nonetheless, population data can assist with our inference of function. Thousands of generations of recombination in contact zones between geographic races have facilitated the identification of narrow genomic regions of high genetic divergence that control variation between distinct colour-pattern forms (Nadeau et al., 2012; Nadeau et al., 2013; Supple et al., 2013; Nadeau et al., 2014; Martin et al., 2013). Strikingly, the same ~100-kb region of divergence has been found between races of both H. melpomene and H. erato within the mapped locus controlling red-colour-pattern variation (H. melpomene B/D and H. erato D). This genomic window does not contain coding sequences demonstrating that cis-regulatory changes (i.e. those which control the expression of adjacent genes) within this region control optix expression to produce convergent phenotypes in both species.

Genes associated with wing-pattern variation in Heliconius are known to play important roles in other aspects of development. For example, optix is known to function in eye and neural development in Drosophila (Seimiya & Gehring, 2000) and is expressed in the optic lobe and medulla of pupal Heliconius (Martin et al., 2014). Coding sequence evolution in such genes is likely constrained, perhaps explaining why regulatory changes are so important in colour-pattern evolution. Precise changes in tissue-specific expression avoid negative pleiotropic effects, essentially disassociating multiple developmental roles. This regulatory subfunctionalization is expected to manifest as discrete enhancer modules within cis-regulatory regions. We now have evidence for three regulatory modules in the B/D region of optix from comparative sequence analysis of multiple Heliconius taxa (R.W.R. Wallbank, S.W. Baxter, C. Pardo-Diaz, J.J. Hanly, S.H. Martin, J. Mallet, K. Dasmahapatra, C. Salazar, M. Joron, N. Nadeau, W.O. McMillan & C.D. Jiggins, submitted). These appear to be specifically associated with different expression patterns of optix resulting in three distinct red pattern elements on the wing (rays, dennis and band – see Fig. 4). The modular nature of these enhancers means that they can be combined to produce considerable phenotypic diversity (see Fig. 4b for examples).

The position of optix in the wing-patterning gene network enhances its potential as a target for wing-pattern evolution in two ways. First, optix lies downstream of numerous genes whose spatial expression orientates wing development. These wing ‘prepatterning’ factors are involved in anteroposterior, dorsoventral and mediolateral axes, vein and scale differentiation, margin determination, etc. and can theoretically be exploited in numerous combinations to drive optix expression in any part of the wing, providing great scope for the evolution of different patterns. Second, optix controls a battery of downstream genes required to produce a multicomponent structure. These include pigment enzymes such as cinnabar, but must also include scale structural factors such as actins. Scanning electron microscopy and damage-inducing experiments have illustrated that colour pattern is developmentally coordinated with scale ultrastructure (Gilbert et al., 1988; Janssen et al., 2001), showing that colour-pattern genes regulate downstream processes controlling both scale pigment and structure. Hence, changes in the expression of a single gene, optix, are sufficient to generate variation in a functioning multicomponent structure, in this case pigmented wing scales.

These findings in Heliconius reinforce those from other systems that particular genes tend to be repeatedly targeted by natural selection (Stern & Orgogozo, 2009; Martin & Orgogozo, 2013). This clearly demonstrates a surprising degree of predictability in genetic architecture. However, in most of these systems, including Heliconius, the phenotypic differences studied were known to result from changes in major-effect loci. This has had major benefits by making both the ecological and genetic foundations of adaptive traits tractable. Other traits with a more polygenic architecture may show less genetic parallelism (Hoekstra & Coyne, 2007; Stern & Orgogozo, 2009; Nadeau & Jiggins, 2010). Nevertheless, in Heliconius there is clearly a restricted set of loci capable of producing major phenotypic switches of ecological importance. Heliconius wing-pattern phenotypes evolve through multiple mutations at a small handful of loci.

What is a Heliconius niche?

Aside from mimicry, two relationships are central to Heliconius biology each of which likely involved bouts of diffuse coevolution: first, their reliance on Passiflora as larval host plants; and second, their reliance as adults on resources that are obtained by systematically collecting pollen (an adaptation unique to Heliconius) (Gilbert, 1972). In response to herbivory, Passiflora have evolved a high diversity of cyanogenic compounds (Spencer, 1988; Engler et al., 2000; Engler-Chaouat & Gilbert, 2007; Hay-Roe & Nation, 2007). Heliconius larvae not only detoxify these cyanogens, but can also disable the plant β-glucosidase enzymes to prevent further release of cyanide, and sequester them for their own defence as adults. Other Passiflora defences include trichomes to disrupt larval locomotion and feeding (Gilbert, 1971), structures that mimic butterfly eggs to deter Heliconius oviposition (Williams & Gilbert, 1981; Gilbert, 1982), variable leaf morphology to disrupt visual searching by gravid females (Gilbert, 1982), and extrafloral nectar production attracting larval predators, in particular ants (Smiley, 1985, 1986). In contrast, relationships with adult food plants are generally mutualistic. Preferred pollen sources, notably Gurania and Psiguria, are those that provide reliable food stations for the duration of an individual's lifespan (Gilbert, 1975). Their exploitation heavily influences individual fitness as pollen feeding supports a prolonged reproductive lifespan (Gilbert, 1972; Dunlap-Pianka et al., 1977; O'Brien et al., 2003), as well as the maintenance of chemical defences (Gilbert, 1972; Dunlap-Pianka et al., 1977; O'Brien et al., 2003; Cardoso & Gilbert, 2013). Members of the Gurania (~40 species) and, in particular, the Psiguria (~16 species) show many adaptations to attract Heliconius (Gilbert, 1975; Murawski & Gilbert, 1986; Condon & Gilbert, 1988). For example, the inflorescence of P. warscewiczii produces just one flower daily over a period of several months, perfectly adapted for pollination by long-lived Heliconius, which learn the location and return daily.

Sensory adaptation and phenotypic diversification

The evolution of animal signals will be shaped by the environment through which the signal is transmitted and the receivers’ sensory abilities. Signal receivers may include both con- and heterospecific individuals, as well as individuals at different life stages and of either sex, resulting in complex fitness trade-offs. To fully understand how signalling trade-offs are resolved in Heliconius, we would like to know how these butterflies perceive their environment and how their perception differs from their predators. Heliconius have perhaps the largest head of any Neotropical butterfly genus, predominantly due to an investment in visual neuropile (Gilbert, 1975). Visual sensitivity is shaped by an organism's ecology (Stevens, 2013), and vision in Hymenoptera, for example, is tuned to maximize perception of variation in flower colour (Chittka & Menzel, 1992). We expect Heliconius vision to be similarly optimized. Heliconius discriminate visual cues including both shape (Gilbert, 1982; Corrêa et al., 2001) and colour (Swihart & Swihart, 1970; Swihart, 1972). Four opsins – light-sensitive proteins – have been identified in the compound eye of H. erato, with sensitivity peaks at ~355 nm (ultraviolet 1), ~398 nm (ultraviolet 2), ~470 (blue) and ~555 nm (longwave) (Zaccardi et al., 2006; Briscoe et al., 2010; Yuan et al., 2010). Red filtering pigments in the ommatidia and retina further increase the range of colour discrimination by changing receptor sensitivity (Zaccardi et al., 2006).

These visual adaptations may have have evolved for intraspecific communication related to the evolution of mimicry (Bybee et al., 2012). As Heliconius butterflies have strong aposematic colour signals serving both predation avoidance and mate recognition, they may be selected to use channels of communication that are not detected by predators (Bybee et al., 2012; Llaurens et al., 2014). In contrast to other Lepidoptera that only have one, the two Heliconius opsins in the UV range confer a greater sensitivity to UV reflectance (Briscoe et al., 2010; Bybee et al., 2012). Heliconius yellow wing pigments appear to have co-evolved with the additional UV opsin, as they have a much higher UV reflectance than yellow pigments in other Lepidoptera (Briscoe et al., 2010; Llaurens et al., 2014). Models of animal vision suggest that birds are less effective in discriminating these UV-yellows as compared to the butterflies, consistent with this being a cryptic channel of communication for Heliconius mate-finding that has evolved to compensate for similarities due to mimicry (Bybee et al., 2012). However, the key experimental question is whether Heliconius can discriminate different species-specific UV signals and therefore avoid interspecific mating and hybridization, and this remains to be demonstrated.

Does behavioural plasticity facilitate ecological adaptation and diversification?

The diversification of Heliconius is often associated with concordant shifts in colour pattern and habitat use (e.g. Mallet, 1993; Estrada & Jiggins, 2002; Arias et al., 2008). The exploitation of novel environments may require secondary adaptations, but can be facilitated in the short term by behavioural plasticity (West-Eberhard, 2003; Pfennig et al., 2010; Dukas, 2013; Lister, 2013; Snell-Rood, 2013). Behavioural flexibility may further accelerate divergence by exposing organisms to new selection regimes. In other taxa, behavioural plasticity has been linked to range and niche expansion (Sol et al., 2008; Gonda et al., 2009), variation in host use (Snell-Rood & Papaj, 2009; Nylin et al., 2014) and mate choice (Svensson et al., 2010; Westerman et al., 2012).

Indirect evidence for the importance of learning in Heliconius ecology comes from the apparent expansion of the mushroom bodies in the central brain of Heliconius (Fig. 5; Sivinski, 1989; Montgomery et al., 2015), which have repeatedly been associated with learning in other insects (Zars, 2000; Farris, 2005; Snell-Rood et al., 2009). Indeed, these structures are larger (in both relative and absolute terms) in Heliconius than in any other Lepidoptera surveyed (Sivinski, 1989; Montgomery et al., 2015). Patterns of interspecific variation and intraspecific plasticity in Hymenoptera strongly link mushroom body expansion with spatial learning in a foraging context (Withers et al., 1993; Farris & Schulmeister, 2011). In Heliconius, the maturation of the mushroom bodies shows strong experience-dependent effects further supporting their role in learning (Montgomery et al., 2015).

Mark–release–recapture studies have demonstrated temporally and spatially faithful trap-lining behaviour in Heliconius, and an effect of experience on foraging efficiency and host plant visitation (Gilbert, 1975; Mallet, 1986; Murawski & Gilbert, 1986; Finkbeiner et al., 2012). These data imply that Heliconius invest heavily in the capacity to learn distributed resources, an adaptation that facilitates efficient foraging in the complex tropical rainforest environment. Heliconius are also able to associate information with colour and shape (Swihart & Swihart, 1970; Gilbert, 1975), which may permit a flexible response to identifying preferred host plants or pollen resources. For the very few Heliconius taxa where data exist, shifts in host use appear innate (Kerpel & Moreira, 2005; Salcedo, 2011; Merrill et al., 2013).

However, regions of the brain involved in primary processing of visual and olfactory information also show evidence of developmental plasticity (Montgomery et al., 2015). These structures are associated with olfactory and visual learning in other taxa (Hammer & Menzel, 1998; Paulk et al., 2009), and their relative size co-evolves with microhabitat and diel pattern (Montgomery & Ott, 2015). Whether behavioural plasticity facilitates diversification in Heliconius for now remains an open question. Nevertheless, the evidence available suggests that behavioural flexibility plays a key role in the ecology of these butterflies and may well promote shifts in habitat use.

A kind of magic

Speciation with gene flow is greatly facilitated if traits under divergent ecological selection also contribute to nonrandom mating (Gavrilets, 2004; Weissing et al., 2011). These so-called magic traits (Gavrilets, 2004; Servedio et al., 2011) evade the homogenizing effects of recombination, which impede the evolution of behavioural isolation when gene flow persists (Felsenstein, 1981). Although the epithet ‘magic’ was perhaps intended to suggest these types of trait were rare in nature, accumulating evidence suggests that this might not be the case (Servedio et al., 2011). However, finding convincing examples has proved difficult because different traits can be strongly correlated making it difficult to distinguish their individual effects, especially with respect to ecological selection. Servedio et al. (2011) propose two experimental criteria: first, the putative magic trait, rather than any other correlated trait, must be subject to divergent selection; and second, the same trait, not a correlated trait, must generate nonrandom mating. The amenability of Heliconius colour patterns to experimental manipulation has provided an excellent opportunity to test their role in reproductive isolation, and it has been argued that Heliconius provide the strongest empirical support for ‘magic traits’ (Servedio et al., 2011). Specifically, experiments using paper models of H. cydno and H. melpomene and their first-generation hybrids, in addition to experiments with live butterflies and captive birds, reveal that colour pattern is under strong disruptive selection due to predation (Merrill et al., 2012). In the same species, Jiggins et al. (2001a) have shown that individuals prefer to mate with live butterflies and court paper models of the same colour pattern. Thus, in contrast to other potential examples of magic traits, both parts of the required evidence are present.

The fact that a combination of colour (hue) and movement stimulates courtship by Heliconius males was demonstrated as early as the 1950s (Crane, 1955). In a battery of subsequent insectary experiments involving fluttering models, made from either dissected female wings or printed paper butterflies (Jiggins et al., 2001a, 2004; Kronforst et al., 2006c; Estrada & Jiggins, 2008; Melo et al., 2009; Muñoz et al., 2010; Merrill et al., 2011b, 2014), males almost invariably show a preference for their own wing pattern over that of conspecific races or other closely related species. Data from the wild are more limited (but see Chamberlain et al., 2009). However, assortative mating has been shown between Heliconius himera and H. erato, which share a narrow hybrid zone (~5 km) in southern Ecuador (Mallet et al., 1998b). Given male reproductive investment, selection against poorly adapted hybrids may contribute to divergent male ‘preferences’ through a reinforcement-like mechanism. Nonmimetic hybrids suffer increased predation (Merrill et al., 2012), as well as reduced mating success (Naisbit et al., 2001) and fertility (Jiggins et al., 2001b; Naisbit et al., 2002; Salazar et al., 2005; Muñoz et al., 2010), and populations not in contact with closely related taxa are less choosy (Jiggins et al., 2001a; Kronforst et al., 2007). However, divergent preferences are also observed between taxa where reinforcement seems unlikely (Jiggins et al., 2004). The ability to visually locate potential mates at long distances may be a considerable advantage, and it seems likely that competition between males for mate location drives local preferences.

It is perhaps often overlooked that the existence of magic traits does not make speciation automatic or inevitable. Shifts in colour pattern must be accompanied by their corresponding mate preferences. In contrast to other Heliconius behaviours, learning appears unimportant for developing mate preferences – although these experiments are by no means exhaustive. Neither isolating males from conspecific colour patterns (either in other butterflies or on themselves) nor exposure to females of a different wing-pattern race for several days prior to testing affects courtship response (Crane, 1955; Jiggins et al., 2004; Merrill et al., 2011b). As offspring inherit alleles underlying both preference and signal, a positive genetic correlation between preference and signalling loci is expected (Fisher, 1930); however, this association may be unstable if assortative mating is weak or incomplete, as might be expected during the early stages of divergence. We now know from genetic crosses that the red/white forewing switch gene distinguishing H. melpomene and H. cydno (B locus), as well as the white/yellow forewing switch gene distinguishing H. cydno and H. pachinus and different morphs of the polymorphic H. cydno alithea (K locus), is physically associated with major loci underlying the corresponding male preferences (Fig. 4) (Kronforst et al., 2006c; Chamberlain et al., 2009; Merrill et al., 2011b). Whether these associations reflect pleiotropy or tight physical linkage between pattern and preference loci remains to be seen. Whatever the cause, they generate robust genetic associations that impede recombination between wing pattern and preference loci, further facilitating the evolution of reproductive isolation. These results additionally suggest a mechanism by which the introgression of colour-pattern elements, linked to the corresponding preference alleles, would directly lead to assortative mating and offer a route towards rapid hybrid speciation. Several putative examples of hybrid species in Heliconius are proposed to have derived from interbreeding between H. cydno and H. melpomene, and notably, the forewing of the best documented example, the Colombian species Heliconius heurippa (Mavárez et al., 2006), involves both red and yellow pattern elements (controlled by the B and N loci, respectively). Indeed, heurippa-like males reconstructed by backcrossing F1 hybrids into H. cydno are more likely to approach and court models of their own colour pattern than either of the parental species (Melo et al., 2009), showing that hybridization can very rapidly lead to premating isolation.

Are multiple components of reproductive isolation necessary to complete speciation?

The number of traits subject to divergent selection can influence speciation (Nosil et al., 2009). Studies of experimental evolution in Drosophila suggest that in contrast to multifarious trait scenarios, selection on a single trait will typically lead only to incomplete reproductive isolation (Rice & Hostert, 1993). Similarly, in Timema walking sticks and Rhagoletis flies, the degree of reproductive isolation appears to correlate with the number of traits subject to divergent selection (Dambroski & Feder, 2007; Nosil & Sandoval, 2008). In Heliconius, potential isolating barriers have been considered across a range of taxon pairs and, in general support of the multifarious hypothesis, strong selection on a single trait (colour pattern) does not seem sufficient to complete speciation. Additional barriers, including additional components of sexual isolation, ecological differences (Jiggins et al., 1997; Estrada & Jiggins, 2002; Merrill et al., 2013) and intrinsic post-mating barriers (Jiggins et al., 2001b; Naisbit et al., 2002; Salazar et al., 2005; Muñoz et al., 2010), clearly contribute to reproductive isolation in Heliconius. Indeed, the recent discovery of ‘cryptic’ taxa, such as H. melpomene malleti and H. timareta florencia, which show strong assortative mating despite almost indistinguishable colour patterns, necessitates this (Giraldo et al., 2008; Mérot et al., 2013).

An obvious, but to date largely unexplored, reproductive barrier is divergence in chemical signals. Estrada and Jiggins (Estrada & Jiggins, 2008) report that H. erato males can distinguish between wings dissected from conspecific and heterospecific (although comimetic H. melpomene) females, but that this effect disappears after wings have been washed in hexane. In addition, comimetic populations of H. melpomene and H. timareta share wing patterns but show strong premating isolation, presumably mediated by chemical cues (Giraldo et al., 2008; Mérot et al., 2013). Nonetheless, the chemical signals used have yet to be identified. Indeed, that female choice (in general) contributes to reproductive isolation between Heliconius taxa has only been shown very indirectly. In choice trials, the mating probability of female backcross hybrids between the sympatric species H. melpomene and H. cydno segregates with a major colour-pattern locus (Merrill et al., 2011b). These results suggest that a locus underlying female rejection segregates with colour pattern, although the cues involved remain elusive. Overall, studies of chemical signalling in Heliconius are in their early stages as compared to those of visual signalling.