Introduction

During ecological speciation, reproductive isolation evolves as a result of divergent natural selection [1]. Although ecological barriers can reduce gene flow between divergent populations, speciation normally requires the evolution of assortative mating [1,2]. This is made easier if traits under divergent ecological selection also contribute to assortative mating, as this couples ecological and behavioral barriers [3–6]. Ecologically relevant mating cues (sometimes known as “magic traits” [2,6]) are now predicted to be widespread in nature [6,7], and the last few years have seen considerable progress in our understanding of their genetic basis. For example, studies have explored the genetic basis of beak shape in Darwin’s finches [8], body shape in sticklebacks [9,10], cuticular hydrocarbons in Drosophila [11], and wing patterns in Heliconius butterflies [12–14]. However, the extent to which these traits contribute to assortative mating depends on the evolution of corresponding preference behaviors and the underlying genetic architecture.

We still know little about the process by which ecological traits are co-opted as mating cues and, in particular, how matching cues and preference behaviors are controlled genetically (but see [15]). Both the substitution of large-effect preference alleles and physical linkage will strengthen linkage disequilibrium (LD, i.e., the nonrandom association of alleles at different loci) [16]) between cue and preference. Strong LD between barrier loci is expected to both maintain and facilitate the evolution of new species in the face of gene flow. This is the result of two key but related processes. First, LD between barrier loci will result in the coupling of barrier effects, and where these effects coincide, the overall barrier to gene flow is increased [4,16,17]. Second, LD between pre- and postmating barrier loci will facilitate an increase in premating isolation in response to selection against hybridization (i.e., reinforcement, sensu [18]) by transferring the effects of selection from the latter to the former [19].

In central Panama, the butterfly H. melpomene rosina is a precise Müllerian mimic of H. erato and normally occurs in forest-edge habitats, whereas the closely related species H. cydno chioneus mimics H. sapho and is more common in closed-forest habitats, although H. melpomene and H. cydno are often seen flying together (Fig 1A and 1B) [20]. Hybrids are viable but occur at very low frequencies in the wild (estimated at approximately 0.1%), consistent with strong assortative mating shown in insectary experiments. Specifically, heterospecific mating was not observed in 50 choice and no-choice trials between Panamanian H. melpomene and H. cydno ([21,22]; see also [23]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Divergence in warning pattern cue and corresponding preference in sympatric Heliconius butterflies.

(A) Wing pattern phenotypes of top, H. cydno chioneus (left), H. melpomene rosina (right), their nonmimetic first-generation hybrid (center); and bottom, their sympatric comimics H. sapho sapho (left) and H. erato demophoon (right). (B) Distribution of H. cydno (blue) and H. melpomene (orange). Individuals were collected and experiments performed in Panama (black circle), where the two species co-occur in Central and northern South America. (C) Overview of crossing design. Colored boxes represent segregating H. cydno (blue) and H. melpomene (orange) alleles; Z and W refer to the alleles on the sex chromosomes and A to those on autosomes. (D) Proportion of courtships directed toward H. melpomene (as opposed to H. cydno) females for CYD, MEL, their F1, and BC and BM. Values in parentheses indicate total number of individuals with behavioral data. Solid colored boxes represent expected average genome contribution of each generation. Note that a further 11 BC individuals were tested but performed no courtship behaviors. Underlying data can be found in S1 Table. BC, backcross hybrid to H. cydno; BM, backcross hybrid to H. melpomene; CYD, H. cydno; MEL, H. melpomene; F1, first-generation hybrid.

https://doi.org/10.1371/journal.pbio.2005902.g001

The amenability of Heliconius color patterns to experimental manipulation has led to the demonstration that color pattern is both under strong disruptive selection due to predation [24] and also that males prefer live females and paper models with the same color pattern as themselves [24]. These results led Servedio and colleagues [6] to conclude that, unlike other putative examples, both criteria for a magic trait have been confirmed with manipulative experiments in H. melpomene rosina and H. cydno chioneus. Although female preferences undoubtedly contribute to assortative mating [25–27], male preferences act first in these species such that strong observed male discrimination against heterospecific females will have a disproportionate contribution to overall reproductive isolation [28]. As highlighted by Coyne and Orr [29], the order in which reproductive isolation acts influences their relative contribution to overall isolation. In this case, the ordering of behavioral decisions is likely predetermined by their sensory systems: Heliconius lack specialized olfactory structures to support long range detection of chemical signals so are only likely to use these in close proximity, whereas they have very good long-range vision [30]. As such, not only is male preference in Heliconius butterflies experimentally more tractable than other components of behavioral isolation, but it is also an important barrier to gene flow.

Crossing experiments have shown that the shift in mimetic warning pattern between H. melpomene rosina and H. cydno chioneus is largely controlled by just three major effect loci [31]. Genes underlying these loci have now been identified: the transcription factor optix controls red patterns [12], the WntA gene controls forewing band shape [13], and yellow patterns map to the gene cortex [14]. In addition, a further locus K segregates in crosses between H. melpomene rosina and H. cydno chioneus with more minor effect [31]. Further modularity occurs within these loci. For example, different regulatory elements of optix each result in distinct red pattern elements [32]. The modular nature of individual color pattern loci and their functionally sufficient enhancers means that they can be combined to produce considerable phenotypic diversity [32,33]. These loci are large-effect “speciation genes,” in that they control traits that generate strong reproductive isolation [34].

Two of these color pattern loci, optix and K, have previously been associated with Heliconius courtship behaviors [25,35,36]; however, these studies do not provide evidence for tight physical linkage (<20 centiMorgans [cM]) between warning pattern and preference loci. Our own previous study tested for an association between Mendelian color pattern loci and preference behaviors [25] but did not correct for the segregation of alleles across the genome so that reported levels of support are likely inflated [37], and an earlier study of the parapatric taxa H. cydno and H. pachinus [35] is limited by small sample size [37]. Regardless of the level of statistical support for preference loci, these studies both lack the resolution to demonstrate the degree of tight physical linkage between loci contributing to reproductive isolation that would be expected to aid speciation. Perhaps the best evidence comes from a study of wild H. cydno alithea [36]. This population is polymorphic for a yellow or white forewing (due to the segregation of alleles at the K locus), and males with a yellow forewing prefer yellow females. These results are important because they suggest a key component of speciation: specifically, coupling between potential behavioral and ecological barriers. However, because they rely on segregation within a wild population rather than laboratory crosses, it is not possible to distinguish physical linkage from genetic associations between cue and preference alleles due to nonrandom mating. The extent to which warning pattern and behavioral loci are physically linked in Heliconius as well as the existence of major preference loci elsewhere in the genome remains unknown. To address this and to complement our extensive knowledge of the genetics of their color pattern cues, here, we use a genome-wide quantitative trait locus (QTL) approach to explore the genetics of male preference behaviors between the sympatric species H. melpomene rosina and H. cydno chioneus.

Results

We studied male mating preference among first-generation hybrid (F1) and backcross hybrid families between H. melpomene rosina and H. cydno chioneus in standardized choice trials [25,38] (Fig 1). We introduced individual males into an experimental cage and recorded courtship directed toward two virgin females, one of each species. In total, we collected data from 1,347 behavioral trials, across 292 individuals. Multiple trials were performed for each backcross male, from which we determined the relative courtship time, as defined in the methods, directed toward H. melpomene and H. cydno females.

Three loci contribute to species differences in preference behavior

As reported previously [25], F1 males have a strong preference for the red H. melpomene females, and little segregation in mate preference is observed among the backcross to melpomene (and whose mean preference does not differ significantly from that of pure H. melpomene males: 2ΔlnL = 1.33, degrees of freedom [d.f.] = 1, P > 0.2), implying that melpomene mate preference alleles are dominant. In contrast, courtship behavior segregates among H. cydno backcross males, permitting analysis of the genetic basis for this mating behavior (Fig 1D). Consequently, all subsequent analyses were performed on backcross-to-cydno males. We used a genome-wide QTL mapping approach to identify the genomic regions underlying divergence in mate attraction. Linkage maps were constructed from genotype data of 331 backcross-to-cydno individuals and their associated parents [39], including 146 individual males for which we had recorded attraction behaviors.

We identified three unlinked QTLs on Chromosomes 1, 17, and 18 associated with variation in the relative amount of time males spent courting red H. melpomene and white H. cydno females (Fig 2A). Of these, one is tightly linked to the optix locus on Chromosome 18, which controls the presence/absence of a red forewing band. Specifically, the QTL peak for the behavioral QTL on Chromosome 18 (at 0 cM) is just 1.2 cM from optix. The associated 1.5-log odds ratio (LOD) support interval is between 0 and 6.0 cM, suggesting that the true location of the QTL is no more than 4.8 cM from the optix coding region (whose genetic position is at 1.2 cM) (Fig 3); however, given that the peak support (i.e., highest LOD score) for our behavioral QTL is at 0 cM and that this rapidly drops off, physical linkage between wing patterning cue and preference loci is likely much tighter than a strict 1.5-LOD interval might suggest. In contrast, the QTL on Chromosome 1 is at least 30 cM from the gene wingless, which, although unlikely to be a color pattern gene itself, has previously been associated with the K wing pattern locus between taxa within the cydno clade [35]. No known wing pattern loci reside on Chromosome 17.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. QTL analysis of variation in mate preference.

(A) QTLs for relative time males court H. melpomene (as opposed to H. cydno) females on Chromosomes 1, 17, and 18 (n = 139). Scale on right axis depicts genome-wide significance, determined through permutation, corresponding to the LOD score as shown on the left axis. Dotted red line represents LOD significance threshold (genome-wide alpha = 0.05, LOD = 2.99). Dashes indicate position of genetic markers (SNPs), and red arrows indicate the position of the max LOD score for each QTL (used in B). Vertical blue lines represent the position of major color pattern loci and their phenotypic effects. Note that the K locus only has limited phenotypic effects in crosses between H. cydno chioneus and H. melpomene rosina but is responsible for the switch from yellow to white color pattern elements between other taxa within the melpomene–cydno clade. Underlying data can be found in the online Dryad repository doi.10.5061/dryad.4b240j4, specifically raw_data/data_for_Rqtl.csv, derived_data/genome_scan_lod_score.csv, and derived_data/permutations_courtship_prop.csv. (B) Proportion of time males court H. melpomene (as opposed to H. cydno) females for each of the two genotypes for respective QTLs (homozygous = CYD:CYD and heterozygous = CYD:MEL). Error bars represent 95% confidence intervals. Lower dashed blue and upper orange bars represent mean phenotypes measured in H. cydno and H. melpomene, respectively. Circle size depicts total number of “courtship minutes” for each male. Vertical black bars indicate the percentage of the difference measured in the parental species explained. Underlying data can be found in the online Dryad repository doi.10.5061/dryad.4b240j4, specifically raw_data/data_for_Rqtl.csv and derived_data/qtl_data.csv. LOD, log odds ratio; QTL, quantitative trait locus, CYD, cydno allele; MEL, melpomene allele.

https://doi.org/10.1371/journal.pbio.2005902.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Genetic and physical positions of behavioral QTL and the warning pattern loci and localized levels of admixture (f_d).

Vertical blue lines represent the position of major color pattern loci and orange lines represent the position of peak LOD score for each behavioral QTL. Gray boxes indicate the 1.5-LOD support interval for each QTL. Top panel: Dashes along the x-axis indicate position of genetic markers (SNPs). Bottom panel: Blue points represent f_d values for 100-kb windows. f_d was measured between H. melpomene rosina and H. cydno chioneus individuals from population samples in Panama; H. melpomene melpomene from French Guiana, which is allopatric with respect to H. cydno, was the “control” population. Underlying data can be found in the online Dryad repository doi.10.5061/dryad.4b240j4, specifically raw_data/data_for_Rqtl.csv, derived_data/genome_scan_lod_score.csv, and raw_data/fd_values_SM_2MAY18/ /bar92.DP8MP4BIMAC2HET75.fourPopPol_melG_melW_cyd_num.w100m1s20.merged.csv. LOD, log odds ratio; QTL, quantitative trait locus.

https://doi.org/10.1371/journal.pbio.2005902.g003

Statistical modeling supports additive effects of all three detected loci (Table 1), and in our mapping population, these three QTLs together explain approximately 60% of the difference in male preference behavior between the parental species (Fig 2B). Given the sample sizes feasible in Heliconius, our analysis lacks the power to resolve smaller effect QTLs. We also found no evidence of pairwise interactions between individual QTLs in our model of relative courtship time, which again is unsurprising given achievable sample sizes. However, genome scans considering individuals with alternative genotypes at the QTL on Chromosome 18 separately revealed a significant QTL on Chromosome 17 (LOD = 3.52, P = 0.016) for heterozygous (i.e., LG18@0cM = cydno allele:melpomene allele [CYD:MEL]) but not for homozygous (i.e., LG18@0cM = CYD:CYD) males (S1 Fig), though this result is not supported by nonparametric interval mapping (LOD = 2.4, P = 0.132). Nevertheless, these results perhaps suggest that alleles on Chromosomes 17 and 18, or the specific behaviors they influence, may interact.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Individual and combined QTLs for differences in relative courtship time.

https://doi.org/10.1371/journal.pbio.2005902.t001

Preference QTL are of large effect

Individually, the measured effect of each of the three QTLs we identified was large, explaining between 23% and 31% of the difference between males of two parental species (Fig 2B). However, in studies with relatively small sample sizes such as ours (n = 139), estimated effects of QTL are routinely overestimated (a phenomenon known as the “Beavis effect,” after [40]). This is because effect sizes are determined only after significance has been determined, and QTL with artificially high effect sizes—due to variation in sampling—are more likely to achieve “significance.”

To determine the extent to which the effects of our QTL may be overestimated, we simulated QTL across a range of effect sizes and compared the distribution of measured effects for all simulations to those which would be significant in our analysis (S2 Fig). Our simulations suggest that the reported effects of our QTL are not greatly overestimated. We first considered what proportion of “significant” simulations would be smaller than our empirically measured effects (Fig 4A). A highly conservative threshold of 95% would suggest that the QTLs on Chromosome 1 and 18 explain at least 10% and 20% of the difference in behavior between the parental species, respectively. Adopting the median values, our simulations would suggest true effects of 25%, 15%, and 30% or greater for the QTLs on Chromosomes 1, 17, and 18, respectively. Given simulated effect sizes similar to those measured empirically, there was little bias among simulation runs that achieved the genome-wide significance threshold (S2 Fig). This suggests that the true effect sizes of our QTL are likely to be large, with somewhat less support for the QTL on Chromosome 17.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. QTL effects in consideration of the Beavis effect.

(A) Proportion of “significant” simulations that would be smaller than our empirically measured effects for preference QTLs on Chromosome 1 (blue), Chromosome 17 (black), and Chromosome 18 (orange). 10,000 simulations were run for effect sizes corresponding to between 5% and 40% of the difference in male preference behavior between the parental species. In each case, the distribution of sample effect sizes was determined for those simulations that reached the genome-wide significance threshold determined through permutation (Fig 2). Underlying data can be found in the online Dryad repository doi.10.5061/dryad.4b240j4, specifically derived_data/QTL_1_sim_results.csv, derived_data/QTL_17_sim_results.csv, and derived_data/QTL_18_sim_results.csv. (B) Proportion of time males court H. melpomene (as opposed to H. cydno) females for each of the two genotypes for white (homozygous = CYD:CYD) and red (heterozygous = CYD:MEL) hybrid males for which we were unable to generate RAD-seq data (so which were not included in our initial QTL analysis). Error bars represent 95% confidence intervals. Lower dashed blue and upper orange bars represent mean phenotypes measured in H. cydno and H. melpomene, respectively. Circle size depicts total number of “courtship minutes” for each male. Vertical black bars indicate the percentage of the difference measured in the parental species explained. Underlying data can be found in the online Dryad repository doi.10.5061/dryad.4b240j4, specifically raw_data/IDs_with_pheno_no_RAD.csv. CYD, cydno allele; MEL, melpomene allele; QTL, quantitative trait locus; RAD-seq, restriction site–associated DNA sequencing.

https://doi.org/10.1371/journal.pbio.2005902.g004

Although our simulations suggest that the effects we have measured are reasonable, ideally, we would estimate effect sizes from a population of individuals that were not used to determine significance. In evolutionary biology, follow-up experiments such as this are uncommon; collecting phenotypic data across a large number of hybrid individuals is often a considerable undertaking, and this is similarly true for Heliconius behaviors. Nevertheless, we were able to follow up our results for the QTL on Chromosome 18, using a sample of a further 35 backcross males for which preference behavior was measured but for which we were unable to generate genotype data (and so were not included in our initial QTL analysis). As reported above, the QTL peak (at 0 cM) on Chromosome 18 is in very tight linkage with the optix color pattern locus (at 1.2 cM), which controls the presence and absence of the red forewing band. Presence of the red forewing band is dominant over its absence so that segregation of the red forewing band can be used to perfectly infer genotype at the optix locus, even without sequence data. This analysis supports our previous result that the QTL on Chromosome 18 is of large effect (Fig 4B); among these 35 hybrid males, the optix locus explains 27% of the difference in behavior between the parental species (compared to 31% for the larger mapping population).

Different preference QTLs affect different aspects of behavior

The male preference QTLs we have identified may influence differences in male attraction toward red H. melpomene females, white H. cydno females, or toward both female types. To further explore the influence of segregating alleles at these loci, we considered the influence of all three QTLs on courtship time directed toward each female type separately (Fig 5). Overall, the proportion of time butterflies spent courting females in our trials was low (median = 0.11, S4 Fig). We have already robustly established a significant effect of these loci on variation in the relative amount of time males spent courting each female type (see Fig 2A). Consequently, although we corrected for multiple testing arising from considering three QTLs across the two data sets [37], in these post hoc analyses, we did not account for multiple segregating loci across the entire genome (in contrast to the results reported above). This greatly increases our power to detect any influence of the QTLs on attraction toward the two species individually but also increases the likelihood of false positives. The QTL on Chromosome 1 influenced the number of courtships directed toward H. cydno females (F_1,145 = 10.85, P < 0.01) but had no significant effect on how males behaved toward H. melpomene females (F_1,145 = 1.35, P > 0.2). In contrast, the QTL on Chromosome 17 influenced the degree of courtship directed toward H. melpomene (F_1,145 = 10.08, P = 0.011) but not H. cydno females (F_1,145 = 0.41, P > 0.2). Similarly, the QTL on Chromosome 18 had a significant effect on courtships directed toward H. melpomene (F_1,145 = 9.93, P = 0.012) females (though we note that prior to Bonferroni correction, there is also some support for an effect on courtships directed toward H. cydno females: F_1,145 = 6.56, P = 0.01).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Different QTLs affect different aspects of behavior.

The QTL on chromosome 1 influences courtship toward H. cydno but not H. melpomene females. The opposite is the case for the QTLs on Chromosomes 17 and 18, for which there is little evidence that either QTLs influence courtships directed toward H. cydno females. Data presented are for number of courtship events corrected by the total number of trials. Blue circles and boxplots represent data for individuals homozygous at each QTL (i.e., CYD:CYD); orange circles and boxplots represent data for individuals heterozygous at each QTL (i.e., CYD:MEL). Underlying data can be found in the online Dryad repository doi.10.5061/dryad.4b240j4, specifically derived_data/qtl_data.csv. QTL, quantitative trait locus.

https://doi.org/10.1371/journal.pbio.2005902.g005

Discussion

Here, we reveal a genetic architecture that will strengthen genetic associations (i.e., LD) between key components of reproductive isolation and so facilitate ecological speciation in the face of gene flow. Specifically, we demonstrate that just three QTLs are largely responsible for an important component of behavioral isolation between two sympatric species of Heliconius butterfly. One of these resides only 1.2 (0–4.8) cM from a major color pattern gene. Our results also suggest that all three preference loci are of large phenotypic effect. Because LD between cue and preference loci will arise as a natural consequence of mating preferences [43], these large-effect preference loci will further increase LD between ecological and behavioral components of reproductive isolation. Additional smaller effect loci undoubtedly also contribute to variation in male preference, which we would be unlikely to detect without very large sample sizes (a caveat shared with many QTL studies of ecologically relevant behaviors, e.g., [15,44,45]). Regardless, our results suggest that during speciation, divergence between populations in both mating cue and the corresponding preference behaviors can have a surprisingly simple genetic architecture.

By ensuring robust genetic associations between components of reproductive isolation, physical linkage between loci for traits influencing pre- and postmating isolation is expected to facilitate speciation with gene flow [19]. Two of the behavioral QTLs we have identified are situated on chromosomes with major color pattern loci (Chromosome 1 includes the K locus, and Chromosome 18 includes the optix locus). Both optix and the K locus have previously been associated with variation in Heliconius courtship behaviors [25,35,36]. Nevertheless, we have not previously been able to robustly estimate the position of QTLs along the chromosome. The QTL we identify on Chromosome 1 is not tightly linked to the K locus. It remains to be seen whether this QTL underlies the association between male preference behavior and the K locus phenotype (a shift between white and yellow color pattern elements) previously observed in crosses between H. cydno and H. pachinus [35] and within a polymorphic population of H. cydno [36]. (Although the K locus phenotype segregating in crosses between H. cydno and H. melpomene [39] has not been mapped, it is very likely that it is the same locus as that observed in H. cydno and H. pachinus.) In contrast, our results reveal that the QTL for male attraction on Chromosome 18 is tightly linked to the optix locus, which controls presence/absence of a red forewing. The mechanistic basis for linkage of trait and preference loci remains unclear. There is no evidence for an inversion at this locus [39]; it also seems unlikely that the same mutations control both wing pattern and the corresponding attraction behavior. However, optix is known to function during eye and neural development in Drosophila [46] and is expressed in the optic lobe and medulla of pupal Heliconius [47], so it is plausible—if unlikely [48]—that the two traits could be controlled by different regulatory elements of the same gene.

Our work joins a small collection of studies in animals in which physical linkage is reported to couple loci contributing to preference behaviors and ecological barriers [15,25,35,36,49], as predicted by Felsenstein [19], and more broadly between loci for cue and preference between incipient species [50–55]. In a seminal study published almost 20 years ago, Hawthorn and Via [49] showed that QTL for preference and performance for different host plants cosegregate in pea aphids. These insects mate on their host, providing a rapid path to speciation. The resolution of molecular markers available at the time allowed linkage to be confirmed to no more than approximately 10 cM, but even this could substantially impede the breakdown of LD: whereas LD between unlinked loci declines by 50% in one generation of random mating, LD between two loci that are 10 cM apart would decline by only approximately 9% per generation (cyclical parthenogenesis would further reduce recombination in these aphid species). Extending the same logic to our results, LD between the preference locus and optix on Chromosome 18 would be expected to decline by 1.2% (0–4.6) per generation (S5 Fig), assuming random mating. However, alleles at the behavioral locus result in a preference for the trait controlled by optix: LD will be further maintained by nonrandom mating because warning pattern is a magic trait. As such, LD is likely to decline much more slowly than this simple model would suggest.

More recently, Bay and colleagues [15] have reported widespread physical linkage between loci for divergent mate choice and ecological phenotypes in benthic and limnetic populations of three-spine sticklebacks. Two lines evidence support this. First, individual QTLs for mate choice and morphology map to Chromosome 14. Second, a polygenic QTL model predicting hybrid position along the benthic–limnetic morphological axis, generated by a previous study [10], explains a significant proportion of variance in mate choice, consistent with physical linkage of ecological and mate choice loci. Our results complement this previous work by explicitly demonstrating tight linkage between assortative mating and ecological traits. In addition, our study shows a much simpler genetic architecture, which should further facilitate the maintenance of LD between traits and which is predicted to facilitate speciation [2].

When mate choice is based on a preference for divergent ecological traits, this will inevitably couple ecological and behavioral components of reproductive isolation. Furthermore, the strength of LD generated will be proportional to the strength of the mating preference, so a genetic architecture with large-effect loci controlling assortative mating will generate stronger LD than a more polygenic architecture. Both our simulations and replication analysis support the existence of large-effect QTLs controlling an important interspecific difference in preference behavior. Even if we adopt an especially cautious approach, the QTLs on Chromosomes 1 and 18 would explain at least 10% and 20% of the difference in male preference behavior, respectively. However, our follow-up analysis, exploiting individuals that were not used to determine significance (thereby evading the Beavis effect), suggests that these estimates are overly conservative; these data explicitly reinforce our initial estimate for the QTL on Chromosome 18, which explains approximately 30% of the difference between parents. One potential caveat is that the position of the putative QTL and that of optix are not the same but are 1.2 cM apart; however, any recombination between these loci in the individuals tested will be rare (we expect just 0.42 recombination events between these two loci across 35 individuals) and likely has very limited impact on our estimates of effect size.

We observed a dramatic reduction in admixture (estimated using f_d) at the proximal end of Chromosome 18 and specifically on the distal side of optix coincident with our QTL. It is tempting to ascribe this to the combined effects of the major preference locus we have identified and the color pattern gene optix. However, in the populations studied here, the phenotypic effect of optix is more striking than the other color pattern loci, and selection against introgression is likely to be stronger at this locus. As a result, tight linkage with optix makes it impossible to determine any effects of the preference locus alone. Similarly, it is difficult to infer a signal of reduced admixture due to the behavioral QTLs on Chromosomes 1 and 17. Levels of F_st are high across the genome between H. cydno and H. melpomene, and patterns of admixture across the genome suggest widespread selection against introgression [42]. At this point, the patterns of divergence between H. cydno and H. melpomene are so heterogenous that it is difficult to disentangle the many processes that could be driving reduced admixture.

A general caveat of our results, alongside other studies of the genetics of assortative mating in Heliconius [35,36] and elsewhere (e.g., [15,56]), is that it is hard to distinguish between loci affecting preference behaviors, per se, from other traits that influence the behavior of the opposite sex. Here, we measured the time Heliconius males spend courting a particular female, which may depend not only on male attraction but also on the female’s response to male behavior (and in turn the male’s response to the female’s behavior). Recent work suggests that H. cydno females respond differently to H. pachinus than conspecific males [27]. Although there is currently no evidence that female Heliconius use color pattern as an interspecific mating (or rejection) cue (but see [57]), it is not inconceivable, and this could perhaps account for the apparent linkage between male interest and forewing color observed in our study and elsewhere [35,36]. In addition, it is possible that either of these QTLs we identified might influence male pheromones, which has been shown to influence female acceptance behaviors within H. melpomene [58]. Nevertheless, using the same hybrids as studied here, we previously demonstrated that individuals that have inherited the red band allele from H. melpomene are more likely to court artificial females with the red melpomene pattern, implying that the QTL on Chromosome 18, at least, influences male response to a visual cue [25]. Regardless of the exact proximate mechanisms involved, the QTLs we identify here influence an important component of male assortative mating behavior.

Overall, the scenario we describe reflects one that modeling broadly predicts will generate a strong overall barrier to gene flow through reinforcement [4]: specifically, the effects of barrier loci on prezygotic isolation are strong, recombination between pre- and postmating isolation barrier loci is reduced, and hybridization imposes high costs. Indeed, experimental evidence shows that nonmimetic hybrids between H. melpomene and H. cydno suffer not only increased predation [24] but also reduced mating success [22] and fertility [59]. In addition, males make a considerable reproductive investment by donating a nutrient-rich spermatophore during mating [60,61], so indirect selection against poorly adapted hybrids could strengthen divergent male preferences. Consistent with a role of reinforcement, H. melpomene males from French Guiana, outside the range of H. cydno, are less choosy than males from Panama, where the species co-occur and are known to occasionally hybridize [21], and similar patterns of reproductive character displacement have been observed elsewhere in the melpomene–cydno clade [62].

Reinforcement is further promoted when indirect selection, resulting from coupling of prezygotic and postzygotic barrier effects, is supplemented by direct selection [4,63]. In Heliconius, divergence in male preferences is likely initiated by divergence in wing pattern, and male preferences are observed between populations with few opportunities for hybridization, e.g., [64]. Female remating is a rare event [65], and males must compete to find virgin females within a visually complex environment [26]. Divergence in female (and male) wing patterns is driven primarily by strong selection for mimicry and is likely to impose divergent sexual selection on male preferences to improve their ability to find receptive females. This is similar to examples of assortative mating driven by sensory drive, such as in cichlid fishes [66], but it is perhaps less well appreciated that morphological traits under ecological selection (such as Heliconius wing patterns) might impose divergent sexual selection on male preferences in a similar fashion.

In addition to a simple genetic architecture, different QTLs appear to control different aspects of preference behavior. Our post hoc analyses suggest that differences associated with QTL1 and QTL17 in the relative amount of time spent courting each female type are driven by differences in attraction to either H. cydno or H. melpomene, respectively, rather than both species. QTL18 also seems to influence attraction to H. melpomene much more strongly than to H. cydno females. This genetic modularity, where discrete, independently segregating loci appear to affect different aspects of behavior, may facilitate evolutionary change and innovation by providing a route for rapid evolution of novel behavioral phenotypes [44,67]. In Heliconius, this might allow different aspects of mating behavior to evolve independently. It might also allow novel composite behavioral preferences to arise through hybridization and recombination. There is some evidence that this has occurred during hybrid speciation in Heliconius. The wing pattern of the hybrid species H. heurippa includes both red and yellow pattern elements, which are believed to have originated from the putative parental species H. melpomene and H. cydno, respectively (local Colombian races of H. cydno have a yellow, as opposed to white, forewing band) [23]. Not only do H. heurippa males prefer this combined pattern over the pure red or yellow patterns of H. melpomene and H. cydno [23], but “recreated H. heurippa,” obtained in first-generation backcrosses between H. melpomene and H. cydno, prefer the pattern of H. heurippa over that of the two putative parents [68]. This is consistent with a hypothesis in which introgression and subsequent recombination of preference alleles are responsible for novel behavioral phenotypes, although further work would be needed to confirm this.

In conclusion, the genetic architecture we demonstrate here will promote the evolution of behavioral isolation by strengthening genetic associations between cue and preference. Disassociation of alleles at loci that are physically close on the chromosome is slower compared to that between alleles at more distant loci (due to reduced crossing over) or at loci on different chromosomes. Similarly, the substitution of large-effect alleles will also increase linkage disequilibrium between cue and preference, even if they are not physically linked, because preference alleles of larger effect will more often find themselves in the same genome as alleles for the corresponding cue compared to preference alleles with smaller effects. We cannot currently distinguish whether preference QTLs result from single adaptive mutations or represent multiple functional loci that have built up during the course of speciation. Nevertheless, the genetic basis of Heliconius mate preferences is remarkably similar to that for differences in the wing pattern cue. Differences in individual color pattern elements probably do involve multiple, sequential mutations (which target the same genes), but “ready-made” alleles of large phenotypic effect can be brought together in new combinations through adaptive introgression. The existence of large-effect preference loci, potentially influencing different aspects of behavior, could similarly facilitate the origin of novel phenotypes through introgression and further facilitate rapid speciation.

Methods

Supporting information

Acknowledgments

We are grateful to Moisés Abanto, Josephine Dessmann, Janet Scott, and Bas van Schooten for help in the insectaries; to Justin Touchon for advice regarding statistics; and to Stephen Montgomery, Markus Möst, Ricardo Pereira, Alexander Hausmann, and Vera Warmuth for comments on the manuscript. We thank the Smithsonian Tropical Research Institute for support and the Ministerio del Ambiente for permission to collect butterflies in Panama. We are also grateful to Edinburgh Genomics for sequencing support.