Phylogeography and demography of the H. melpomene clade

We obtained approximately 5.2 Mb of sequence distributed across 8 chromosomes from 473 individuals and 39 populations representing 10 species from the H. melpomene clade (S1 and S2 Tables). Phylogenetic reconstructions confirmed that H. cydno populations, with the sole exception of H. c. cordula found east of the Andes and in the Magdalena Valley, and H. timareta populations from east of the Andes cluster as separate lineages from the H. melpomene clade (Fig 1B and 1D). Phylogenetic inferences including all sequenced regions agreed with previous multilocus phylogenies, in which H. cydno and H. timareta form a sister clade to H. melpomene (Figs 1D and S1) [44,50]. The tree built using only neutral background data (i.e., regions a priori not suspected to be under mimicry selection, see Materials and methods) largely clustered populations according to geography, i.e., H. cydno with western H. melpomene and H. timareta with eastern H. melpomene subspecies (Fig 1B and 1D). The neutral topology is consistent with ongoing gene flow between sympatric populations resulting in highly heterogeneous relatedness patterns along the genome [71,72]. Six out of nine individuals with the dennis-ray pattern, sampled from the H. melpomene vicina population in the Colombian Amazon (Fig 1A and 1C), consistently clustered within H. timareta. This suggests the presence of a lowland population of H. timareta considerably further from the Andes than has been detected previously, hereafter referred to as H. timareta ssp. nov. (Colombia).

Download:

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

Fig 1. Distribution, phylogenetic relations, major colour pattern loci, and sequence capture targets of the H. melpomene, H. cydno, and H. timareta clade species.

(A) Broad distributions of the H. melpomene, H. cydno, and H. timareta colour pattern races and species (based on all known sampling localities; for details, see S2 Fig). Distribution colours match the shadings around the phylogeny and butterfly images in panel B. The dashed line indicates the Andes. Note the distinct clusters formed by individuals sampled from the H. m. vicina population. The cluster grouping with H. timareta is referred to as H. timareta ssp. nov. (Colombia) (B) FastTree cladogram inferred using capture sequence from putatively neutral loci. Colours in the tree indicate the H. melpomene (pink), H. cydno (green), and H. timareta (blue) clades and match the boxes of the distribution maps in panel A. (C) Sequence information was obtained for 4 putatively neutral regions (green) and 4 regions to which functional variation has been mapped to a yellow/white colour switch (chr 1); forewing band shape (chr 10); yellow/white fore- and hindwing bars, band margins, and ventral colour (chr 15); and red colour pattern elements (chr 18). The various phenotypes controlled by the respective colour pattern loci are depicted. Note that whereas most phenotypes have descriptive names, the red blotch at the base of the forewing was termed ‘dennis’. (D) Phylogenetic relations obtained when building a tree from all captured regions compared to the neutral regions.

https://doi.org/10.1371/journal.pbio.3000597.g001

To assess demographic events, which may affect selection tests, we estimated effective population size across time for all populations with whole-genome data (S1 and S3 Tables). In line with previous studies [51,70], we found that bottlenecks were rare across those populations with the exception of a recent decline in population size in H. heurippa and older, moderate dips in H. besckei and H. m. nanna (S3 Fig).

Signatures and limits of detection of classic sweeps assessed by simulations

We used forward-in-time simulations to investigate differences in the signals produced by classic as compared to introgressed selective sweeps in genome scan data, which have been relatively unexplored [27]. Our simulation results are intended to demonstrate qualitative patterns, but we also parameterise the simulations according to the Heliconius populations. This allows us to assess the time period over which sweeps can be detected in real data and place bounds on the timing of selection in natural populations. In our analysis, we primarily use SweepFinder2 (SF2), which is appropriate for our genomic data because it is able to identify the sweep site. This method is also robust to demographic processes [73,74], because these are incorporated in the null model used by SF2 (for more details, see Materials and methods). However, to more qualitatively explore patterns of diversity at sites undergoing selection, we here also present results for Tajima’s D.

The time over which we can expect to detect sweep signals is determined by the time to coalescence and is thus determined by N, the (effective) population size. We therefore here report time since the sweep in generations, scaled by 4N [75]. Sweep signals are expected to decay rapidly because of the joint effects of mutation, recombination, and drift. Indeed, SF2, which uses the predicted effect of a selective sweep on the local site frequency spectrum (SFS) to infer the probability and location of sweeps [73,74,76], has low power to detect even hard selective sweeps that occurred over 0.25 (scaled) generations ago and cannot localise sweeps older than 0.4 (scaled) generations [74]. Consequently, any detected sweep signals in H. melpomene are likely under 0.8 Mya, assuming an effective population size of 2 million [70,77] and a generation time of 3 months [78]. These estimates vary with N, so the time limit for sweep detection varies among species, from only 0.2 Mya for H. besckei (N approximately 0.5 million) to 1.4 Mya for H. erato (N approximately 3.5 million). We used simulations to further interpret the empirical signatures of selection and explore the limits of detection (Fig 2).

Download:

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

Fig 2.

SFS signatures of selection for simulated classic hard sweeps (left) and introgressed sweeps (right). (A) CLR statistics (upper panel, [73,74]) and Tajima’s D (lower panel) across a simulated chromosome for different time points (0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, and 1 in units of scaled generations, i.e., 4N generations) after a classic hard (left) or introgressed (right) sweep (effective migration rate M = 0.2). The sweep occurs in the centre of the simulated chromosome. Different colours indicate time since sweep. Full, dashed, and dotted vertical black lines in the lower panel indicate positions at different distances from the sweep centre for which time series of CLR and Tajima’s D statistics are depicted in panel B in the same style. (B) CLR (upper panel) and Tajima’s D (lower panel) statistics over time at 3 positions relative to the sweep centre as shown in panel A. Also shown are neutral background values, BG, calculated over neutral simulations, either without migration (left hand panels, for classic sweeps) or with migration at M = 0.2 (right hand panels, for introgressed sweeps). Time is given in units of scaled generations. Data are available from https://github.com/markusmoest/SelectionHeliconius.git. BG, background; CLR, composite likelihood ratio.

https://doi.org/10.1371/journal.pbio.3000597.g002

We initially simulated the case of a hard sweep, such that s = 0.5, which is appropriate to the very strong selection pressure experienced by the colour pattern loci in Heliconius (Table 1). We found that SF2 signals broke down rapidly after the sweep (Fig 2). The magnitude of the CLR peak decreased by an order of magnitude after just 0.1 scaled generations, corresponding to 0.2 Mya for H. melpomene, and was not distinguishable from background values after 0.2 scaled generations, i.e., 0.4 Mya in H. melpomene (Welch t test, p = 0.065). Similarly, the estimated strength of selection calculated with SF2 from our simulations declined rapidly with time. Although the magnitude of the SF2 peak is affected, we find that the time for which we can detect selective sweeps does not change if we vary either the strength of selection (using alternative values of s = 0.1 and s = 0.25), or the mutation rate, which was scaled up such that levels of neutral diversity in our simulations are equivalent to those seen in our Heliconius populations (S4 Fig and S4 Table). Levels of linkage disequilibrium were in the range of the empirical data for all simulated scenarios (S4 and S15–S18 Tables).

Signatures and limits of detection of introgressed sweeps assessed by simulations

We extended our simulations to explore the expected SFS signature left by an allele undergoing adaptive introgression, by simulating a second population which exchanged migrants with the first, leading to an introgressed sweep in the second population. Adaptive introgression produces a highly distinctive SFS signature. At and very close to the selected site itself there was a reduction in diversity and an excess of rare alleles, similar to the pattern observed for a classic sweep. However, this reduction was narrow and flanked by broad genomic regions with high diversity and an excess of intermediate frequency variants. This is due to variants that have hitchhiked into the recipient population along with the beneficial variant and subsequently recombined before reaching fixation [20,27]. The overall SFS signature covered a considerably wider genomic area than that of a classic sweep (Fig 2).

The introgression signature we observe at the sweep site itself was very similar to that for a classical sweep, and we could detect it for a similar length of time. SF2 managed to detect introgressed sweeps, although it detected only the central region of lowered diversity, producing a high but very narrow CLR peak at the sweep site itself; this contrasts with the peaks for classic selective sweeps, which extended over a wider genomic area (Fig 2). The distribution of CLR values at the sweep site was significantly different from values calculated over neutral regions for up to 0.1 generations after the sweep (p = 0.0041). However, as for a classical sweep, the magnitude of the peak decreased rapidly.

In the simulations described above, we used an effective migration rate of M = 0.2. Estimates of M between hybridising Heliconius species vary from 0.08 to 10 migrants per generation [47–49], and so we also explored a broad range of values of M, from 0.02 to 200, in order to cover the estimated range for Heliconius (S5 Fig). We find that the the reduction of diversity at the introgression site itself is strongly affected by migration rate. As M increases, the central reduction in diversity becomes less pronounced, representing an increasingly ‘soft’ introgressed sweep (S5 Fig) [21,79]. Therefore, detecting introgressed sweeps from this central region will be difficult in populations in which M is high. However, for values of M below 2, varying M had little effect on the regions of increased diversity and excess of intermediate frequency variants that flank the sweep locus (S5 Fig).

Strong signatures of selection across Heliconius colour pattern regions

In our empirical data, SF2 found strong support for positive selection acting across multiple populations and species for all 4 colour pattern loci (Fig 3). In contrast, our background regions, as well as regions flanking the colour pattern associated loci, showed little evidence of sweeps, apart from a few isolated examples (S6 Fig).

Download:

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

Fig 3. Signature of selection across colour pattern regions in the H. melpomene clade.

The regions containing the tandem copies of aristaless, al1 and al2, WntA, cortex, and optix (left to right) are depicted. Colour pattern genes are annotated in red in the gene annotation panel. On the y-axis Sweepfinder2’s CLR statistics is shown (peaks capped at 1,000). The colour gradient indicates the estimated intensity of selection α [73] (black = high α values, weak selection; red = low α values, strong selection). Grey shadings indicate annotated colour pattern CREs [30,36,37,39] (S7–S10 Figs). Blue horizontal bars indicate regions with CLR values above threshold. Top panel shows colour pattern phenotypes and symbols indicate distinct colour pattern elements, and their presence is annotated in population panels. Note that the yellow hindwing bar controlled by the cortex region can be expressed on the dorsal and ventral side (yellow/yellow square symbol) or on the ventral side only (black/yellow square symbol) [39]. Moreover, the actual shape of the forewing band can depend on the allelic state of WntA. Full, grey lines connect colour pattern elements with annotated CREs. Phenotypes are depicted on the right. Data are available from https://github.com/markusmoest/SelectionHeliconius.git. CLR, composite likelihood ratio; CRE, cis-regulatory element.

https://doi.org/10.1371/journal.pbio.3000597.g003

This is consistent with previous genome-wide selection scans in H. melpomene that detected only a few strong sweep signatures [70]. These results therefore lend support to the long-standing assertion that wing patterning loci are among the most strongly selected loci in the genome and have a distinctive evolutionary history [80], without excluding the potential presence of other local sweeps in the respective populations.

Broadly, signals of selection were stronger and more widespread in regions near cortex and optix and weaker near WntA and aristaless. For example, all 31 populations showed sweep signals above threshold near cortex, 26 near optix, 24 near WntA, albeit less pronounced in most cases, and only 7 near aristaless (Fig 3 and S5–S8 Tables). A similar pattern was reflected in our estimates for strength of selection (s) calculated from α estimates (Table 2; see Materials and methods for a detailed description and formula for this calculation) with the highest selection strength at colour pattern loci being s = 0.141 for the cortex (H. m. nanna), s = 0.036 for the optix (H. m. plesseni), s = 0.049 for the WntA (H. m. xenoclea), and s = 0.01 (H. t. florencia) for the aristaless region (H. t. florencia). These patterns are broadly concordant with the expected phenotypic effects of these loci. For example, in H. cydno, which has primarily yellow and/or white patterns associated with the cortex region [39,81], significant peaks were mostly found at this locus, whereas in H. melpomene, which has red, yellow, and white patterns, strong signals were seen at both cortex and optix regions. Consistently, a lower strength of selection was found for the aristaless region (s < 0.01), which controls a modification of pale patterns from yellow to white that is putatively less salient to predators [82] and may contain fewer potential targets of selection.

Download:

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

Table 2. Position, CLR statistics, and estimates for strength of selection (α, 2N_es, and s) for populations and sweeps discussed in detail.

Annotated colour pattern genes and CREs that overlap with peaks are given. Positions are given in Hmel2 scaffold coordinates (see S5 and S7 Tables).

https://doi.org/10.1371/journal.pbio.3000597.t002

There were also differences seen across the sampled populations. Widely distributed colour patterns (e.g., H. m. melpomene and H. m. malleti) tended to show only modest evidence for selective sweeps (Figs 3 and S11). Comparisons with our simulated data nonetheless suggest selective events that occurred no more than 400,000 years ago. Although there was no significant general correlation between distributional ranges of populations and evidence for selection (S12 and S13 Figs), the strongest signatures of selection were found in geographically localised patterns and likely reflect sweeps within the last 100,000 years (Fig 4 and Table 2). For example, H. m. plesseni is exclusively found in the upper Pastaza valley in Ecuador and shows a unique split red-white forewing band (Figs 1 and 4). This population showed strong selection at 3 colour pattern regions—optix, cortex, and WntA—suggesting recent selection acting on the entire pattern (s_cortex = 0.074, s_WntA = 0.035, and s_optix = 0.035), and patterns of both nucleotide diversity and Tajima’s D are consistent with strong classic sweeps (Figs 3, 4 and S11 and S5 Table). H. m. xenoclea, also found on the eastern slopes of the Andes but further south in Peru, shows the same split forewing band associated with the WntA region and again a very strong selection signal at this locus (s_WntA = 0.049), as well as weaker signatures at cortex (s_cortex = 0.04) and optix (s_optix = 0.022; Figs 3 and S11 and S5 Table). The clear signatures of recent and strong selection pressure perhaps indicate that the split forewing band is a novel and highly salient signal. Additionally, H. m. meriana from the Guiana shield revealed a striking signature of selection at optix (s_optix = 0.023). Its dennis-only pattern (Fig 4) has previously been shown to have arisen through recombination between adjacent dennis and ray regulatory modules at optix, and the signature of selection at this locus, which encompasses both of these regulatory modules, implies a recent sweep of this recombinant allele [36] (Figs 3, 4 and S11 and S5 Table).

Download:

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

Fig 4. Selected examples of sweeps.

The 3 examples show the split forewing band (WntA region) in H. m. plessini, the yellow and white patterns (cortex region) in H. cydno weymeri f. weymeri and the red dennis patch (optix region) in H. m. meriana (left to right). The respective colour pattern elements are indicated with red and grey arrows. Colour patterns and gene annotations in the colour pattern regions are depicted in the top panel. Colour pattern genes are annotated in red. Nucleotide diversity π, Tajima’s D, and SweepFinder2’s CLR statistics (peaks capped at 1,000) show the signatures of a selective sweep (bottom panels). Loess smoother lines are depicted in yellow. The colour gradient in the CLR panel indicates the estimated intensity of selection α [73] (black = high α values, weak selection; red = low α values, strong selection). Grey shadings indicate annotated CREs, and red and grey arrows depict associations with the respective colour pattern elements in the H. melpomene clade. Data are available from https://github.com/markusmoest/SelectionHeliconius.git. CLR, composite likelihood ratio; CRE, cis-regulatory element.

https://doi.org/10.1371/journal.pbio.3000597.g004

In light of our simulations of introgressed sweeps, there were cases in our data in which previously well-documented adaptive introgression events showed signatures characteristic of introgressed sweeps. The hindwing yellow bar pattern was shown to have introgressed from H. melpomene into H. c. weymeri and then back again into the races H. m. vulcanus and H. m. cythera [39]. Accordingly, we found narrow SF2 peaks and an increase in Tajima’s D at surrounding sites at these modules in the cortex region in H. m. cythera, H. m. vulcanus, and H. c. weymeri, consistent with introgressed sweeps (Fig 3 and S11 Fig). H. c. weymeri f. weymeri also had a second, striking signature further upstream more typical of a classic sweep (Figs 3 and 4), at a region associated with the yellow forewing band in H. melpomene and H. timareta [30]. This is consistent with evidence for a role of cortex in controlling the white forewing band in H. cydno [81] and the presence of this band in the weymeri morph, which could therefore represent a recent evolutionary innovation. Other loci previously implicated as having introgressed include the optix region in H. heurippa and H. elevatus, which both showed signals coinciding with regions previously associated with the respective phenotypes [36,37]. In contrast, there was a lack of clear introgressed sweep signals in dennis-ray H. timareta, which is one of the best documented examples of introgression. This could be explained by the age of the sweeps and/or high rates of migration, which our simulations show can reduce the sweep signal in the recipient population (S5 Fig). We also performed scans with VolcanoFinder, a new method designed to detect SFS signatures created by introgressed sweeps [27]. Similar to SF2, VolcanoFinder detected strong signatures of selection in colour pattern regions in the respective populations but not in the neutral background regions (S14–S16 and S19 Figs). However, the estimated divergence values (D) did not allow for a clear distinction of introgressed from classic sweeps in our data.

Novel targets of selection in colour pattern regions

Many of the signals of selection we detected overlap with previously identified regulatory regions associated with colour pattern variation. However, our analysis also found additional nearby regions showing consistent signals of selection that may also be involved in colour pattern evolution (Figs 3 and S17). For example, in the first intron of the WntA gene, we found a consistent signal across several H. melpomene, H. timareta, and H. cydno populations (S17B Fig). Within this region (Hmel210004:1806000–1833000), phylogenetic clustering of the 2 split forewing band races H. m. plesseni and H. m. xenoclea indicates a common origin of the split band in these currently disjunct populations (S7 Fig). Additionally, 2 strong selection signatures are frequently found in a region approximately 200 kb upstream of WntA (S17B Fig; Hmel210004:1550000–1650000), which suggests additional loci involved in colour pattern regulation.

Near cortex, selection signatures at closely linked genes support findings from previous studies. Several populations show distinct peaks upstream and downstream of cortex and broadly coincide with a wider region, possibly containing several genes involved in colour pattern regulation [30,84] (S17C Fig). Multiple peaks are located upstream of cortex within an array of genes that all showed significant associations with yellow colour pattern variation [30] (S9 Table). A particular concentration of signals fell near the serine/threonine-protein kinase gene LMTK1 (HMEL000033; Hmel215006:1,418,342–1,464,802) and close to washout. The latter gene is involved in actin cytoskeleton organization in Drosophila [85] and previously showed a strong association with the yellow forewing band [30] as well as differential expression patterns between different H. numata morphs [84]. Likewise, selection signals clustered downstream of cortex in a region containing additional candidate genes identified previously (S9 Table). In the optix region, consistent signals across several populations indicated that several as yet uncharacterised elements may be under mimicry selection. Intriguingly, a kinesin motorprotein gene, which shows an association of expression with the red forewing band [86,87], was among these (S17D Fig).

Parallel selective sweep signatures between mimetic species

There has been considerable interest in whether the H. erato and H. melpomene co-mimics have co-diverged and simultaneously converged onto the same colour pattern [88–91] or whether one species evolved towards diverse phenotypes of the other, i.e., advergence [67,92–94]. Homologous genes control corresponding phenotypes [30,35,95,96], but there is no allele sharing between the melpomene and erato clade [67,68]. We used published genomic data for H. erato (Van Belleghem and colleagues, 2017; S10 Table) to obtain 8.9 Mb of sequence homologous to the regions studied in the H. melpomene clade for 103 individuals from 13 populations and 3 species in the H. erato radiation and scanned for selective sweeps. Generally, a comparison of the location of selection peaks between H. melpomene and H. erato across several co-mimetic races suggests a rather simple and concordant regulatory architecture in the 2 species at the WntA locus. However, in the cortex and optix regions, this architecture appears to be more complex and differs more strongly between the 2 clades (Figs 5, S17 and S18).

Download:

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

Fig 5.

Signatures of selection in the co-mimic populations of H. melpomene (upper panels) and H. erato (lower panels). The regions containing WntA, cortex, and optix are shown (left to right). Co-mimics in H. melpomene and H. erato are depicted in the same order with phenotypes on the left. The y-axis indicates CLR statistics across each region (capped at 1,000). The colour gradient indicates the estimated intensity of selection α [73] (black = high α values, weak selection; red = low α values, strong selection). Grey shadings indicate annotated colour pattern CREs [30,36,37,39] (S7–S10 Figs) and blue horizontal bars indicate regions with CLR statistics above threshold. The central panel shows an alignment of the respective regions in H. melpomene and H. erato and gene annotations with colour pattern genes in red. Top and bottom panel show colour pattern phenotypes, and symbols indicate distinct colour pattern elements and their presence in each population panel. Note that the yellow hindwing bar controlled by the cortex region can be expressed on the dorsal and ventral side (yellow/yellow square symbol) or on the ventral side only (black/yellow square symbol) [39]. Full, grey lines connect colour pattern elements with annotated CREs. Note that the genetics of the yellow forewing band differs between H. erato¸ in which it involves the WntA and optix locus, and H. melpomene, in which the band is controlled by the cortex and its shape by the WntA region. Data are available from https://github.com/markusmoest/SelectionHeliconius.git. CLR, composite likelihood ratio; CRE, cis-regulatory element.

https://doi.org/10.1371/journal.pbio.3000597.g005

Similar to the melpomene clade radiation, we found strong signatures of selection across the optix, cortex, and WntA regions (Figs 5 and S20–S22 and Tables 2 and S11–S14). Most notably, H. e. notabilis from Ecuador showed strong signals of selection at 3 colour pattern loci (s_optix = 0.06, s_cortex = 0.015, s_WntA = 0.015) similar to its co-mimic H. m. plesseni (Table 2). In both cases, selection across the 3 major loci represented some of the strongest signals in both species. Additionally, H. e. amalfreda, co-mimic with the red dennis-only race H. m. meriana, showed one of the strongest selection signals at optix. This suggests that these phenotypes are recent innovations in both species, consistent with co-divergence. Other geographically localised variants controlled by WntA also showed strong signals of selection, indicating a recent origin. For example, H. e. etylus, like H. m. ecuadoriensis, has a restricted forewing band shape that corresponds to the more distal element of the notabilis forewing band (s_WntA = 0.015). Clear, narrow, and very similar selection signals were found near WntA in H. e. amalfreda and H. e. erato (s_WntA = 0.006 in each), both with a broken forewing band, as well as H. e. emma (s_WntA = 0.003) and H. e. lativitta (s_WntA = 0.004), both with a narrow forewing band (S11 Table).

More broadly across the H. erato populations, there was a clear difference between the Amazonian dennis-ray races (i.e., H. e. amalfreda, H. e. erato, H. e. emma, H. e. etylus, and H. e. lativitta), all exhibiting a similar selection pattern at optix, and red forewing band races (H. e. favorinus, H. e. venus, H. e. cyrbia and H. e. hydara in Panama, and H. e. demophoon) which showed little or no signature of selection. This is in agreement with the hypothesis that the widespread dennis-ray phenotype at optix has a more recent origin as compared with the red band phenotype [67]. One notable exception to this pattern was H. e. hydara in French Guiana, the only red banded H. erato form with a strong signal at optix (s_optix = 0.09). There are slight variations across the range in the band phenotype, and perhaps a recent modification of the band phenotype swept in this population. The pattern in H. melpomene is less clear, possibly due to the age of the alleles and the considerably lower effective population size in H. melpomene.

At the cortex locus, there was a consistent peak centred on lethal (2) just next to the cytokine receptor gene domeless, which in Drosophila is essential for the JAK/STAT signalling pathway controlling embryonic segmentation and trachea specification [97], and washout (annotated in S18 Fig). However, surprisingly, the signal is almost identical across populations with a variety of different yellow colour pattern phenotypes (H. e. amalfreda, H. e. erato, H. e. hydara in French Guiana, H. e. emma, H. e. etylus, H. e. lativitta, H. e. notabilis, H. e. favorinus, H. himera) and completely absent in Northwestern populations (H. e. cyrbia, H. e. venus, H. e. hydara in Panama, H. e. demophoon; S20 Fig). The sweep signal therefore shows little obvious association with any particular wing pattern phenotype but may still indicate a locus involved in the colour pattern pathway. In addition, we detected very distinct signals between H. e. favorinus (Cr1) and H. e. demophoon (Cr2) consistent with previous studies [30,38,98] that found evidence for independent evolution of the yellow hindwing bar on either side of the Andes. Although H. e. favorinus lacks any signature at Cr2 and shows a weak signal at Cr1, a clear peak was found for H. e. demophoon at Cr2 indicating that this allele may be more recent (Figs 5, S18 and S20).

Ethics statement

Panamanian specimens were collected under permit SE/AP-14-18 issued by the Ministerio de Ambiente de Panamá. Samples from Ecuador were collected with permission of the Ministerio del Ambiente under permits number 006-2012-IC-FAU-DPL-MA, 002-16-IC_FLO_FAU_DNB/MA, 033-10-IC_FAU/FLO_DPN/MA, and 0007-IC-FAU/FLO-DPPZ/MA. Colombian specimens were collected under the permit IDB0199/No16 and permit 530 granted to Universidad del Rosario by the Autoridad Nacional de Liencias Ambientales (ANLA-Colombia). Samples from Peru were collected under permit N°0148-2011-AG-DGFFS-DGEFFS and N°0236-2012-AG-DGFFS-DGEFFS from the Ministerio de la Agricultura, Peru. Samples from Suriname were collected and exported under a permit (No. 10865) from the Nature Conservation Division of the Suriname Forest Service. Field collections in Brazil were made under IBAMA/ICMBio license number 2024629 granted to GRPM. Recommendations of Animal Care and Use Committee (CEUA) of the Federal University of Rio Grande do Sul (UFRGS) were followed during laboratory procedures, including DNA extractions.

Sampling and DNA extraction

Our sampling covers most of the distribution and colour pattern variation of the Heliconius radiation in South and Central America. Specimens were sampled or provided by collaborators with the respective sampling permissions and stored in salt saturated DMSO or ethanol at −20°C until further processing. For DNA extractions, thorax muscle tissue was dissected, disrupted and digested, and DNA was extracted using a TissueLyser II bead mill together with the DNeasy Blood and Tissue Kit (Qiagen, Hilden, Germany) following supplier recommendations.

Targeted capture and sequencing

For hybridisation-based target enrichment a NimbleGen SeqCap EZ Library SR capture probes library was designed and synthesized by the provider (Roche NimbleGen Inc, United States). The templates for designing probes for 4 colour pattern regions (approximately 3.2 Mb) and 4 genomic background regions (approximately 2 Mb) were assembled and curated using the H. melpomene genome assembly Hmel1 [44], available BAC walks [31,115], fosmid data [69], and alignments from Wallbank and colleagues [36]. The neutral background regions were chosen to represent the average genome. We therefore excluded regions with extended stretches of extreme values for diversity and/or divergence, and we only considered regions located on a single, well-assembled scaffold.

Sample DNA was sheared with an ultrasonicator (Covaris Inc, Massachusetts, United States) and adapter-ligated libraries with insert sizes of 200 to 250 bp were generated using the Custom NEXTflex-96 Pre-Capture Combo Kit (Bioo Scientific Corporation, United States). For sequence capture, 24 libraries each were pooled into a capture library, hybridised with blocking oligos and the biotinylated capture library probes, and subsequently captured with streptavidin-coated magnetic capture beads using the NimbleGen SeqCap EZ Kits (Roche NimbleGen Inc, Wisconsin, United States). After capture and clean-up, 3 capture library pools were combined each. For the resulting sequencing pools of 72 samples, Illumina 100 or 150 bp paired-end short read data were generated on Illumina’s HiSeq 2000 (BGI, China) and HiSeq 4000 (Novogene Co. Ltd, China), respectively (S1 Table).

Whole-genome data

Whole-genome resequencing data available for the melpomene clade from previously published work were also included [30,39,42,44,45,51,70–72]. For a few additional samples, 100 to 150 bp paired-end whole-genome resequencing data were generated on an Illumina X Ten platform (Novogene Co. Ltd, China; S1 Table). In addition, we downloaded, processed, and analysed a publicily available data set for H. cydno galanthus [49] with a more moderate depth of coverage (for results see S14 Fig). For the erato clade already published, whole-genome-resequencing data were used [38] (S10 Table).

The whole-genome data were mainly used for demographic reconstructions, whereas, for other analyses, the regions matching the capture regions were used.

Genotyping

For melpomene clade data, sequenced reads were aligned to the H. melpomene version 2 reference genome (Hmel2, [137]), using BWA-mem version 0.7 [116]. PCR duplicated reads were removed using Picard version 2.2.4 (http://picard.sourceforge.net), and reads were sorted using SAMtools version 1.3.1 [117]. Genotypes for variant and invariant sites were called using the Genome Analysis Tool Kit’s (GATK) Haplotypecaller version 3.5 [118]. Individual genomic VCF records (gVCF) were jointly genotyped per population using GATK’s genotypeGVCFs version 3.5 [118]. Genotype calls were only considered in downstream analyses if they had a minimum depth (DP) ≥ 10, and for variant calls, a minimum genotype quality (GQ) ≥ 30, and indels were removed. Filtering was done with bcftools version 1.4 [117], and for downstream calculations of summary statistics and creating SF2 input, VCF files were parsed into tab delimited genotype files (scripts available at https://github.com/simonhmartin). For the erato clade, read data were mapped to the H. erato demophoon version 1 genome reference [38] and further processed as described above.

Phasing

SHAPEIT2 [119] was used to phase haplotypes using both population information and paired read information. First, monomorphic and biallelic sites were filtered with GQ ≥ 30 and DP ≥ 10, and sites with less than 20% of sample genotypes were removed.

Next, phase informative reads (PIRs) with a minimum base-quality and read quality of 20 were extracted from individual BAM files using the extractPIRs tool. These BAM files were obtained from BWA-mem [116] mappings to the H. melpomene version 2 genome, with duplicates removed.

Finally, SHAPEIT2 was run with PIR information and default parameters on each scaffold using samples from single populations, which resulted in a haplotype file that was transformed into VCF format. Sites with no genotype information were imputed.

Phylogenetic reconstruction

FastTree2 [120] was run using default parameters to infer approximate maximum likelihood phylogenies. Separate phylogenies for a concatenated SNP data set comprising neutral background regions only and for the full data set including the colour pattern regions to account for the effect of including regions putatively under strong selection were produced.

Population historical demography

Changes in the historical population size were inferred from individual consensus whole-genome sequences (S3 Table) using Pairwise Sequentially Markovian Coalescent (PSMC’) analyses as implemented in MSMC [121]. This method fits a model of changing population size by estimating the distribution of times to the most recent common ancestor along diploid genomes. When used on single diploid genomes, this method is similar to PSMC analyses [122]. Genotypes were inferred from BWA version 0.7 [116] mapped reads separately from previous genotyping analysis using SAMtools version 0.1.19 [117]. This involved a minimum mapping (-q) and base (-Q) quality of 20 and adjustment of mapping quality (-C) 50. A mask file was generated for regions of the genome with a minimum coverage depth of 30× and was provided together with heterozygosity calls to the MSMC tool. MSMC was run on heterozygosity calls from all contiguous scaffolds longer than 500 kb, excluding scaffolds on the Z chromosome. We scaled the PSMC’ estimates using a generation time of 0.25 years and a mutation rate of 2×10⁻⁹ estimated for H. melpomene [47,77].

SLiM simulations

Simulations were conducted to compare the genomic signatures of classical selective sweeps and sweeps that occur via adaptive introgression using SLiM (version 2) forward-in-time population simulation software [123,124]. Because SLiM tracks mutations and individuals through time, we were able to track individual beneficial alleles going to fixation and post sweep; however, it is computationally intractable to simulate very large populations with SLiM, and so we instead simulated smaller populations and rescaled population genetic parameters, N and μ, such that our results are applicable to Heliconius (as is commonly done [124,125]). Two populations of N = 1,000 were simulated with a neutral mutation rate μ of 6×10⁻⁷ such that the expected level of neutral diversity in the population was 0.0024, which is within an order of magnitude of that observed in our Heliconius populations [38,70] (S15–S18 Tables). Each individual in our simulated populations was represented by a single diploid recombining chromosome (recombination rate was also scaled such that NR is within the values of those observed in Heliconius, 4×10⁻⁷, or 40 cM/Mb) of length 750,000 bp. We also ran simulations on a shorter length of chromosome (50,000 bp) with an higher value of μ, raising levels of neutral diversity to those observed within Heliconius, to ensure our results are consistent for higher values of μ.

Our simulations were first allowed to equilibrate for a burn-in phase of 10N generations, after which we introduced a single strongly advantageous mutation of s = 0.5 in the centre of the chromosome in order to simulate a ‘classical’ hard selective sweep in the population (which we will refer to as p1). We also ran our simulations with 2 lower values of s (s = 0.1 and s = 0.25). Only those simulations in which the mutation went to fixation were kept; if the beneficial mutation was lost during the course of a simulation, the simulation was reset to a point just after the burn-in phase and the mutation was reintroduced. The simulations were then allowed to run for a further 5N generations. During this time, p1 does not experience any migration or population size change. In order to simulate an introgressed sweep, we simulated an additional neutrally evolving population, p2, which exchanges migrants with population p1 at a constant rate of 0.0001 migrants per generation, which allowed the beneficial mutation fixed in p1 to introgress into p2. The simulations were then allowed to run for a further 10N generations with a constant migration rate. For each set of parameters, we ran our simulations 100 times.

For both populations, a complete sample of the segregating neutral mutations was taken every 100 generations after the burn-in phase and prior to the introduction of the beneficial mutation, and every 50 generations after the introduction of the beneficial mutation. We also tracked the change in frequency over time of the beneficial mutation during the simulations. From these results we calculated 2 summary statistics, Tajima’s D and π, in windows of 10,000 bp across our simulated chromosomes for a range of time points. Time points are as follows, in 4N generations post sweep: 0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 1, and 2 background rates: one post burn-in, during which populations are not experiencing any migration, and one post sweep, during which the populations are exchanging migrants. Values were then averaged across simulations. Additionally, to model the effect of changing effective migration rates on the introgression sweep signal, we ran simulations with different levels of migration, using the following 4 values of M: 200, 2, 0.2, and 0.02, with recombination rate = 4cM/Mb and s = 0.1. The simulations were otherwise set up as before, with 30 simulation runs generated for each set of parameters.

We used these results to generate SF2 [76] input files, after first subsampling the number of mutations down, such that our simulated SF2 files for each population represent a sample of 500 simulated individuals. This step is necessary because SweepFinder has an upper limit on the number of sequences that can be included per sample [126]. We then ran SF2 using mode–lg 100 for each simulation for each of the time points, using 1 of 2 precomputed site frequency spectra as appropriate: one calculated across multiple neutral simulations without migration and one calculated across multiple neutral simulations with migration (these neutral simulations correspond to the 2 background rates described above). Further details of SF2 and its various run modes are included in the ‘SF2’ section.

Phylogenetic weighting

A phylogenetic weighting approach was used to evaluate the support for alternative phylogenetic hypotheses across colour pattern loci using Twisst [127]. Given a tree and a set of predefined groups, in this case Heliconius populations sharing specific colour pattern elements, Twisst determines a weighting for each possible topology describing the relationship of the groups. The weightings thus represent to what extent loci cluster according to phenotype, rather than geographic relatedness of populations. Topology weightings are determined by sampling a single member of each group and identifying the topology matched by the resulting subtree. This process is iterated over a large number of subtrees, and weightings are calculated as the frequency of occurrence of each topology. Weightings were estimated from 1,000 sampling iterations over trees produced by RAxML version 8.0.2681 [128] for 50 SNP windows with a stepping size of 20 SNPs. For phylogenetic weighting along the WntA interval, weightings of topologies that grouped populations with the split forewing band phenotype or, alternatively, the hourglass shape were assessed (S7 Fig). For the region containing the aristaless genes, we focussed on topologies that clustered populations with white or yellow colour phenotypes (S8 Fig). For the cortex region, we focussed on topologies grouping populations showing the ventral and dorsal yellow hindwing bar, respectively (S9 Fig). Finally, for the optix interval, we assessed topologies grouping populations according to the absence or presence of the red dennis patch, the red hindwing rays, or the red forewing band and repeated the analysis for different geographic settings (S10 Fig). To obtain weightings for hypothesized phylogenetic groupings of specific colour pattern forms, we summed the counts of all topologies that were consistent with the hypothesized grouping.

Inference of selection and summary statistics in sliding windows

Summary statistics informative on diversity and selection patterns were calculated. From the unphased data, nucleotide diversity, Kelly’s Z_nS, Tajima’s D, and number of sites genotyped for each population were calculated in 1 kb nonoverlapping sliding windows with at least 100 sites genotyped for at least 75% of all individuals within that population using custom python scripts and the EggLib library version 3[129]. Scans for selection using signals of extended haploptype homozygosity and calculation of the pooled integrated haplotype homozygosity score (iHH12) [11,130] were performed using the program selscan1.2 [131] and our phased data set.

SF2

To detect local distortions of the SFS that are indicative of selective sweeps, SF2, an extension of Nielsen and colleagues’ [73] SweepFinder program, with increased sensitivity and robustness [74,76] was used. The SweepFinder framework builds on a composite likelihood ratio test using the SFS to compare the likelihood for a model with a selective sweep versus the likelihood for a model without a sweep. Huber and colleagues [74] showed that including substitutions, i.e., fixed differences relative to an outgroup, increases power while maintaining robustness to variation in mutation rate. SF2 also permits the use of recombination maps. The use of polarised sites increases power and we therefore polarised sites when possible.

We filtered our data set for biallelic sites only and initially tested different input data sets and parameter settings and created 2 types of data sets for this purpose; one using polymorphic sites only with both polarised and unpolarised sites and one with polymorphic sites and substitutions that contained only polarised sites. As an outgroup, H. numata was used for the melpomene clade and H. hermathena for the erato clade. We used biallelic sites only that were present in ≥75% of the focal populations and polarised sites by randomly drawing an outgroup allele from sites with a minimum number of outgroup samples with genotype data of either one (−OM1) or 3 (−OM3) of 4 for the melpomene clade and one (−OM1) or 2 (−OM2) of 3 for the erato clade.

SF2 was then run in 2 modes for each data set: with flag -s, calculating the likelihoods from the SFS of the respective region and with flag -l, using a SFS precalculated either from the background regions only or from background regions and colour pattern regions combined. These precalculated SFSs are used by SF2 as null models that incorporate the underlying demography of the populations of interest, making SF2 sensitive to selective sweeps even in populations that are not at equilibrium [132]. For the melpomene clade, recombination rate information from a fine scale recombination map was included (flag -r) [133]. To create a recombination file, recombination map coordinates were transferred to Hmel2 coordinates, and between sites recombination rates were calculated.

SF2 test runs for different grid spaces (flag–g; tested values: −g1, −g5, −g50, −g100, −g1000) were performed to find a setting allowing for reasonable runtimes without loss of accuracy and based on these test CLR and α were calculated for every 50th site (−g50) across all populations and regions.

Generally, the results were largely consistent among the different runs and data sets. As expected, power to detect sweeps was higher when including substitutions [74], and the minimum number of outgroup samples had only marginal effects. We therefore focussed on the results for data sets with outgroup minimum 1 (−OM1) and background SFS calculated from background regions and background regions and colour pattern regions combined, respectively. Including the colour pattern regions inflates the estimated background SFS with regions affected by selective sweeps which results in slightly lower CLR and higher α estimates. Because selective sweeps across the genome have been found to be rare in H. melpomene [70], these estimates represent a lower bound, and the estimates derived with background SFS from the background regions only are most likely a better approximation. Only CLR peaks exceeding a threshold defined as the 99.9th percentile of the distribution of CLR values across all background regions were considered as evidence for selection.

To obtain estimates for strength of selection (s) we calculated s as s = r×ln(2N_e)/α [132,134] with region- and population-specific estimates of effective population size (N_e) estimated from the data using the mutation rate given in Keightley and colleagues [77] and per chromosome recombination rate estimates (r) from Davey and colleagues [133] and Van Belleghem and colleagues [38].

VolcanoFinder

We also tested the new software VolcanoFinder on our data, described in a recent preprint, which is specifically designed to detect introgression sweeps but can also detect classic sweeps [27]. As for the SF2 runs, we used data sets with outgroup minimum 1 (−OM1) and background SFS calculated from background regions to generate the allele frequency files and the required unnormalized site frequency spectrum. We then ran VolcanoFinder with the following specifications: Model 1 and P = 0.