Forewing yellow band in Ithomiini butterflies: repeated reuse of ivory

Using whole-genome sequences of 285 wild-caught individuals, GWA was used to find genotypic associations with the presence or absence of the forewing yellow band in the ithomiine species Melinaea mothone (49 specimens), Melinaea menophilus (64 specimens), Mechanitis messenoides (111 specimens), and Hypothyris anastasia (61 specimens). In all comparisons, clusters of significantly associated SNPs were identified in the long noncoding RNA ivory (near the gene cortex) which also controls melanisation patterns in other Lepidoptera (Figs 2 and S2; [29,31–35]). In all cases, alleles associated with the presence of the yellow band are recessive (S2–S6 Figs). A small number of SNPs in narrow 1,155–2,140 bp genomic intervals are perfectly associated with the phenotype (Figs 2 and S2–S6) in all species apart from Melinaea menophilus, where the peak of association is relatively broad (7,373 bp). We find surprising concordance in the location of the genomic intervals controlling an identical mimetic phenotype across these four species (Figs 2 and 3). While there is no clear sequence homology in the identified regions, in all four species they lie within the first intron of ivory, 25,800−33,500 bp downstream of the ivory promoter and a short distance upstream of the conserved E230 element, an ivory cis-regulatory region in the butterfly Junonia coenia [29]. Tian and colleagues [34] demonstrated that the microRNA mir-193 derived from ivory is likely the main effector gene, repressing multiple pigmentation genes. The concordant GWA peaks likely indicate conservation of transcriptional control of ivory over ~28 million years of evolution in the Ithomiini.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Ivory and optix control convergent phenotypes across multiple species.

Zoomed Manhattan plots of genome-wide association for wing pattern variation involving (A) ivory and (B) optix. SNPs above the Bonferroni-corrected significance threshold (orange dashed line) are colored according to the strength of correlation (ρ², squared Spearman’s rank correlation coefficient) between genotype and the phenotypes compared as shown in the wing images. Black points represent SNPs fully associated with the phenotypes. The bottom Manhattan plot shows the wider ~1 Mb region of high association in Chetone histrio around the ivory region. The results for Heliconius numata were retrieved from [28]. Heliconius pardalinus results are based on QTL mapping (S16 Fig). E230: cis-regulatory element 230 [29]; ivory pro.: ivory promoter; hyd. like: hydrolyze like; LRR1: Leucine Rich Repeat Protein 1. Association plots across the whole-genome are shown in S2 Fig. The ~ 400 kb P1 inversion (orange segment) is associated with color pattern variation in Heliconius numata [28,30] and corresponds closely in location to the Chetone histrio inversion as shown by the red segment (S24 Fig).

https://doi.org/10.1371/journal.pbio.3003742.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Concordant locations of SNPs associated with convergent phenotypes at ivory and optix.

Blue-shaded blocks represent intervals containing SNPs significantly associated with convergent wing phenotypes (Fig 2). Homologous regions (1,000 bp segments) are shown with gray lines between species pairs. Purple lines connect regions of homology within GWA intervals that are shared across adjacent species pairs, highlighting conserved areas. Within each species, the SNPs most strongly associated with the phenotype are colored according to the strength of correlation with the phenotype (ρ², squared Spearman’s rank correlation coefficient). Partial ivory annotations are shown in pink for three of the species. Hatched boxes mark annotated genomic features: E230: cis-regulatory element 230 [29]; hyd. like: hydrolyze like; LRR1: Leucine Rich Repeat Protein 1.

https://doi.org/10.1371/journal.pbio.3003742.g003

Hindwing melanisation in Ithomiini butterflies: repeated reuse of optix

We next used GWA to uncover genotypic associations with variation in black vs orange hindwing patterning in Mel. marsaeus, Mel. menophilus, Mec. messenoides and Hyp. anastasia. In all comparisons, we found peaks of association upstream of the known color patterning gene optix, with genotype-phenotype correlations ranging from 0.54 to 1 (Fig 2). In most cases, these were the SNPs with the strongest associations across the genomes (S2 Fig). Additional associated SNPs in some species are likely a result of population structure correlated with the phenotypes. The regions of peak association in the two Melinaea species correspond closely, and they are near but not overlapping the region identified in Hyp. anastasia. All three genomic intervals fall within the wider associated region identified in Mec. messenoides (Fig 3).

In Mec. messenoides we additionally investigated genetic associations with black vs orange patterning in the forewing base and tip, which also yielded associated SNPs near optix. The SNPs showing the strongest genotype-phenotype correlations for the black vs orange patterns in different wing regions fall in three adjacent genomic regions (Figs 4 and S10–S13). These may correspond to separate cis-regulatory modules of optix controlling different aspects of wing melanisation, similar to those proposed for different red/black phenotypes in Heliconius [24,36]. The two regions of association we find upstream of optix in Mel. marsaeus (Fig 2) may also represent separate cis-regulatory modules, but our samples lack the phenotypic variation required to test this hypothesis. In contrast, black vs orange patterning in the forewing tip of Mel. menophilus was strongly associated (genotype-phenotype correlations of 0.87) with SNPs located between the Hox genes Antennapedia (Antp) and Ultrabithorax (Ubx) (S14, S15 Figs). As Ubx expression in butterflies is restricted to the hindwing, this phenotype likely arises through modulation of Antp [37]. This is similar to Bicyclus anynana, where Antp and Ubx promote eyespot development in fore- and hind-wings respectively [38].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Detailed phenotypic and functional analysis in Mechanitis messenoides.

(A) Genome-wide association analysis for separate forewing base, forewing tip and hindwing black vs orange phenotypes (red-circled regions on wings) indicate three distinct cis-regulatory modules of optix. SNPs above the significance threshold (dashed orange line) are colored according to the strength of association between genotype and phenotype (squared Spearman’s rank correlation coefficient, ρ²). hyd. like (hydrolase-like) and LRR1 (leucine rich repeat protein 1). Genome-wide association plots are shown in S10 Fig. (B) CRISPR mutagenesis of optix and ivory in Mec. messenoides. Wings of mosaic knockout individuals are shown: orange scales turn black in optix mutants (blue dashed lines highlighting mutant patches), orange and black scales turn yellow in ivory mutants. Additional mutants are shown in S33 Fig. (C) HCR in situ hybridization for ivory in Mec. messenoides messenoides. The top shows an adult forewing, and the bottom, a day 3 pupal forewing stained for ivory. Regions of ivory expression appear as white staining. The colored dots indicate vein-based wing landmarks demarcating the yellow band region which lacks ivory expression. S34 Fig shows Mec. messenoides deceptus.

https://doi.org/10.1371/journal.pbio.3003742.g004

Genetic architecture of color patterning in Heliconius pardalinus

While the genetic basis of mimicry ring switches is known for multiple Heliconius species, the Heliconius species joining the tigerwing Ithomiini-dominated mimicry rings have received relatively little attention, with the exception of Hel. numata [25,28,39]. We thus investigated the loci controlling color patterning in Hel. pardalinus by mapping variation segregating in 82 backcross individuals between subspecies belonging to different tiger sub-mimicry rings. The only significant quantitative trait locus (QTL) for forewing yellow patterning is on chromosome 15 and contains ivory (Figs 2 and S16). A QTL for forewing orange patterning is found on chromosome 18, and QTLs for hindwing orange patterning are found on chromosomes 13 [40] and 18 (Figs 2 and S16). Both of the chromosome 18 QTLs encompass optix. These results are consistent with our findings from Ithomiini, showing that both ivory and optix predictably control convergent phenotypes across 80 My of divergence time.

Genetic architecture of color patterning in the moth Chetone histrio

Unlike most tiger mimicry ring species in which different mimetic forms within-species have broadly parapatric distributions (S17–S22 Figs), the moth Chetone histrio is locally polymorphic in Peru with individuals belonging to either the striped (C. histrio histrio) or orange-black (C. histrio hydra) sub-mimicry rings, which consistently differ in multiple color pattern elements across both fore- and hind-wings (S23 Fig), akin to the similarly locally polymorphic species Heliconius numata. Comparing striped and orange-black C. histrio, we again find a GWA peak around ivory. However, unlike in the previous GWA results, the C. histrio GWA shows a ~1 Mb block of SNPs in perfect association with the phenotype (Fig 2). This genomic interval includes ivory and several other flanking genes. Examination of Illumina read-pair orientation and mapped insert sizes demonstrates that this block corresponds to a 1.018 Mb inversion (S24 Fig). The striped C. histrio histrio are homozygous for the ancestral gene order, and the orange-black C. histrio hydra are either heterozygous or homozygous for the derived inversion.

Tiger color pattern polymorphism in Hel. numata is also maintained via an inversion architecture, and in both species the inversions likely maintain allelic combinations (“supergenes”) that simultaneously control multiple color pattern elements [25]. One of the breakpoints of the C. histrio inversion closely matches the location of the Hel. numata P1 inversion, falling between the sucrose-6-phosphate hydrolase and glutaminyl-peptide cyclotransferase genes [28]. The other inversion breakpoint of C. histrio is located within eight genes of that of Hel. numata (S24 Fig). Similar to C. histrio, the derived P1 inversion in Hel. numata is dominant and controls black vs orange patterning. The similarities in genomic architecture between C. histrio and Hel. numata demonstrate striking parallel evolution not only in gene usage but also in genetic architecture, location and dominance of the inversion haplotypes between lineages that diverged ~120 MYA.

Functional characterization in Mechanitis messenoides

To confirm that the genes nearest to the GWA peaks, ivory and optix, are causally involved in color pattern formation in ithomiine butterflies, we performed CRISPR-Cas9 gene knock-outs in Mec. messenoides. In mosaic ivory knockout individuals, black and orange scales turn yellow, and optix-knockout causes orange scales to turn black (Figs 4B and S33), consistent with findings in other butterflies [29,41,42]. In situ hybridization of pupal forewing discs (50 hours post pupation) using ivory probes demonstrates that while ivory expression occurs across the entire forewing in the orange-black Mec. messenoides deceptus, in the yellow-banded Mec. messenoides messenoides, a region lacking ivory expression prefigures the adult forewing yellow band phenotype (Figs 4C and S34). In contrast, staining with antibodies against cortex, a gene overlapping with ivory that has been reported to affect color patterns in Lepidoptera [35], shows no spatial association with the forewing yellow band phenotype, further supporting the role of ivory rather than cortex in wing melanisation as recently reported for other butterflies [42] (S35 Fig). Comparing yellow-banded with nonyellow-banded individuals, we do not find that ivory is differentially expressed in forewing pupal wing discs of Mec. messenoides or Mel. menophilus (S36 Fig). However, this is not unexpected as whole forewing tissues were used and the anticipated difference in ivory expression is ~20% (proportional to the area of the yellow band region compared to the whole wing) and thus not detectable in these experiments.

In Mec. messenoides, only 5 SNPs located in a ~1.5 kb interval (CHROM6:6,877,302−6,878,798) are fully associated with the yellow band phenotype. These are candidate binding sites for transcription factors controlling ivory expression. In this interval, we find eight sequence motifs which contain a fixed SNP and are present in all homozygotes of one subspecies and absent in all homozygotes of the other subspecies. Of these, four correspond closely to binding sites of known transcription factors that are also expressed in pupal wing discs of Mec. messenoides (Sox15, Ftz-F1, ttk, and br-Z4; S37 Fig). Sox15 is potentially implicated in the differentiation of lepidopteran scale cells [43]. In Junonia butterflies, Ftz-F1 has been shown to bind to the ivory promoter [29]. These SNPs and transcription factors are strong candidates for future investigation of transcriptional control of ivory.

Limited evidence for introgressive allele sharing

Convergent evolution among very closely related Heliconius species often results from the sharing of adaptive alleles among species via occasional hybridization [22–24]. Both the forewing yellow band and the hindwing black mimetic phenotypes are found in multiple Melinaea and Hypothyris species (and to a limited extent in Mechanitis) allowing us to test if allele sharing via introgression facilitated color pattern convergence. In addition to the species used in the GWAs, we analyzed the genomes of 222 individuals of 24 species (8 Melinaea, 9 Hypothyris, and 7 Mechanitis species; S3–S9 Figs, S4 Table). Although we detect ongoing interspecific genome-wide gene flow among many species in all three genera [44; S25 Fig], inspection of all significantly associated GWA SNPs at both ivory and optix regions failed to detect instances where the GWA SNP alleles were shared with congeneric species with mimetic phenotypes (S3–S9 Figs). Twisst [45] and Relate [46] analyses also failed to detect signals of allele sharing (S26–S28 Figs). The only exception is in Melinaea where narrow signals of introgression near optix are present among multiple species sharing melanic hindwing patterns (S28, S29 Figs). Despite the ivory and optix regions being repeated targets of selection across multiple species, we do not find evidence of long-term balancing selection maintaining diversity at these regions in Mechanitis, Melinaea, or Hypothyris (S30–S32 Figs). We cannot fully rule out allele sharing because it is possible that our top GWA SNPs do not include causative genetic variants such as indels and structural variants. However, in the ithomiine genera examined, we do not find the relatively long introgressed haplotype blocks that characterize mimetic convergence among closely related Heliconius species, despite ongoing interspecific genome-wide gene flow in both groups [22,23]. While most Heliconius species have a highly conserved karyotype [47], the frequent karyotypic differences present even among closely related ithomiine species may generate intrinsic postzygotic isolation which limits introgression [44,48,49].

Repeatable and predictable evolution

The neotropical tiger mimicry ring is exceptionally species-rich, comprising over 100 species from divergent lepidopteran lineages. These butterflies and moths were recognized as a classic example of adaptation by Bates [50], but the genetic basis of their color pattern changes represents a longstanding debate. While Bateson and Punnett [50] argued for large-effect mutations controlling mimicry [51], Fisher [51] predicted that these large-effect loci were likely due to polygenic modifiers scattered across the genome [52]. Nicholson proposed a “two-step” model with a large-effect mutation followed by the fine-tuning of mimicry by polygenic modifiers [53]. Our study clarifies this longstanding dispute: across widely divergent lineages we find two major switch loci, ivory and optix, with modifiers located next to these switches, rather than scattered across the genome. These modifiers are likely involved in cis-regulation of ivory and optix, and control where on the wing pigmentation changes occur. Phenotypic convergence in all species studied in this mimicry ring is characterized by a simple genetic architecture where a few large-effect genes, primarily ivory and optix, are reused repeatedly. Even among the most closely related species of the same genus, convergent phenotypic switches occur via independent mutations in the regulatory regions of these genes rather than reuse of standing genetic variation [11] or alleles shared via introgression from other species [13,22]. Similar patterns of gene reuse are found in vertebrate pigmentation, though not necessarily for convergent traits, potentially suggesting that developmental pathways underlying coloration could be more constrained than other traits frequently involved in local adaptation [54,55].

Ivory and optix are known to control color patterning across Lepidoptera. When these genes control highly similar phenotypes, evolution is surprisingly predictable, with convergence caused by recurrent mutations at very similar regions of these genes. These regions could be mutation hotspots that enable rapid adaptation [56]. The repeatability of evolution also extends to the similar inversion architectures that maintain different allelic combinations in the locally polymorphic butterfly and moth species Hel. numata and C. histrio. These results suggest that developmental pathways controlling the convergent phenotypes are highly constrained. The different tiger sub-mimicry rings represent locally adaptive fitness peaks. Our results show that not only are the paths to reach these peaks constrained, but also that the steps along these paths tend to be few and large in size, i.e., using large-effect loci. The limited number of paths leading to these fitness peaks may enable diverse taxa to join this species-rich mimicry ring more easily. Once occupying a particular fitness peak, species may then jump via regulatory changes with few pleiotropic effects to alternative peaks representing other locally prevalent color patterns. The outcome of “replaying life’s tape” has been a longstanding question in evolutionary biology [3]. Our repeated discovery of convergent adaptation via narrow and repeatable pathways over 120 million years suggests that the running of this tape may be more predictable than expected.

Methods

BUSCO phylogeny and divergence times

To infer the phylogeny and divergence time between the major groups analyzed in this study, we downloaded additional reference genomes from ENA and built a dated phylogenetic tree of 18 species: Napeogenes inachia (GCA_959347435.2), Hypothyris anastasia (this study), Ithomia salapia (GCA_028829205.1), Scada zibia (GCA_964345855.1), Methona curvifascia (GCA_982264085.1), Mechanitis messenoides (GCA_959347415.1), Melinaea mothone (GCA_965197345.1), Melinaea menophilus (GCA_918358695.1), Melinaea marsaeus (GCA_918358865.1), Danaus plexippus (GCA_018135715.1), Heliconius pardalinus (GCA_001486225.1), Heliconius melpomene (Hmel2.5 [74,75]), Heliconius numata (GCA_016802625.1), Heliconius erato [76], Bicyclus anynana (GCA_947172395.1), Chetone histrio (this study), Biston betularia (GCA_905404145.2), and Plutella xylostella (GCF_932276165.1) using BUSCO genes.

For each reference genome, we ran BUSCO v5.4.3 [66] and extracted the BUSCO genes common to all species. Sequences for each gene were translated into amino acids and aligned using MUSCLE v3.8.31 [77]. The alignment was then reverse-translated to nucleotides using PAL2NAL v14 [78], retaining only genes with fewer than 2% gaps, resulting in a dataset of 257 genes.

We inferred the maximum likelihood (ML) tree based on these concatenated common genes using RAxML v8.2.12 [79,80]. We generated 100 bootstrap alignments using the -f j option in RAxML and optimized the model parameters and branch lengths of these bootstrapped trees based on the previously inferred ML tree using the -f e option. All trees were rooted using pxrr v1.3.1 [80] with Plutella xylostella as the outgroup.

Divergence time estimates were obtained using a penalized-likelihood-based approach implemented in TreePL v1.0 [81]. The node separating Papilionidae from the moths was used as a calibration node, constraining it to range from 100 to 120 Mya [82]. TreePL was run on each bootstrapped tree to obtain age estimate ranges for each node. The priming step was performed on each of the 100 bootstrapped trees, cross-validation was run 10 times, and finally, for the dating step, the best smoothing parameters for each run were chosen based on the lowest χ² value and the most common value out of the 10 runs. We used the TreeAnnotator utility from the BEAST package [83] to calculate the 95% highest posterior density for the node ages using a burn-in of 10%.

Assessing population structure

Population structure within the taxa involved in each genome-wide association analysis was assessed via Principal Component Analysis (PCA) in PLINK v1.9 [84] using an LD pruned SNP dataset. LD pruning was performed in PLINK using a window size of 100 SNPs, a window shift of 10 SNPs and an r² value of 0.1.

Genome-wide association mapping

We investigated the genetic basis of phenotypic differences using a genome-wide association (GWA) approach [85]. To identify single-nucleotide polymorphisms (SNPs) associated with each trait, we applied linear univariate mixed models in GEMMA v0.98.5 [86]. SNPs were filtered to retain only those with a minor allele frequency ≥10% and missingness <25%. To account for multiple testing, we applied a Bonferroni correction. Sample relatedness was controlled for by incorporating a pairwise relatedness matrix as a covariate in the model. All other parameters were set to default values.

To annotate the association peak regions, we predicted genes within a 250 kb interval around each peak using AUGUSTUS v3.5.0 [87], trained on the Hel. melpomene annotation. We then performed BLASTP [63] searches against the UniProt database [88] to identify and annotate the genes. For SNPs exceeding the significance threshold, we calculated squared Spearman’s rank correlation coefficients (ρ²) to quantify the strength of association between SNP genotypes and phenotypic traits, as both were encoded categorically.

GWA analyses were performed for the presence/absence of a forewing yellow band in Mel. mothone (N = 49), Mel. menophilus (N = 64), Mec. messenoides (N = 111), and Hyp. anastasia (N = 61). For the hindwing, GWA was conducted in Mel. marsaeus (N = 40), Mec. messenoides (N = 102), and Hyp. anastasia (N = 61), comparing solid black versus striped-black phenotypes. For Mel. menophilus (N = 67), we compared three hindwing phenotypes: solid black (encoded as 1), striped-black (0.5), and completely orange (0). For Mec. messenoides, hindwing black pattern variation is somewhat more continuous rather than strictly discrete. We also performed an additional GWA using quantitative phenotype values obtained using Patternize [59], producing results consistent with the manual classification (S38 Fig). We also conducted targeted GWA in Mec. messenoides to investigate black vs orange pattern variation at both the forewing base (N = 111) and tip (N = 111; S10 Fig), as well as in Mel. menophilus (N = 67) to assess the presence/absence of an apical spot on the forewing (S14 Fig).

GWA peak alignments

For each pair of species, we aligned the regions around the GWA peaks to assess whether these association peaks fall within homologous genomic regions. We limited the alignment to the two genes flanking the GWA peaks. The alignments were performed using Nucmer from the MUMmer package v3.23 [89]. For each genome pair, one genome was divided into nonoverlapping sliding windows of 1,000 bp. These windows were then individually aligned to the alternative genome. Due to the high divergence between genomes, we ran nucmer using the following flags: --mum -c 20 -b 500 -l 10 --maxgap 500.

QTL mapping in Heliconius pardalinus

Crosses and sequencing of hybrids between Hel. pardalinus butleri and Hel. pardalinus sergestus are described in [90]. Dorsal surfaces of wings from 82 backcross hybrids were photographed in a standardized light box against a white background using a Canon EOS D1000 together with an X-rite ColorChecker Mini to enable color calibration. Yellow and orange forewing patterning, along with orange hindwing patterning were quantified using a standardized patternize workflow as follows. A reference image was selected, and the RGB color signature of a key pattern element was extracted using the sampleRGB() function. Images were aligned to this reference to standardize spatial orientation using the patRegRGB() function, with a color offset (colOffset = 0.15) and background removal threshold (removebg = 100) to isolate the focal pattern. To quantify variation in color pattern distribution among individuals, we performed a PCA on the aligned pattern rasters using patPCA(). The resulting PCA scores were subsequently used for quantitative trait locus (QTL) mapping as described in [40].

Butterfly husbandry

Wild Mec. messenoides messenoides and Mec. messenoides deceptus individuals were caught with hand nets in the Napo province of Ecuador, and used to establish breeding stocks in outdoor insectaries at Ikiam Regional Universidad Amazonica. The adults were fed a solution of sucrose and pollen and provided with Lantana and Asteraceae flowers. Solanum quitoense was used for oviposition and rearing larvae.

CRISPR-Cas9 genetic modification

Ivory and optix were annotated in the reference genome of Mec. messenoides (ilMecMess1.1.primary.fa [44]) and ivory in Mel. mothone (ilMelMoth8.1.primary.fa [44]) based on manual curation of BLAST-hits with the corresponding genes from Danaus plexippus and Hel. erato. RNA-guides (sgRNA) were designed against the annotated ivory- and optix-genes using Geneious (www.geneious.com) (S7 Table). Eggs were collected and arranged with nontoxic glue on a microscope slide. The eggs were injected with a 1:1 mixture of the sgRNA (Sigma-Aldrich) and Cas9-protein (TrueCut Cas9 Protein V2, Invitrogen) at 1 μg/μl within 3–4 hours of laying following established protocols [91].

In situ hybridization with HCR

Wing tissues were dissected at different developmental timepoints (5th instar caterpillar, day 1−4 after pupation). Caterpillars and pupae were anesthetized on ice before dissection in cold phosphate-buffered saline (PBS). Dissected wing tissue was fixed for 30−40 min with formaldehyde (0.25 ml 37% formaldehyde with 0.75 ml PBS 2mM ethylene glycol tetraacetic acid), and subsequently dehydrated and stored in methanol at −20 °C following the protocol of [92] until the ‘post-fixation’ step. Subsequently, the HCR in situ protocol of Molecular Instruments (MI-Protocol-RNAFISH-GenericSolution) was followed. The wings were also stained with DAPI/HOECHST antibody to visualize the nucleus (Sigma-Aldrich; 10236276001), mounted on a slide in 60% glycerol, and imaged with a Leica SP8 confocal microscope.

Cortex antibody staining

Dissected wing tissues were collected with a fixation (~30−40 min) in 4% paraformaldehyde in PBS 2 mM EGTA. The wings were stored in PT-BSA (PBS 0.1% Triton X-100 with 0.1% sodium azide and Bovine Serum Albumin (0.05 g in 10 ml). The samples were subsequently washed and stained according to a protocol adapted from [93,94]. The primary antibodies against cortex were made for Heliconius (the same antibodies as [94]; rabbit), with secondary antibodies goat anti-rabbit AlexaFluor-555 (ThermoFisher; A-21428). The wings were also stained with Wheat Germ Agglutinin (WGA; plasma membrane; Cambridge BioScience BT29022-1; CF488A Conjugate) at 488 nm and DAPI/HOECHST for the nuclear DNA at 405 nm (Sigma-Aldrich; 10236276001). The tissues were imaged as for the HCR.

RNA sequencing for annotation of ivory

RNA-seq nanopore long reads were generated for 12 Mel. menophilus wing discs (Day 2 after pupation), comprising three forewing and three hindwing disc samples each from two subspecies: ssp. nov. 1 and hicetas. Nanopore sequences were also generated from wing disc tissues (Day 2 after pupation) for three samples each of Mec. messenoides messenoides, Mec. messenoides deceptus, and Mel. mothone mothone. Wing discs were dissected out and either flash-frozen (Mec. messenoides and Mel. mothone) or stored in RNAlater (Mel. menophilus) at −70 °C until processing. Following extraction, RNA quality was assessed using the Agilent Bioanalyzer. Oxford Nanopore Technologies full-length cDNA sequencing libraries were prepared using the ONT cDNA-PCR Sequencing V14 Barcoding kit (SQK-PCB114.24). Barcoded libraries, with cDNA pooled at equimolar concentrations, were sequenced on R10 flow cells (FLO-PRO114) using an ONT PromethION sequencer running MinKNOW version 24.02.10. Super-accuracy basecalling was performed with ONT’s Dorado software version 7.3.9. The reads from each sample were aligned to their respective reference genomes using minimap2 v2.26 [95] with the command “minimap2 -ax splice -uf -k14”. The resulting BAM files were visualized in IGV [96] with the junction track option, focusing on the ivory region. Splicing events supported by at least five reads and present in at least two individuals were retained to define the exons and isoforms of ivory.

Differential gene expression in Melinaea menophilus

Nanopore reads were mapped to the Mel. menophilus reference genome using Minimap2 v2.26 [95] with parameters optimized for long-read spliced alignment. The resulting BAM files were then used to assemble transcripts for each sample individually with StringTie v3.0.0 [97], using the -L option to account for long reads. These individual transcript assemblies were subsequently merged into a unified transcriptome to generate a consensus annotation. Read quantification, with parameters optimized for long-read data (-L -s 0 -M --fraction -O), was performed using featureCounts v2.0.4 [98] to count reads assigned to genes and isoforms inferred by StringTie. Because automated quantification of long noncoding RNA is often unreliable, the automatically assigned read counts for ivory were manually removed and replaced with curated counts obtained using IGV. Differential gene expression in Mel. menophilus was analyzed by comparing forewing and hindwing datasets for both subspecies (ssp versus hicetas). Three replicates were included for each condition. Read count data were processed using DESeq2 v1.38.3 [99]. Raw count data were normalized and differential expression was assessed using the Wald test with a Benjamini–Hochberg correction for multiple testing. Genes with an adjusted p-value (padj) < 0.05 were considered significantly differentially expressed. We then specifically examined whether ivory was differentially expressed between the yellow-banded and nonyellow-banded forewings.

Differential gene expression in Mechanitis messenoides

Fore- and hind-wings were dissected at six different timepoints (early 5th instar, late 5th instar, Day 1 after pupation to Day 4 after pupation), for three forewings and hindwings each for both Mec. messenoides deceptus and messenoides (6 × 3 × 3 × 2 = 72 samples). Each wing was collected separately and flash-frozen in liquid nitrogen. RNA was extracted with a MagMAX Mirvana Total RNA isolation kit, following the manufacturer’s protocol (Thermo Fisher Scientific: AM1830). Libraries were 150 bp paired-end Illumina sequenced on two Novaseq X 10B lanes. Illumina reads were trimmed with FastP v0.23.2 [100] and Rcorrector [101], before using STAR without an annotation in ‘two pass mode’ v2.7.9a [102] to align them to the Mec. messenoides reference genome. Based on the STAR-alignments, a de novo transcriptome was assembled using Trinity v2.15.1 [103] with the ‘genome_guided_bam’ option. Read quantification was performed using Salmon v1.10.2 [104]. Differential expression in the forewing and hindwing datasets was then assessed for both subspecies (messenoides versus deceptus) using the same approaches as for Mel. menophilus (see previous paragraph).

Measuring genome-wide introgression using f₄ statistics

To investigate potential recent introgression events between sympatric Melinaea, Mechanitis, and Hypothyris species, we employed the f₄-statistics framework implemented in ADMIXTOOLS v2.0.8 [105]. f₄-statistics identify deviations in allele frequency correlations across populations or species, which can provide evidence for admixture or introgression. Specifically, we computed f₄-statistics between pairs of populations of different species in the same location versus between pairs of populations of the same species in different locations. We designed the test so that a positive f₄-statistic indicates excess allele sharing between sympatric species, consistent with gene flow. Statistical significance was assessed using a block-jackknife approach, dividing the genome into nonoverlapping 500 kb blocks.

Measuring introgression at color genes

We tested for evidence of introgression at wing color pattern genes through inference of local genealogies consistent with introgression, and complemented this approach with visualization of genotypes at top GWA SNPs. These analyses were applied independently to Mechanitis, Melinaea, and Hypothyris, testing whether species with similar phenotypes (i.e., co-mimetic species) showed evidence of gene flow at color loci. F_d and f_dM statistics [106,107] did not reveal any evidence of introgression at candidate color loci. We therefore focused on alternative methods that provide more localized or topology-based insights into gene flow in these regions.

To assess introgression at GWA peaks, we employed two complementary approaches to infer local genealogies: (1) marginal trees at individual SNPs using Relate [46], and (2) Neighbor-Joining (NJ) trees built from nonoverlapping 100-SNP sliding windows, using the ape R package [108]. Both methods were applied to a core set of quartets, and some additional quartets were analyzed using the NJ approach (S26–S28 Figs). The two analyses followed a common framework for summarizing topologies and assessing significance.

Quartet Definition and Taxon Selection: Analyses were conducted across four different taxon combinations (see S4 Table). Each quartet ideally consisted of two species, each represented by two morphologically distinct subspecies (spA: sspA1/sspA2 and spB: sspB1/sspB2). We specifically focused on introgression between two distant species with similar wing phenotypes, particularly between sspA1 and sspB1. When this configuration was not possible, we modified the arrangement to include one taxon (sspB1) for introgression testing with sspA1 and a third outgroup (sspC1). In both quartet types, we interpreted clustering of sspA1 and sspB1 near the GWA peak as evidence of potential introgression between these species.

Relate based inference: Relate was run on 4–10 Mb regions centered on each GWA peak, comparing species involved in the GWA and closely related, phenotypically similar species (S26–S28 Figs). VCFs were imputed and phased with SHAPEIT v4.2 [73], and ancestral alleles were inferred using an outgroup species—Hyalyris antea and Hyalyris lactea for Hypothyris; Forbestra equicola, F. proceris, and F. olivencia for Mechanitis; and Melinaea ludovica for Melinaea. For SNPs missing in the outgroup, we used the most frequent allele across the genus. Relate was run with an effective population size of 1 × 10⁷ and a mutation rate of 2.9 × 10⁻⁹ per site per generation, based on estimates from Heliconius studies [109–111]. Because Relate requires a genetic map, we used a uniform recombination rate of 6 cM/Mb, generated with a custom script available at: https://github.com/joanam/scripts/blob/master/createuniformrecmap.r.

Sliding window NJ tree inference: As an alternative to Relate, we inferred genealogies by constructing NJtrees from nonoverlapping 100 SNP sliding windows using the ape R package. This window-based approach offered a complementary view of local genealogies by summarizing phylogenetic signal across short genomic regions. For this analysis, we used separate VCFs that retained multiallelic SNPs. NJ tree inference was applied to the same taxon quartets used in the Relate analysis. As this test can be run using a small sample size, additional quartets based on non-GWA species were also analyzed.

Summarizing tree topologies and assessing significance: The results from both Relate and the sliding window NJ tree inference were summarized using Twisst [45]. To assess whether a region exhibited a significant excess of introgression-compatible topologies, we performed a permutation test using block-shuffling [112]. This method disrupts potential signals of introgression whilst preserving the local genomic structure, by randomly shuffling 100 kb blocks across the entire region. The null distribution for introgression topology was derived by counting the number of windows with a Twisst weight of at least 0.95 for the introgressed topology (referred to hereafter as intro95). A p-value was calculated by comparing the observed intro95 count in the GWA peak region to the null distribution, based on 50,000 permutations. This approach was applied to both the Relate and sliding window NJ tree analyses to assess introgression significance consistently across both methods.

Genotypes matrices: We also created a genotype matrix for the most strongly associated GWA SNPs (S2 Table) to determine whether the same SNPs associated with wing color patterns in one species are also linked to similar phenotypic traits in co-mimetic species. This approach allows us to detect subtle signals of introgression, particularly in cases where the local genealogy analyses might not identify introgression due to the signal being present in a very narrow window.

Balancing selection

To assess whether color pattern genes are targets of long-term balancing selection, we examined patterns of polymorphism and allele frequency across Mechanitis, Melinaea, and Hypothyris species. To test for balancing selection at the color genes of interest across species, we used three approaches: multispecies nucleotide diversity, detection of trans-species polymorphisms, and the MuteBaSS method [113]. For Melinaea, analyses were conducted at ivory, optix, and antp, whereas in Mechanitis and Hypothyris, analyses focused on ivory and optix.

Multispecies nucleotide diversity: To quantify polymorphism at color genes and surrounding regions, we calculated multispecies nucleotide diversity by pooling data across species and subspecies. We used Pixy v1.2.5 [114] to compute nucleotide diversity in nonoverlapping 50 kb sliding windows. Analyses were performed on repeat-masked VCF files including invariant sites as well as bi- and multiallelic SNPs. Repetitive elements in the Lepidoptera genomes were identified using RepeatModeler v2.0.4 [115], and subsequently masked with RepeatMasker v4.1.2 [116] using the generated repeat library. We calculated nucleotide diversity under two conditions:

1. Analysis including all species.
2. Pairwise comparisons between individual species/subspecies to identify cases where balancing selection signals are limited to a subset of taxa.

Detailed species groupings for Melinaea, Mechanitis, and Hypothyris are provided in S6 Table.

Trans-species polymorphisms: To test for trans-species polymorphisms across multiple species, we analyzed genomic variation shared across multiple species using a custom python pipeline based on pysam v0.21.0 (https://github.com/pysam-developers/pysam). We used the same repeat-masked VCF files as described in the ‘multispecies nucleotide diversity’ section.

A site was classified as transpolymorphic if it met the following criteria: i) at least four species were polymorphic (when more than three species were present); ii) all species were polymorphic (if three or fewer species were present); iii) each allele was represented by at least six total copies across species.

MuteBaSS: To further test for local signals of balancing selection, we used MuteBaSS v1.0 [113], which detects balancing selection without relying on trans-species polymorphic sites. We used MuteBaSS to compute the multispecies NCD, NCDsub, and NCDopt statistics, as well as the trans-HKA test. To capture the potentially narrow signals of balancing selection, all statistics were calculated in 1 kb windows with a 500 bp sliding step. Ancestral states were inferred as described in the Relate based inference section. For Melinaea and Mechanitis, we assumed SNPs followed the phylogeny proposed by Van der Heijden and colleagues [44]. For Hypothyris, we used the phylogeny from Chazot and colleagues [117].

Transcription factor binding site analysis

In all cases excluding the inversion in Chetone histrio, the color pattern associated regions identified by GWA fall outside of the genes of interest. This may suggest that the associated regions control color pattern by influencing gene regulation, potentially by affecting transcription factor binding. As the peaks of association with the yellow band phenotype are narrow and contain only a few fixed SNPs, it may be possible to associate these SNPs with a specific transcription factor binding site, by identifying binding motifs which are enriched consequences of one color pattern form compared to the other. To test this, we focused on our best sampled species, Mec. messenoides. Multiple analyses were performed to uncover transcription factor (TF) binding sites within and around the ivory GWA peak region, defined here as the ~1.5 kb region between the weakly-associated SNPs which flank the associated region identified through GWA (S2 Fig). Only individuals homozygous for the fixed SNPs identified in the GWA analysis were selected. A fasta file for the region of interest was generated for each individual using the consensus commands in either samtools or bcftools. Homer [118] was used to identify de novo DNA motifs which are enriched in one set of sequences compared to a control group, using the findMotifs.pl script and the insect known motif collection. The two forms (yellow-banded versus nonyellow-banded) were alternately used as both control and target sequences.

To locate and identify known TF binding sites within the peak region, the FIMO algorithm in MEME Suite [119] was run on all sequences, including both forms, for each species. A background file was generated based on the entire chromosome containing the peak region (CHROM6), using the fasta-get-markov command. The expression of detected motifs was checked by examining expression levels of the corresponding transcription factors within the Mec. messenoides RNAseq data.

Ethics statement

We thank SERFOR, the Peruvian Ministry of Agriculture and the Área de Conservación Regional Cordillera Escalera (0289-2014-MINAGRI-DGFFS/DGEFFS, 020-014/GRSM/PEHCBM/DMA/ACR-CE and 040–2015/GRSM/PEHCBM/DMA/ACR-CE) for collecting permits; the Ministerio del Ambiente and Museo Ecuatoriano de Ciencias Naturales in Ecuador (MAATE-DBI-CM-2021-176 and MAATE-DBI-CM-2023-0286) for collecting and butterfly rearing permits. Field collections in Colombia were conducted under permit no. 530 issued by the Autoridad Nacional de Licencias Ambientales (ANLA) of Colombia.

Supporting information

Acknowledgments

We thank Mathieu Joron and Stephen Montgomery for providing samples, Ismael Aldas for support with fieldwork, Nicol Rueda for help with the Relate analysis and Chris Thomas and Jane Hill for comments on the manuscript. ONT library preparation and sequencing were carried out by the Genomics Laboratory at the Bioscience Technology Facility (University of York). The University of York Viking2 HPC and the Wellcome Sanger farm HPC were used for bioinformatics analyses.