The origins of diverse vertebrate body plans have fascinated comparative anatomists and evolutionary biologists for centuries1,2. Although studies over the last 40 years have identified many cellular pathways required for body axis formation and development using induced mutations in model organisms, it is challenging to identify specific changes in genes and regulatory regions that underlie the diversity of body forms and traits in wild species3,4.

Hox genes were one of the first classes of major developmental genes to be identified and analysed in comparative studies across animals. They were initially discovered by linked clusters of mutations in Drosophila that could transform particular body segments into others5. Molecular studies revealed that Hox loci consist of clustered homeodomain transcription factor genes, whose expression patterns along the anterior–posterior body axis were correlated with their physical positions along the chromosome6,7,8,9,10.

In an early review of genetic work on homeotic loci, Lewis5 hypothesized that regulatory mutations in Hox genes might underlie classic anterior–posterior patterning differences between species, such as four-winged versus two-winged insects. Although subsequent studies showed that Hox expression patterns are conserved between two-winged fruit flies and four-winged butterflies11,12, the important role of Hox genes in controlling body patterning has led to speculation that mutations in these genes underlie key morphological differences in nature10,13. Variation in Hox cluster number and structure across taxa support this idea, and intriguing correlations can be drawn between morphological differences and Hox expression changes10,14,15. On the other hand, much of the diversification and expansion of Hox clusters occurred before well-known morphological changes among animal phyla16. Furthermore, many laboratory mutations in Hox genes lead to reduced viability or fertility, and prominent evolutionary biologists17,18 and critics of evolutionary biology19 have suggested that natural mutations in Hox genes would lead to ‘hopeless monsters’ rather than adaptive changes in wild species. Natural differences in leg trichomes and abdominal pigmentation have been linked to genetic variation in Hox loci in insects, with regulatory mutations providing a possible mechanism for bypassing the broader deleterious consequences seen with many laboratory mutations20,21. However, few detailed examples exist for the long-postulated idea that genetic changes in Hox loci may also be the basis for major changes in skeletal structures along the anterior–posterior body axis of wild vertebrates15,22.

Almost a third of extant vertebrate species fall in the large and diverse Acanthomorpha group of spiny-rayed fish23, many of which show dramatic changes in the size or number of axial skeletal structures. A key evolutionary innovation of this group is the development of stiff, unsegmented bony spines anterior to the median dorsal and anal fins. These dorsal spines can be raised to protect against predators or lowered to facilitate swimming24. The number, length and morphology of bony spines differ substantially among species25. Recent studies have begun to reveal how spines form and grow within the median fin fold of developing fish26,27,28. However, little is known about the molecular changes that underlie the diverse spine patterns seen in different species.

Sticklebacks form a diverse clade of fish within Acanthomorpha. Multiple genera of sticklebacks live in northern marine and freshwater environments and diverged over 16 million years ago29,30,31. The most well studied of these species, Gasterosteus aculeatus, also known as the three-spine stickleback, colonized new freshwater postglacial habitats from the oceans after widespread melting of glaciers approximately 12,000 years ago32. In new freshwater environments containing different food sources and predators, Gasterosteus populations evolved substantial differences in craniofacial structures, vertebrae and the number of defensive bony plates and spines along the anterior–posterior body axis32. Many recently evolved populations show major reductions of structures, including armour plate loss, pelvic hind fin loss, spine length reduction and reduced body pigmentation28,33,34,35. However, recently derived populations can also evolve increases in size or number of structures, including increased body size, increased number of teeth, increased spine length and increased dorsal spine number27,36,37,38.

In this study, we used genetic and genomic approaches in two different stickleback genera to study the molecular mechanisms involved in spine patterning changes in natural populations. Our studies provide new evidence to support the long-standing hypothesis that mutations in the cis-regulatory regions of Hox genes underlie the evolution of new skeletal patterns along the anterior–posterior body axis of wild vertebrate species.


Quantitative trait locus mapping of spine number and length in Gasterosteus

To study the genetics of spine number in Gasterosteus aculeatus, we generated an F2 cross by crossing a wild-caught female freshwater stickleback from Boulton Lake, British Columbia, Canada and a wild-caught male marine stickleback from Bodega Bay, California, USA. The marine fish had the three dorsal spines typically seen in Gasterosteus. The freshwater fish had 2 dorsal spines (Extended Data Fig. 1), as is true for 80% of the Gasterosteus found in Boulton; the other 20% of fish have 3 dorsal spines39. We intercrossed F1 males and females with 3 dorsal spines and raised 590 F2 offspring (Fig. 1a). Most F2 individuals had 3 dorsal spines (n = 563), but 6 had 2 spines and 21 had 4 spines (Extended Data Fig. 1). We numbered spines from anterior to posterior, with the posterior-most spine immediately in front of the dorsal fin called dorsal spine last (DSL). Therefore, a four-spine Gasterosteus has dorsal spine 1 (DS1), dorsal spine 2 (DS2), dorsal spine 3 (DS3) and DSL, which we refer to as high-spine. A typical three-spine Gasterosteus has DS1, DS2 and DSL, which we refer to as low-spine in this study (Extended Data Fig. 2a).

Fig. 1: Genetic mapping, expression and role of HOXD11B in stickleback dorsal spine development.
figure 1

a, Gasterosteus mapping cross. b, QTL scan results for spine number and spine length. x axis: Gasterosteus chromosomes; y axis: LOD score for three- versus four-spine trait (top), length of DS2 (bottom). The QTL peak on chromosome 6 includes the HOXDB cluster (gene diagram at the bottom, scale bar, 1 kb). The peak on chromosome 4 includes the EDA-MSX2A-STC2A cluster described elsewhere27,28. Dashed lines: genome-wide significance thresholds from permutation testing. c, Integration of GFP reporter using CRISPR–Cas9 upstream of the endogenous HOXD11B locus of low-spine Gasterosteus. Plasmid: grey; eGFP: green; basal hsp70 promoter: blue; chromosomal locus: black. Scale bar, 100 bp. TSS, transcription start site. d, eGFP expression in posterior half of fish at the stage when the dorsal spines are forming (Swarup stage 31). Scale bar, 1 mm. e, Note expression in fin fold between DS2 and DSL, DSL and dorsal fin (DF). Scale bar, 1 mm. f, X-ray of uninjected Gasterosteus (top) and Gasterosteus injected at the single-cell stage with Cas9 and sgRNA targeting the coding region of HOXD11B (bottom). Arrows: two blank pterygiophores are often located between DS2 and DSL but only in uninjected fish (insets: two blank pterygiophores in n = 5 out of 18 control and n = 0 out of 23 injected F0 mutants, two-tailed Fisher’s exact test P = 0.01). Scale bar, 5 mm. g, Length comparisons of dorsal and anal spines. Box and whisker plot: centre line, median; box limits, interquartile range (IQR); whiskers, 1.5× IQR; individual measurements shown as single points (circles: WT; triangles: mutant). y axis: residuals after accounting for standard length of fish (Extended Data Fig. 2a). DSL and AS were significantly longer in injected than uninjected fish (two-tailed t-test Bonferroni-corrected at α = 0.05, n = 18 control and n = 23 injected, DSL Padj  = 3 × 10−5, AS Padj = 0.02). DS1 and DS2 lengths were not significantly different.

Source data

To examine the genetic basis of morphological phenotypes along the anterior–posterior body axis, we genotyped 340 F2 individuals with a custom single-nucleotide polymorphism (SNP) array (Methods)40 and phenotyped fish for the number and length of dorsal spines, number of flat bony plates that form in the dorsal midline or at the base of spines, known as pterygiophores, and the number of abdominal, caudal and total vertebrae (Fig. 1b and Extended Data Fig. 2). There were not enough two-spine fish for mapping the two- versus three-spine trait. When mapping three- versus four-spine as a categorical trait, we detected a significant quantitative trait locus (QTL) on the distal end of chromosome 6 (percentage variance explained: 5.8%). The same chromosome region showed a significant QTL for DS2 length (percentage variance explained: 8.2%). The allele linked to both increased spine number and decreased DS2 length was inherited from the freshwater Boulton parent. None of the vertebral traits mapped to the distal end of chromosome 6 (Extended Data Fig. 2), suggesting that the effect of this chromosome region was specific to patterning dorsal spines but not to axial patterning as a whole. Previous studies of other populations have identified other loci controlling vertebral number41,42.

HOXD11B is in the candidate interval and expressed in spines

The distal end of chromosome 6 contains the HOXDB locus in Gasterosteus. While not annotated in the original reference genome (gasAcu1 (ref. 43)), previous studies of Hox clusters across multiple species suggest that the Gasterosteus locus includes three genes (HOXD11B, HOXD9B and HOXD4B) and one microRNA (miR-10d)44. Hox genes are known to be expressed in the neural tube and somites as the body axis forms45,46. To investigate HOXDB gene expression in sticklebacks, we used in situ hybridization during embryonic axis formation (stage 19/20)47. HOXD4B was expressed in the hindbrain, neural tube and anterior-most somites; HOXD9B was expressed more posteriorly in the somites and neural tube, and HOXD11B was expressed in the most posterior somites and tailbud (Extended Data Fig. 3), which is consistent with similar colinear patterns in other organisms10,15.

Dorsal spines form weeks after early embryonic patterning within a median fin that encircles the developing stickleback (stages 28–31)47. To examine post-embryonic expression, we designed a knock-in strategy to introduce an enhanced green fluorescent protein (eGFP) reporter gene upstream of the endogenous HOXD11B locus using CRISPR–Cas9 (Fig. 1c). The reporter line was generated in an anadromous Gasterosteus background from the Little Campbell River (LITC), British Columbia, Canada, a typical three-spine population that migrates between marine and freshwater environments48. At stage 19–20, we saw GFP expression in the posterior somites and tailbud, a pattern that recapitulated the HOXD11B in situ hybridization results at this embryonic stage (Extended Data Fig. 3c,d). When dorsal spines later form (stage 31), we saw expression in the posterior half of the fish (Fig. 1d), in the dorsal fin fold between the DS2 and DSL, DSL, dorsal fin (DF) (Fig. 1e), the anal fin (AF) and the anal spine (AS). This reporter expression suggests that HOXD11B is expressed both in early development and later during dorsal spine formation (a conclusion also supported by RNA sequencing (RNA-seq); see below and Extended Data Fig. 5).

To determine if HOXDB genes are functionally important for stickleback spine patterning, we used CRISPR–Cas9 to target the coding region of HOXD11B in typical anadromous low-spine Gasterosteus (Little Campbell). Fish in the F0 generation that were mosaic for different mutations in the coding region of HOXD11B showed significantly longer DSL compared to their uninjected control siblings (Fig. 1f,g). The anal spine was also significantly longer (Fig. 1f,g). We also saw an effect on the number of pterygiophores, along the dorsal midline (Fig. 1f and Extended Data Fig. 2a). While low-spine Gasterosteus develop one or two blank (non-spine-bearing) pterygiophores between DS2 and DSL, all CRISPR–Cas9 targeted fish developed only one. To further validate these results, we also tested the effect of HOXD11B targeting in a second population (Rabbit Slough (RABS), Alaska). Again, we observed a significant effect on the length of the DSL and AS (Extended Data Fig. 4). There was no effect on spine number in either population. These results show that HOXD11B is functionally important for dorsal skeletal development.

HOXDB expression is expanded in high-spine Gasterosteus

To examine whether four-spine/high-spine Gasterosteus fish have cis-acting regulatory changes in HOXDB gene expression, we generated F1 hybrids between low-spine and high-spine stocks and used RNA-seq to look for allele-specific expression patterns detectable even when both alleles were present in the same trans-acting environment. The hybrids were generated by crossing LITC anadromous fish, which predominantly have three dorsal spines (‘low-spine’), with a stock descended from the QTL progeny that carry the Boulton HOXDB allele and predominantly show four or five spines (‘high-spine’; Methods). In this cross, 77% of 57 F1 hybrids had 3 dorsal spines, 21% had 4, and 1 fish had 5. RNA was isolated from micro-dissections of each dorsal spine (DS1, DS2, DS3 (if present), DSL), blank pterygiophore (Pter), DF and AF at the developing fin fold stage (Fig. 2a), and separately from whole embryos at embryonic stage 19/20.

Fig. 2: HOXDB genes show cis-acting expression differences in Gasterosteus spines.
figure 2

a, F1 progeny were generated in a cross between a low-spine and high-spine Gasterosteus and tissues were isolated from up to seven indicated locations (DS1, DS2, DS3 (if present), blank pterygiophore (Pter), DSL, DF and AF) to measure allele-specific gene expression in the fin fold stage. Note DS3 only developed in some F1 progeny, so this location has fewer samples (n = 6 for DS3; n = 18 for all other tissues). b, The box plots show the ratios of high-spine to low-spine allele expression at each of three HOXDB genes. The y axis is the log2 of the high-spine versus low-spine read ratio at a SNV (black line: equal expression at a log2 ratio of 0). The x axis shows the seven tissues collected from fin fold stage fish arranged from anterior to posterior, as well as the sample collected from earlier whole embryos (embryo). Centre line, median; box limits, IQR; whiskers, 1.5× IQR; each measurement is represented by a single point. SNVs scored for each dorsal tissue compared to the anal fin: HOXD11B, chrVI:17,756,571; HOXD9B, chrVI:17,764,664; and HOXD4B, chrVI:17,783,616 (gasAcu1-4). Differences were significant in the anterior spines for all three HOXDB genes (DS1: HOXD11B (chrVI:17,756,571) P = 3 × 10−7; HOXD9B (chrVI:17,764,664) P = 6 × 10−6; HOXD4B (chrVI:17,783,616) P = 1 × 10−5; DS2: HOXD11B (chrVI:17,756,571) P = 3 × 10−7; HOXD9B (chrVI:17,764,664) P = 4 × 10−7; HOXD4B (chrVI:17,783,616) P = 4 × 10−7; DS3: HOXD11B (chrVI:17,756,571) P = 4 × 10−4; HOXD9B (chrVI:17,764,664) P = 8 × 10−4; HOXD4B (chrVI:17,783,616) P = 6 × 10−4; all P values by two-tailed Mann–Whitney U-test). **P ≤ 1 × 10−3, ***P ≤ 1 × 10−6. All alleles with 0 reads have been replaced with 0.5 for graphical representation purposes and statistical analysis. NS, not significant.

All three HOXDB genes were expressed in the whole embryo samples from stages 19–20 (Fig. 2b). Reads from RNA-seq were assigned to low- or high-spine HOXDB alleles using exonic single-nucleotide variants (SNVs) that differ between the Little Campbell and Boulton haplotypes (Extended Data Fig. 6). HOXD9B showed no significant allele-specific expression differences at 5 different informative SNVs. HOXD11B showed differences at 3 of 8 SNVs (binomial test P < 0.01), and HOXD4B showed differences at 3 of 6 informative SNVs (binomial test P < 0.001) (Fig. 3b). Different results for different SNVs may reflect the heterogeneity of expression locations and gene isoforms present in whole embryos. Overall, there were no striking expression differences between the two alleles in embryos.

At the later fin fold stage, we sequenced dissected tissues from 12 three-spined and 6 four-spined F1 individuals. We compared the expression in the dorsal spines and fin to anal fin expression as a control. DS1, DS2 and DS3 showed allele-specific expression differences of all three HOXDB genes. Higher expression was seen from the high-spine parent allele (Fig. 2b). Expression differences were seen in all F1 hybrid siblings, regardless of whether they had a three- or four-spined phenotype. Elevated expression of the high-spine allele was not seen at the Pter, DSL or DF locations (Fig. 2b). In DS1 and DS2, almost all detectable sequence reads for all three HOXDB genes came from the high-spine Gasterosteus allele. This is consistent with the previous patterns observed with the HOXD11B low-spine GFP reporter line, which showed low-spine allele expression at posterior, but not anterior, fin fold locations (Fig. 1e). Elevated expression from the high-spine allele led to a significant positive log2 ratio of high-spine to low-spine expression in each of the dorsal spines when compared to the anal fin (Fig. 2b). Similar results were seen for all SNVs that were scoreable in the 3 HOXDB genes (HOXD11B, 10 SNVs; HOXD9B, 10 SNVs; HOXD4B, 8 SNVs).

HOXDB is associated with spine number and length in Apeltes

To determine if other stickleback genera use the same locus to control dorsal spine patterning, we conducted an association mapping study in Apeltes quadracus. As their scientific name suggests, Apeltesquadracus” typically has four dorsal spines. However, multiple wild populations in Canada show more or fewer spines49 (see Fig. 3 and Extended Data Figs. 7 and 8 for further details of anatomy). Apeltes spine number differences are heritable and correlated with ecological conditions50,51. We sampled 2 populations in Nova Scotia: ‘Louisbourg Fortress’ with predominantly 5 dorsal spines (range 3–6; Fig. 3a) and ‘Tidnish River 3’ with predominantly 4 dorsal spines (range 2–6, Extended Data Fig. 8). We genotyped roughly equal numbers of low-spine (2–4 spines) and high-spine (5–6 spines) individuals across the HOXDB locus (Louisbourg n = 211 total, 1 three-spine, 104 four-spine, 99 five-spine, 7 six-spine; Tidnish n = 121 total, 1 two-spine, 1 three-spine, 59 four-spine, 59 five-spine, 1 six-spine). We observed a highly significant association between spine number and genotypes at two markers located between HOXD9B and HOXD11B (Fig. 3b, black line). At the peak marker (AQ-HOXDB_6), wild fish homozygous for the AA allele had an average of 5.1 spines (s.d. = 0.4) while fish homozygous for the GG allele had an average of 4.2 spines (s.d. = 0.5).

Fig. 3: Spine number and DS3 length are associated with the HOXDB locus in Apeltes.
figure 3

a, X-rays of Apeltes quadracus from Louisbourg Fortress with 3–6 dorsal spines. Scale bar, 5 mm. b, Association mapping of Apeltes from Louisbourg Fortress (n = 211) and Tidnish River 3 (n = 121). Both populations show a significant association between spine number and the HOXDB locus. Apeltes from Louisbourg Fortress were also phenotyped for dorsal spine length and showed a significant association between DS3 length and the HOXDB cluster. The markers used are displayed across the top (1–12) and the peak marker (6) is highlighted in red. The dotted line represents the Bonferroni-corrected significance threshold at α = 0.05. The 95% confidence intervals (CIs) (2-LOD) for spine number are denoted by the bars at the bottom (grey for Louisbourg, black for Tidnish). The overlapping 95% CI for spine number is approximately 5,750 bp from the third exon of HOXD11B to the first exon of HOXD9B. The smallest interval shared by both spine number and spine length intervals is approximately 2 kb, including HOXD11B exon 3 and part of the intergenic region between HOXD9B and HOXD11B. Additional anatomical details and association plots for other spine lengths are shown in Extended Data Fig. 7.

Source data

We also measured spine length to test whether, as in Gasterosteus, the Apeltes HOXDB cluster was associated with spine length changes in the Louisbourg fish. We numbered spines similarly to Gasterosteus, with a three-spine Apeltes having from anterior to posterior DS1, DS2 and DSL, and a six-spine Apeltes having DS1, DS2, DS3, DS4, DS5 and DSL (Extended Data Fig. 7a). Only DS3 length was strongly associated with genotypes in the HOXDB region (Fig. 3b, blue line and Extended Data Fig. 7). The genotype at the peak marker (AQ-HOXDB_6) explained 22% of the overall variance in DS3 length of wild-caught fish.

The minimal genomic interval shared by both spine number and length associations was approximately 2 kilobases (kb), including HOXD11B exon 3 and part of the intergenic region between HOXD9B and HOXD11B (Fig. 3b). Based on whole-genome DNA sequencing from Louisbourg (n = 2) and RNA-seq data (n = 14) (Methods), no sequence variation was found in the protein-coding regions of HOXD11B or HOXD9B. The peak marker for both associations was a change of two adjacent intergenic bases from GG to AA. Together, these results suggest that increased spine number and increased DS3 length in some Apeltes probably arise from a regulatory difference in the non-coding interval between HOXD9B and HOXD11B.

Apeltes HOXDB genes show cis-regulatory spine differences

To test for cis-acting regulatory differences in Apeltes HOXDB genes, we generated F1 hybrids with contrasting Apeltes haplotypes in the key genomic interval and carried out RNA-seq on spines, blank pterygiophores and dorsal and anal fins at the fin fold stage (Fig. 4a). While there were no sequence differences in the protein-coding portions of the HOXDB genes, the 3′ untranslated regions (UTRs) of HOXD9B and HOXD11B had variants that could be used to determine the expression level coming from the genotypes associated with low-spine (L) or high-spine number (H) in the association study (Fig. 4b). In F1 fish carrying one L haplotype and one H haplotype, the L haplotype HOXD9B and HOXD11B genes had significantly higher expression (Fig. 4c). Differences were most pronounced in DS3, the same spine whose overall length was associated with HOXDB genotypes (Fig. 4c).

Fig. 4: HOXDB genes show cis-acting expression differences in Apeltes spines.
figure 4

a, Outline of Apeltes fin fold stage fry. Tissues were isolated from DS1, DS2, DS3, DS4, Pter, DSL, DF and AF to measure allele-specific gene expression in F1 hybrids. (DS4 only developed in some progeny.) b, Three HOXDB haplotypes segregating in cross. Black lines: association mapping markers. Pink: regions with genotypes associated with low-spine phenotypes in Fig. 3b. Yellow: regions with genotypes associated with high-spine phenotypes in Fig. 3b. Lighter shading: regions where marker association is unknown but DNA variants are shared between haplotypes of the same colour. c, Box plots showing allele-specific expression ratios in all tissues dissected from fry heterozygous for the H and L haplotypes (n = 4). All box and whisker plots: centre line, median; box limits, IQR; whiskers, 1.5× IQR; each measurement is represented by a single point. Reads from DS3, DS4, Pter, DSL and DF were compared to reads from AF to determine significance (DS3: HOXD9B (chr06:16,028,519) P = 3 × 10−7; HOXD11B (chr06:16,020,516) P = 9 × 10−4, two-tailed Fisher’s exact test). DS1 and DS2 were not assessed because read counts were too low (Extended Data Fig. 5). d, Box plots showing allele-specific expression ratios in all the tissues dissected from fry heterozygous for LHR and H haplotypes (n = 3). Allele-specific expression was seen in DS3 compared to anal fin for both HOXD11B and HOXD9B (DS3: HOXD9B (chr06:16,027,923) P = 9 × 10−8; HOXD11B (chr06:16,020,516) P = 1 × 10−3, two-tailed Fisher’s exact test). e, Box plot showing allele-specific expression ratios in all tissues dissected from fry heterozygous for L and LHR haplotypes (n = 4). Only HOXD9B is shown because HOXD11B lacks informative variants between L and LHR haplotypes (DS3: HOXD9B (chr06:16,027,923) P = 0.08, two-tailed Fisher’s exact test). **P ≤ 1 × 10−3, ***P ≤ 1 × 10−6.

Some F1 fish generated for the allele-specific expression experiment carried both an H haplotype and a recombinant haplotype that we termed the low-high-recombinant (LHR) haplotype (Fig. 4b). These fish showed an allele-specific expression pattern similar to fish heterozygous for an H and L haplotype, with more expression of HOXD9B and HOXD11B coming from the LHR haplotype (Fig. 4d). In contrast, fish heterozygous for the LHR and L haplotypes showed no significant difference in HOXD9B expression (Fig. 4e). Thus, at a gene expression level, the LHR haplotype behaved more like the L haplotype than the H haplotype. Similarly, at the phenotypic level, F1 individuals heterozygous for the L and LHR haplotypes typically had low spine numbers, resembling fish homozygous for L haplotypes, while fish homozygous for H haplotypes had higher spine numbers (L/L fish: 16 out of 16 with 3 or 4 spines; L/LHR fish: 15 out of 18 with 3 or 4 spines, 3 out of 18 with 5 spines; H/H fish: 16 out of 16 with 5 or 6 spines; two-tailed Fisher’s exact test, P = 9 × 10−11; post-hoc pairwise Fisher’s exact test, Bonferroni-corrected at α = 0.05, L/L versus L/LHR Padj = 0.7; L/LHR versus H/H Padj = 1 × 10−6; L/L versus H/H Padj = 1 × 10−8). These results suggest that the key genomic interval controlling both gene expression differences and phenotypic differences between the L/LHR and H haplotypes maps to the minimal approximate 5 kb region shared between the L and LHR haplotypes (pink region on the left of Fig. 4b).

Genomic changes in a Gasterosteus and Apeltes spine enhancer

To search for cis-regulatory sequences contributing to HOXDB expression variation, we looked for conserved non-coding sequences and open chromatin domains located in the minimal interval defined by the association and gene expression studies. This identified 1 approximately 500 base pair (bp) region (Fig. 5a) found in both Apeltes and Gasterosteus that is conserved by phastCons alignment to Tetraodon, Medaka, and Fugu52. This small conserved region contained the peak scoring marker in the Apeltes association study (two adjacent bases changed from GG to AA) (Fig. 5c). This conserved non-coding region also corresponds to a region of open chromatin in Medaka embryos at stages equivalent to those where we see embryonic HOXDB expression in sticklebacks53.

Fig. 5: The genomic region between HOXD11B and HOXD9B contains a conserved AxE showing sequence changes in both Gasterosteus and Apeltes.
figure 5

a, The protein-coding exons of HOXD11B and HOXD9B are shown in Gasterosteus (gasAcu1) genomic coordinates. Sequence conservation: phastCons conserved sequence regions identified in exons and in an approximately 500 bp intergenic region from comparisons between fish genomes. The conserved non-coding region overlaps an assay for transposase-accessible chromatin using sequencing (ATAC-seq) peak from Medaka embryonic stage 19 (ref. 53) and partially overlaps the genomic intervals defined by spine phenotype and RNA expression changes in Apeltes. In high-spine Gasterosteus, the conserved region AxE is deleted (as indicated by a black line between the two grey boxes) and ERV and LINE sequences are inserted (in red and yellow, respectively). b, Approximately 600 bp AxE regions from low-spine and high-spine Apeltes were cloned into a Tol2 GFP expression construct and injected into Gasterosteus embryos. Both versions drove expression in the tailbud of embryos (left) and the fin fold, spines, and dorsal, anal and caudal fins of stage 31 fry (right), confirming that the region acts as an enhancer. Scale bar, 1 mm. c, There are four sequence differences in the AxE region of high- and low-spine Apeltes alleles: one microsatellite variation; an 18 bp indel; a SNP; and 2 adjacent SNPs. Only the single and two adjacent SNPs are within the region implicated by Apeltes recombination and RNA-seq differences (pink bar).

We cloned the Apeltes region from both the L and H haplotypes (611 bp in L; 587 bp in H) and tested whether the sequences could drive GFP reporter expression in transgenic enhancer assays. Because Apeltes have very small clutch sizes, constructs were injected into Gasterosteus to obtain sufficient transgenic embryos for analysis. The approximately 600 bp non-coding constructs both drove expression at embryonic time points in the tail of transgenic embryos in a similar pattern to that seen in the in situ hybridizations for HOXD9B and HOXD11B (Fig. 5b left and Extended Data Fig. 3b,c). At later time points, the conserved non-coding regions drove expression in the dorsal, caudal and anal fins; dorsal and pelvic spines; and the posterior of the fish (Fig. 5b, right). Similar patterns were driven by both the L- and H-type Apeltes constructs (Fig. 5c), although we note that differences in both strength and patterns of expression can be difficult to detect in mosaic transgenic fish resulting from random Tol2 integration (L 611 bp region: n = 22 transgenics with bilateral green eyes; n = 6 out of 22 pectoral fin, n = 8 out of 22 pelvis, n = 9 out of 22 dorsal spines, n = 11 out of 22 dorsal fin, n = 9 out of 22 anal fin, n = 10 out of 22 posterior muscle; H 587 bp region: n = 19 with bilateral green eyes; n = 2 out of 19 pectoral fin, n = 4 out of 19 pelvis, n = 9 out of 19 dorsal spines, n = 11 out of 19 dorsal fin, n = 11 out of 19 anal fin, n = 10 out of 19 posterior muscle (Fisher’s exact test at all sites: Padj = 1)). Given the consistent expression patterns seen in both tail buds and later axial structures of transgenic fish, we refer to the approximately 600 bp conserved intergenic sequence as an axial enhancer (AxE) of the HOXDB locus.

Although AxE sequences are conserved between Apeltes and typical Gasterosteus, we were unable to amplify the AxE region from the Boulton high-spine allele in the Gasterosteus QTL cross. We used PacBio long-read sequencing to identify the intergenic region between HOXD9B and HOXD11B from the Boulton high-spine allele. The sequenced region shows major structural changes, including a deletion that removes almost all of the AxE and the presence of two transposable elements not present at this location in the low-spine reference genome43: a long interspersed nuclear element (LINE) (L2–5_GA) element and an endogenous retrovirus (ERV1-6_GA-I) (Fig. 5a). The LINE element is approximately 1 kb and also present in additional Gasterosteus populations in the Pacific Northwest (sequencing data from54). When the LINE element was detected in other populations, it was not associated with the AxE sequence deletion seen in Boulton. The ERV insertion was approximately 11 kb containing open reading frames for an envelope and Gag-Pol proteins, flanked by approximately 1 kb long terminal repeats (LTRs). Junction sequences for this retroviral insertion near AxE were not found in the sequenced genomes of over 200 Gasterosteus from different populations43,54. Thus, the Boulton high-spine allele shows both the nearly complete loss of AxE and the addition of new sequences.

To determine if loss of the AxE sequence alone was sufficient to recapitulate the phenotypic effect of a higher spine number and shorter DS2 length in Gasterosteus, we deleted the region in low-spine Gasterosteus using CRISPR targeting. In mosaic F0 founder fish, no significant effects on spine number were detected. However, DSL and AS were significantly longer in the F0 injected mutants compared to their control siblings (Extended Data Fig. 9). Both the spines affected and the direction of phenotypic effects resembled the phenotypes seen when targeting the HOXD11B protein-coding region. These results suggest that the AxE region is required for normal spine length patterning in Gasterosteus but that additional sequence changes probably contribute to the spine number phenotypes linked to the region.

To test whether any of the additional transposable element sequences in the Boulton high-spine allele might contribute new enhancer activities, we tested whether the approximately 1 kb LINE or the approximately 1 kb LTR of the ERV could drive GFP reporter gene expression in transgenic enhancer assays compared to an empty vector control. While the empty vector and the LINE did not drive expression at an early fin fold stage (Swarup stage 29), the LTR drove expression in the dorsal fin fold, anal fin, posterior muscle, heart and gills (empty vector: n = 17 transgenics with bilateral green eyes; n = 2 out of 17 whole-body, n = 1 out of 17 heart, n = 0 out of 17 posterior muscle, n = 0 out of 17 dorsal fin fold, n = 0 out of 17 anal fin fold, n = 0 out of 17 gill, n = 0 out of 17 caudal fin; LINE: n = 13 transgenics with bilateral green eyes; n = 0 out of 13 whole-body, n = 0 out of 13 heart, n = 0 out of 13 posterior muscle, n = 0 out of 13 dorsal fin fold, n = 0 out of 13 anal fin fold, n = 0 out of 13 gill, n = 0 out of 13 caudal fin; LTR n = 28 transgenics with bilateral green eyes; n = 3 out of 28 whole-body, n = 10 out of 28 heart, n = 16 out of 28 posterior muscle, n = 13 out of 28 dorsal fin fold, n = 12 out of 28 anal fin fold, n = 4 out of 28 gill, n = 16 out of 28 caudal fin; Fisher’s exact test empty vector versus LTR, whole-body Padj = 1, heart Padj = 0.2, posterior muscle Padj = 4 × 10−4, dorsal fin fold Padj = 4 × 10−3, anal fin fold Padj = 7 × 10−3, gill Padj = 1 and caudal fin Padj = 4 × 10−4). Thus, the ERV insertion that is unique to the Boulton allele includes new cis-regulatory enhancer sequences. Additional regulatory sequences may be present in the rest of the insertion or surrounding region.


Vigorous historical debates have existed about whether mutations in homeotic genes are the likely basis of common morphological changes seen in wild animals. Most laboratory or human-selected mutations in genes are deleterious. In addition, transposable element insertions are strikingly depleted at Hox loci, an effect attributed to the likely deleterious consequences of making substantial regulatory changes in genes essential for development and survival55,56. On the other hand, the diversity of Hox cluster number, composition and expression patterns, and the powerful effects of Hox genes on many phenotypes in laboratory models, have made the genes often-cited candidates for the molecular basis of phenotypic differences between wild species, including sticklebacks45,46. Cis-acting regulatory differences at Hox loci clearly underlie evolutionary differences in trichome and pigmentation patterns in insects, but the underlying molecular changes are not known20,21. Our studies show that independent regulatory changes have occurred in the HOXDB locus of Gasterosteus and Apeltes, providing a compelling example of cis-acting variation in Hox genes linked to the evolution of new axial skeletal patterns in wild vertebrate species.

Adaptive significance of dorsal spine number and length

Dorsal spines in sticklebacks play an important role in predator defence. Long spines increase the effective cross-sectional diameter of sticklebacks57 and can provide a survival advantage against gape-limited predators58. Prominent spines also provide holdfasts for grappling insect predators and may therefore increase the risk of predation by macroinvertebrates39,59. Different predation regimes may thus favour either increased or decreased spine lengths and numbers. The intensity of bird, fish and insect predation varies across locations, years and seasons, contributing to a range of spine phenotypes in natural stickleback populations60,61,62,63;.

Boulton is an extensively studied population where fish typically show two or three dorsal spines39. Detailed seasonal and longitudinal surveys showed that lower spine numbers in Boulton fish are correlated with a higher intensity of insect predation and higher spine numbers with a higher intensity of bird predation60,64. Because fish with four dorsal spines have not been seen in over 20,000 wild-caught Boulton fish, the occurrence of four-spine sticklebacks in the Boulton × Bodega Bay F2 laboratory cross is a transgressive phenotype65 that emerges when Boulton alleles are inherited on a mixed genetic background. We note, however, that the Boulton HOXDB region is also linked to shortening of DS2 in the QTL cross. We hypothesize that a HOXDB allele probably evolved in this population for its contributions to reduced DS2 length via a posteriorizing mechanism (Fig. 6). In the Boulton genetic background, this posteriorizing effect may be sufficient to change DS length without inducing formation of a new spine on a blank pterygiophore, while in the mixed genetic background of the QTL cross, the posteriorizing tendency of the Boulton allele leads to both spine length and number changes.

Fig. 6: Model of dorsal spine pattern changes resulting from coding and expression differences in laboratory and wild stickleback populations.
figure 6

Typical Gasterosteus aculeatus (three-spine sticklebacks) have a convex pattern of spine lengths where the middle/DS2 spine is the longest). All three HOXDB genes have no or low expression in the anterior, DS1 and the strongest expression in the posterior DSL regions. CRISPR-induced coding region mutations in HOXD11B lengthen DSL. This is expected since the loss of a posterior Hox gene typically results in the anteriorization of structures; in this case, DSL is longer and thus looks more like DS2 (left). In contrast, the naturally occurring Boulton high-spine allele has expanded HOXDB gene expression anteriorly, such that all the HOXDB genes are expressed in all of the spines. Corresponding morphological effects seen in the QTL cross (shortening of DS2 and addition of a new spine on a blank pterygiophore between DS2 and DSL) are both consistent with posteriorization of dorsal midline structures (middle bottom). These morphological and gene expression changes are linked to cis-regulatory changes at the HOXDB locus, including the deletion of an enhancer (AxE) and insertion of two transposable elements. Typical Apeltes quadracus fish have a concave pattern of spine lengths (middle/DS3 spine the shortest). In low-spine Apeltes, HOXDB genes are strongly expressed from DS3 to the posterior. In high-spine Apeltes, the anterior boundary of HOXDB expression is shifted to the posterior and thus DS3 has lower HOXDB expression. Loss of expression of HOXDB genes is associated with lengthening of DS3 and the formation of a new spine on a pterygiophore between DS3 and DSL, both consistent with anteriorization of dorsal midline identities.

Although Apeltes sticklebacks typically develop four dorsal spines, many “quadracus” populations in Canada have predominantly three or five spines49,50. In an extensive comparison of Apeltes spine numbers and environmental variables across 570 locations, Blouw and Hagen66,67,68 found that increased spine number was correlated with the presence of predatory fish, while decreased spine number was correlated with more and denser vegetation. To study the possible adaptive value of spine number differences, Blouw and Hagan exposed mixed populations of four- and five-spine Apeltes to predatory fish and measured differential survival of spine morphs when half of the sticklebacks had been eaten67. When vegetation was present, predation was non-selective; however, when vegetation was absent, five-spine fish were less likely to be eaten by perch and trout. Consistent with the experimental predation experiments and known ecological correlations, the stomachs of wild-caught trout contained more four- than five-spine fish, while the stomachs of herons contained more five- than four-spine fish67. Thus, the dorsal spines in sticklebacks provide an excellent example of a prominent adaptive structure in vertebrates that evolves in response to different predation regimes in natural environments and diversifies in part through repeated regulatory changes in Hox genes.

Although spine phenotypes are clearly adaptive in sticklebacks, we note that we have not yet established whether the specific HOXDB alleles identified in the Boulton and Nova Scotia populations have been subjected to positive selection. Multiple dorsal spine phenotypes have been present in the Boulton, Louisbourg and Tidnish populations for decades, probably due to long-term balancing selection in the face of fluctuating predation regimes49,60,67. Long-term balancing selection is more difficult to detect at the molecular level than selective sweeps of a favoured allele69. However, recent experiments have successfully monitored short-term changes in the frequency of alleles of interest in sticklebacks after changes in environmental conditions54,70 or experimental exposure to predators59. In the future, it will be interesting to extend such studies to differential survival of particular Hox alleles and dorsal spine phenotypes, using either the alternative alleles we have identified in this study from natural Gasterosteus and Apeltes populations or CRISPR-edited sticklebacks engineered to carry particular HOXDB sequence variants.

Hox genes and dorsal midline skeletal patterns

Hox genes are well known for controlling the identity of structures in repeating series, including body segments in insects, somite fates in vertebrates, digit identities in limbs and rhombomere segments in the hindbrain10. We propose that changes in dorsal spines and pterygiophores of fish also represent identity transformations within the dorsal midline. Many Hox changes are governed by ‘posterior prevalence’, where the posterior-most, highest-numbered Hox gene expressed in a given region generally controls the fate of that region71,72. Therefore, when posterior gene expression expands, the regions with expanded expression generally acquire a more posterior fate. Conversely, when activity of a Hox gene is lost from a region, that region typically acquires a more anterior fate.

Our RNA-seq studies show that many Hox genes in both Gasterosteus and Apeltes are expressed in the dorsal spines or fins of developing sticklebacks, with the exception of Hox PG1, some Hox PG13 genes, and all HOXBB cluster genes (Extended Data Fig. 5). Several Hox genes show strong differential expression across different spines and pterygiophores (including HOXDB genes), suggesting that morphological fates in the dorsal midline are probably influenced by the combined expression of multiple genes.

The Gasterosteus high-spine allele that causes expanded expression of the HOXDB genes would be predicted to result in a posteriorization of the regions gaining expression. As expected, this allele is associated with both a blank pterygiophore developing a spine (consistent with a shift to a more posterior identity like DSL) and the shortening of DS2 (consistent with what is typically the longest spine becoming more like the shortest, DSL). Conversely, knocking down HOXD11B expression by CRISPR–Cas9 is predicted to result in anteriorization of structures; the increased length of the DSL that we observe is consistent with a shift of DSL to a more anterior and therefore longer spine fate (Fig. 6).

The Apeltes H allele that shows reduced HOXD9B and HOXD11B gene expression in DS3 is associated with both increased spine number and a longer DS3. The number and length phenotypes can both be interpreted as transformations to a more anterior fate. In this model, the appearance of a fifth spine on a normally blank pterygiophore could be explained by partial transformation to a more anterior, spine-bearing fate. Anterior spines are normally longer than posterior spines in Apeltes, so the increased length of DS3 is also consistent with an anterior transformation (Fig. 6).

Independent changes in HOXDB cis-regulatory elements

Our allele-specific expression experiments show that the changes in HOXDB expression we see in sticklebacks are due to cis-acting regulatory differences linked to the Hox genes themselves, rather than secondary consequences of changes in unknown trans-regulatory factors. Our mapping, association, and transgenic experiments have identified a particular cis-acting enhancer region located between HOXD9B and HOXD11B that can recapitulate axial expression patterns and shows independent sequence changes in Gasterosteus and Apeltes with different spine numbers. In Apeltes, the most likely sequence difference mediating changes of HOXD9B and HOXD11B expression are two adjacent SNPs (marker: AQ-6) in AxE. These SNPs convert a TGGT sequence in the L allele to TAAT in the H allele. They represent the peak marker scored by association mapping and also map within the minimal recombination interval that controls H versus L and LHR cis-acting expression differences seen in F1 hybrids (Fig. 5c). Both the TAAT change and a nearby 18 bp indel, which represents the second highest scoring marker (AQ-7) in the association mapping, are found in the high-spine fish of two populations on opposite coasts of Nova Scotia. Thus, repeated evolution of different high-spine Apeltes populations probably takes place through a shared underlying molecular haplotype at the Hox locus rather than independent mutations in these different populations. A similar allele sharing process underlies recurrent evolution of a variety of other phenotypic traits in both sticklebacks and other organisms35,73,74. We note that the derived TAAT sequence in the Apeltes high-spine allele creates a predicted core binding motif for a homeodomain protein. Previous studies in Drosophila and other organisms have shown that Hox genes can autoregulate in positive and negative feedback loops75,76,77. We hypothesize that the creation of a new putative homeodomain binding site located between the HOXD9B and HOXD11B genes may contribute to the decreased HOXDB expression observed with the H allele. Note that we have not been able to recapitulate the altered expression patterns using AxE transgenic reporter constructs integrated at random locations in the genome. Because endogenous Hox expression patterns are likely controlled by interactions between multiple long-distance control elements and surrounding topological domains78,79, the most accurate functional tests of the phenotypic effects of mutations will come from scoring those changes at their correct genomic position. Future advances in genome editing may eventually make it possible to recreate or revert the TGGT and TAAT sequence change at the endogenous HOXDB locus in sticklebacks and further test whether these two adjacent base pair changes are sufficient to alter Hox gene expression and spine length or number.

In Gasterosteus, the AxE enhancer has been deleted from the Boulton HOXDB high-spine allele and replaced with two transposable elements, one ERV and one LINE. Removing the endogenous AxE enhancer by CRISPR–Cas9 does not lead to spine number and DS2 phenotypes but it does recapitulate the DSL length changes seen by targeting the HOXD11B coding region. The LTRs found in ERVs can act as enhancers80, including the LTR sequences inserted in the Boulton allele. We hypothesize that the large insertion in the Boulton allele underlies broader expression in the dorsal spines and the other phenotypic consequences of the Boulton high-spine allele.

Repeated regulatory changes in morphological evolution

A long-standing question in evolutionary biology is whether the same genetic mechanisms are used repeatedly to evolve similar traits in different populations and species. Although Gasterosteus and Apeltes last shared a common ancestor over 16 million years ago30, our data show that both stickleback groups have independent cis-regulatory changes in HOXDB, which are linked to new dorsal spine patterns in recently evolved populations. The types of mutations made in the AxE regulatory region are clearly distinct and the naturally occurring Gasterosteus and Apeltes H alleles lead to contrasting HOXDB expression changes. Interestingly, the HOXD locus is also used repeatedly during horn evolution in mammals. The HOXD region shows accelerated evolution and insertion of a new retroviral element in the diverse clade of species with horns and antlers81. In addition, rare four-horned sheep and goats have recently been shown to have independent mutations in the HOXD locus, ranging from a 4 bp mutation that alters splicing to a large (approximately 500 kb) deletion that is lethal when homozygous82,83,84. The fish and mammalian results support a growing body of literature that has found repeated use of the same loci underlying similar traits, even though the direction of effect of gene expression and mutational mechanism are often different74.

While both of our examples of spine variation in recently diverged populations of Gasterosteus and Apeltes involve cis-regulatory changes, Hox coding region mutations may also contribute to diversification of spine patterns over a wider phylogenetic scale. For example, the Gasterosteidae family can be separated into five different genera of predominantly three-spine, four-spine, five-spine, nine-spine and fifteen-spine sticklebacks (Gasterosteus, Apeltes, Culaea, Pungitius and Spinachia, respectively). We note that the coding region of HOXD11B shows a high rate of non-synonymous to synonymous substitutions across the stickleback family; the dN/dS ratio is greater than 1.0 for comparisons between Apeltes and Gasterosteus (Extended Data Fig. 10). This suggests that changes in HOXD11B coding regions have likely been under positive selection during the divergence of Apeltes and Gasterosteus, perhaps contributing to the distinctive patterns of spine length and number that are characteristic of these two genera.

Spiny-rayed fish are among the most successful vertebrates, making up nearly a third of extant vertebrate species. The lengths and numbers of spines show remarkable diversity across Acanthomorpha, including elaborate modifications that have evolved for defence, camouflage, luring prey or swimming biomechanics24. Our results show that changes in the dorsal spine patterns of wild fish species have evolved in part through genetic changes in Hox genes. Based on the recurrent use of the same locus for spine evolution in different stickleback species, we hypothesize that repeated mutations in Hox genes may also underlie other interesting changes that have evolved in the axial skeletal patterns of many other wild fish and animal species.


Ethical compliance

We complied with all relevant ethical regulations during this study.

Stickleback care

Sticklebacks were captured using minnow traps or dip and seine nets. Populations and their Global Positioning System (GPS) coordinates are in Supplementary Table 1. All sticklebacks were treated in accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health (NIH), using protocols approved by the Institutional Animal Care and Use Committee of Stanford University (protocol no. 13834), in animal facilities accredited by the Association for Assessment and Accreditation of Laboratory Animal Care International.

DNA extraction

DNA was isolated from fins by incubating in lysis buffer (10 mM Tris, pH 8, 100 mM NaCl, 10 mM EDTA, 0.5% SDS, Proteinase K (333 μg ml−1); catalogue no. P8107S; New England Biolabs)) at 55 °C for 4 h to overnight, extracting with phenol:cholorform:isoamyl alcohol 25:24:1 (catalogue no. P3803; Sigma-Aldrich) in phase lock tubes (MaXtract High Density, catalogue no. 129056; QIAGEN), ethanol precipitating overnight and resuspending in TE Buffer (10 mM Tris, 1 mM EDTA, pH 8).

QTL mapping

A wild-caught female from Boulton was crossed by in vitro fertilization to a marine male stickleback from Bodega Bay. F1 progeny were raised to adulthood in the laboratory in 29 gal aquariums in reverse osmosis-purified water with 3.5 ppt Instant Ocean salt and intercrossed to generate multiple F2 families. Sperm from F1 males were cryopreserved85, so single males could be crossed multiple times. F2 progeny fish were raised in the laboratory for 1 year, euthanized with 200 mg l−1 tricaine methanesulfonate (Abbreviated New Animal Drug Application no. 200-226; Western Chemical) buffered to pH 7 with sodium bicarbonate and preserved in 70% ethanol.

Fish DNA samples were genotyped using an Illumina GoldenGate genotyping array with 1,536 features40. Intensity data were processed using GenomeStudio v.2011. Genotype clusters were inspected and adjusted manually and uninformative or low-intensity SNPs were excluded from downstream analysis. Phasing and linkage map construction were performed with TMAP version 0.686. The linkage map and phased genotype data were then loaded into R/qtl version 1.26.087 and filtered to remove fish with fewer than 600 genotype calls and markers with fewer than 300 calls. A final map was generated with 343 F2s and 452 markers.

Gasterosteus anatomical traits and landmarks are diagrammed in Extended Data Fig. 2. Pterygiophore numbers and abdominal, caudal, and total vertebrae were counted from X-rays taken on a Faxitron UltraFocus X-ray cabinet (settings: 38 kV, 4.8 s). Spine lengths were measured on X-rays using Fiji version 2.0.088 and adjusted by taking residuals from multiple regression against standard length and sex. Presence or absence of a fourth spine and Pterygiophore numbers (6 or more than 6) were coded as binary traits (0 or 1). Phenotypes were analysed in R/qtl using Haley–Knott regression via the scanone function, with a normal model for the length traits and a binary model for the spine and pterygiophore number87. For the vertebral counts, a non-parametric scanone analysis was done. Permutation tests (n = 1,000) were used to establish the logarithm of the odds (LOD) significance thresholds (α = 0.05) for each trait. The analysis is based on 340 F2 fish and a set of 452 SNP markers.

In situ hybridization probes

RNA was extracted by homogenizing 10–20 stage 19/20 embryos in TRIzol using a FastPrep-24 machine (MP Biomedicals) and lysing matrix M. RNA was washed once with chloroform, precipitated with isopropanol and resuspended in diethyl pyrocarbonate water. RNA was treated with on-column DNase and was cleaned up using the QIAGEN RNeasy Mini (catalogue no. 74104). Complementary DNA (cDNA) was made with SuperScript VILO cDNA Synthesis (catalogue no. 11754050; Thermo Fisher Scientific). For each riboprobe, PCR with reverse transcription amplification was done with the primers shown in Supplementary Table 2. The HOXD11B probe was cloned into pCR2.1-TOPO (catalogue no. K450001; Invitrogen) in both orientations; the HOXD4B and HOXD9B probes were cloned into pCRII-Blunt II-TOPO (catalogue no. K280020; Invitrogen) in both orientations. The vectors containing the probe sequences were linearized with BamHI (catalogue no. FD0054; Thermo Fisher Scientific), and the sense and antisense probes were in vitro transcribed with T7 RNA Polymerase (catalogue no. P2075; Promega Corporation).

Whole-mount in situ hybridization

Whole-mount in situ hybridizations at Swarup stages 19–20 were done as described by Thisse & Thisse89 with the following modifications. Embryos were manually dechorionated with forceps (catalogue no. 11251-10; Fine Science Tools) after overnight fixation in 4% paraformaldehyde in PBS. To remove the pigmentation, they were bleached for 10 min in 0.8% KOH, 3% hydrogen peroxide (30%) and 0.1% Tween 20. Finally, embryos were permeabilized with proteinase K for 10 s at 10 µg ml−1 in PBS with 0.1% Tween 20.

GFP knock-in

CRISPR–Cas9 was used to generate GFP reporter lines for HOXD11B, as described elsewhere90. Cas9 protein (QB3 MacroLab, University of California at Berkeley), a donor plasmid (pTia1l-hspGFP, deposited at Addgene, containing hsp70, GFP and a single guide RNA (sgRNA) target site), and two sgRNAs were injected. One sgRNA (HOXD11B-GFP-sgRNA; Supplementary Table 3) targeted the region 346 bp upstream of the endogenous HOXD11B start codon, and one targeted the donor plasmid. The HOXD11B-GFP-sgRNA was designed as described previously91. Tia1l sgRNA was used to cut the plasmid92 and has a sequence not present in the Gasterosteus aculeatus genome. The injection mix contained a final concentration of 1 μg μl−1 Cas9 protein, 31 ng μl−1 Tia1l sgRNA, 31 ng μl−1 HOXD11B-GFP-sgRNA, and 0.05% phenol red; it was adjusted to the final concentration with 10 mM Tris, pH 7.5.

Fertilized eggs from Gasterosteus LITC fish were injected at the single-cell stage and embryos were screened at stage 20 (approximately 84 hours post-fertilization) for GFP expression. GFP+ fish were imaged again at stages 29–31 (18 days post-fertilization). Fry were anaesthetized with 3 mg −1 tricaine. Imaging was done with an MZ FLIII fluorescence stereomicroscope (Leica Microsystems) using GFP2 filters and a ProgResCF camera (Jenoptik). GFP+ fish were grown to adulthood and crossed to wild-type LITC fish. Progeny were screened at stage 20 for GFP expression. To confirm integration and orientation of the GFP construct, primers were designed upstream and downstream of the HOXD11B-GFP-sgRNA site and in the plasmid on either side of the sgRNA cut site within the hsp70 promoter or the TOPO backbone (Supplementary Table 2). All combinations of primers were tested by PCR. The presence and absence of bands were used to determine orientation. Sanger sequencing was used to determine exact integration sites.

HOXD11B coding and regulatory mutations using CRISPR–Cas9

Mutations in the HOXD11B coding regions were generated by injecting Cas9 protein and an sgRNA targeting the first exon (HOXD11B-coding-sgRNA; Supplementary Table 3). The sgRNA was designed and synthesized as described previously91 and injected at 300 ng μl−1 with 1 μg μl−1 of Cas9 protein and 0.05% phenol red in 10 mM Tris, pH 7.5, into fertilized eggs from two anadromous Gasterosteus populations (LITC and RABS) at the single-cell stage. Mutations were confirmed by PCR (using Phusion High-Fidelity DNA Polymerase (catalogue no. F-530L; Thermo Fisher Scientific), GC buffer and 3% dimethylsulfoxide (DMSO) with HOXD11B-coding_1F and 1R; Supplementary Table 2) amplified at 98 °C (3 min), then 35 cycles at 98 °C (10 s)/60 °C (30 s)/ 72 °C (30 s), and a final extension at 72 °C for 10 min.

Two strategies were used to delete AxE, the conserved enhancer (466 bp) between HOXD9B and HOXD11B: (1) 3 sgRNAs (AxE-sgRNA_1, 2 and 3; Supplementary Table 3) and a 60 bp repair phosphorothioate modified oligonucleotide (Integrated DNA Technologies) with 30 bp of homology to either side of the enhancer93; or (2) a total of 6 sgRNAs (AxE-sgRNA_1 through 6; Supplementary Table 3) targeting the edges and middle of the enhancer. The sequence for the repair oligo was G*A*A CGT AAA AGG ATT CAG GAG CTC AAG CGA GTC GGT TCC AAA CGT GTC GTT GCC CAG C*A*G (the asterisks indicate phosphorothioate bonds). If the first two bases of the sgRNA target sequences were not Gs, then they were replaced to aid in the transcription of the sgRNA. The injection mix included 1 μg μl−1 of Cas9 protein, 300 ng μl−1 of total of the sgRNAs (100 ng μl−1 of each for strategy 1, 50 ng μl−1 of each for strategy 2), 1.5 pmol μl−1 repair oligonucleotide (strategy 2 only), 300 mM KCl94, 0.05% phenol red. Mutations were confirmed by PCR as described above, except that the extension time was 1 min, and the annealing temperature was 64 °C. Two sets of primers were used, with the first amplifying only the 571 bp region including the enhancer and the second including approximately 3.6 kb around the enhancer (Supplementary Table 2).

Apeltes quadracus association mapping

Apeltes quadracus were collected in May 2018 and May and July 2019 using minnow traps and dip nets from Fortress Louisbourg (site 325) and Tidnish River Site 3 (site 171)50 (GPS coordinates in Supplementary Table 1). Sticklebacks were euthanized as described above and fixed in 70% ethanol or Alfred Lamb’s Navy Dark Rum 151 Proof. Apeltes anatomical traits and landmarks are diagrammed in Extended Data Fig. 7. Fish were phenotyped for spine number by X-ray as described above. The spine lengths and standard lengths of Louisbourg fish were measured in triplicate using digital callipers, averaged, and used to calculate residuals for each spine against the standard length of the fish.

To identify Apeltes genotyping markers, HOXDB sequences were amplified by PCR using primers (PUNG-GAC_1-11; Supplementary Table 2) conserved between the Gasteosteus aculeatus43 and Pungitius pungitius genomes (GenBank assembly accession nos. GCA_003399555.1, GCA_003935095.1, GCA_902500615.2)95,96 (Supplementary Table 2). PCR products were cloned into pCRII-Blunt II-TOPO, miniprepped and Sanger-sequenced from 2–4 individuals with differing spine numbers to identify variable regions. Additional regions were filled by designing primers spanning the initial products. These PCR products were also cloned and Sanger-sequenced using primers shown in Supplementary Table 2.

Twelve markers were identified throughout the Apeltes HOXDB cluster and scored in 211 fish from Louisbourg Fortress (7 six-spine, 99 five-spine, 104 four-spine, 1 three-spine) and 121 fish from Tidnish River 3 (1 six-spine, 59 five-spine, 59 four-spine, 1 three-spine, 1 two-spine).

Microsatellite markers were amplified using the universal fluorescent primer system97. A 20 µl PCR reaction mixture contained 2× Master Mix (catalogue no. K0171; Thermo Fisher Scientific), 0.5 µM 6FAM M13 forward universal primer, 0.125 µM forward primer, 0.5 µM reverse primer and 10 ng genomic DNA. The PCR programme was 94 °C (5 min), 30 cycles at 94 °C (30 s)/58 °C (45 s)/72 °C (45 s), 8 cycles at 94 °C (30 s)/53 °C (45 s)/72 °C (45 s) and a final extension at 72 °C for 10 min. For AQ-HOXDB_2, the cycle number was reduced from 30 to 27. PCR was cleaned using ExoSAP-IT PCR Product Cleanup Reagent (catalogue no. 78205.1.ML; Applied Biosystems), fragment sizes were analysed on an Applied Biosystems 3730xl Genetic Analyzer, and peaks were called using the Microsatellite plugin for Geneious version 10.2.3.

Indel and SNP markers were scored using PCR with 2× Master Mix, 0.5 µM forward primer, 0.5 µM reverse primer and 10 ng genomic DNA. The PCR programme was 95 °C (5 min), 35 cycles at 95 °C (30 s)/54 °C (45 s)/72 °C (30 s) and a final extension at 72 °C for 10 min. The one exception was AQ-HOXDB_6, where PCR was done with Phusion High-Fidelity DNA Polymerase, GC buffer and 3% DMSO; the PCR programme was 98 °C (3 min), 35 cycles at 98 °C (10 s)/60 °C (30 s)/72 °C (10 s) and a final extension at 72 °C for 10 min. The AQ-HOXDB_5 PCR product was digested with BssSI-v2 (catalogue no. R0680L; New England Biolabs); the AQ-HOXDB_6 PCR product was digested with NdeI (catalogue no. FD0583; New England Biolabs). PCR products were run on a 2% agarose gel to score size differences.

Allele frequencies in low-spine (two- to four-spine) and high-spine (five- to six-spine) fish were compared using CLUMP version 2.498, which performs a modified chi-squared analysis to determine the significance of allele frequency differences. For microsatellite markers, the negative log P values of the chi-squared value (T4) from the 2 × 2 contingency table generated by CLUMP are shown in Fig. 3. For indel or SNP markers, a chi-squared test was performed in R v.3.6.1. Associations between residual spine lengths and genotypes were quantified using an analysis of variance performed in R.

Apeltes genome assembly

Whole-genome sequencing (WGS) using 10X Genomics Chromium-linked read technology was performed on two Apeltes quadracus from the Louisbourg Fortress population (one four-spine and one five-spine). Genomic DNA was extracted from brains and prepared using the QIAGEN MagAttract HMW DNA kit. Linked read data from each fish were assembled using Supernova v.2.1.1 with default settings99. The four-spine Apeltes assembly had 16,216 scaffolds with 416,290,932 bases (scaffold N50: 393,888 bp; L50: 247; N90: 7,174 bp; L90: 3,684). The five-spine Apeltes assembly had 24,175 scaffolds with 397,678,333 bases (scaffold N50: 69,128 bp; L50:1,629; N90: 4,805 bp; L90: 10,192).

To use GATK version in the allele-specific RNA-seq pipeline, the genome needed to be on fewer scaffolds than generated by linked read data. To achieve this, we started with the four-spine Apeltes assembly and assumed that the chromosome structure of Apeltes and Gasterosteus are similar. We used a reference-guided scaffold approach by generating global genome to genome alignments with minimap2 (ref. 100) and MUMmer version 4.0.0101. The alignment information was processed by RaGOO version 1.1102 to order and orient contigs into scaffolds, which resulted in the Apeltes genome reference used in the GATK allele-specific RNA-seq pipeline.

High-spine Gasterosteus genome assembly

WGS using 10X Genomics Chromium-linked read technology was performed on two four-spine Gasterosteus aculeatus from the F5 generation of the Boulton-Bodgea Bay QTL cross. Genomic DNA was extracted from the brains of the fish and prepared using the QIAGEN MagAttract HMW DNA kit. The linked read data of each fish were assembled using Supernova v.2.1.1 with default settings99.

WGS using PacBio HiFi was also performed on one four-spine Gasterosteus aculeatus from the F5 generation of the Boulton-Bodega Bay QTL cross. Genomic DNA was extracted from the testes of the fish and prepared using the QIAGEN MagAttract HMW DNA kit. The genome was assembled using Canu version 2.1.1. The purge haplotigs pipeline ( was used to phase the alleles and identify the contigs that appeared twice in the assembly. The final assembly had 483 scaffolds with a total of 489,328,730 bases (scaffold N50: 3,689,351 bp; L50: 37; N90: 633,554 bp; L90: 166).

Transgenic enhancer assays

To identify and confirm sequence variants in the intergenic region between HOXD9B and HOXD11B, the approximately 6 kb intergenic region from Apeltes was amplified from a three-spine and a six-spine Apeltes from the Louisbourg population and a three-spine and a six-spine from the Tidnish population with Phusion High-Fidelity DNA Polymerase in GC buffer and 3% DMSO using the primers in Supplementary Table 2. The resulting products were TOPO-cloned into pCRII-Blunt II-TOPO. Colonies were miniprepped and Sanger-sequenced. To generate the plasmids for the enhancer assay, the low- and high-spine versions of the approximately 600 bp region that contains AxE were then amplified with primers that included overhangs homologous to the PT2HE GFP reporter vector28,103. The reporter vector was cut with EcoRV (catalogue no. ER0201; Thermo Fisher Scientific), and the insert and vector were joined using Gibson Cloning (catalogue no. E2611S; New England Biolabs). The resulting plasmids were screened by SacI (catalogue no. ER1131; Thermo Fisher Scientific) restriction digest and further Sanger-sequenced to check for mutations. The 587 bp high-spine and 611 bp low-spine Apeltes AxE sequences are available in GenBank at OK383404 and OK383405, respectively.

To test the enhancer activity of portions of the transposable element insertions found in the high-spine Boulton allele, the LTR from the ERV and LINE were amplified with the primers BOUL-HOXDB_LTR and BOUL-HOXDB_LINE, respectively. The resulting products were inserted into the PT2HE GFP reporter vector as described above.

Transgenic Gasterosteus aculeatus sticklebacks were generated by microinjection at the single-cell stage from LITC Gasterosteus aculeatus as described in Chan et al.34. Plasmids (25 ng µl−1) were injected with Tol2 transposase messenger RNA (36 ng µl−1) and 0.1% phenol red as described in Hosemann et al.104. Tol2 mRNA was synthesized by in vitro transcription using the mMessage mMachine SP6 kit (catalogue no. AM1340; Invitrogen) from the pCS-TP plasmid105 cut with Bsp120I (catalogue no. ER0131; Thermo Fisher Scientific). Transgenics were imaged at stage 20 (approximately 84 hours post-fertilization) and stage 29/31 (approximately 18–30 days post-fertilization) as described in the GFP knock-in section above. The hsp70 promoter drives expression in the lens of the eye by 9 d post-fertilization106. At stage 29/31, bilateral lens GFP expression was used to identify less mosaic fish. For all statistics, to compare expression patterns between different constructs, a Fisher’s exact test with a Bonferroni-corrected significance threshold at α = 0.05 was used.

dN-dS calculation

dN-dS calculations were performed in R using ape v.5.3 (ref. 107). Sequence alignments for each gene (HOXD11B, HOXD9B, and HOXD4B) were generated in Geneious using translation alignment. Gasterosteus transcripts were based on splicing patterns validated from cDNA. Sequences for P. pungitius were determined by BLASTN version 2.7.1108 of Gasterosteus exons against Pungitius: GCA_003935095.1 (ref. 95). Sequences for Apeltes were identified from our genome assembly. Sequences for Gasterosteus wheatlandi, Culaea inconstans and Spinachia spinachia were identified by BLAST version 2.7.1 of Gasterosteus exons against unassembled short reads from WGS of the respective species109 (C. Peichel and M. Hiltbrunner, personal communication).


For Gasterosteus RNA-seq, a lab-raised LITC anadromous female with three dorsal spines was crossed to a high-spine Gasterosteus male with five dorsal spines. The high-spine Gasterosteus line is the F5 generation of the original QTL cross between Boulton and Bodega Bay. The fish were selected for high spine number; by F5, more than 80% of fish have 4 or more dorsal spines. To confirm that the fish carried the Boulton allele at the HOXDB locus, the allele was amplified using BOUL-HOXDB_1F and 1R (Supplementary Table 2) with Phusion High-Fidelity DNA Polymerase in GC buffer and 3% DMSO using a 2-step PCR programme (94 °C (1 min), 30 cycles at 98 °C (10 s)/68 °C (15 min) and a final extension at 72 °C for 10 min) and run on an agarose gel. The Boulton allele is approximately 15 kb, and the BDGB allele is approximately 1.9 kb. The sequences of the two alleles are available at OK383406 (BDGB) and OK383407 (Boulton) in GenBank.

The resulting clutch was raised to 11–13 mm. Fry were euthanized as described above and dissected on a 2% agarose plate with size 00 insect pins and Spring Scissors (catalogue no. 15000-08; Fine Science Tools). Spine, fin and pterygiophore tissues (Fig. 2a) were collected and flash-frozen in liquid nitrogen in FastPrep Tubes (catalogue no. MP115076200; MP Biomedicals). DNA from tails was amplified and genotyped to ensure fish had informative SNPs in the coding regions of HOXD11B and HOXD9B using HOXD11B-coding_1F and 1R primers, HOXD9B-coding_1F and 1R primers (Supplementary Table 2) and the PCR conditions described above.

Based on genotyping, 12 three-spine progeny and 6 four-spine progeny were chosen for RNA extraction, library preparation and sequencing. Samples for RNA extraction were homogenized using MP FastPrep 2 ×20 s with Matrix M with a 5 min rest in between. RNA extractions were performed using NucleoSpin RNA XS (Takara Bio) and resuspended in 20 µl RNase-free water. RNA was quantified by Qubit (Invitrogen) using the HS Assay Kit (catalogue no. Q32851; Invitrogen). A subset of samples was quality-controlled to check the RNA integrity number (RIN) values by BioAnalyzer using the RNA 6000 Pico Kit (catalogue no. 5067-1513; Agilent Technologies). The RINs were between 8.2 and 10, with most higher than 9.6. Sequencing libraries were generated with the Illumina Stranded mRNA Prep kit (catalogue no. 20040532) and 20–100 ng RNA. PCR cycle numbers were determined by quantitative PCR (generally 12 cycles for embryo samples with 200 ng RNA, 14 cycles for dorsal spine and pterygiophore samples with 100 ng RNA, 13 cycles for dorsal and anal fin samples with 100 ng RNA and 15 cycles for samples with less than 100 ng RNA input). Quality control of libraries was done by Qubit with a dsDNA HS Assay Kit to check concentrations and by BioAnalyzer with a high-sensitivity kit (catalogue no. 5067-4626; Agilent Technologies) to check sizes. Libraries were sequenced to approximately 30 million read coverage on a NovaSeq 6000 (2 × 150 bp) by NovoGene. Reads were trimmed with Cutadapt version 2.4110 using the TrimGalore version 0.6.6 wrapper ( and mapped to the gasAcu1-4 reference genome ( with STAR version 2.7.10a two-pass mapping111. For allele-specific expression analysis, base quality was adjusted and variants were called using GATK version as recommended by the Broad Institute112,113. Reads at each site were counted using GATK ASEReadCounter version We required that SNPs be called as heterozygous in at least 1 tissue of each fish, that the number of reads at a given site be greater than 12 (three-spine) or 14 (four-spine) for each fish and that the overall minor allele frequency be greater than 5%. To quantify allele-specific expression differences seen between the dorsal tissues and anal fin (control), we took log2 ratios of reference reads to alternate reads in each sample and compared each dorsal tissue to the anal fin using a Mann–Whitney U-test.

To improve gene predictions and recover new transcripts for differential gene expression analysis, StringTie version 2.1.4 was used along with the existing Ensembl annotations114. Given the large number of reads, BAM files were filtered by quality, downsampled to 20% and merged into one file that was used as the input for StringTie. The merge function was used to add genes from the Ensembl annotations not present in the sequenced samples. All Hox genes were manually checked. In some cases, the two genes were merged into one due to their close proximity; these were manually separated in the GTF file. FeatureCounts version 1.6.0 was then used with the new GTF file to assign reads to genes115. Differential gene expression between different tissues was performed in DESeq2 version 1.26.0 (ref. 116).

Apeltes RNA-seq followed similar protocols as Gasterosteus, with the following differences. Spines, blank pterygiophores, dorsal fin and anal fin were dissected from Louisbourg Apeltes clutches raised in the lab to 11–13 mm. The fry were genotyped for the 2 peak association mapping markers (AQ-HOXDB_6 and 7). For allele-specific expression analysis, four fish with the L/LHR genotype, four fish with the H/L genotype and three fish with the H/LHR genotype were sequenced. Three four-spine L/L genotypes were also sequenced to examine expression differences between tissues. To generate gene predictions for the Apeltes genome, StringTie was used; the Hox genes were identified by BLAST108 and manually named in the GTF file. For high-spine fish, allele-specific analysis was performed as described above for Gasterosteus. The Apeltes 10X-linked read data were used as input for 10X Long Ranger v.2.2.2 to generate a VCF file of known variants for GATK BaseRecalibator version Because the clutch size of Apeltes is smaller and thus the number of replicates was lower than for Gasterosteus, we used a two-tailed Fisher’s exact test to compare expression in dorsal tissues to the anal fin. We summed the references and alternate reads within each tissue to generate a 2 × 2 contingency table. For the analysis shown in Extended Data Fig. 5, differential gene expression between tissues was performed in DESeq2 (ref. 116).

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.